Code-generate QU8 GEMM and IGEMM microkernels for SSE2/SSSE3/SSE4.1
PiperOrigin-RevId: 382681546
diff --git a/src/qs8-igemm/MRx4c2-sse.c.in b/src/qs8-igemm/MRx4c2-sse.c.in
index 9a103a2..208a884 100644
--- a/src/qs8-igemm/MRx4c2-sse.c.in
+++ b/src/qs8-igemm/MRx4c2-sse.c.in
@@ -7,7 +7,7 @@
$assert not XOP or AVX
$assert not AVX or SSE == 4
$assert REQUANTIZATION in ["GEMMLOWP", "FP32"]
-$assert DATATYPE in ["QC8", "QS8"]
+$assert DATATYPE in ["QC8", "QS8", "QU8"]
$assert DATATYPE != "QC8" or REQUANTIZATION == "FP32"
$assert SSE != 3 or REQUANTIZATION != "FP32"
$assert VARIANT in ["LD64", "LD128"]
@@ -29,21 +29,22 @@
#include <xnnpack/math.h>
-$PARAMS_UNION = "xnn_qs8_minmax_params" if DATATYPE == "QC8" else "xnn_qs8_conv_minmax_params"
-$PARAMS_STRUCT = ("" if DATATYPE == "QC8" else REQUANTIZATION.lower() + "_") + ("sse4" if SSE >= 4 else "sse2")
+$PARAMS_UNION = "xnn_qs8_minmax_params" if DATATYPE == "QC8" else "xnn_%s_conv_minmax_params" % DATATYPE.lower()
+$PARAMS_STRUCT = ("" if DATATYPE == "QC8" else REQUANTIZATION.lower() + "_") + ("sse4" if SSE >= 4 and DATATYPE != "QU8" else "sse2")
+$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
$ISA = "xop" if XOP else "avx" if AVX else {2: "sse2", 3: "ssse3", 4: "sse41"}[SSE]
void xnn_${DATATYPE.lower()}_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x4c2__${ISA}_${VARIANT.lower()}(
size_t mr,
size_t nc,
size_t kc,
size_t ks,
- const int8_t** restrict a,
+ const ${XINT8_T}** restrict a,
const void* restrict w,
- int8_t* restrict c,
+ ${XINT8_T}* restrict c,
size_t cm_stride,
size_t cn_stride,
size_t a_offset,
- const int8_t* zero,
+ const ${XINT8_T}* zero,
const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
{
assert(mr != 0);
@@ -52,15 +53,15 @@
assert(kc != 0);
assert(ks != 0);
assert(ks % (${MR} * sizeof(void*)) == 0);
- assert(a_offset % sizeof(int8_t) == 0);
+ assert(a_offset % sizeof(${XINT8_T}) == 0);
assert(a != NULL);
assert(w != NULL);
assert(c != NULL);
kc = round_up_po2(kc, 2);
- int8_t* c0 = c;
+ ${XINT8_T}* c0 = c;
$for M in range(1, MR):
- int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride);
+ ${XINT8_T}* c${M} = (${XINT8_T}*) ((uintptr_t) c${M-1} + cm_stride);
$if M % 2 == 0:
if XNN_UNPREDICTABLE(mr <= ${M}) {
c${M} = c${M-1};
@@ -83,20 +84,30 @@
size_t p = ks;
do {
$for M in range(MR):
- const int8_t* restrict a${M} = a[${M}];
+ const ${XINT8_T}* restrict a${M} = a[${M}];
if XNN_UNPREDICTABLE(a${M} != zero) {
- a${M} = (const int8_t*) ((uintptr_t) a${M} + a_offset);
+ a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} + a_offset);
}
a += ${MR};
size_t k = kc;
- while (k >= 8 * sizeof(int8_t)) {
+ $if DATATYPE == "QU8":
+ const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.kernel_zero_point);
+ $if SSE < 4:
+ const __m128i vzero = _mm_setzero_si128();
+ while (k >= 8 * sizeof(${XINT8_T})) {
$for M in range(MR):
const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M});
- $if SSE == 4:
- const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
+ $if DATATYPE == "QU8":
+ $if SSE == 4:
+ const __m128i vxa${M} = _mm_cvtepu8_epi16(va${M});
+ $else:
+ const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, vzero);
$else:
- const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8);
+ $if SSE == 4:
+ const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
+ $else:
+ const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8);
a${M} += 8;
$if VARIANT == "LD128":
@@ -104,10 +115,14 @@
$if K == 0:
const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) w);
$else:
- const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) ((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}${K+1});
+ const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) ((const ${XINT8_T}*) w + ${K * 8}));
+ $if DATATYPE == "QU8":
+ const __m128i vxb${K} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${K}${K+1}, vzero), vb_zero_point);
+ const __m128i vxb${K+1} = _mm_sub_epi16(_mm_unpackhi_epi8(vb${K}${K+1}, vzero), vb_zero_point);
+ $else:
+ const __m128i vsb${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});
$for M in range(MR):
$if XOP:
@@ -129,11 +144,17 @@
$if K == 0:
const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) w);
$else:
- const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + ${K * 8}));
- $if SSE == 4:
- const __m128i vxb${K} = _mm_cvtepi8_epi16(vb${K});
+ const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) ((const ${XINT8_T}*) w + ${K * 8}));
+ $if DATATYPE == "QU8":
+ $if SSE == 4:
+ const __m128i vxb${K} = _mm_sub_epi16(_mm_cvtepu8_epi16(vb${K}), vb_zero_point);
+ $else:
+ const __m128i vxb${K} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${K}, vzero), vb_zero_point);
$else:
- const __m128i vxb${K} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${K}, vb${K}), 8);
+ $if SSE == 4:
+ const __m128i vxb${K} = _mm_cvtepi8_epi16(vb${K});
+ $else:
+ const __m128i vxb${K} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${K}, vb${K}), 8);
$for M in range(MR):
$if XOP:
@@ -143,21 +164,36 @@
vacc${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*) ((const int8_t*) w + 32);
- k -= 8 * sizeof(int8_t);
+ w = (const void*) ((const ${XINT8_T}*) w + 32);
+ k -= 8 * sizeof(${XINT8_T});
}
if (k != 0) {
$for M in range(MR):
const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M});
- $if SSE == 4:
- const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
+ $if DATATYPE == "QU8":
+ $if SSE == 4:
+ const __m128i vxa${M} = _mm_cvtepu8_epi16(va${M});
+ $else:
+ const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, vzero);
$else:
- const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8);
- a${M} = (const int8_t*) ((uintptr_t) a${M} + k);
+ $if SSE == 4:
+ const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
+ $else:
+ const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8);
+ a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} + k);
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ w = (const void*) ((const ${XINT8_T}*) w + 8);
+ $if DATATYPE == "QU8":
+ $if SSE == 4:
+ const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
+ $else:
