• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-gemm/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/gemm.h>
15 
16 
xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_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 float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_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 float*restrict acc,
28     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(float) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38   assert(acc != NULL);
39 
40   const float* a0 = a;
41   float* c0 = c;
42 
43   do {
44     __m256 vacc0x01234567 = _mm256_load_ps(acc + 0);
45     __m256 vacc0x89ABCDEF = _mm256_load_ps(acc + 8);
46     acc += 16;
47 
48     size_t k = kc;
49     while (k >= 4 * sizeof(float)) {
50       __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
51       a0 += 4;
52 
53 
54       const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
55       const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
56 
57       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);
58       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);
59 
60       va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
61 
62       const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
63       const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
64 
65       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);
66       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);
67 
68       va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
69 
70       const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
71       const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
72 
73       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);
74       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc2, vacc0x89ABCDEF);
75 
76       va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
77 
78       const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
79       const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
80 
81       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);
82       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);
83 
84 
85       w += 64;
86       k -= 4 * sizeof(float);
87     }
88     if XNN_UNLIKELY(k != 0) {
89       do {
90         const __m256 va0 = _mm256_broadcast_ss(a0);
91         a0 += 1;
92 
93         const __m256 vb01234567 = _mm256_load_ps(w);
94         const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
95         w += 16;
96 
97         vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
98         vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF);
99 
100         k -= sizeof(float);
101       } while (k != 0);
102     }
103 
104     const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
105     vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
106     vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
107 
108     const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
109     vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
110     vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
111 
112     if XNN_LIKELY(nc >= 16) {
113       _mm256_storeu_ps(c0, vacc0x01234567);
114       _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
115       c0 = (float*) ((uintptr_t) c0 + cn_stride);
116 
117       a0 = (const float*) ((uintptr_t) a0 - kc);
118 
119       nc -= 16;
120     } else {
121       if (nc & 8) {
122         _mm256_storeu_ps(c0, vacc0x01234567);
123 
124         vacc0x01234567 = vacc0x89ABCDEF;
125 
126         c0 += 8;
127       }
128       __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
129       if (nc & 4) {
130         _mm_storeu_ps(c0, vacc0x0123);
131 
132         vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
133 
134         c0 += 4;
135       }
136       if (nc & 2) {
137         _mm_storel_pi((__m64*) c0, vacc0x0123);
138 
139         vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
140 
141         c0 += 2;
142       }
143       if (nc & 1) {
144         _mm_store_ss(c0, vacc0x0123);
145       }
146 
147       nc = 0;
148     }
149   } while (nc != 0);
150 }
151