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_gemminc_ukernel_5x16__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 float * restrict acc,const union xnn_f32_output_params params[restrict static1])18 void xnn_f32_gemminc_ukernel_5x16__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 float*restrict acc,
29 const union xnn_f32_output_params params[restrict static 1])
30 {
31 assert(mr != 0);
32 assert(mr <= 5);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39 assert(acc != NULL);
40
41 const float* a0 = a;
42 float* c0 = c;
43 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
62 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67
68 do {
69 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0);
70 __m512 vacc1x0123456789ABCDEF = _mm512_load_ps(acc + 16);
71 __m512 vacc2x0123456789ABCDEF = _mm512_load_ps(acc + 32);
72 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48);
73 __m512 vacc4x0123456789ABCDEF = _mm512_load_ps(acc + 64);
74 acc += 80;
75
76 size_t k = kc;
77 do {
78 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w);
79 w += 16;
80
81 vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789ABCDEF);
82 vacc1x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a1), vb0123456789ABCDEF, vacc1x0123456789ABCDEF);
83 vacc2x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a2), vb0123456789ABCDEF, vacc2x0123456789ABCDEF);
84 vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789ABCDEF);
85 vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789ABCDEF);
86
87 a0 += 1;
88 a1 += 1;
89 a2 += 1;
90 a3 += 1;
91 a4 += 1;
92
93 k -= sizeof(float);
94 } while (k != 0);
95
96 const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
97 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax);
98 vacc1x0123456789ABCDEF = _mm512_min_ps(vacc1x0123456789ABCDEF, vmax);
99 vacc2x0123456789ABCDEF = _mm512_min_ps(vacc2x0123456789ABCDEF, vmax);
100 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax);
101 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax);
102
103 const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
104 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin);
105 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin);
106 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin);
107 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin);
108 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin);
109
110 if XNN_LIKELY(nc >= 16) {
111 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF);
112 c4 = (float*) ((uintptr_t) c4 + cn_stride);
113 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF);
114 c3 = (float*) ((uintptr_t) c3 + cn_stride);
115 _mm512_storeu_ps(c2, vacc2x0123456789ABCDEF);
116 c2 = (float*) ((uintptr_t) c2 + cn_stride);
117 _mm512_storeu_ps(c1, vacc1x0123456789ABCDEF);
118 c1 = (float*) ((uintptr_t) c1 + cn_stride);
119 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF);
120 c0 = (float*) ((uintptr_t) c0 + cn_stride);
121
122 a4 = (const float*) ((uintptr_t) a4 - kc);
123 a3 = (const float*) ((uintptr_t) a3 - kc);
124 a2 = (const float*) ((uintptr_t) a2 - kc);
125 a1 = (const float*) ((uintptr_t) a1 - kc);
126 a0 = (const float*) ((uintptr_t) a0 - kc);
127
128 nc -= 16;
129 } else {
130 if (nc & 15) {
131 // Prepare mask for valid 32-bit elements (depends on nc).
132 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1)));
133
134 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF);
135 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF);
136 _mm512_mask_storeu_ps(c2, vmask, vacc2x0123456789ABCDEF);
137 _mm512_mask_storeu_ps(c1, vmask, vacc1x0123456789ABCDEF);
138 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF);
139 }
140
141 nc = 0;
142 }
143 } while (nc != 0);
144 }
145