+ const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
+ $else:
+ $if SSE == 4:
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ $else:
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
$for M in range(MR):
$if XOP:
@@ -167,10 +203,19 @@
vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
- if (k > 2 * sizeof(int8_t)) {
+ if (k > 2 * sizeof(${XINT8_T})) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ w = (const void*) ((const ${XINT8_T}*) w + 8);
+ $if DATATYPE == "QU8":
+ $if SSE == 4:
+ const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
+ $else:
+ const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
+ $else:
+ $if SSE == 4:
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ $else:
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
$for M in range(MR):
$if XOP:
@@ -180,10 +225,19 @@
vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
- if (k > 4 * sizeof(int8_t)) {
+ if (k > 4 * sizeof(${XINT8_T})) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
- w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ w = (const void*) ((const ${XINT8_T}*) w + 8);
+ $if DATATYPE == "QU8":
+ $if SSE == 4:
+ const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
+ $else:
+ const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
+ $else:
+ $if SSE == 4:
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ $else:
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
$for M in range(MR):
$if XOP:
@@ -296,20 +350,29 @@
$for M in range(0, MR, 2):
__m128i vacc${M}${min(M+1, MR-1)}x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc${M}x0123, vacc${min(M+1, MR-1)}x0123), voutput_zero_point);
- $if SSE < 4:
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max);
- $for M in range(0, MR, 2):
- vacc${M}${min(M+1, MR-1)}x0123 = _mm_min_epi16(_mm_max_epi16(vacc${M}${min(M+1, MR-1)}x0123, voutput_min), voutput_max);
+ $if DATATYPE == "QU8":
+ $if MR > 2:
+ __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);
+ $else:
+ __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123);
- $if MR > 2:
- __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);
+ vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min));
+ vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max));
$else:
- __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123);
+ $if SSE < 4:
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max);
+ $for M in range(0, MR, 2):
+ vacc${M}${min(M+1, MR-1)}x0123 = _mm_min_epi16(_mm_max_epi16(vacc${M}${min(M+1, MR-1)}x0123, voutput_min), voutput_max);
- $if SSE == 4:
- vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min));
- vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max));
+ $if MR > 2:
+ __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);
+ $else:
+ __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123);
+
+ $if SSE == 4:
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max));
if (nc >= 4) {
$for M in reversed(range(1, MR)):
@@ -317,11 +380,11 @@
*((uint32_t*) c${M}) = (uint32_t) _mm_extract_epi32(vout, ${M});
$else:
*((uint32_t*) c${M}) = (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(${M}, ${M}, ${M}, ${M})));
- c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);
+ c${M} = (${XINT8_T}*) ((uintptr_t) c${M} + cn_stride);
*((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
- c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c0 = (${XINT8_T}*) ((uintptr_t) c0 + cn_stride);
- a = (const int8_t**restrict) ((uintptr_t) a - ks);
+ a = (const ${XINT8_T}**restrict) ((uintptr_t) a - ks);
nc -= 4;
} else {
@@ -334,11 +397,11 @@
if (nc & 1) {
$if SSE == 4:
$for M in reversed(range(MR)):
- *((int8_t*) c${M}) = (int8_t) _mm_extract_epi8(vout, ${M * 4});
+ *c${M} = (${XINT8_T}) _mm_extract_epi8(vout, ${M * 4});
$else:
$for M in reversed(range(1, MR)):
- *((int8_t*) c${M}) = (int8_t) _mm_extract_epi16(vout, ${M * 2});
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c${M} = (${XINT8_T}) _mm_extract_epi16(vout, ${M * 2});
+ *c0 = (${XINT8_T}) _mm_cvtsi128_si32(vout);
}
nc = 0;
diff --git a/src/qs8-igemm/MRx4c8-sse.c.in b/src/qs8-igemm/MRx4c8-sse.c.in
index 1364e26..2a72587 100644
--- a/src/qs8-igemm/MRx4c8-sse.c.in
+++ b/src/qs8-igemm/MRx4c8-sse.c.in
@@ -7,7 +7,7 @@
$assert not XOP or AVX
$assert not AVX or SSE == 4
$assert REQUANTIZATION in ["GEMMLOWP", "FP32"]
-$assert DATATYPE in ["QC8", "QS8"]
+$assert DATATYPE in ["QC8", "QS8", "QU8"]
$assert DATATYPE != "QC8" or REQUANTIZATION == "FP32"
$assert VARIANT in ["LD64", "LD128"]
$assert MR <= 4
@@ -28,21 +28,22 @@
#include <xnnpack/math.h>
-$PARAMS_UNION = "xnn_qs8_minmax_params" if DATATYPE == "QC8" else "xnn_qs8_conv_minmax_params"
-$PARAMS_STRUCT = ("" if DATATYPE == "QC8" else REQUANTIZATION.lower() + "_") + ("sse4" if SSE >= 4 else "sse2")
+$PARAMS_UNION = "xnn_qs8_minmax_params" if DATATYPE == "QC8" else "xnn_%s_conv_minmax_params" % DATATYPE.lower()
+$PARAMS_STRUCT = ("" if DATATYPE == "QC8" else REQUANTIZATION.lower() + "_") + ("sse4" if SSE >= 4 and DATATYPE != "QU8" else "sse2")
+$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
$ISA = "xop" if XOP else "avx" if AVX else {2: "sse2", 3: "ssse3", 4: "sse41"}[SSE]
void xnn_${DATATYPE.lower()}_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x4c8__${ISA}_${VARIANT.lower()}(
size_t mr,
size_t nc,
size_t kc,
size_t ks,
- const int8_t** restrict a,
+ const ${XINT8_T}** restrict a,
const void* restrict w,
- int8_t* restrict c,
+ ${XINT8_T}* restrict c,
size_t cm_stride,
size_t cn_stride,
size_t a_offset,
- const int8_t* zero,
+ const ${XINT8_T}* zero,
const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
{
assert(mr != 0);
@@ -51,15 +52,15 @@
assert(kc != 0);
assert(ks != 0);
assert(ks % (${MR} * sizeof(void*)) == 0);
- assert(a_offset % sizeof(int8_t) == 0);
+ assert(a_offset % sizeof(${XINT8_T}) == 0);
assert(a != NULL);
assert(w != NULL);
assert(c != NULL);
kc = round_up_po2(kc, 8);
- int8_t* c0 = c;
+ ${XINT8_T}* c0 = c;
$for M in range(1, MR):
- int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride);
+ ${XINT8_T}* c${M} = (${XINT8_T}*) ((uintptr_t) c${M-1} + cm_stride);
$if M % 2 == 0:
if XNN_UNPREDICTABLE(mr <= ${M}) {
c${M} = c${M-1};
