Refactor pre-SSE4 versions of QS8/QC8 GEMM/IGEMM microkernels
- Replace sign-extension from 8 to 16 bits with a more efficient sequence
- Replace casts to uintptr_t to casts to typed pointer types where possible
PiperOrigin-RevId: 382654408
diff --git a/src/qs8-igemm/MRx4c2-sse.c.in b/src/qs8-igemm/MRx4c2-sse.c.in
index 944dfbe..9a103a2 100644
--- a/src/qs8-igemm/MRx4c2-sse.c.in
+++ b/src/qs8-igemm/MRx4c2-sse.c.in
@@ -78,7 +78,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
$for M in range(1, MR):
__m128i vacc${M}x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -96,7 +96,7 @@
$if SSE == 4:
const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
$else:
- const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, _mm_cmpgt_epi8(_mm_setzero_si128(), va${M}));
+ const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8);
a${M} += 8;
$if VARIANT == "LD128":
@@ -104,7 +104,7 @@
$if K == 0:
const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) w);
$else:
- const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + ${K * 8}));
+ const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + ${K * 8}));
const __m128i vsb${K}${K+1} = _mm_cmpgt_epi8(_mm_setzero_si128(), vb${K}${K+1});
const __m128i vxb${K} = _mm_unpacklo_epi8(vb${K}${K+1}, vsb${K}${K+1});
const __m128i vxb${K+1} = _mm_unpackhi_epi8(vb${K}${K+1}, vsb${K}${K+1});
@@ -129,11 +129,11 @@
$if K == 0:
const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) w);
$else:
- const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${K * 8}));
+ const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + ${K * 8}));
$if SSE == 4:
const __m128i vxb${K} = _mm_cvtepi8_epi16(vb${K});
$else:
- const __m128i vxb${K} = _mm_unpacklo_epi8(vb${K}, _mm_cmpgt_epi8(_mm_setzero_si128(), vb${K}));
+ const __m128i vxb${K} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${K}, vb${K}), 8);
$for M in range(MR):
$if XOP:
@@ -143,7 +143,7 @@
vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K}));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -152,12 +152,12 @@
$if SSE == 4:
const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
$else:
- const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, _mm_cmpgt_epi8(_mm_setzero_si128(), va${M}));
+ const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8);
a${M} = (const int8_t*) ((uintptr_t) a${M} + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
$for M in range(MR):
$if XOP:
@@ -169,8 +169,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
$for M in range(MR):
$if XOP:
@@ -182,8 +182,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
$for M in range(MR):
$if XOP:
@@ -281,7 +281,7 @@
$if DATATYPE == "QC8":
const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ w = (const void*) ((const float*) w + 4);
$for M in range(MR):
vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale0123);
$else:
diff --git a/src/qs8-igemm/MRx4c8-sse.c.in b/src/qs8-igemm/MRx4c8-sse.c.in
index 965c2a9..1364e26 100644
--- a/src/qs8-igemm/MRx4c8-sse.c.in
+++ b/src/qs8-igemm/MRx4c8-sse.c.in
@@ -79,7 +79,7 @@
$for M in range(1, MR):
$for N in range(4):
__m128i vacc${M}x${N} = vacc0x${N};
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -97,7 +97,7 @@
$if SSE == 4:
const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
$else:
- const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, _mm_cmpgt_epi8(_mm_setzero_si128(), va${M}));
+ const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8);
a${M} += 8;
$if VARIANT == "LD128":
@@ -105,7 +105,7 @@
$if N == 0:
const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) w);
$else:
- const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) ((uintptr_t) w + ${N * 8}));
+ const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) ((const int8_t*) w + ${N * 8}));
const __m128i vsb${N}${N+1} = _mm_cmpgt_epi8(_mm_setzero_si128(), vb${N}${N+1});
const __m128i vxb${N} = _mm_unpacklo_epi8(vb${N}${N+1}, vsb${N}${N+1});
const __m128i vxb${N+1} = _mm_unpackhi_epi8(vb${N}${N+1}, vsb${N}${N+1});
@@ -122,11 +122,11 @@
$if N == 0:
const __m128i vb${N} = _mm_loadl_epi64((const __m128i*) w);
$else:
- const __m128i vb${N} = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${N * 8}));
+ const __m128i vb${N} = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + ${N * 8}));
$if SSE == 4:
const __m128i vxb${N} = _mm_cvtepi8_epi16(vb${N});
$else:
- const __m128i vxb${N} = _mm_unpacklo_epi8(vb${N}, _mm_cmpgt_epi8(_mm_setzero_si128(), vb${N}));
+ const __m128i vxb${N} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${N}, vb${N}), 8);
$for M in range(MR):
$if XOP:
@@ -134,7 +134,7 @@
$else:
vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}, vxb${N}));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= ${MR} * sizeof(void*);
@@ -238,7 +238,7 @@
$if DATATYPE == "QC8":
const __m128 vscale0123 = _mm_load_ps((const float*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ w = (const void*) ((const float*) w + 4);
$for M in range(MR):
vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale0123);
$else:
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld128.c b/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld128.c
index 37bde25..2a2ede8 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld128.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -71,7 +71,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -82,7 +82,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -91,24 +91,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld64.c b/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld64.c
index 0b84a74..4a815b9 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld64.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -66,23 +66,23 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -91,24 +91,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld128.c b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld128.c
index a5cbc1d..925cfa5 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld128.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -58,7 +58,7 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -71,7 +71,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -82,33 +82,33 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld64.c b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld64.c
index 55e4b0c..d8f5d52 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld64.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -58,57 +58,57 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld128.c b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld128.c
index f9324e3..d363292 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld128.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -71,7 +71,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -82,7 +82,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -91,24 +91,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld64.c b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld64.c
index abed29a..43fdb55 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld64.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -66,23 +66,23 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -91,24 +91,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld128.c b/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld128.c
index f26464a..1e3d6a8 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld128.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -76,7 +76,7 @@
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -87,7 +87,7 @@
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -96,24 +96,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld64.c b/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld64.c
index 3170471..c9cd4c0 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld64.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -71,23 +71,23 @@
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -96,24 +96,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld128.c b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld128.c
index 661d8ba..d84e43a 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld128.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -71,7 +71,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -82,7 +82,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -91,24 +91,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld64.c b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld64.c
index 7d25a0b..0c41602 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld64.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -66,23 +66,23 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -91,24 +91,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld128.c b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld128.c
index c0dcabd..28b48ca 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld128.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -58,7 +58,7 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -71,7 +71,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -82,33 +82,33 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld64.c b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld64.c
