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_4x16__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_4x16__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 <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(float) == 0);
37 assert(ks != 0);
38 assert(ks % (4 * 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
58 do {
59 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w);
60 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF;
61 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF;
62 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF;
63 w += 16;
64
65 size_t p = ks;
66 do {
67 const float* restrict a0 = a[0];
68 assert(a0 != NULL);
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const float*) ((uintptr_t) a0 + a_offset);
71 }
72 const float* restrict a1 = a[1];
73 assert(a1 != NULL);
74 if XNN_UNPREDICTABLE(a1 != zero) {
75 a1 = (const float*) ((uintptr_t) a1 + a_offset);
76 }
77 const float* restrict a2 = a[2];
78 assert(a2 != NULL);
79 if XNN_UNPREDICTABLE(a2 != zero) {
80 a2 = (const float*) ((uintptr_t) a2 + a_offset);
81 }
82 const float* restrict a3 = a[3];
83 assert(a3 != NULL);
84 if XNN_UNPREDICTABLE(a3 != zero) {
85 a3 = (const float*) ((uintptr_t) a3 + a_offset);
86 }
87 a += 4;
88
89 size_t k = kc;
90 do {
91 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w);
92 w += 16;
93
94 vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789ABCDEF);
95 vacc1x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a1), vb0123456789ABCDEF, vacc1x0123456789ABCDEF);
96 vacc2x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a2), vb0123456789ABCDEF, vacc2x0123456789ABCDEF);
97 vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789ABCDEF);
98
99 a0 += 1;
100 a1 += 1;
101 a2 += 1;
102 a3 += 1;
103
104 k -= sizeof(float);
105 } while (k != 0);
106 p -= 4 * sizeof(void*);
107 } while (p != 0);
108
109 const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
110 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax);
111 vacc1x0123456789ABCDEF = _mm512_min_ps(vacc1x0123456789ABCDEF, vmax);
112 vacc2x0123456789ABCDEF = _mm512_min_ps(vacc2x0123456789ABCDEF, vmax);
113 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax);
114
115 const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
116 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin);
117 vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin);
118 vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin);
119 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin);
120
121 if XNN_LIKELY(nc >= 16) {
122 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF);
123 c3 = (float*) ((uintptr_t) c3 + cn_stride);
124 _mm512_storeu_ps(c2, vacc2x0123456789ABCDEF);
125 c2 = (float*) ((uintptr_t) c2 + cn_stride);
126 _mm512_storeu_ps(c1, vacc1x0123456789ABCDEF);
127 c1 = (float*) ((uintptr_t) c1 + cn_stride);
128 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF);
129 c0 = (float*) ((uintptr_t) c0 + cn_stride);
130
131 a = (const float**restrict) ((uintptr_t) a - ks);
132 nc -= 16;
133 } else {
134 if (nc & 15) {
135 // Prepare mask for valid 32-bit elements (depends on nc).
136 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1)));
137
138 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF);
139 _mm512_mask_storeu_ps(c2, vmask, vacc2x0123456789ABCDEF);
140 _mm512_mask_storeu_ps(c1, vmask, vacc1x0123456789ABCDEF);
141 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF);
142 }
143
144 nc = 0;
145 }
146 } while (nc != 0);
147 }
148