@@ -84,20 +85,30 @@
size_t p = ks;
do {
$for M in range(MR):
- const int8_t* restrict a${M} = a[${M}];
+ const ${XINT8_T}* restrict a${M} = a[${M}];
if XNN_UNPREDICTABLE(a${M} != zero) {
- a${M} = (const int8_t*) ((uintptr_t) a${M} + a_offset);
+ a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} + a_offset);
}
a += ${MR};
size_t k = 0;
+ $if DATATYPE == "QU8":
+ const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.kernel_zero_point);
+ $if SSE < 4:
+ const __m128i vzero = _mm_setzero_si128();
while (k < kc) {
$for M in range(MR):
const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M});
- $if SSE == 4:
- const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
+ $if DATATYPE == "QU8":
+ $if SSE == 4:
+ const __m128i vxa${M} = _mm_cvtepu8_epi16(va${M});
+ $else:
+ const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, vzero);
$else:
- const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8);
+ $if SSE == 4:
+ const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M});
+ $else:
+ const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8);
a${M} += 8;
$if VARIANT == "LD128":
@@ -105,10 +116,14 @@
$if N == 0:
const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) w);
$else:
- const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) ((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}${N+1});
+ const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) ((const ${XINT8_T}*) w + ${N * 8}));
+ $if DATATYPE == "QU8":
+ const __m128i vxb${N} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${N}${N+1}, vzero), vb_zero_point);
+ const __m128i vxb${N+1} = _mm_sub_epi16(_mm_unpackhi_epi8(vb${N}${N+1}, vzero), vb_zero_point);
+ $else:
+ const __m128i vsb${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});
$for M in range(MR):
$if XOP:
@@ -122,11 +137,17 @@
$if N == 0:
const __m128i vb${N} = _mm_loadl_epi64((const __m128i*) w);
$else:
- const __m128i vb${N} = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + ${N * 8}));
- $if SSE == 4:
- const __m128i vxb${N} = _mm_cvtepi8_epi16(vb${N});
+ const __m128i vb${N} = _mm_loadl_epi64((const __m128i*) ((const ${XINT8_T}*) w + ${N * 8}));
+ $if DATATYPE == "QU8":
+ $if SSE == 4:
+ const __m128i vxb${N} = _mm_sub_epi16(_mm_cvtepu8_epi16(vb${N}), vb_zero_point);
+ $else:
+ const __m128i vxb${N} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${N}, vzero), vb_zero_point);
$else:
- const __m128i vxb${N} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${N}, vb${N}), 8);
+ $if SSE == 4:
+ const __m128i vxb${N} = _mm_cvtepi8_epi16(vb${N});
+ $else:
+ const __m128i vxb${N} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${N}, vb${N}), 8);
$for M in range(MR):
$if XOP:
@@ -134,8 +155,8 @@
$else:
vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}, vxb${N}));
- w = (const void*) ((const int8_t*) w + 32);
- k += 8 * sizeof(int8_t);
+ w = (const void*) ((const ${XINT8_T}*) w + 32);
+ k += 8 * sizeof(${XINT8_T});
}
p -= ${MR} * sizeof(void*);
} while (p != 0);
@@ -253,20 +274,29 @@
$for M in range(0, MR, 2):
__m128i vacc${M}${min(M+1, MR-1)}x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc${M}x0123, vacc${min(M+1, MR-1)}x0123), voutput_zero_point);
- $if SSE < 4:
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max);
- $for M in range(0, MR, 2):
- vacc${M}${min(M+1, MR-1)}x0123 = _mm_min_epi16(_mm_max_epi16(vacc${M}${min(M+1, MR-1)}x0123, voutput_min), voutput_max);
+ $if DATATYPE == "QU8":
+ $if MR > 2:
+ __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);
+ $else:
+ __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123);
- $if MR > 2:
- __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);
+ vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min));
+ vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max));
$else:
- __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123);
+ $if SSE < 4:
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max);
+ $for M in range(0, MR, 2):
+ vacc${M}${min(M+1, MR-1)}x0123 = _mm_min_epi16(_mm_max_epi16(vacc${M}${min(M+1, MR-1)}x0123, voutput_min), voutput_max);
- $if SSE == 4:
- vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min));
- vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max));
+ $if MR > 2:
+ __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);
+ $else:
+ __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123);
+
+ $if SSE == 4:
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_max));
if (nc >= 4) {
$for M in reversed(range(1, MR)):
@@ -274,11 +304,11 @@
*((uint32_t*) c${M}) = (uint32_t) _mm_extract_epi32(vout, ${M});
$else:
*((uint32_t*) c${M}) = (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(${M}, ${M}, ${M}, ${M})));
- c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);
+ c${M} = (${XINT8_T}*) ((uintptr_t) c${M} + cn_stride);
*((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
- c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c0 = (${XINT8_T}*) ((uintptr_t) c0 + cn_stride);
- a = (const int8_t**restrict) ((uintptr_t) a - ks);
+ a = (const ${XINT8_T}**restrict) ((uintptr_t) a - ks);
nc -= 4;
} else {
@@ -291,11 +321,11 @@
if (nc & 1) {
$if SSE == 4:
$for M in reversed(range(MR)):
- *((int8_t*) c${M}) = (int8_t) _mm_extract_epi8(vout, ${M * 4});
+ *c${M} = (${XINT8_T}) _mm_extract_epi8(vout, ${M * 4});
$else:
$for M in reversed(range(1, MR)):
- *((int8_t*) c${M}) = (int8_t) _mm_extract_epi16(vout, ${M * 2});
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c${M} = (${XINT8_T}) _mm_extract_epi16(vout, ${M * 2});
+ *c0 = (${XINT8_T}) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 2a2ede8..cb6b952 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld128.c
@@ -92,7 +92,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -100,7 +100,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -108,7 +108,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -148,7 +148,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 4a815b9..3a6470c 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-avx-ld64.c
