Optimize QS8 VADD[C] microkernels for SSE4/AVX/XOP/AVX2
Use microkernel-specific parameters to remove some operations
PiperOrigin-RevId: 385914437
diff --git a/src/qs8-vaddc/avx2-mul32-ld64.c.in b/src/qs8-vaddc/avx2-mul32-ld64.c.in
index 69d7d34..736a619 100644
--- a/src/qs8-vaddc/avx2-mul32-ld64.c.in
+++ b/src/qs8-vaddc/avx2-mul32-ld64.c.in
@@ -21,21 +21,19 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m256i va_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.a_multiplier));
- const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding));
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
+ const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier);
+ const __m256i vrounding = _mm256_load_si256((const __m256i*) params->avx2.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->avx2.shift);
$if BATCH_TILE > 8:
- const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
- const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
- const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
+ const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
$else:
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx2.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max);
- __m256i vbias = _mm256_broadcastsi128_si256(_mm_add_epi32(
- _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b)),
- _mm_load_si128((const __m128i*) params->sse2.bias)));
+ __m256i vbias = _mm256_add_epi32(
+ _mm256_broadcastd_epi32(_mm_cvtsi32_si128(params->avx2.b_multiplier[0] * (int32_t) *input_b)),
+ _mm256_load_si256((const __m256i*) params->avx2.bias));
for (; n >= ${BATCH_TILE} * sizeof(int8_t); n -= ${BATCH_TILE} * sizeof(int8_t)) {
const __m256i va${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
$for N in range(8, BATCH_TILE, 8):
@@ -58,18 +56,22 @@
$for N in range(0, BATCH_TILE, 16):
$if N + 8 < BATCH_TILE:
- vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]} = _mm256_min_epi16(_mm256_max_epi16(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}, voutput_min), voutput_max);
- $elif BATCH_TILE > 8:
- vout${ABC[N:N+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[N:N+8]}, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
- $else:
- vout${ABC[N:N+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[N:N+8]}, voutput_min), voutput_max);
-
- $for N in range(0, BATCH_TILE, 16):
- $if N + 8 < BATCH_TILE:
__m128i vout${ABC[N:N+16]} = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}), _mm256_extracti128_si256(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}, 1)), _MM_SHUFFLE(3, 1, 2, 0));
$else:
__m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});
+ $for N in range(0, BATCH_TILE, 16):
+ $if N + 8 < BATCH_TILE:
+ vout${ABC[N:N+16]} = _mm_max_epi8(vout${ABC[N:N+16]}, voutput_min);
+ $else:
+ vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_max_epi8(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_min);
+
+ $for N in range(0, BATCH_TILE, 16):
+ $if N + 8 < BATCH_TILE:
+ vout${ABC[N:N+16]} = _mm_min_epi8(vout${ABC[N:N+16]}, voutput_max);
+ $else:
+ vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_min_epi8(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_max);
+
$if BATCH_TILE >= 16:
_mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
$else:
@@ -93,11 +95,11 @@
$if BATCH_TILE > 8:
__m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), _mm256_castsi256_si128(voutput_zero_point));
- vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
$else:
__m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), voutput_zero_point);
- vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, voutput_min), voutput_max);
__m128i vout${ABC[0:8]}${ABC[0:8]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[0:8]});
+ vout${ABC[0:8]}${ABC[0:8]} = _mm_max_epi8(vout${ABC[0:8]}${ABC[0:8]}, voutput_min);
+ vout${ABC[0:8]}${ABC[0:8]} = _mm_min_epi8(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);
$if BATCH_TILE > 8:
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
diff --git a/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x16.c b/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x16.c
index 4d619dd..14aefc5 100644
--- a/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x16.c
+++ b/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x16.c
@@ -21,17 +21,17 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
- const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
+ const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
+ const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul16.rounding);
+ const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
@@ -62,13 +62,12 @@
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
output += 16;
@@ -93,10 +92,10 @@
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x24.c b/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x24.c
index d9a6211..01b0ba7 100644
--- a/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x24.c
+++ b/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x24.c
@@ -21,17 +21,17 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
- const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
+ const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
+ const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul16.rounding);
+ const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
@@ -72,16 +72,15 @@
__m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
__m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
+
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
@@ -107,10 +106,10 @@
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x32.c b/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x32.c
index 8755b62..a04d2c0 100644
--- a/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x32.c
+++ b/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x32.c
@@ -21,17 +21,17 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
- const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
+ const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
+ const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul16.rounding);
+ const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
@@ -82,18 +82,15 @@
__m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
__m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
- voutOPQRSTUV = _mm_max_epi16(voutOPQRSTUV, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
