FP32 requantization in QS8 GEMM/IGEMM microkernels for SSE/AVX/XOP

PiperOrigin-RevId: 376966195
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld64.c b/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld64.c
new file mode 100644
index 0000000..c19d705
--- /dev/null
+++ b/src/qs8-igemm/gen/1x4c8-minmax-fp32-ssse3-ld64.c
@@ -0,0 +1,132 @@
+// 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 <tmmintrin.h>
+
+#include <xnnpack/igemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__ssse3_ld64(
+    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 <= 1);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(ks != 0);
+  assert(ks % (1 * 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;
+
+  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]);
+    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);
+      }
+      a += 1;
+
+      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));
+        a0 += 8;
+
+        const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+        const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+
+        vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, 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));
+
+        vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, 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));
+
+        vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, 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));
+
+        vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+
+        w = (const void*) ((uintptr_t) w + 32);
+        k += 8 * sizeof(int8_t);
+      }
+      p -= 1 * sizeof(void*);
+    } while (p != 0);
+
+    const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+    const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+
+    __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+
+    __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+    const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
+    vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
+
+    vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+    const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
+    __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+    const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
+    const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max);
+    vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
+
+    __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+
+    if (nc >= 4) {
+      *((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*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+        c0 += 2;
+        vout = _mm_srli_epi32(vout, 16);
+      }
+      if (nc & 1) {
+        *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}