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_6x8__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_6x8__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 <= 6);
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   const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
65   float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
66   if XNN_UNPREDICTABLE(mr != 6) {
67     a5 = a4;
68     c5 = c4;
69   }
70 
71   do {
72     __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
73     __m256 vacc1x01234567 = vacc0x01234567;
74     __m256 vacc2x01234567 = vacc0x01234567;
75     __m256 vacc3x01234567 = vacc0x01234567;
76     __m256 vacc4x01234567 = vacc0x01234567;
77     __m256 vacc5x01234567 = vacc0x01234567;
78     w += 8;
79 
80     size_t k = kc;
81     do {
82       const __m256 va0 = _mm256_broadcast_ss(a0);
83       a0 += 1;
84       const __m256 va1 = _mm256_broadcast_ss(a1);
85       a1 += 1;
86       const __m256 va2 = _mm256_broadcast_ss(a2);
87       a2 += 1;
88       const __m256 va3 = _mm256_broadcast_ss(a3);
89       a3 += 1;
90       const __m256 va4 = _mm256_broadcast_ss(a4);
91       a4 += 1;
92       const __m256 va5 = _mm256_broadcast_ss(a5);
93       a5 += 1;
94 
95       const __m256 vb01234567 = _mm256_load_ps(w);
96       w += 8;
97 
98       vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567));
99       vacc1x01234567 = _mm256_add_ps(vacc1x01234567, _mm256_mul_ps(va1, vb01234567));
100       vacc2x01234567 = _mm256_add_ps(vacc2x01234567, _mm256_mul_ps(va2, vb01234567));
101       vacc3x01234567 = _mm256_add_ps(vacc3x01234567, _mm256_mul_ps(va3, vb01234567));
102       vacc4x01234567 = _mm256_add_ps(vacc4x01234567, _mm256_mul_ps(va4, vb01234567));
103       vacc5x01234567 = _mm256_add_ps(vacc5x01234567, _mm256_mul_ps(va5, vb01234567));
104 
105       k -= sizeof(float);
106     } while (k != 0);
107 
108     const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
109     vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
110     vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
111     vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
112     vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
113     vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
114     vacc5x01234567 = _mm256_min_ps(vacc5x01234567, vmax);
115 
116     const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
117     vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
118     vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
119     vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
120     vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
121     vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
122     vacc5x01234567 = _mm256_max_ps(vacc5x01234567, vmin);
123 
124     if XNN_LIKELY(nc >= 8) {
125       _mm256_storeu_ps(c5, vacc5x01234567);
126       c5 = (float*) ((uintptr_t) c5 + cn_stride);
127       _mm256_storeu_ps(c4, vacc4x01234567);
128       c4 = (float*) ((uintptr_t) c4 + cn_stride);
129       _mm256_storeu_ps(c3, vacc3x01234567);
130       c3 = (float*) ((uintptr_t) c3 + cn_stride);
131       _mm256_storeu_ps(c2, vacc2x01234567);
132       c2 = (float*) ((uintptr_t) c2 + cn_stride);
133       _mm256_storeu_ps(c1, vacc1x01234567);
134       c1 = (float*) ((uintptr_t) c1 + cn_stride);
135       _mm256_storeu_ps(c0, vacc0x01234567);
136       c0 = (float*) ((uintptr_t) c0 + cn_stride);
137 
138       a5 = (const float*) ((uintptr_t) a5 - kc);
139       a4 = (const float*) ((uintptr_t) a4 - kc);
140       a3 = (const float*) ((uintptr_t) a3 - kc);
141       a2 = (const float*) ((uintptr_t) a2 - kc);
142       a1 = (const float*) ((uintptr_t) a1 - kc);
143       a0 = (const float*) ((uintptr_t) a0 - kc);
144 
145       nc -= 8;
146     } else {
147       __m128 vacc5x0123 = _mm256_castps256_ps128(vacc5x01234567);
148       __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
149       __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
150       __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
151       __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
152       __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
153       if (nc & 4) {
154         _mm_storeu_ps(c5, vacc5x0123);
155         _mm_storeu_ps(c4, vacc4x0123);
156         _mm_storeu_ps(c3, vacc3x0123);
157         _mm_storeu_ps(c2, vacc2x0123);
158         _mm_storeu_ps(c1, vacc1x0123);
159         _mm_storeu_ps(c0, vacc0x0123);
160 
161         vacc5x0123 = _mm256_extractf128_ps(vacc5x01234567, 1);
162         vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
163         vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
164         vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
165         vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
166         vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
167 
168         c5 += 4;
169         c4 += 4;
170         c3 += 4;
171         c2 += 4;
172         c1 += 4;
173         c0 += 4;
174       }
175       if (nc & 2) {
176         _mm_storel_pi((__m64*) c5, vacc5x0123);
177         _mm_storel_pi((__m64*) c4, vacc4x0123);
178         _mm_storel_pi((__m64*) c3, vacc3x0123);
179         _mm_storel_pi((__m64*) c2, vacc2x0123);
180         _mm_storel_pi((__m64*) c1, vacc1x0123);
181         _mm_storel_pi((__m64*) c0, vacc0x0123);
182 
183         vacc5x0123 = _mm_movehl_ps(vacc5x0123, vacc5x0123);
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         c5 += 2;
191         c4 += 2;
192         c3 += 2;
193         c2 += 2;
194         c1 += 2;
195         c0 += 2;
196       }
197       if (nc & 1) {
198         _mm_store_ss(c5, vacc5x0123);
199         _mm_store_ss(c4, vacc4x0123);
200         _mm_store_ss(c3, vacc3x0123);
201         _mm_store_ss(c2, vacc2x0123);
202         _mm_store_ss(c1, vacc1x0123);
203         _mm_store_ss(c0, vacc0x0123);
204       }
205 
206       nc = 0;
207     }
208   } while (nc != 0);
209 }
210