Debian Codegraph @

src/x32-packw/gen/x32-packw-x8-gemm-gio-avx-u1-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/x32-packw/gio-avx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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 #include <stddef.h>
12 #include <stdint.h>
13
14 #include <immintrin.h>
15
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/packw.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_x32_packw_gemm_gio_ukernel_x8__avx_u1_prfm(
22 size_t g,
23 size_t nc,
24 size_t kc,
25 size_t nr,
26 size_t kr,
27 size_t sr,
28 size_t k_stride,
29 const uint32_t* weights,
30 const uint32_t* bias,
31 const void* scale,
32 uint32_t* packed_weights,
33 size_t extra_bytes,
34 const void* params)
35 {
36 assert(g != 0);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(nr == 8); // This kernel is for NR=8
40 assert(kr == 1);
41 assert(sr == 1);
42 assert(k_stride != 0);
43 assert(weights != NULL);
44 assert(packed_weights != NULL);
45 static const int32_t mask_table[16] = {
46 -1, -1, -1, -1, -1, -1, -1, -1,
47 0, 0, 0, 0, 0, 0, 0, 0,
48 };
49
50 const __m256 vzero = _mm256_setzero_ps();
51 const float* b = (const float*) bias;
52 float* packed_w = (float*) packed_weights;
53 do {
54 // NC main loop multiple of 8
55 const float* w = (const float*) weights;
56 size_t n = nc;
57
58 for (; n >= 8; n -= 8) {
59 if XNN_LIKELY(b != NULL) {
60 const __m256 vb0 = _mm256_loadu_ps(b + 0);
61 _mm256_store_ps(packed_w + 0, vb0);
62 b += 8;
63 } else {
64 _mm256_store_ps(packed_w + 0, vzero);
65 }
66 packed_w += 8;
67
68 size_t k = kc;
69
70 // KC remainder loop
71 for (; k > 0; --k) {
72 const __m256 v0 = _mm256_loadu_ps(w + 0);
73 xnn_prefetch_to_l1((const int8_t*) w + 960);
74 _mm256_store_ps(packed_w + 0, v0);
75 w += k_stride;
76 packed_w += 8;
77 }
78 w = w - kc * k_stride + 8; // Advance to next column of 8 floats
79 }
80
81 // NC remainder (1..7)
82 if XNN_UNLIKELY(n != 0) {
83 assert(n >= 1);
84 assert(n <= 7);
85 const __m256i vmask0 = _mm256_loadu_si256((const __m256i*) &mask_table[8 - n]);
86
87 if XNN_LIKELY(b != NULL) {
88 const __m256 vb0 = _mm256_maskload_ps(b + 0, vmask0);
89 _mm256_store_ps(packed_w + 0, vb0);
90 b += n;
91 } else {
92 _mm256_store_ps(packed_w + 0, vzero);
93 }
94 packed_w += 8;
95
96 // KC main loop
97 for (size_t k = kc; k > 0; --k) {
98 const __m256 v0 = _mm256_maskload_ps(w + 0, vmask0);
99 _mm256_store_ps(packed_w + 0, v0);
100 w += k_stride;
101 packed_w += 8;
102 }
103 }
104 weights += nc * kc;
105 } while (--g != 0);
106 }