@@ -92,7 +92,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -100,7 +100,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -108,7 +108,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -148,7 +148,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 925cfa5..02159fa 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld128.c
@@ -149,7 +149,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 d8f5d52..99a6611 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse2-ld64.c
@@ -149,7 +149,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 d363292..8003b56 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld128.c
@@ -92,7 +92,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -100,7 +100,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -108,7 +108,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -148,7 +148,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 43fdb55..bbce5a9 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-sse41-ld64.c
@@ -92,7 +92,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -100,7 +100,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -108,7 +108,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -148,7 +148,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 1e3d6a8..dd3c288 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld128.c
@@ -97,7 +97,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -105,7 +105,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -113,7 +113,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -153,7 +153,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 c9cd4c0..0499b23 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-fp32-xop-ld64.c
@@ -97,7 +97,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -105,7 +105,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -113,7 +113,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -153,7 +153,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 d84e43a..fbe2503 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld128.c
@@ -92,7 +92,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -100,7 +100,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -108,7 +108,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -164,7 +164,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 0c41602..3c7cc20 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-avx-ld64.c
@@ -92,7 +92,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -100,7 +100,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -108,7 +108,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -164,7 +164,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 28b48ca..e426312 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld128.c
@@ -181,7 +181,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 2b65df2..7835c4d 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld64.c
@@ -181,7 +181,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 1f8b868..d1ead47 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld128.c
@@ -92,7 +92,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -100,7 +100,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -108,7 +108,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -164,7 +164,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 6f97176..c6f5af2 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse41-ld64.c
@@ -92,7 +92,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -100,7 +100,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -108,7 +108,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -164,7 +164,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 d006ea0..ec18bb0 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld128.c
@@ -181,7 +181,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 2d323e1..0b57f70 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-ssse3-ld64.c
@@ -181,7 +181,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 a8fb8da..734dca9 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld128.c
@@ -97,7 +97,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -105,7 +105,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -113,7 +113,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -169,7 +169,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 7cdc634..26b8ceb 100644
--- a/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-xop-ld64.c
@@ -97,7 +97,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -105,7 +105,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -113,7 +113,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -169,7 +169,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 c9be575..22ad639 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld128.c
@@ -120,7 +120,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 0568e12..2fd9087 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-avx-ld64.c
@@ -122,7 +122,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 6711743..6cc6702 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld128.c
@@ -121,7 +121,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 f16ea89..d62f339 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse2-ld64.c
@@ -123,7 +123,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 4bd1f41..c68b7cc 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld128.c
@@ -120,7 +120,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 2a31258..23fd0eb 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-sse41-ld64.c
@@ -122,7 +122,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 6f85867..a53437f 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld128.c
@@ -121,7 +121,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 db23866..e61ae95 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld64.c
@@ -123,7 +123,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 3fdbf3a..f4aa440 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld128.c
@@ -125,7 +125,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 6e56155..8226e3b 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-xop-ld64.c
@@ -127,7 +127,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 17b8f27..74f69f7 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld128.c
@@ -136,7 +136,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 69009c7..ca23fea 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-avx-ld64.c
