Refactor pre-SSE4 versions of QS8/QC8 GEMM/IGEMM microkernels
- Replace sign-extension from 8 to 16 bits with a more efficient sequence
- Replace casts to uintptr_t to casts to typed pointer types where possible
PiperOrigin-RevId: 382654408
diff --git a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c
index 8f98520..82e5115 100644
--- a/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c
+++ b/src/qs8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c
@@ -56,7 +56,7 @@
__m128i vacc1x1 = vacc0x1;
__m128i vacc1x2 = vacc0x2;
__m128i vacc1x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -73,34 +73,34 @@
size_t k = 0;
while (k < kc) {
const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
- const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
a0 += 8;
const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
- const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
a1 += 8;
const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
- const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
- const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
- const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
+ const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
- const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
- const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
+ const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
- const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
- const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
+ const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 2 * sizeof(void*);