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_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_gemm_minmax_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 5);
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
66 do {
67 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w);
68 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
69 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
70 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
71 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF;
72 w += 16;
73
74 size_t k = kc;
75 do {
76 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w);
77 w += 16;
78
79 vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789ABCDEF);
80 vacc1x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a1), vb0123456789ABCDEF, vacc1x0123456789ABCDEF);
81 vacc2x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a2), vb0123456789ABCDEF, vacc2x0123456789ABCDEF);
82 vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789ABCDEF);
83 vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789ABCDEF);
84
85 a0 += 1;
86 a1 += 1;
87 a2 += 1;
88 a3 += 1;
89 a4 += 1;
90
91 k -= sizeof(float);
92 } while (k != 0);
93
94 const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
95 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax);
96 vacc1x0123456789ABCDEF = _mm512_min_ps(vacc1x0123456789ABCDEF, vmax);
97 vacc2x0123456789ABCDEF = _mm512_min_ps(vacc2x0123456789ABCDEF, vmax);
98 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax);
99 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax);
100
101 const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
102 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin);
103 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin);
104 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin);
105 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin);
106 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin);
107
108 if XNN_LIKELY(nc >= 16) {
109 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF);
110 c4 = (float*) ((uintptr_t) c4 + cn_stride);
111 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF);
112 c3 = (float*) ((uintptr_t) c3 + cn_stride);
113 _mm512_storeu_ps(c2, vacc2x0123456789ABCDEF);
114 c2 = (float*) ((uintptr_t) c2 + cn_stride);
115 _mm512_storeu_ps(c1, vacc1x0123456789ABCDEF);
116 c1 = (float*) ((uintptr_t) c1 + cn_stride);
117 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF);
118 c0 = (float*) ((uintptr_t) c0 + cn_stride);
119
120 a4 = (const float*) ((uintptr_t) a4 - kc);
121 a3 = (const float*) ((uintptr_t) a3 - kc);
122 a2 = (const float*) ((uintptr_t) a2 - kc);
123 a1 = (const float*) ((uintptr_t) a1 - kc);
124 a0 = (const float*) ((uintptr_t) a0 - kc);
125
126 nc -= 16;
127 } else {
128 if (nc & 15) {
129 // Prepare mask for valid 32-bit elements (depends on nc).
130 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1)));
131
132 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF);
133 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF);
134 _mm512_mask_storeu_ps(c2, vmask, vacc2x0123456789ABCDEF);
135 _mm512_mask_storeu_ps(c1, vmask, vacc1x0123456789ABCDEF);
136 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF);
137 }
138
139 nc = 0;
140 }
141 } while (nc != 0);
142 }
143