@@ -138,7 +138,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 ec61f42..1d309a0 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld128.c
@@ -153,7 +153,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 50c2d49..66beb10 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse2-ld64.c
@@ -155,7 +155,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 d406f4e..4cfa48f 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld128.c
@@ -136,7 +136,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 e2b9bef..43a3161 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-sse41-ld64.c
@@ -138,7 +138,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 faaf3f5..c8ce0c8 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld128.c
@@ -153,7 +153,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 b8bd4c5..2601753 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-ssse3-ld64.c
@@ -155,7 +155,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 c8d1f8e..e05b5f6 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld128.c
@@ -141,7 +141,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 0d1cb95..9922c8b 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-xop-ld64.c
@@ -143,7 +143,7 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 6d12bd6..9a2b742 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld128.c
@@ -115,7 +115,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -125,7 +125,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -135,7 +135,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -184,8 +184,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 e89696a..a2800d9 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-avx-ld64.c
@@ -115,7 +115,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -125,7 +125,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -135,7 +135,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -184,8 +184,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 b7f7f8d..3de5bc2 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld128.c
@@ -185,8 +185,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 398ee81..666d22b 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse2-ld64.c
@@ -185,8 +185,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 791ba21..dd270a9 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld128.c
@@ -115,7 +115,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -125,7 +125,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -135,7 +135,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -184,8 +184,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 1b839e6..56b1069 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-sse41-ld64.c
@@ -115,7 +115,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -125,7 +125,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -135,7 +135,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -184,8 +184,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 d80dda7..47ec5ca 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld128.c
@@ -120,7 +120,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -130,7 +130,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -140,7 +140,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -189,8 +189,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 f4f8b9a..24fa887 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-fp32-xop-ld64.c
@@ -120,7 +120,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -130,7 +130,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -140,7 +140,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -189,8 +189,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 81f4c32..3f24371 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld128.c
@@ -115,7 +115,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -125,7 +125,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -135,7 +135,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -207,8 +207,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 9d164eb..cd7dd70 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-avx-ld64.c
@@ -115,7 +115,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -125,7 +125,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -135,7 +135,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -207,8 +207,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 39aea8e..f4f6283 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld128.c
@@ -232,8 +232,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 128ac34..5e5e497 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse2-ld64.c
@@ -232,8 +232,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 9245e7e..abd2c8e 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld128.c
@@ -115,7 +115,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -125,7 +125,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -135,7 +135,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -207,8 +207,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 f97b73c..65cf168 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-sse41-ld64.c
@@ -115,7 +115,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -125,7 +125,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -135,7 +135,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -207,8 +207,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 d3c9714..f744b3a 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld128.c
@@ -232,8 +232,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 bf47976..7e9f2a7 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-ssse3-ld64.c
@@ -232,8 +232,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 aefa9c1..6291abf 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld128.c
@@ -120,7 +120,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -130,7 +130,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -140,7 +140,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -212,8 +212,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 f88b0f1..80929e3 100644
--- a/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/2x4c2-minmax-gemmlowp-xop-ld64.c