- voutOPQRSTUV = _mm_min_epi16(voutOPQRSTUV, voutput_max);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
+
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
@@ -119,10 +116,10 @@
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x8.c b/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x8.c
index 6ff8a0e..2fbbc2d 100644
--- a/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x8.c
+++ b/src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x8.c
@@ -21,17 +21,17 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
- const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
+ const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
+ const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul16.rounding);
+ const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
input_a += 8;
@@ -52,11 +52,12 @@
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
- const __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
output += 8;
@@ -80,10 +81,10 @@
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if (n & (4 * sizeof(int8_t))) {
*((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x16.c b/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x16.c
index 514d9e5..4a9784e 100644
--- a/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x16.c
+++ b/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x16.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -50,16 +50,14 @@
vacc89AB = _mm_sra_epi32(_mm_add_epi32(vacc89AB, vrounding), vshift);
vaccCDEF = _mm_sra_epi32(_mm_add_epi32(vaccCDEF, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
output += 16;
@@ -76,11 +74,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x24.c b/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x24.c
index 55b3e67..7df2578 100644
--- a/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x24.c
+++ b/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x24.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -56,20 +56,18 @@
vaccGHIJ = _mm_sra_epi32(_mm_add_epi32(vaccGHIJ, vrounding), vshift);
vaccKLMN = _mm_sra_epi32(_mm_add_epi32(vaccKLMN, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
+ const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
@@ -87,11 +85,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x32.c b/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x32.c
index b507e4f..72954c0 100644
--- a/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x32.c
+++ b/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x32.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -62,23 +62,19 @@
vaccOPQR = _mm_sra_epi32(_mm_add_epi32(vaccOPQR, vrounding), vshift);
vaccSTUV = _mm_sra_epi32(_mm_add_epi32(vaccSTUV, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
- __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
+ const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
+ const __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
- voutOPQRSTUV = _mm_max_epi16(voutOPQRSTUV, voutput_min);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
- voutOPQRSTUV = _mm_min_epi16(voutOPQRSTUV, voutput_max);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
@@ -96,11 +92,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x8.c b/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x8.c
index a544bc3..4b79e1c 100644
--- a/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x8.c
+++ b/src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x8.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -44,13 +44,13 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
+ __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
- const __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
output += 8;
@@ -66,11 +66,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if (n & (4 * sizeof(int8_t))) {
*((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c
index 21bc59f..3c00442 100644
--- a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c
+++ b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m256i va_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.a_multiplier));
- const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding));
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
- const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
- const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
+ const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier);
+ const __m256i vrounding = _mm256_load_si256((const __m256i*) params->avx2.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->avx2.shift);
+ const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max);
- __m256i vbias = _mm256_broadcastsi128_si256(_mm_add_epi32(
- _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b)),
- _mm_load_si128((const __m128i*) params->sse2.bias)));
+ __m256i vbias = _mm256_add_epi32(
+ _mm256_broadcastd_epi32(_mm_cvtsi32_si128(params->avx2.b_multiplier[0] * (int32_t) *input_b)),
+ _mm256_load_si256((const __m256i*) params->avx2.bias));
for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
const __m256i va89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
@@ -45,10 +45,12 @@
__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
- vout012389AB4567CDEF = _mm256_min_epi16(_mm256_max_epi16(vout012389AB4567CDEF, voutput_min), voutput_max);
-
__m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
output += 16;
}
@@ -62,8 +64,9 @@
vacc01234567 = _mm256_sra_epi32(_mm256_add_epi32(vacc01234567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
- vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c
index 3abdfca..68e236a 100644
--- a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c
+++ b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m256i va_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.a_multiplier));
- const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding));
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
- const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
- const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
+ const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier);
+ const __m256i vrounding = _mm256_load_si256((const __m256i*) params->avx2.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->avx2.shift);
+ const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max);
- __m256i vbias = _mm256_broadcastsi128_si256(_mm_add_epi32(
