1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/avx-shuffle4.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_minmax_ukernel_1x16s4__fma3_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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
45 do {
46 __m256 vacc0x01234567 = _mm256_load_ps(w);
47 __m256 vacc0x89ABCDEF = _mm256_load_ps(w + 8);
48 w += 16;
49
50 size_t p = ks;
51 do {
52 const float* restrict a0 = a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const float*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 while (k >= 4 * sizeof(float)) {
61 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
62 a0 += 4;
63
64
65 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
66 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
67
68 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);
69 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);
70
71 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
72
73 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
74 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
75
76 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);
77 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);
78
79 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
80
81 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
82 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
83
84 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);
85 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc2, vacc0x89ABCDEF);
86
87 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
88
89 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
90 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
91
92 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);
93 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);
94
95
96 w += 64;
97 k -= 4 * sizeof(float);
98 }
99 if XNN_UNLIKELY(k != 0) {
100 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
101 a0 = (const float*) ((uintptr_t) a0 + k);
102
103 const __m256 vzero = _mm256_setzero_ps();
104
105 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
106 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
107
108 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc0x01234567);
109 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc0x89ABCDEF);
110
111 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
112
113 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
114 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
115
116 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc0x01234567);
117 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc0x89ABCDEF);
118
119 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
120
121 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
122 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
123
124 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc0x01234567);
125 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc0x89ABCDEF);
126
127 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
128
129 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
130 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
131
132 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc0x01234567);
133 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc0x89ABCDEF);
134
135
136 w += 64;
137 }
138 p -= 1 * sizeof(void*);
139 } while (p != 0);
140
141 const __m256 vmin = _mm256_load_ps(params->avx.min);
142 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
143 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
144
145 const __m256 vmax = _mm256_load_ps(params->avx.max);
146 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
147 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
148
149 if XNN_LIKELY(nc >= 16) {
150 _mm256_storeu_ps(c0, vacc0x01234567);
151 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
152 c0 = (float*) ((uintptr_t) c0 + cn_stride);
153
154 a = (const float**restrict) ((uintptr_t) a - ks);
155 nc -= 16;
156 } else {
157 if (nc & 8) {
158 _mm256_storeu_ps(c0, vacc0x01234567);
159
160 vacc0x01234567 = vacc0x89ABCDEF;
161
162 c0 += 8;
163 }
164 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
165 if (nc & 4) {
166 _mm_storeu_ps(c0, vacc0x0123);
167
168 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
169
170 c0 += 4;
171 }
172 if (nc & 2) {
173 _mm_storel_pi((__m64*) c0, vacc0x0123);
174
175 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
176
177 c0 += 2;
178 }
179 if (nc & 1) {
180 _mm_store_ss(c0, vacc0x0123);
181 }
182
183 nc = 0;
184 }
185 } while (nc != 0);
186 }
187