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_6x16__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_6x16__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 <= 6);
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
72 do {
73 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w);
74 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
75 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
76 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
77 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
78 __m512 vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF;
79 w += 16;
80
81 size_t k = kc;
82 do {
83 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w);
84 w += 16;
85
86 const __m512 va0 = _mm512_set1_ps(*a0);
87 vacc0x0123456789ABCDEF = _mm512_fmadd_ps(va0, vb0123456789ABCDEF, vacc0x0123456789ABCDEF);
88 const __m512 va1 = _mm512_set1_ps(*a1);
89 vacc1x0123456789ABCDEF = _mm512_fmadd_ps(va1, vb0123456789ABCDEF, vacc1x0123456789ABCDEF);
90 const __m512 va2 = _mm512_set1_ps(*a2);
91 vacc2x0123456789ABCDEF = _mm512_fmadd_ps(va2, vb0123456789ABCDEF, vacc2x0123456789ABCDEF);
92 const __m512 va3 = _mm512_set1_ps(*a3);
93 vacc3x0123456789ABCDEF = _mm512_fmadd_ps(va3, vb0123456789ABCDEF, vacc3x0123456789ABCDEF);
94 const __m512 va4 = _mm512_set1_ps(*a4);
95 vacc4x0123456789ABCDEF = _mm512_fmadd_ps(va4, vb0123456789ABCDEF, vacc4x0123456789ABCDEF);
96 const __m512 va5 = _mm512_set1_ps(*a5);
97 vacc5x0123456789ABCDEF = _mm512_fmadd_ps(va5, vb0123456789ABCDEF, vacc5x0123456789ABCDEF);
98
99 a0 += 1;
100 a1 += 1;
101 a2 += 1;
102 a3 += 1;
103 a4 += 1;
104 a5 += 1;
105
106 k -= sizeof(float);
107 } while (k != 0);
108
109 const __m512 vmin = _mm512_set1_ps(params->scalar.min);
110 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin);
111 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin);
112 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin);
113 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin);
114 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin);
115 vacc5x0123456789ABCDEF = _mm512_max_ps(vacc5x0123456789ABCDEF, vmin);
116
117 const __m512 vmax = _mm512_set1_ps(params->scalar.max);
118 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax);
119 vacc1x0123456789ABCDEF = _mm512_min_ps(vacc1x0123456789ABCDEF, vmax);
120 vacc2x0123456789ABCDEF = _mm512_min_ps(vacc2x0123456789ABCDEF, vmax);
121 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax);
122 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax);
123 vacc5x0123456789ABCDEF = _mm512_min_ps(vacc5x0123456789ABCDEF, vmax);
124
125 if XNN_LIKELY(nc >= 16) {
126 _mm512_storeu_ps(c5, vacc5x0123456789ABCDEF);
127 c5 = (float*) ((uintptr_t) c5 + cn_stride);
128 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF);
129 c4 = (float*) ((uintptr_t) c4 + cn_stride);
130 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF);
131 c3 = (float*) ((uintptr_t) c3 + cn_stride);
132 _mm512_storeu_ps(c2, vacc2x0123456789ABCDEF);
133 c2 = (float*) ((uintptr_t) c2 + cn_stride);
134 _mm512_storeu_ps(c1, vacc1x0123456789ABCDEF);
135 c1 = (float*) ((uintptr_t) c1 + cn_stride);
136 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF);
137 c0 = (float*) ((uintptr_t) c0 + cn_stride);
138
139 a5 = (const float*) ((uintptr_t) a5 - kc);
140 a4 = (const float*) ((uintptr_t) a4 - kc);
141 a3 = (const float*) ((uintptr_t) a3 - kc);
142 a2 = (const float*) ((uintptr_t) a2 - kc);
143 a1 = (const float*) ((uintptr_t) a1 - kc);
144 a0 = (const float*) ((uintptr_t) a0 - kc);
145
146 nc -= 16;
147 } else {
148 if (nc & 15) {
149 // Prepare mask for valid 32-bit elements (depends on nc).
150 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1)));
151
152 _mm512_mask_storeu_ps(c5, vmask, vacc5x0123456789ABCDEF);
153 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF);
154 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF);
155 _mm512_mask_storeu_ps(c2, vmask, vacc2x0123456789ABCDEF);
156 _mm512_mask_storeu_ps(c1, vmask, vacc1x0123456789ABCDEF);
157 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF);
158 }
159
160 nc = 0;
161 }
162 } while (nc != 0);
163 }
164