- _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b)),
- _mm_load_si128((const __m128i*) params->sse2.bias)));
+ __m256i vbias = _mm256_add_epi32(
+ _mm256_broadcastd_epi32(_mm_cvtsi32_si128(params->avx2.b_multiplier[0] * (int32_t) *input_b)),
+ _mm256_load_si256((const __m256i*) params->avx2.bias));
for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
const __m256i va89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
@@ -49,12 +49,15 @@
__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
__m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vaccGHIJKLMN), _mm256_extracti128_si256(vaccGHIJKLMN, 1)), _mm256_castsi256_si128(voutput_zero_point));
- vout012389AB4567CDEF = _mm256_min_epi16(_mm256_max_epi16(vout012389AB4567CDEF, voutput_min), voutput_max);
- voutGHIJKLMN = _mm_min_epi16(_mm_max_epi16(voutGHIJKLMN, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
-
__m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
__m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
+
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
+
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
output += 24;
@@ -69,8 +72,9 @@
vacc01234567 = _mm256_sra_epi32(_mm256_add_epi32(vacc01234567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
- vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c
index 64c6933..fbbffe5 100644
--- a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c
+++ b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m256i va_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.a_multiplier));
- const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding));
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
- const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
- const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
+ const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier);
+ const __m256i vrounding = _mm256_load_si256((const __m256i*) params->avx2.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->avx2.shift);
+ const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max);
- __m256i vbias = _mm256_broadcastsi128_si256(_mm_add_epi32(
- _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b)),
- _mm_load_si128((const __m128i*) params->sse2.bias)));
+ __m256i vbias = _mm256_add_epi32(
+ _mm256_broadcastd_epi32(_mm_cvtsi32_si128(params->avx2.b_multiplier[0] * (int32_t) *input_b)),
+ _mm256_load_si256((const __m256i*) params->avx2.bias));
for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
const __m256i va89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
@@ -52,12 +52,15 @@
__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
__m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
- vout012389AB4567CDEF = _mm256_min_epi16(_mm256_max_epi16(vout012389AB4567CDEF, voutput_min), voutput_max);
- voutGHIJOPQRKLMNSTUV = _mm256_min_epi16(_mm256_max_epi16(voutGHIJOPQRKLMNSTUV, voutput_min), voutput_max);
-
__m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
__m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
+
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
+
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
output += 32;
@@ -72,8 +75,9 @@
vacc01234567 = _mm256_sra_epi32(_mm256_add_epi32(vacc01234567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
- vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c
index ea00cc7..46782a6 100644
--- a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c
+++ b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m256i va_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.a_multiplier));
- const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding));
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier);
+ const __m256i vrounding = _mm256_load_si256((const __m256i*) params->avx2.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->avx2.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx2.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max);
- __m256i vbias = _mm256_broadcastsi128_si256(_mm_add_epi32(
- _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b)),
- _mm_load_si128((const __m128i*) params->sse2.bias)));
+ __m256i vbias = _mm256_add_epi32(
+ _mm256_broadcastd_epi32(_mm_cvtsi32_si128(params->avx2.b_multiplier[0] * (int32_t) *input_b)),
+ _mm256_load_si256((const __m256i*) params->avx2.bias));
for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
input_a += 8;
@@ -42,10 +42,12 @@
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
- vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
-
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
+
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
output += 8;
}
@@ -58,8 +60,9 @@
vacc01234567 = _mm256_sra_epi32(_mm256_add_epi32(vacc01234567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
- vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if (n & (4 * sizeof(int8_t))) {
*((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c b/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c
index 283dbbd..927e3ca 100644
--- a/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c
+++ b/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c
@@ -21,6 +21,9 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse2.bias));
const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
@@ -29,9 +32,6 @@
const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
__m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
__m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
@@ -70,7 +70,8 @@
vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
output += 16;
diff --git a/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c b/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c
index 388f01d..00d833c 100644
--- a/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c
+++ b/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c
@@ -21,6 +21,9 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse2.bias));
const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
@@ -29,9 +32,6 @@
const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
__m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
__m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
@@ -83,8 +83,9 @@
vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
+
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