index 05f8476..2b65df2 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld64.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -58,57 +58,57 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld128.c b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld128.c
index 4ea067c..1f8b868 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld128.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -71,7 +71,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -82,7 +82,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -91,24 +91,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld64.c b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld64.c
index 27eebfa..6f97176 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld64.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -66,23 +66,23 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -91,24 +91,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld128.c b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld128.c
index 7cd1c20..d006ea0 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld128.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -58,7 +58,7 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -71,7 +71,7 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -82,33 +82,33 @@
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld64.c b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld64.c
index 4ac0e38..2d323e1 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld64.c
@@ -45,7 +45,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -58,57 +58,57 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld128.c b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld128.c
index 93e6feb..a8fb8da 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld128.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -76,7 +76,7 @@
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -87,7 +87,7 @@
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -96,24 +96,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld64.c b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld64.c
index 86570d8..7cdc634 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld64.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -71,23 +71,23 @@
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -96,24 +96,24 @@
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld128.c
index a81f6a4..c9be575 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld128.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -71,7 +71,7 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -79,7 +79,7 @@
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld64.c
index ead06e3..0568e12 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld64.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -68,20 +68,20 @@
const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld128.c
index 4360fb7..6711743 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld128.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -61,7 +61,7 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -71,7 +71,7 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -79,7 +79,7 @@
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld64.c
index b1cffaf..f16ea89 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld64.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -61,27 +61,27 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld128.c
index ef57058..4bd1f41 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld128.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -71,7 +71,7 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -79,7 +79,7 @@
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld64.c
index dfc49bf..2a31258 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld64.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -68,20 +68,20 @@
const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld128.c
index cc9b9c2..6f85867 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld128.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -61,7 +61,7 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -71,7 +71,7 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -79,7 +79,7 @@
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld64.c
index c19d705..db23866 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld64.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -61,27 +61,27 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld128.c
index 406550e..3fdbf3a 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld128.c
@@ -53,7 +53,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -76,7 +76,7 @@
vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -84,7 +84,7 @@
vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld64.c
index 2894376..6e56155 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld64.c
@@ -53,7 +53,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,20 +73,20 @@
const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld128.c
index 9bad1ab..17b8f27 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld128.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -71,7 +71,7 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -79,7 +79,7 @@
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld64.c
index 5c58a64..69009c7 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld64.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -68,20 +68,20 @@
const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld128.c
index 8429156..ec61f42 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld128.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -61,7 +61,7 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -71,7 +71,7 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -79,7 +79,7 @@
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld64.c
index b7a1222..50c2d49 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld64.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -61,27 +61,27 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld128.c
index be97585..d406f4e 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld128.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -71,7 +71,7 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -79,7 +79,7 @@
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld64.c
index f5d3dcf..e2b9bef 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld64.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -68,20 +68,20 @@
const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld128.c
index 8e085ec..faaf3f5 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld128.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -61,7 +61,7 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -71,7 +71,7 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -79,7 +79,7 @@
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld64.c
index 83d2a34..b8bd4c5 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld64.c
@@ -48,7 +48,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -61,27 +61,27 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld128.c
index d9d0c90..c8d1f8e 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld128.c
@@ -53,7 +53,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -76,7 +76,7 @@
vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -84,7 +84,7 @@
vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld64.c
index 11011b1..0d1cb95 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld64.c
@@ -53,7 +53,7 @@
__m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
__m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
__m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,20 +73,20 @@
const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 1 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld128.c b/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld128.c
index 2a6314a..6d12bd6 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld128.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -87,7 +87,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -102,7 +102,7 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -114,8 +114,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld64.c b/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld64.c
index 4ade5aa..e89696a 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld64.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -80,21 +80,21 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -102,7 +102,7 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -114,8 +114,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld128.c b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld128.c
index 2cebe9c..b7f7f8d 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld128.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -67,10 +67,10 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -87,7 +87,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -102,20 +102,20 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld64.c b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld64.c
