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))) {