diff --git a/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c b/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c
index b17aa07..41dd72b 100644
--- a/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c
+++ b/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c
@@ -21,6 +21,9 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse2.bias));
const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
@@ -29,9 +32,6 @@
const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
__m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
__m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
@@ -96,8 +96,9 @@
voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
voutOPQRSTUV = _mm_min_epi16(voutOPQRSTUV, voutput_max);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
+
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
diff --git a/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c b/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c
index 28bcd07..36b0671 100644
--- a/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c
+++ b/src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c
@@ -21,6 +21,9 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse2.bias));
const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
@@ -29,9 +32,6 @@
const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
__m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
input_a += 8;
@@ -57,7 +57,8 @@
vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- const __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
output += 8;
diff --git a/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c b/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c
index 38f12ba..97e9333 100644
--- a/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c
+++ b/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c
@@ -21,17 +21,17 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
- const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
+ const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
+ const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul16.rounding);
+ const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
@@ -62,13 +62,12 @@
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
output += 16;
@@ -93,10 +92,10 @@
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c b/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c
index 082650f..e4e1548 100644
--- a/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c
+++ b/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c
@@ -21,17 +21,17 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
- const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
+ const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
+ const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul16.rounding);
+ const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
@@ -72,16 +72,15 @@
__m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
__m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
+
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
@@ -107,10 +106,10 @@
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c b/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c
index ff6c6ed..00619ea 100644
--- a/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c
+++ b/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c
@@ -21,17 +21,17 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
- const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
+ const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
+ const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul16.rounding);
+ const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
@@ -82,18 +82,15 @@
__m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
__m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
- voutOPQRSTUV = _mm_max_epi16(voutOPQRSTUV, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
- voutOPQRSTUV = _mm_min_epi16(voutOPQRSTUV, voutput_max);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
+
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
@@ -119,10 +116,10 @@
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c b/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c
index b4613b2..85ab3f0 100644
--- a/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c
+++ b/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c
@@ -21,17 +21,17 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
- const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
+ const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
+ const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul16.rounding);
+ const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
input_a += 8;
@@ -52,11 +52,12 @@
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
- const __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
output += 8;
@@ -80,10 +81,10 @@
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if (n & (4 * sizeof(int8_t))) {
*((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c b/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c
index 12812c9..ab1ccfa 100644
--- a/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c
+++ b/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -50,16 +50,14 @@
vacc89AB = _mm_sra_epi32(_mm_add_epi32(vacc89AB, vrounding), vshift);
vaccCDEF = _mm_sra_epi32(_mm_add_epi32(vaccCDEF, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
output += 16;
@@ -76,11 +74,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c b/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c
index 2dfbeae..2dae9ff 100644
--- a/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c
+++ b/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -56,20 +56,18 @@
vaccGHIJ = _mm_sra_epi32(_mm_add_epi32(vaccGHIJ, vrounding), vshift);
vaccKLMN = _mm_sra_epi32(_mm_add_epi32(vaccKLMN, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
+ const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
@@ -87,11 +85,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c b/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c
index 55eb457..635700b 100644
--- a/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c
+++ b/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -62,23 +62,19 @@
vaccOPQR = _mm_sra_epi32(_mm_add_epi32(vaccOPQR, vrounding), vshift);
vaccSTUV = _mm_sra_epi32(_mm_add_epi32(vaccSTUV, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