index a373cc6..398ee81 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld64.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -67,55 +67,55 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld128.c b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld128.c
index b2232e7..791ba21 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld128.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -87,7 +87,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -102,7 +102,7 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -114,8 +114,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld64.c b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld64.c
index 074b5d0..1b839e6 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld64.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -80,21 +80,21 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -102,7 +102,7 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -114,8 +114,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld128.c b/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld128.c
index a53f14d..d80dda7 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld128.c
@@ -55,7 +55,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -92,7 +92,7 @@
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -107,7 +107,7 @@
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -119,8 +119,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -129,8 +129,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -139,8 +139,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld64.c b/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld64.c
index a5349ee..f4f8b9a 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld64.c
@@ -55,7 +55,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,21 +85,21 @@
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_maddd_epi16(
@@ -107,7 +107,7 @@
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -119,8 +119,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -129,8 +129,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -139,8 +139,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld128.c b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld128.c
index 33a075a..81f4c32 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld128.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -87,7 +87,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -102,7 +102,7 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -114,8 +114,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld64.c b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld64.c
index 2edacd9..9d164eb 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld64.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -80,21 +80,21 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -102,7 +102,7 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -114,8 +114,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld128.c b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld128.c
index 8dada86..39aea8e 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld128.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -67,10 +67,10 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -87,7 +87,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -102,20 +102,20 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld64.c b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld64.c
index 402b6d4..128ac34 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld64.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -67,55 +67,55 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld128.c b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld128.c
index cce3b5f..9245e7e 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld128.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -87,7 +87,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -102,7 +102,7 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -114,8 +114,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld64.c b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld64.c
index 9d7733e..f97b73c 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld64.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -80,21 +80,21 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -102,7 +102,7 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -114,8 +114,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld128.c b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld128.c
index c91fcd9..d3c9714 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld128.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -67,10 +67,10 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -87,7 +87,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -102,20 +102,20 @@
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld64.c b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld64.c
index f790aef..bf47976 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld64.c
@@ -50,7 +50,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -67,55 +67,55 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
vacc1x0123 = _mm_add_epi32(vacc1x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -124,8 +124,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -134,8 +134,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld128.c b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld128.c
index 51cc3cf..aefa9c1 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld128.c
@@ -55,7 +55,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -92,7 +92,7 @@
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -107,7 +107,7 @@
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -119,8 +119,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -129,8 +129,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -139,8 +139,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld64.c b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld64.c
index 251dded..f88b0f1 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld64.c
@@ -55,7 +55,7 @@
do {
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,21 +85,21 @@
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_maddd_epi16(
@@ -107,7 +107,7 @@
vacc1x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -119,8 +119,8 @@
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -129,8 +129,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -139,8 +139,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld128.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld128.c
index 00d798f..aa4745d 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld128.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -88,7 +88,7 @@
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -98,7 +98,7 @@
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld64.c
index 56827ed..26f50da 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld64.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -84,23 +84,23 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld128.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld128.c
index 834cbdf..e46e6ed 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld128.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,10 +73,10 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -88,7 +88,7 @@
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -98,7 +98,7 @@
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c
index 8f98520..82e5115 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,34 +73,34 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld128.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld128.c
index e42f500..0cacb43 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld128.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -88,7 +88,7 @@
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -98,7 +98,7 @@
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld64.c
index de772cb..69e9ef8 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld64.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -84,23 +84,23 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld128.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld128.c
index d3f20da..dc0a620 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld128.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,10 +73,10 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -88,7 +88,7 @@
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -98,7 +98,7 @@
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld64.c
index 88a422b..e8e086e 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld64.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,34 +73,34 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld128.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld128.c
index 5c7af79..ada4f99 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld128.c
@@ -61,7 +61,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -93,7 +93,7 @@
vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -103,7 +103,7 @@
vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld64.c
index a725181..0f3b693 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld64.c
@@ -61,7 +61,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -89,23 +89,23 @@
vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld128.c b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld128.c
index ceed36c..987dcb0 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld128.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -88,7 +88,7 @@
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -98,7 +98,7 @@
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld64.c
index ca30e42..3206354 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld64.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -84,23 +84,23 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld128.c b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld128.c
index bc54486..4f88e7c 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld128.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,10 +73,10 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -88,7 +88,7 @@
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -98,7 +98,7 @@
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld64.c
index c16c117..29f6bb9 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld64.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,34 +73,34 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld128.c b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld128.c
index cd2d9f2..fc1ce4d 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld128.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -88,7 +88,7 @@
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -98,7 +98,7 @@
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld64.c
index 6eb3515..696aea2 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld64.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -84,23 +84,23 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld128.c b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld128.c
index 727aac5..ce4dbf1 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld128.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,10 +73,10 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -88,7 +88,7 @@
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -98,7 +98,7 @@
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld64.c
index 5bfabbd..e7c1e34 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld64.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,34 +73,34 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld128.c b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld128.c
index d3a2605..c690081 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld128.c
@@ -61,7 +61,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -93,7 +93,7 @@
vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -103,7 +103,7 @@
vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld64.c
index 717b1ae..f69db77 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld64.c
@@ -61,7 +61,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -89,23 +89,23 @@
vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld128.c b/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld128.c
index 76801ec..39e1fc4 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld128.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -122,7 +122,7 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -137,8 +137,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld64.c b/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld64.c
index 60b05d1..d939ab6 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld64.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -94,7 +94,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -112,7 +112,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -122,7 +122,7 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -137,8 +137,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld128.c b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld128.c
index 1ec2f1e..a2ecb0f 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld128.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -76,13 +76,13 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -122,23 +122,23 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld64.c b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld64.c
index 90828cd..4642e48 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld64.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -76,17 +76,17 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -94,8 +94,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -103,8 +103,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
@@ -112,8 +112,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
@@ -122,23 +122,23 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld128.c b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld128.c
index 72b509c..020f028 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld128.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -122,7 +122,7 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -137,8 +137,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld64.c b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld64.c
index 0fba4e3..9e3ff00 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld64.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -94,7 +94,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -112,7 +112,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -122,7 +122,7 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -137,8 +137,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld128.c b/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld128.c
index cfc480e..d14a6c9 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld128.c
@@ -60,7 +60,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -108,7 +108,7 @@
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -127,7 +127,7 @@
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -142,8 +142,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -154,8 +154,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -166,8 +166,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld64.c b/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld64.c
index cf86f08..a2b6c23 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld64.c
@@ -60,7 +60,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -99,7 +99,7 @@
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
@@ -108,7 +108,7 @@
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
@@ -117,7 +117,7 @@
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_maddd_epi16(
@@ -127,7 +127,7 @@
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -142,8 +142,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -154,8 +154,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -166,8 +166,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld128.c b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld128.c
index 0c941bd..abed9b6 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld128.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -122,7 +122,7 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -137,8 +137,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld64.c b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld64.c
index e98b019..37cc36d 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld64.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -94,7 +94,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -112,7 +112,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -122,7 +122,7 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -137,8 +137,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld128.c b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld128.c
index 51b1271..3836866 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld128.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -76,13 +76,13 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -122,23 +122,23 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld64.c b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld64.c
index 25b9a16..5d1886d 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld64.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -76,17 +76,17 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -94,8 +94,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -103,8 +103,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
@@ -112,8 +112,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
@@ -122,23 +122,23 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld128.c b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld128.c
index 02850f4..1ffbff8 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld128.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -122,7 +122,7 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -137,8 +137,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld64.c b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld64.c
