1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/avx-broadcast.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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
16
xnn_f32_gemm_minmax_ukernel_5x8__avx_broadcast(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_minmax_ukernel_5x8__avx_broadcast(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float*restrict a,
22 size_t a_stride,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 5);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 const float* a0 = a;
39 float* c0 = c;
40 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
41 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr < 2) {
43 a1 = a0;
44 c1 = c0;
45 }
46 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
47 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 a2 = a1;
50 c2 = c1;
51 }
52 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 a3 = a2;
56 c3 = c2;
57 }
58 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
59 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 a4 = a3;
62 c4 = c3;
63 }
64
65 do {
66 __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
67 __m256 vacc1x01234567 = vacc0x01234567;
68 __m256 vacc2x01234567 = vacc0x01234567;
69 __m256 vacc3x01234567 = vacc0x01234567;
70 __m256 vacc4x01234567 = vacc0x01234567;
71 w += 8;
72
73 size_t k = kc;
74 do {
75 const __m256 va0 = _mm256_broadcast_ss(a0);
76 a0 += 1;
77 const __m256 va1 = _mm256_broadcast_ss(a1);
78 a1 += 1;
79 const __m256 va2 = _mm256_broadcast_ss(a2);
80 a2 += 1;
81 const __m256 va3 = _mm256_broadcast_ss(a3);
82 a3 += 1;
83 const __m256 va4 = _mm256_broadcast_ss(a4);
84 a4 += 1;
85
86 const __m256 vb01234567 = _mm256_load_ps(w);
87 w += 8;
88
89 vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567));
90 vacc1x01234567 = _mm256_add_ps(vacc1x01234567, _mm256_mul_ps(va1, vb01234567));
91 vacc2x01234567 = _mm256_add_ps(vacc2x01234567, _mm256_mul_ps(va2, vb01234567));
92 vacc3x01234567 = _mm256_add_ps(vacc3x01234567, _mm256_mul_ps(va3, vb01234567));
93 vacc4x01234567 = _mm256_add_ps(vacc4x01234567, _mm256_mul_ps(va4, vb01234567));
94
95 k -= sizeof(float);
96 } while (k != 0);
97
98 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
99 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
100 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
101 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
102 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
103 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
104
105 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
106 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
107 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
108 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
109 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
110 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
111
112 if XNN_LIKELY(nc >= 8) {
113 _mm256_storeu_ps(c4, vacc4x01234567);
114 c4 = (float*) ((uintptr_t) c4 + cn_stride);
115 _mm256_storeu_ps(c3, vacc3x01234567);
116 c3 = (float*) ((uintptr_t) c3 + cn_stride);
117 _mm256_storeu_ps(c2, vacc2x01234567);
118 c2 = (float*) ((uintptr_t) c2 + cn_stride);
119 _mm256_storeu_ps(c1, vacc1x01234567);
120 c1 = (float*) ((uintptr_t) c1 + cn_stride);
121 _mm256_storeu_ps(c0, vacc0x01234567);
122 c0 = (float*) ((uintptr_t) c0 + cn_stride);
123
124 a4 = (const float*) ((uintptr_t) a4 - kc);
125 a3 = (const float*) ((uintptr_t) a3 - kc);
126 a2 = (const float*) ((uintptr_t) a2 - kc);
127 a1 = (const float*) ((uintptr_t) a1 - kc);
128 a0 = (const float*) ((uintptr_t) a0 - kc);
129
130 nc -= 8;
131 } else {
132 __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
133 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
134 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
135 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
136 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
137 if (nc & 4) {
138 _mm_storeu_ps(c4, vacc4x0123);
139 _mm_storeu_ps(c3, vacc3x0123);
140 _mm_storeu_ps(c2, vacc2x0123);
141 _mm_storeu_ps(c1, vacc1x0123);
142 _mm_storeu_ps(c0, vacc0x0123);
143
144 vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
145 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
146 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
147 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
148 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
149
150 c4 += 4;
151 c3 += 4;
152 c2 += 4;
153 c1 += 4;
154 c0 += 4;
155 }
156 if (nc & 2) {
157 _mm_storel_pi((__m64*) c4, vacc4x0123);
158 _mm_storel_pi((__m64*) c3, vacc3x0123);
159 _mm_storel_pi((__m64*) c2, vacc2x0123);
160 _mm_storel_pi((__m64*) c1, vacc1x0123);
161 _mm_storel_pi((__m64*) c0, vacc0x0123);
162
163 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
164 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
165 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
166 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
167 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
168
169 c4 += 2;
170 c3 += 2;
171 c2 += 2;
172 c1 += 2;
173 c0 += 2;
174 }
175 if (nc & 1) {
176 _mm_store_ss(c4, vacc4x0123);
177 _mm_store_ss(c3, vacc3x0123);
178 _mm_store_ss(c2, vacc2x0123);
179 _mm_store_ss(c1, vacc1x0123);
180 _mm_store_ss(c0, vacc0x0123);
181 }
182
183 nc = 0;
184 }
185 } while (nc != 0);
186 }
187