@@ -120,7 +120,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -130,7 +130,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -140,7 +140,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -212,8 +212,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 aa4745d..6a24f81 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld128.c
@@ -149,8 +149,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 26f50da..4ca40e1 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-avx-ld64.c
@@ -151,8 +151,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 e46e6ed..17d7560 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld128.c
@@ -150,8 +150,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 82e5115..49e69bf 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c
@@ -152,8 +152,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 0cacb43..5f0e258 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld128.c
@@ -149,8 +149,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 69e9ef8..e6c77fd 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse41-ld64.c
@@ -151,8 +151,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 dc0a620..59090cb 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld128.c
@@ -150,8 +150,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 e8e086e..f6d6210 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-ssse3-ld64.c
@@ -152,8 +152,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 ada4f99..187b8bd 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld128.c
@@ -154,8 +154,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 0f3b693..c82921f 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-xop-ld64.c
@@ -156,8 +156,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 987dcb0..d6477d3 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld128.c
@@ -172,8 +172,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 3206354..337df9a 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-avx-ld64.c
@@ -174,8 +174,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 4f88e7c..f24fc1a 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld128.c
@@ -197,8 +197,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 29f6bb9..04fec0c 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse2-ld64.c
@@ -199,8 +199,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 fc1ce4d..a999775 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld128.c
@@ -172,8 +172,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 696aea2..b30c7eb 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-sse41-ld64.c
@@ -174,8 +174,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 ce4dbf1..630f5c0 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld128.c
@@ -197,8 +197,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 e7c1e34..2ccb6b0 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-ssse3-ld64.c
@@ -199,8 +199,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 c690081..14be109 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld128.c
@@ -177,8 +177,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 f69db77..e2e3ed8 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-xop-ld64.c
@@ -179,8 +179,8 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 39e1fc4..304a4f7 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld128.c
@@ -138,7 +138,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -150,7 +150,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -162,7 +162,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -221,9 +221,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 d939ab6..c1cec78 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-avx-ld64.c
@@ -138,7 +138,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -150,7 +150,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -162,7 +162,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -221,9 +221,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 a2ecb0f..59ba39d 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld128.c
@@ -223,9 +223,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 4642e48..83ee714 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse2-ld64.c
@@ -223,9 +223,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 020f028..210f54a 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld128.c
@@ -138,7 +138,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -150,7 +150,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -162,7 +162,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -221,9 +221,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 9e3ff00..fbac5d1 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-sse41-ld64.c
@@ -138,7 +138,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -150,7 +150,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -162,7 +162,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -221,9 +221,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 d14a6c9..6190ade 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld128.c
@@ -143,7 +143,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -155,7 +155,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -167,7 +167,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -226,9 +226,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 a2b6c23..31575f4 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-fp32-xop-ld64.c
@@ -143,7 +143,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -155,7 +155,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -167,7 +167,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -226,9 +226,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 abed9b6..fef31cf 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld128.c
@@ -138,7 +138,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -150,7 +150,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -162,7 +162,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -251,9 +251,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 37cc36d..bf89939 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-avx-ld64.c
