xref: /aosp_15_r20/external/XNNPACK/src/f16-gemm/gen/1x8-minmax-avx2-broadcast.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-gemm/avx2-broadcast.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2022 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <immintrin.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_f16_gemm_minmax_ukernel_1x8__avx2_broadcast(size_t mr,size_t nc,size_t kc,const void * restrict a,size_t a_stride,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_gemm_minmax_ukernel_1x8__avx2_broadcast(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const void*restrict a,
23     size_t a_stride,
24     const void*restrict w,
25     void*restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(uint16_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   const uint16_t* a0 = a;
40   uint16_t* c0 = c;
41 
42   do {
43     __m256 vacc0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
44     w = (const uint16_t*) w + 8;
45 
46     size_t k = kc;
47     do {
48       const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
49       a0 += 1;
50 
51       const __m256 vb01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
52       w = (const uint16_t*) w + 8;
53 
54       vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb01234567, vacc0x01234567), _MM_FROUND_NO_EXC));
55 
56       k -= sizeof(uint16_t);
57     } while (k != 0);
58 
59     const __m256 vmin = _mm256_load_ps(params->avx.min);
60     vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
61 
62     const __m256 vmax = _mm256_load_ps(params->avx.max);
63     vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
64 
65     if XNN_LIKELY(nc >= 8) {
66       _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
67       c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
68 
69       a0 = (const uint16_t*) ((uintptr_t) a0 - kc);
70 
71       nc -= 8;
72     } else {
73       __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
74       if (nc & 4) {
75         _mm_storel_epi64((__m128i*) c0, vh0x01234567);
76 
77         vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);
78 
79         c0 += 4;
80       }
81       if (nc & 2) {
82         _mm_storeu_si32(c0, vh0x01234567);
83 
84         vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);
85 
86         c0 += 2;
87       }
88       if (nc & 1) {
89         *c0 = (uint16_t) _mm_extract_epi16(vh0x01234567, 0);
90       }
91 
92       nc = 0;
93     }
94   } while (nc != 0);
95 }
96