index b886943..84d280b 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld64.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -94,7 +94,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -112,7 +112,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -122,7 +122,7 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -137,8 +137,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld128.c b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld128.c
index c9743bc..92e9447 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld128.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -76,13 +76,13 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -103,7 +103,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -122,23 +122,23 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld64.c b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld64.c
index 5b6f8e6..75700c3 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld64.c
@@ -55,7 +55,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -76,17 +76,17 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -94,8 +94,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -103,8 +103,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
@@ -112,8 +112,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
@@ -122,23 +122,23 @@
vacc2x0123 = _mm_add_epi32(vacc2x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -149,8 +149,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -161,8 +161,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld128.c b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld128.c
index 7df63c6..6356e76 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld128.c
@@ -60,7 +60,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -108,7 +108,7 @@
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -127,7 +127,7 @@
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -142,8 +142,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -154,8 +154,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -166,8 +166,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld64.c b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld64.c
index 197a6af..4912f55 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld64.c
@@ -60,7 +60,7 @@
__m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -99,7 +99,7 @@
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
@@ -108,7 +108,7 @@
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
@@ -117,7 +117,7 @@
_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_maddd_epi16(
@@ -127,7 +127,7 @@
vacc2x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -142,8 +142,8 @@
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -154,8 +154,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -166,8 +166,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld128.c
index 4116936..f198a7d 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld128.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -105,7 +105,7 @@
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -117,7 +117,7 @@
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld64.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld64.c
index aa33476..fb028a4 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld64.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -100,26 +100,26 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld128.c
index bd5324f..8dc7fdb 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld128.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,13 +85,13 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -105,7 +105,7 @@
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -117,7 +117,7 @@
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld64.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld64.c
index 2a81f95..079ee69 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld64.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,41 +85,41 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
index 6531f50..d262753 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -105,7 +105,7 @@
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -117,7 +117,7 @@
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld64.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld64.c
index 500f75d..a3f8f88 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld64.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -100,26 +100,26 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld128.c
index 8831628..8ffc6fc 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld128.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,13 +85,13 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -105,7 +105,7 @@
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -117,7 +117,7 @@
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld64.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld64.c
index 036b585..6c712bf 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld64.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,41 +85,41 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld128.c
index 883264f..55639ef 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld128.c
@@ -69,7 +69,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -110,7 +110,7 @@
vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -122,7 +122,7 @@
vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld64.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld64.c
index 439051b..19f5904 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld64.c
@@ -69,7 +69,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -105,26 +105,26 @@
vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld128.c
index b49c379..a13dd17 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld128.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -105,7 +105,7 @@
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -117,7 +117,7 @@
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld64.c b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld64.c
index a7d9926..fbad091 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld64.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -100,26 +100,26 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld128.c
index a59b40a..e4200cd 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld128.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,13 +85,13 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -105,7 +105,7 @@
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -117,7 +117,7 @@
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld64.c b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld64.c
index ca33a86..a5f3732 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld64.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,41 +85,41 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld128.c
index ca1f14f..15de5b5 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld128.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -105,7 +105,7 @@
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -117,7 +117,7 @@
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld64.c b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld64.c
index d658603..610c200 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld64.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -100,26 +100,26 @@
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld128.c
index 1ec5500..b5919d7 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld128.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,13 +85,13 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb01 = _mm_load_si128((const __m128i*) w);
@@ -105,7 +105,7 @@
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -117,7 +117,7 @@
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld64.c b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld64.c
index ebf492b..3f6204f 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld64.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,41 +85,41 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld128.c
index 510b01b..a866141 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld128.c
@@ -69,7 +69,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -110,7 +110,7 @@
vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -122,7 +122,7 @@
vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld64.c b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld64.c
index 3c38797..c297cc2 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld64.c
@@ -69,7 +69,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -105,26 +105,26 @@
vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld128.c b/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld128.c
index 20e55bd..9575adb 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld128.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -142,7 +142,7 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -160,8 +160,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld64.c b/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld64.c