@@ -138,7 +138,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -150,7 +150,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -162,7 +162,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -251,9 +251,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 3836866..35b26a3 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld128.c
@@ -285,9 +285,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 5d1886d..7d345ea 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse2-ld64.c
@@ -285,9 +285,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 1ffbff8..fc27bef 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld128.c
@@ -138,7 +138,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -150,7 +150,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -162,7 +162,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -251,9 +251,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 84d280b..975ce10 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-sse41-ld64.c
@@ -138,7 +138,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -150,7 +150,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -162,7 +162,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -251,9 +251,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 92e9447..8074fcb 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld128.c
@@ -285,9 +285,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 75700c3..ff2e060 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-ssse3-ld64.c
@@ -285,9 +285,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 6356e76..91e7c30 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld128.c
@@ -143,7 +143,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -155,7 +155,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -167,7 +167,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -256,9 +256,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 4912f55..983aa01 100644
--- a/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/3x4c2-minmax-gemmlowp-xop-ld64.c
@@ -143,7 +143,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -155,7 +155,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -167,7 +167,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -256,9 +256,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 f198a7d..b4a8597 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld128.c
@@ -179,9 +179,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 fb028a4..9cc0420 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-avx-ld64.c
@@ -181,9 +181,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 8dc7fdb..b1095e5 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld128.c
@@ -181,9 +181,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 079ee69..4dfa033 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse2-ld64.c
@@ -183,9 +183,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 d262753..5a51a7c 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
@@ -179,9 +179,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 a3f8f88..303b7f8 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld64.c
@@ -181,9 +181,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 8ffc6fc..118cdd9 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld128.c
@@ -181,9 +181,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 6c712bf..e9bf06c 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-ssse3-ld64.c
@@ -183,9 +183,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 55639ef..1a4e7db 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld128.c
@@ -184,9 +184,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 19f5904..1314632 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-xop-ld64.c
@@ -186,9 +186,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 a13dd17..e3cf40a 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld128.c
@@ -209,9 +209,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 fbad091..dc1751e 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-avx-ld64.c
@@ -211,9 +211,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 e4200cd..8b65958 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld128.c
@@ -243,9 +243,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 a5f3732..b26e36d 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse2-ld64.c
@@ -245,9 +245,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 15de5b5..6627108 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld128.c
@@ -209,9 +209,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 610c200..7217bd9 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-sse41-ld64.c
@@ -211,9 +211,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 b5919d7..0adb317 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld128.c
@@ -243,9 +243,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 3f6204f..00811cf 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-ssse3-ld64.c
@@ -245,9 +245,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 a866141..2c25542 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld128.c
@@ -214,9 +214,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 c297cc2..f2f08b8 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-xop-ld64.c
@@ -216,9 +216,9 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 9575adb..96dbae3 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld128.c
@@ -161,7 +161,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -175,7 +175,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -189,7 +189,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -257,10 +257,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 86dbc98..e0fd57e 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-avx-ld64.c
@@ -161,7 +161,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -175,7 +175,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -189,7 +189,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -257,10 +257,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 31c84c2..06be3e4 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld128.c
@@ -259,10 +259,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi16(vout, 6);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c3 = (int8_t) _mm_extract_epi16(vout, 6);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 0d6a7da..958b7a9 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse2-ld64.c
@@ -259,10 +259,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi16(vout, 6);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c3 = (int8_t) _mm_extract_epi16(vout, 6);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 112e243..06d0d8a 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld128.c
@@ -161,7 +161,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -175,7 +175,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -189,7 +189,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -257,10 +257,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 c8a3cca..ae535c4 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-sse41-ld64.c
@@ -161,7 +161,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -175,7 +175,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -189,7 +189,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -257,10 +257,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 efa3ada..ab8e9bb 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld128.c
@@ -166,7 +166,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -180,7 +180,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -194,7 +194,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -262,10 +262,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 9f9ad98..b1f20e1 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-fp32-xop-ld64.c
@@ -166,7 +166,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -180,7 +180,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -194,7 +194,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -262,10 +262,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 681ea5f..a5f5764 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld128.c
@@ -161,7 +161,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -175,7 +175,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -189,7 +189,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -294,10 +294,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 7e66cc1..9a4da26 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-avx-ld64.c
@@ -161,7 +161,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -175,7 +175,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -189,7 +189,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -294,10 +294,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 5875e0e..0c0d4d9 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld128.c
@@ -336,10 +336,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi16(vout, 6);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c3 = (int8_t) _mm_extract_epi16(vout, 6);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 89ba44a..b586305 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse2-ld64.c
@@ -336,10 +336,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi16(vout, 6);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c3 = (int8_t) _mm_extract_epi16(vout, 6);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 64e1ff5..3d24064 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld128.c
@@ -161,7 +161,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -175,7 +175,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -189,7 +189,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -294,10 +294,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 8571512..0585add 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-sse41-ld64.c
@@ -161,7 +161,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_add_epi32(vacc0x0123,
_mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
@@ -175,7 +175,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 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));
@@ -189,7 +189,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ 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));
@@ -294,10 +294,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 35d07a0..6f78980 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld128.c
@@ -336,10 +336,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi16(vout, 6);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c3 = (int8_t) _mm_extract_epi16(vout, 6);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 5d141e0..341314f 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-ssse3-ld64.c
@@ -336,10 +336,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi16(vout, 6);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
- *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *c3 = (int8_t) _mm_extract_epi16(vout, 6);
+ *c2 = (int8_t) _mm_extract_epi16(vout, 4);
+ *c1 = (int8_t) _mm_extract_epi16(vout, 2);
+ *c0 = (int8_t) _mm_cvtsi128_si32(vout);
}
nc = 0;
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 49717e2..1d28198 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c
@@ -166,7 +166,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -180,7 +180,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -194,7 +194,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -299,10 +299,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;
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 d4b022a..61e46f2 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld64.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld64.c
@@ -166,7 +166,7 @@
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
@@ -180,7 +180,7 @@
if (k > 2 * sizeof(int8_t)) {
const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
@@ -194,7 +194,7 @@
if (k > 4 * sizeof(int8_t)) {
const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
w = (const void*) ((const int8_t*) w + 8);
- const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
vacc0x0123 = _mm_maddd_epi16(
_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
@@ -299,10 +299,10 @@
vout = _mm_srli_epi32(vout, 16);
}
if (nc & 1) {
- *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
- *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
- *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
- *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *c3 = (int8_t) _mm_extract_epi8(vout, 12);
+ *c2 = (int8_t) _mm_extract_epi8(vout, 8);
+ *c1 = (int8_t) _mm_extract_epi8(vout, 4);
+ *c0 = (int8_t) _mm_extract_epi8(vout, 0);
}
nc = 0;