1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_f32_gemm_minmax_ukernel_7x16__avx512f_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)])18 void xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const float*restrict a,
23 size_t a_stride,
24 const float*restrict w,
25 float*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 7);
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
39 const float* a0 = a;
40 float* c0 = c;
41 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
42 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
54 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 a3 = a2;
57 c3 = c2;
58 }
59 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
60 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 a4 = a3;
63 c4 = c3;
64 }
65 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
66 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
67 if XNN_UNPREDICTABLE(mr < 6) {
68 a5 = a4;
69 c5 = c4;
70 }
71 const float* a6 = (const float*) ((uintptr_t) a5 + a_stride);
72 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
73 if XNN_UNPREDICTABLE(mr <= 6) {
74 a6 = a5;
75 c6 = c5;
76 }
77
78 do {
79 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w);
80 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
81 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
82 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
83 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
84 __m512 vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
85 __m512 vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF;
86 w += 16;
87
88 size_t k = kc;
89 do {
90 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w);
91 w += 16;
92
93 vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789ABCDEF);
94 vacc1x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a1), vb0123456789ABCDEF, vacc1x0123456789ABCDEF);
95 vacc2x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a2), vb0123456789ABCDEF, vacc2x0123456789ABCDEF);
96 vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789ABCDEF);
97 vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789ABCDEF);
98 vacc5x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a5), vb0123456789ABCDEF, vacc5x0123456789ABCDEF);
99 vacc6x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a6), vb0123456789ABCDEF, vacc6x0123456789ABCDEF);
100
101 a0 += 1;
102 a1 += 1;
103 a2 += 1;
104 a3 += 1;
105 a4 += 1;
106 a5 += 1;
107 a6 += 1;
108
109 k -= sizeof(float);
110 } while (k != 0);
111
112 const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
113 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax);
114 vacc1x0123456789ABCDEF = _mm512_min_ps(vacc1x0123456789ABCDEF, vmax);
115 vacc2x0123456789ABCDEF = _mm512_min_ps(vacc2x0123456789ABCDEF, vmax);
116 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax);
117 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax);
118 vacc5x0123456789ABCDEF = _mm512_min_ps(vacc5x0123456789ABCDEF, vmax);
119 vacc6x0123456789ABCDEF = _mm512_min_ps(vacc6x0123456789ABCDEF, vmax);
120
121 const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
122 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin);
123 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin);
124 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin);
125 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin);
126 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin);
127 vacc5x0123456789ABCDEF = _mm512_max_ps(vacc5x0123456789ABCDEF, vmin);
128 vacc6x0123456789ABCDEF = _mm512_max_ps(vacc6x0123456789ABCDEF, vmin);
129
130 if XNN_LIKELY(nc >= 16) {
131 _mm512_storeu_ps(c6, vacc6x0123456789ABCDEF);
132 c6 = (float*) ((uintptr_t) c6 + cn_stride);
133 _mm512_storeu_ps(c5, vacc5x0123456789ABCDEF);
134 c5 = (float*) ((uintptr_t) c5 + cn_stride);
135 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF);
136 c4 = (float*) ((uintptr_t) c4 + cn_stride);
137 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF);
138 c3 = (float*) ((uintptr_t) c3 + cn_stride);
139 _mm512_storeu_ps(c2, vacc2x0123456789ABCDEF);
140 c2 = (float*) ((uintptr_t) c2 + cn_stride);
141 _mm512_storeu_ps(c1, vacc1x0123456789ABCDEF);
142 c1 = (float*) ((uintptr_t) c1 + cn_stride);
143 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF);
144 c0 = (float*) ((uintptr_t) c0 + cn_stride);
145
146 a6 = (const float*) ((uintptr_t) a6 - kc);
147 a5 = (const float*) ((uintptr_t) a5 - kc);
148 a4 = (const float*) ((uintptr_t) a4 - kc);
149 a3 = (const float*) ((uintptr_t) a3 - kc);
150 a2 = (const float*) ((uintptr_t) a2 - kc);
151 a1 = (const float*) ((uintptr_t) a1 - kc);
152 a0 = (const float*) ((uintptr_t) a0 - kc);
153
154 nc -= 16;
155 } else {
156 if (nc & 15) {
157 // Prepare mask for valid 32-bit elements (depends on nc).
158 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1)));
159
160 _mm512_mask_storeu_ps(c6, vmask, vacc6x0123456789ABCDEF);
161 _mm512_mask_storeu_ps(c5, vmask, vacc5x0123456789ABCDEF);
162 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF);
163 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF);
164 _mm512_mask_storeu_ps(c2, vmask, vacc2x0123456789ABCDEF);
165 _mm512_mask_storeu_ps(c1, vmask, vacc1x0123456789ABCDEF);
166 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF);
167 }
168
169 nc = 0;
170 }
171 } while (nc != 0);
172 }
173