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/3x4c8-minmax-fp32-sse41-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
index 6531f50..d262753 100644
--- a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
@@ -64,7 +64,7 @@
__m128i vacc2x1 = vacc0x1;
__m128i vacc2x2 = vacc0x2;
__m128i vacc2x3 = vacc0x3;
- w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+ w = (const void*) ((const int32_t*) w + 4);
size_t p = ks;
do {
@@ -105,7 +105,7 @@
vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
- const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
@@ -117,7 +117,7 @@
vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
- w = (const void*) ((uintptr_t) w + 32);
+ w = (const void*) ((const int8_t*) w + 32);
k += 8 * sizeof(int8_t);
}
p -= 3 * sizeof(void*);