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_ukernel_1x16__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_output_params params[restrict static1])18 void xnn_f32_gemm_ukernel_1x16__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_output_params params[restrict static 1])
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
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
42 do {
43 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w);
44 w += 16;
45
46 size_t k = kc;
47 do {
48 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w);
49 w += 16;
50
51 vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789ABCDEF);
52
53 a0 += 1;
54
55 k -= sizeof(float);
56 } while (k != 0);
57
58 const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
59 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax);
60
61 const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
62 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin);
63
64 if XNN_LIKELY(nc >= 16) {
65 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF);
66 c0 = (float*) ((uintptr_t) c0 + cn_stride);
67
68 a0 = (const float*) ((uintptr_t) a0 - kc);
69
70 nc -= 16;
71 } else {
72 if (nc & 15) {
73 // Prepare mask for valid 32-bit elements (depends on nc).
74 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1)));
75
76 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF);
77 }
78
79 nc = 0;
80 }
81 } while (nc != 0);
82 }
83