- __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
+ const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
+ const __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
- voutOPQRSTUV = _mm_max_epi16(voutOPQRSTUV, voutput_min);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
- voutOPQRSTUV = _mm_min_epi16(voutOPQRSTUV, voutput_max);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
@@ -96,11 +92,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c b/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c
index bb1d54e..7532fea 100644
--- a/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c
+++ b/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c
@@ -22,16 +22,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -44,13 +44,13 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
+ __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
- const __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
output += 8;
@@ -66,11 +66,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if (n & (4 * sizeof(int8_t))) {
*((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c b/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c
index 5ffd432..698085b 100644
--- a/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c
+++ b/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c
@@ -27,16 +27,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -55,16 +55,14 @@
vacc89AB = _mm_sra_epi32(_mm_add_epi32(vacc89AB, vrounding), vshift);
vaccCDEF = _mm_sra_epi32(_mm_add_epi32(vaccCDEF, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
output += 16;
@@ -81,11 +79,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c b/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c
index a688ddf..8beb90a 100644
--- a/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c
+++ b/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c
@@ -27,16 +27,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -61,20 +61,18 @@
vaccGHIJ = _mm_sra_epi32(_mm_add_epi32(vaccGHIJ, vrounding), vshift);
vaccKLMN = _mm_sra_epi32(_mm_add_epi32(vaccKLMN, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
+ const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
@@ -92,11 +90,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c b/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c
index b5ccf0a..2b48ee6 100644
--- a/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c
+++ b/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c
@@ -27,16 +27,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -67,23 +67,19 @@
vaccOPQR = _mm_sra_epi32(_mm_add_epi32(vaccOPQR, vrounding), vshift);
vaccSTUV = _mm_sra_epi32(_mm_add_epi32(vaccSTUV, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
- __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
- __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
+ const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
+ const __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
- voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
- voutOPQRSTUV = _mm_max_epi16(voutOPQRSTUV, voutput_min);
+ __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
- vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
- voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
- voutOPQRSTUV = _mm_min_epi16(voutOPQRSTUV, voutput_max);
+ vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+ voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
- const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
- const __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
+ vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+ voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
_mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
@@ -101,11 +97,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
diff --git a/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c b/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c
index 3d83113..ce4ab97 100644
--- a/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c
+++ b/src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c
@@ -27,16 +27,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -49,13 +49,13 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
+ __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
- const __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
_mm_storel_epi64((__m128i*) output, vout0123456701234567);
output += 8;
@@ -71,11 +71,11 @@
vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
- __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
- vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
- vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+ const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
__m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+ vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+ vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
if (n & (4 * sizeof(int8_t))) {
*((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
diff --git a/src/qs8-vaddc/sse-mul16-ld64.c.in b/src/qs8-vaddc/sse-mul16-ld64.c.in
index 6915333..be7a57e 100644
--- a/src/qs8-vaddc/sse-mul16-ld64.c.in
+++ b/src/qs8-vaddc/sse-mul16-ld64.c.in
@@ -16,6 +16,7 @@
#include <xnnpack/vadd.h>
+$PARAMS_STRUCT = "sse4_mul16" if SSE == 4 else "sse2"
$ISA = "avx" if AVX else {2: "sse2", 4: "sse41"}[SSE]
void xnn_qs8_vaddc_minmax_ukernel__${ISA}_mul16_ld64_x${BATCH_TILE}(
size_t n,
@@ -24,17 +25,17 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
- const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i vbias = _mm_add_epi32(
+ _mm_shuffle_epi32(_mm_cvtsi32_si128(params->${PARAMS_STRUCT}.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
+ _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.bias));
+ const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.a_multiplier_lo);
+ const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.a_multiplier_hi);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.rounding);
+ const __m128i vshift = _mm_cvtsi32_si128((int) params->${PARAMS_STRUCT}.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_zero_point);
+ 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);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
- vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
for (; n >= ${BATCH_TILE} * sizeof(int8_t); n -= ${BATCH_TILE} * sizeof(int8_t)) {
