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_ukernel_1x16__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_output_params params[restrict static1])18 void xnn_f32_igemm_ukernel_1x16__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_output_params params[restrict static 1])
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(float) == 0);
37 assert(ks != 0);
38 assert(ks % (1 * 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
46 do {
47 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w);
48 w += 16;
49
50 size_t p = ks;
51 do {
52 const float* restrict a0 = a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const float*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 do {
61 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w);
62 w += 16;
63
64 vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789ABCDEF);
65
66 a0 += 1;
67
68 k -= sizeof(float);
69 } while (k != 0);
70 p -= 1 * sizeof(void*);
71 } while (p != 0);
72
73 const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
74 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax);
75
76 const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
77 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin);
78
79 if XNN_LIKELY(nc >= 16) {
80 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF);
81 c0 = (float*) ((uintptr_t) c0 + cn_stride);
82
83 a = (const float**restrict) ((uintptr_t) a - ks);
84 nc -= 16;
85 } else {
86 if (nc & 15) {
87 // Prepare mask for valid 32-bit elements (depends on nc).
88 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1)));
89
90 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF);
91 }
92
93 nc = 0;
94 }
95 } while (nc != 0);
96 }
97