xref: /aosp_15_r20/external/XNNPACK/src/f16-gemm/gen/4x8-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_4x8__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_4x8__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 <= 4);
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   const uint16_t* a1 = (const uint16_t*) ((uintptr_t) a0 + a_stride);
42   uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr < 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47   const uint16_t* a2 = (const uint16_t*) ((uintptr_t) a1 + a_stride);
48   uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     a2 = a1;
51     c2 = c1;
52   }
53   const uint16_t* a3 = (const uint16_t*) ((uintptr_t) a2 + a_stride);
54   uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr != 4) {
56     a3 = a2;
57     c3 = c2;
58   }
59 
60   do {
61     __m256 vacc0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
62     __m256 vacc1x01234567 = vacc0x01234567;
63     __m256 vacc2x01234567 = vacc0x01234567;
64     __m256 vacc3x01234567 = vacc0x01234567;
65     w = (const uint16_t*) w + 8;
66 
67     size_t k = kc;
68     do {
69       const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
70       a0 += 1;
71       const __m256 va1 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a1));
72       a1 += 1;
73       const __m256 va2 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a2));
74       a2 += 1;
75       const __m256 va3 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a3));
76       a3 += 1;
77 
78       const __m256 vb01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
79       w = (const uint16_t*) w + 8;
80 
81       vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb01234567, vacc0x01234567), _MM_FROUND_NO_EXC));
82       vacc1x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va1, vb01234567, vacc1x01234567), _MM_FROUND_NO_EXC));
83       vacc2x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va2, vb01234567, vacc2x01234567), _MM_FROUND_NO_EXC));
84       vacc3x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va3, vb01234567, vacc3x01234567), _MM_FROUND_NO_EXC));
85 
86       k -= sizeof(uint16_t);
87     } while (k != 0);
88 
89     const __m256 vmin = _mm256_load_ps(params->avx.min);
90     vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
91     vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
92     vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
93     vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
94 
95     const __m256 vmax = _mm256_load_ps(params->avx.max);
96     vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
97     vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
98     vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
99     vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
100 
101     if XNN_LIKELY(nc >= 8) {
102       _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
103       c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
104       _mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC));
105       c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
106       _mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC));
107       c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
108       _mm_storeu_si128((__m128i*) c3, _mm256_cvtps_ph(vacc3x01234567, _MM_FROUND_NO_EXC));
109       c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
110 
111       a0 = (const uint16_t*) ((uintptr_t) a0 - kc);
112       a1 = (const uint16_t*) ((uintptr_t) a1 - kc);
113       a2 = (const uint16_t*) ((uintptr_t) a2 - kc);
114       a3 = (const uint16_t*) ((uintptr_t) a3 - kc);
115 
116       nc -= 8;
117     } else {
118       __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
119       __m128i vh1x01234567 = _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC);
120       __m128i vh2x01234567 = _mm256_cvtps_ph(vacc2x01234567, _MM_FROUND_NO_EXC);
121       __m128i vh3x01234567 = _mm256_cvtps_ph(vacc3x01234567, _MM_FROUND_NO_EXC);
122       if (nc & 4) {
123         _mm_storel_epi64((__m128i*) c0, vh0x01234567);
124         _mm_storel_epi64((__m128i*) c1, vh1x01234567);
125         _mm_storel_epi64((__m128i*) c2, vh2x01234567);
126         _mm_storel_epi64((__m128i*) c3, vh3x01234567);
127 
128         vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);
129         vh1x01234567 = _mm_unpackhi_epi64(vh1x01234567, vh1x01234567);
130         vh2x01234567 = _mm_unpackhi_epi64(vh2x01234567, vh2x01234567);
131         vh3x01234567 = _mm_unpackhi_epi64(vh3x01234567, vh3x01234567);
132 
133         c0 += 4;
134         c1 += 4;
135         c2 += 4;
136         c3 += 4;
137       }
138       if (nc & 2) {
139         _mm_storeu_si32(c0, vh0x01234567);
140         _mm_storeu_si32(c1, vh1x01234567);
141         _mm_storeu_si32(c2, vh2x01234567);
142         _mm_storeu_si32(c3, vh3x01234567);
143 
144         vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);
145         vh1x01234567 = _mm_srli_epi64(vh1x01234567, 32);
146         vh2x01234567 = _mm_srli_epi64(vh2x01234567, 32);
147         vh3x01234567 = _mm_srli_epi64(vh3x01234567, 32);
148 
149         c0 += 2;
150         c1 += 2;
151         c2 += 2;
152         c3 += 2;
153       }
154       if (nc & 1) {
155         *c0 = (uint16_t) _mm_extract_epi16(vh0x01234567, 0);
156         *c1 = (uint16_t) _mm_extract_epi16(vh1x01234567, 0);
157         *c2 = (uint16_t) _mm_extract_epi16(vh2x01234567, 0);
158         *c3 = (uint16_t) _mm_extract_epi16(vh3x01234567, 0);
159       }
160 
161       nc = 0;
162     }
163   } while (nc != 0);
164 }
165