$if SSE == 4:
const __m128i va${ABC[0:8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
@@ -70,17 +71,31 @@
$for N in range(0, BATCH_TILE, 8):
__m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(vacc${ABC[N:N+4]}, vacc${ABC[N+4:N+8]}), voutput_zero_point);
- $for N in range(0, BATCH_TILE, 8):
- vout${ABC[N:N+8]} = _mm_max_epi16(vout${ABC[N:N+8]}, voutput_min);
+ $if SSE < 4:
+ $for N in range(0, BATCH_TILE, 8):
+ vout${ABC[N:N+8]} = _mm_max_epi16(vout${ABC[N:N+8]}, voutput_min);
- $for N in range(0, BATCH_TILE, 8):
- vout${ABC[N:N+8]} = _mm_min_epi16(vout${ABC[N:N+8]}, voutput_max);
+ $for N in range(0, BATCH_TILE, 8):
+ vout${ABC[N:N+8]} = _mm_min_epi16(vout${ABC[N:N+8]}, voutput_max);
$for N in range(0, BATCH_TILE, 16):
$if N + 8 < BATCH_TILE:
- const __m128i vout${ABC[N:N+16]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N+8:N+16]});
+ __m128i vout${ABC[N:N+16]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N+8:N+16]});
$else:
- const __m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});
+ __m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});
+
+ $if SSE == 4:
+ $for N in range(0, BATCH_TILE, 16):
+ $if N + 8 < BATCH_TILE:
+ vout${ABC[N:N+16]} = _mm_max_epi8(vout${ABC[N:N+16]}, voutput_min);
+ $else:
+ vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_max_epi8(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_min);
+
+ $for N in range(0, BATCH_TILE, 16):
+ $if N + 8 < BATCH_TILE:
+ vout${ABC[N:N+16]} = _mm_min_epi8(vout${ABC[N:N+16]}, voutput_max);
+ $else:
+ vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_min_epi8(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_max);
$if BATCH_TILE >= 16:
_mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
@@ -119,10 +134,14 @@
vacc${ABC[4:8]} = _mm_sra_epi32(_mm_add_epi32(vacc${ABC[4:8]}, vrounding), vshift);
__m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(vacc${ABC[0:4]}, vacc${ABC[4:8]}), voutput_zero_point);
- vout${ABC[0:8]} = _mm_max_epi16(vout${ABC[0:8]}, voutput_min);
- vout${ABC[0:8]} = _mm_min_epi16(vout${ABC[0:8]}, voutput_max);
+ $if SSE < 4:
+ vout${ABC[0:8]} = _mm_max_epi16(vout${ABC[0:8]}, voutput_min);
+ vout${ABC[0:8]} = _mm_min_epi16(vout${ABC[0:8]}, voutput_max);
__m128i vout${ABC[0:8]}${ABC[0:8]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[0:8]});
+ $if SSE == 4:
+ vout${ABC[0:8]}${ABC[0:8]} = _mm_max_epi8(vout${ABC[0:8]}${ABC[0:8]}, voutput_min);
+ vout${ABC[0:8]}${ABC[0:8]} = _mm_min_epi8(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);
$if BATCH_TILE > 8:
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
diff --git a/src/qs8-vaddc/sse-mul32-ld32.c.in b/src/qs8-vaddc/sse-mul32-ld32.c.in
index 570178a..bf1dadb 100644
--- a/src/qs8-vaddc/sse-mul32-ld32.c.in
+++ b/src/qs8-vaddc/sse-mul32-ld32.c.in
@@ -32,16 +32,16 @@
int8_t* output,
const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
- const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
- const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
- const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
- const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
- const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
- const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+ const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+ const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
- __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+ __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
- vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
+ vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
for (; n >= ${BATCH_TILE} * sizeof(int8_t); n -= ${BATCH_TILE} * sizeof(int8_t)) {
const __m128i va${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
$for N in range(4, BATCH_TILE, 4):
@@ -60,19 +60,25 @@
vacc${ABC[N:N+4]} = _mm_sra_epi32(_mm_add_epi32(vacc${ABC[N:N+4]}, vrounding), vshift);
$for N in range(0, BATCH_TILE, 8):
- __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(vacc${ABC[N:N+4]}, vacc${ABC[N+4:N+8]}), voutput_zero_point);
-
- $for N in range(0, BATCH_TILE, 8):
- vout${ABC[N:N+8]} = _mm_max_epi16(vout${ABC[N:N+8]}, voutput_min);
-
- $for N in range(0, BATCH_TILE, 8):
- vout${ABC[N:N+8]} = _mm_min_epi16(vout${ABC[N:N+8]}, voutput_max);
+ const __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(vacc${ABC[N:N+4]}, vacc${ABC[N+4:N+8]}), voutput_zero_point);
$for N in range(0, BATCH_TILE, 16):
$if N + 8 < BATCH_TILE:
- const __m128i vout${ABC[N:N+16]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N+8:N+16]});
+ __m128i vout${ABC[N:N+16]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N+8:N+16]});
$else:
- const __m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});
+ __m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});
+
+ $for N in range(0, BATCH_TILE, 16):
+ $if N + 8 < BATCH_TILE:
+ vout${ABC[N:N+16]} = _mm_max_epi8(vout${ABC[N:N+16]}, voutput_min);
+ $else:
+ vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_max_epi8(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_min);
+
+ $for N in range(0, BATCH_TILE, 16):
+ $if N + 8 < BATCH_TILE:
+ vout${ABC[N:N+16]} = _mm_min_epi8(vout${ABC[N:N+16]}, voutput_max);
+ $else:
+ vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_min_epi8(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_max);
$if BATCH_TILE >= 16:
_mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
@@ -102,11 +108,11 @@
vacc${ABC[0:4]} = _mm_sra_epi32(_mm_add_epi32(vacc${ABC[0:4]}, vrounding), vshift);
vacc${ABC[4:8]} = _mm_sra_epi32(_mm_add_epi32(vacc${ABC[4:8]}, vrounding), vshift);
- __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(vacc${ABC[0:4]}, vacc${ABC[4:8]}), voutput_zero_point);
- vout${ABC[0:8]} = _mm_max_epi16(vout${ABC[0:8]}, voutput_min);
- vout${ABC[0:8]} = _mm_min_epi16(vout${ABC[0:8]}, voutput_max);
+ const __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(vacc${ABC[0:4]}, vacc${ABC[4:8]}), voutput_zero_point);
__m128i vout${ABC[0:8]}${ABC[0:8]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[0:8]});
+ vout${ABC[0:8]}${ABC[0:8]} = _mm_max_epi8(vout${ABC[0:8]}${ABC[0:8]}, voutput_min);
+ vout${ABC[0:8]}${ABC[0:8]} = _mm_min_epi8(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);
$if BATCH_TILE > 8:
if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {