FP32 requantization in QS8 GEMM/IGEMM microkernels for SSE/AVX/XOP
PiperOrigin-RevId: 376966195
diff --git a/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
new file mode 100644
index 0000000..6531f50
--- /dev/null
+++ b/src/qs8-igemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
@@ -0,0 +1,190 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-igemm/MRx4c8-sse.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <smmintrin.h>
+
+#include <xnnpack/igemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld128(
+ size_t mr,
+ size_t nc,
+ size_t kc,
+ size_t ks,
+ const int8_t** restrict a,
+ const void* restrict w,
+ int8_t* restrict c,
+ size_t cm_stride,
+ size_t cn_stride,
+ size_t a_offset,
+ const int8_t* zero,
+ const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ assert(nc != 0);
+ assert(kc != 0);
+ assert(ks != 0);
+ assert(ks % (3 * sizeof(void*)) == 0);
+ assert(a_offset % sizeof(int8_t) == 0);
+ assert(a != NULL);
+ assert(w != NULL);
+ assert(c != NULL);
+
+ kc = round_up_po2(kc, 8);
+ int8_t* c0 = c;
+ int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
+ if XNN_UNPREDICTABLE(mr < 2) {
+ c1 = c0;
+ }
+ int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
+ if XNN_UNPREDICTABLE(mr <= 2) {
+ c2 = c1;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ __m128i vacc2x0 = vacc0x0;
+ __m128i vacc2x1 = vacc0x1;
+ __m128i vacc2x2 = vacc0x2;
+ __m128i vacc2x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t p = ks;
+ do {
+ const int8_t* restrict a0 = a[0];
+ if XNN_UNPREDICTABLE(a0 != zero) {
+ a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
+ }
+ const int8_t* restrict a1 = a[1];
+ if XNN_UNPREDICTABLE(a1 != zero) {
+ a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
+ }
+ const int8_t* restrict a2 = a[2];
+ if XNN_UNPREDICTABLE(a2 != zero) {
+ a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
+ }
+ a += 3;
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ 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 vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+ 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);
+ k += 8 * sizeof(int8_t);
+ }
+ p -= 3 * sizeof(void*);
+ } while (p != 0);
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+ const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
+ const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+ __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+ c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a = (const int8_t**restrict) ((uintptr_t) a - ks);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}