index d5e1122..86dbc98 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld64.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -108,7 +108,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -130,7 +130,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -142,7 +142,7 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -160,8 +160,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld128.c b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld128.c
index e8144fe..31c84c2 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld128.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,16 +85,16 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -142,26 +142,26 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld64.c b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld64.c
index b305c8d..0d6a7da 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld64.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,20 +85,20 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -108,8 +108,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -119,8 +119,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
@@ -130,8 +130,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
@@ -142,26 +142,26 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld128.c b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld128.c
index 48ff03e..112e243 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld128.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -142,7 +142,7 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -160,8 +160,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld64.c b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld64.c
index cb3166d..c8a3cca 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld64.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -108,7 +108,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -130,7 +130,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -142,7 +142,7 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -160,8 +160,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld128.c b/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld128.c
index 48d99be..efa3ada 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld128.c
@@ -65,7 +65,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -124,7 +124,7 @@
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -147,7 +147,7 @@
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -165,8 +165,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -179,8 +179,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -193,8 +193,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld64.c b/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld64.c
index 9387431..9f9ad98 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld64.c
@@ -65,7 +65,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -113,7 +113,7 @@
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
@@ -124,7 +124,7 @@
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
@@ -135,7 +135,7 @@
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_maddd_epi16(
@@ -147,7 +147,7 @@
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -165,8 +165,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -179,8 +179,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -193,8 +193,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld128.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld128.c
index de5da17..681ea5f 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld128.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -142,7 +142,7 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -160,8 +160,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld64.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld64.c
index d4f6be1..7e66cc1 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld64.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -108,7 +108,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -130,7 +130,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -142,7 +142,7 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -160,8 +160,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld128.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld128.c
index 3c7af53..5875e0e 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld128.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,16 +85,16 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -142,26 +142,26 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld64.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld64.c
index e40ce90..89ba44a 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld64.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,20 +85,20 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -108,8 +108,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -119,8 +119,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
@@ -130,8 +130,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
@@ -142,26 +142,26 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld128.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld128.c
index e577a64..64e1ff5 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld128.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -142,7 +142,7 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -160,8 +160,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld64.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld64.c
index 2ab9bc8..8571512 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld64.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -108,7 +108,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -130,7 +130,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
@@ -142,7 +142,7 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -160,8 +160,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld128.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld128.c
index f844b67..35d07a0 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld128.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,16 +85,16 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 += 8;
const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
@@ -119,7 +119,7 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -142,26 +142,26 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld64.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld64.c
index d8765d9..5d141e0 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld64.c
@@ -60,7 +60,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -85,20 +85,20 @@
size_t k = kc;
while (k >= 8 * sizeof(int8_t)) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 += 8;
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -108,8 +108,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -119,8 +119,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
@@ -130,8 +130,8 @@
_mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
@@ -142,26 +142,26 @@
vacc3x0123 = _mm_add_epi32(vacc3x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 = (const int8_t*) ((uintptr_t) a0 + k);
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 = (const int8_t*) ((uintptr_t) a1 + k);
const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
- const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
a2 = (const int8_t*) ((uintptr_t) a2 + k);
const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
- const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -174,8 +174,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
@@ -188,8 +188,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c
index 61b2117..49717e2 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c
@@ -65,7 +65,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -124,7 +124,7 @@
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
- const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -147,7 +147,7 @@
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -165,8 +165,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -179,8 +179,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -193,8 +193,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld64.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld64.c
index fc69a9c..d4b022a 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld64.c
@@ -65,7 +65,7 @@
__m128i vacc1x0123 = vacc0x0123;
__m128i vacc2x0123 = vacc0x0123;
__m128i vacc3x0123 = vacc0x0123;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -113,7 +113,7 @@
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
@@ -124,7 +124,7 @@
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
@@ -135,7 +135,7 @@
_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
vacc0x0123 = _mm_maddd_epi16(
@@ -147,7 +147,7 @@
vacc3x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k -= 8 * sizeof(int8_t);
}
if (k != 0) {
@@ -165,8 +165,8 @@
a3 = (const int8_t*) ((uintptr_t) a3 + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -179,8 +179,8 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -193,8 +193,8 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((uintptr_t) w + 8);
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((const int8_t*) w + 8);
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);