1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/avx512-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 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_f32_igemm_minmax_ukernel_8x16__avx512f_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)])18 void xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const float**restrict a,
24 const float*restrict w,
25 float*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const float* zero,
30 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 8);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(float) == 0);
37 assert(ks != 0);
38 assert(ks % (8 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(float) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 float* c0 = c;
45 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 c3 = c2;
56 }
57 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
62 if XNN_UNPREDICTABLE(mr < 6) {
63 c5 = c4;
64 }
65 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 6) {
67 c6 = c5;
68 }
69 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
70 if XNN_UNPREDICTABLE(mr != 8) {
71 c7 = c6;
72 }
73
74 do {
75 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w);
76 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
77 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
78 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
79 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
80 __m512 vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
81 __m512 vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
82 __m512 vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF;
83 w += 16;
84
85 size_t p = ks;
86 do {
87 const float* restrict a0 = a[0];
88 assert(a0 != NULL);
89 if XNN_UNPREDICTABLE(a0 != zero) {
90 a0 = (const float*) ((uintptr_t) a0 + a_offset);
91 }
92 const float* restrict a1 = a[1];
93 assert(a1 != NULL);
94 if XNN_UNPREDICTABLE(a1 != zero) {
95 a1 = (const float*) ((uintptr_t) a1 + a_offset);
96 }
97 const float* restrict a2 = a[2];
98 assert(a2 != NULL);
99 if XNN_UNPREDICTABLE(a2 != zero) {
100 a2 = (const float*) ((uintptr_t) a2 + a_offset);
101 }
102 const float* restrict a3 = a[3];
103 assert(a3 != NULL);
104 if XNN_UNPREDICTABLE(a3 != zero) {
105 a3 = (const float*) ((uintptr_t) a3 + a_offset);
106 }
107 const float* restrict a4 = a[4];
108 assert(a4 != NULL);
109 if XNN_UNPREDICTABLE(a4 != zero) {
110 a4 = (const float*) ((uintptr_t) a4 + a_offset);
111 }
112 const float* restrict a5 = a[5];
113 assert(a5 != NULL);
114 if XNN_UNPREDICTABLE(a5 != zero) {
115 a5 = (const float*) ((uintptr_t) a5 + a_offset);
116 }
117 const float* restrict a6 = a[6];
118 assert(a6 != NULL);
119 if XNN_UNPREDICTABLE(a6 != zero) {
120 a6 = (const float*) ((uintptr_t) a6 + a_offset);
121 }
122 const float* restrict a7 = a[7];
123 assert(a7 != NULL);
124 if XNN_UNPREDICTABLE(a7 != zero) {
125 a7 = (const float*) ((uintptr_t) a7 + a_offset);
126 }
127 a += 8;
128
129 size_t k = kc;
130 do {
131 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w);
132 w += 16;
133
134 vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789ABCDEF);
135 vacc1x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a1), vb0123456789ABCDEF, vacc1x0123456789ABCDEF);
136 vacc2x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a2), vb0123456789ABCDEF, vacc2x0123456789ABCDEF);
137 vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789ABCDEF);
138 vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789ABCDEF);
139 vacc5x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a5), vb0123456789ABCDEF, vacc5x0123456789ABCDEF);
140 vacc6x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a6), vb0123456789ABCDEF, vacc6x0123456789ABCDEF);
141 vacc7x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a7), vb0123456789ABCDEF, vacc7x0123456789ABCDEF);
142
143 a0 += 1;
144 a1 += 1;
145 a2 += 1;
146 a3 += 1;
147 a4 += 1;
148 a5 += 1;
149 a6 += 1;
150 a7 += 1;
151
152 k -= sizeof(float);
153 } while (k != 0);
154 p -= 8 * sizeof(void*);
155 } while (p != 0);
156
157 const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
158 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax);
159 vacc1x0123456789ABCDEF = _mm512_min_ps(vacc1x0123456789ABCDEF, vmax);
160 vacc2x0123456789ABCDEF = _mm512_min_ps(vacc2x0123456789ABCDEF, vmax);
161 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax);
162 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax);
163 vacc5x0123456789ABCDEF = _mm512_min_ps(vacc5x0123456789ABCDEF, vmax);
164 vacc6x0123456789ABCDEF = _mm512_min_ps(vacc6x0123456789ABCDEF, vmax);
165 vacc7x0123456789ABCDEF = _mm512_min_ps(vacc7x0123456789ABCDEF, vmax);
166
167 const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
168 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin);
169 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin);
170 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin);
171 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin);
172 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin);
173 vacc5x0123456789ABCDEF = _mm512_max_ps(vacc5x0123456789ABCDEF, vmin);
174 vacc6x0123456789ABCDEF = _mm512_max_ps(vacc6x0123456789ABCDEF, vmin);
175 vacc7x0123456789ABCDEF = _mm512_max_ps(vacc7x0123456789ABCDEF, vmin);
176
177 if XNN_LIKELY(nc >= 16) {
178 _mm512_storeu_ps(c7, vacc7x0123456789ABCDEF);
179 c7 = (float*) ((uintptr_t) c7 + cn_stride);
180 _mm512_storeu_ps(c6, vacc6x0123456789ABCDEF);
181 c6 = (float*) ((uintptr_t) c6 + cn_stride);
182 _mm512_storeu_ps(c5, vacc5x0123456789ABCDEF);
183 c5 = (float*) ((uintptr_t) c5 + cn_stride);
184 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF);
185 c4 = (float*) ((uintptr_t) c4 + cn_stride);
186 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF);
187 c3 = (float*) ((uintptr_t) c3 + cn_stride);
188 _mm512_storeu_ps(c2, vacc2x0123456789ABCDEF);
189 c2 = (float*) ((uintptr_t) c2 + cn_stride);
190 _mm512_storeu_ps(c1, vacc1x0123456789ABCDEF);
191 c1 = (float*) ((uintptr_t) c1 + cn_stride);
192 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF);
193 c0 = (float*) ((uintptr_t) c0 + cn_stride);
194
195 a = (const float**restrict) ((uintptr_t) a - ks);
196 nc -= 16;
197 } else {
198 if (nc & 15) {
199 // Prepare mask for valid 32-bit elements (depends on nc).
200 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1)));
201
202 _mm512_mask_storeu_ps(c7, vmask, vacc7x0123456789ABCDEF);
203 _mm512_mask_storeu_ps(c6, vmask, vacc6x0123456789ABCDEF);
204 _mm512_mask_storeu_ps(c5, vmask, vacc5x0123456789ABCDEF);
205 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF);
206 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF);
207 _mm512_mask_storeu_ps(c2, vmask, vacc2x0123456789ABCDEF);
208 _mm512_mask_storeu_ps(c1, vmask, vacc1x0123456789ABCDEF);
209 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF);
210 }
211
212 nc = 0;
213 }
214 } while (nc != 0);
215 }
216