1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
15
16
xnn_f32_igemm_ukernel_5x8__avx_broadcast(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_igemm_ukernel_5x8__avx_broadcast(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_output_params params[restrict static 1])
30 {
31 assert(mr != 0);
32 assert(mr <= 5);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (5 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr < 4) {
54 c3 = c2;
55 }
56 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 4) {
58 c4 = c3;
59 }
60
61 do {
62 __m256 vacc0x01234567 = _mm256_load_ps(w);
63 __m256 vacc1x01234567 = vacc0x01234567;
64 __m256 vacc2x01234567 = vacc0x01234567;
65 __m256 vacc3x01234567 = vacc0x01234567;
66 __m256 vacc4x01234567 = vacc0x01234567;
67 w += 8;
68
69 size_t p = ks;
70 do {
71 const float* restrict a0 = a[0];
72 assert(a0 != NULL);
73 if XNN_UNPREDICTABLE(a0 != zero) {
74 a0 = (const float*) ((uintptr_t) a0 + a_offset);
75 }
76 const float* restrict a1 = a[1];
77 assert(a1 != NULL);
78 if XNN_UNPREDICTABLE(a1 != zero) {
79 a1 = (const float*) ((uintptr_t) a1 + a_offset);
80 }
81 const float* restrict a2 = a[2];
82 assert(a2 != NULL);
83 if XNN_UNPREDICTABLE(a2 != zero) {
84 a2 = (const float*) ((uintptr_t) a2 + a_offset);
85 }
86 const float* restrict a3 = a[3];
87 assert(a3 != NULL);
88 if XNN_UNPREDICTABLE(a3 != zero) {
89 a3 = (const float*) ((uintptr_t) a3 + a_offset);
90 }
91 const float* restrict a4 = a[4];
92 assert(a4 != NULL);
93 if XNN_UNPREDICTABLE(a4 != zero) {
94 a4 = (const float*) ((uintptr_t) a4 + a_offset);
95 }
96 a += 5;
97
98 size_t k = kc;
99 do {
100 const __m256 vb01234567 = _mm256_load_ps(w);
101 w += 8;
102
103 const __m256 va0 = _mm256_broadcast_ss(a0);
104 a0 += 1;
105 const __m256 va1 = _mm256_broadcast_ss(a1);
106 a1 += 1;
107 const __m256 va2 = _mm256_broadcast_ss(a2);
108 a2 += 1;
109 const __m256 va3 = _mm256_broadcast_ss(a3);
110 a3 += 1;
111 const __m256 va4 = _mm256_broadcast_ss(a4);
112 a4 += 1;
113
114 vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567));
115 vacc1x01234567 = _mm256_add_ps(vacc1x01234567, _mm256_mul_ps(va1, vb01234567));
116 vacc2x01234567 = _mm256_add_ps(vacc2x01234567, _mm256_mul_ps(va2, vb01234567));
117 vacc3x01234567 = _mm256_add_ps(vacc3x01234567, _mm256_mul_ps(va3, vb01234567));
118 vacc4x01234567 = _mm256_add_ps(vacc4x01234567, _mm256_mul_ps(va4, vb01234567));
119 k -= sizeof(float);
120 } while (k != 0);
121 p -= 5 * sizeof(void*);
122 } while (p != 0);
123
124 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
125 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
126 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
127 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
128 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
129 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
130
131 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
132 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
133 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
134 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
135 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
136 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
137
138 if XNN_LIKELY(nc >= 8) {
139 _mm256_storeu_ps(c4, vacc4x01234567);
140 c4 = (float*) ((uintptr_t) c4 + cn_stride);
141 _mm256_storeu_ps(c3, vacc3x01234567);
142 c3 = (float*) ((uintptr_t) c3 + cn_stride);
143 _mm256_storeu_ps(c2, vacc2x01234567);
144 c2 = (float*) ((uintptr_t) c2 + cn_stride);
145 _mm256_storeu_ps(c1, vacc1x01234567);
146 c1 = (float*) ((uintptr_t) c1 + cn_stride);
147 _mm256_storeu_ps(c0, vacc0x01234567);
148 c0 = (float*) ((uintptr_t) c0 + cn_stride);
149
150 a = (const float**restrict) ((uintptr_t) a - ks);
151 nc -= 8;
152 } else {
153 __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
154 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
155 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
156 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
157 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
158 if (nc & 4) {
159 _mm_storeu_ps(c4, vacc4x0123);
160 _mm_storeu_ps(c3, vacc3x0123);
161 _mm_storeu_ps(c2, vacc2x0123);
162 _mm_storeu_ps(c1, vacc1x0123);
163 _mm_storeu_ps(c0, vacc0x0123);
164
165 vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
166 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
167 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
168 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
169 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
170
171 c4 += 4;
172 c3 += 4;
173 c2 += 4;
174 c1 += 4;
175 c0 += 4;
176 }
177 if (nc & 2) {
178 _mm_storel_pi((__m64*) c4, vacc4x0123);
179 _mm_storel_pi((__m64*) c3, vacc3x0123);
180 _mm_storel_pi((__m64*) c2, vacc2x0123);
181 _mm_storel_pi((__m64*) c1, vacc1x0123);
182 _mm_storel_pi((__m64*) c0, vacc0x0123);
183
184 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
185 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
186 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
187 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
188 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
189
190 c4 += 2;
191 c3 += 2;
192 c2 += 2;
193 c1 += 2;
194 c0 += 2;
195 }
196 if (nc & 1) {
197 _mm_store_ss(c4, vacc4x0123);
198 _mm_store_ss(c3, vacc3x0123);
199 _mm_store_ss(c2, vacc2x0123);
200 _mm_store_ss(c1, vacc1x0123);
201 _mm_store_ss(c0, vacc0x0123);
202 }
203
204 nc = 0;
205 }
206 } while (nc != 0);
207 }
208