xref: /aosp_15_r20/external/XNNPACK/src/f16-igemm/gen/1x8-minmax-avx2-broadcast.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-igemm/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/igemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_f16_igemm_minmax_ukernel_1x8__avx2_broadcast(size_t mr,size_t nc,size_t kc,size_t ks,const void ** restrict a,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const void * zero,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_igemm_minmax_ukernel_1x8__avx2_broadcast(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const void**restrict a,
24     const void*restrict w,
25     void*restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const void* zero,
30     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32   assert(mr != 0);
33   assert(mr <= 1);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(kc % sizeof(uint16_t) == 0);
37   assert(ks != 0);
38   assert(ks % (1 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(uint16_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   uint16_t* c0 = c;
45 
46   do {
47     __m256 vacc0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
48     w = (const uint16_t*) w + 8;
49 
50     size_t p = ks;
51     do {
52       const uint16_t* restrict a0 = (const uint16_t*) a[0];
53       assert(a0 != NULL);
54       if XNN_UNPREDICTABLE(a0 != zero) {
55         a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset);
56       }
57       a += 1;
58 
59       size_t k = kc;
60       do {
61         const __m256 vb01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
62         w = (const uint16_t*) w + 8;
63 
64         const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
65         a0 += 1;
66 
67         vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb01234567, vacc0x01234567), _MM_FROUND_NO_EXC));
68 
69         k -= sizeof(uint16_t);
70       } while (k != 0);
71       p -= 1 * sizeof(void*);
72     } while (p != 0);
73 
74     const __m256 vmin = _mm256_load_ps(params->avx.min);
75     vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
76 
77     const __m256 vmax = _mm256_load_ps(params->avx.max);
78     vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
79 
80     if XNN_LIKELY(nc >= 8) {
81       _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
82       c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
83 
84       a = (const void**restrict) ((uintptr_t) a - ks);
85       nc -= 8;
86     } else {
87       __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
88       if (nc & 4) {
89         _mm_storel_epi64((__m128i*) c0, vh0x01234567);
90 
91         vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);
92 
93         c0 += 4;
94       }
95       if (nc & 2) {
96         _mm_storeu_si32(c0, vh0x01234567);
97 
98         vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);
99 
100         c0 += 2;
101       }
102       if (nc & 1) {
103         *c0 = _mm_extract_epi16(vh0x01234567, 0);
104       }
105 
106       nc = 0;
107     }
108   } while (nc != 0);
109 }
110