Debian Codegraph @

src/x32-packw/gen/x32-packw-x8-gemm-gio-avx-u8-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_u8_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 // KC main loop 8x8
70 for (; k >= 8; k -= 8) {
71 const __m256 v0_0 = _mm256_loadu_ps(w + 0 + 0 * k_stride);
72 const __m256 v0_1 = _mm256_loadu_ps(w + 0 + 1 * k_stride);
73 const __m256 v0_2 = _mm256_loadu_ps(w + 0 + 2 * k_stride);
74 const __m256 v0_3 = _mm256_loadu_ps(w + 0 + 3 * k_stride);
75 const __m256 v0_4 = _mm256_loadu_ps(w + 0 + 4 * k_stride);
76 const __m256 v0_5 = _mm256_loadu_ps(w + 0 + 5 * k_stride);
77 const __m256 v0_6 = _mm256_loadu_ps(w + 0 + 6 * k_stride);
78 const __m256 v0_7 = _mm256_loadu_ps(w + 0 + 7 * k_stride);
79 xnn_prefetch_to_l1((const int8_t*) w + 960);
80 xnn_prefetch_to_l1((const int8_t*) w + 960);
81 xnn_prefetch_to_l1((const int8_t*) w + 960);
82 xnn_prefetch_to_l1((const int8_t*) w + 960);
83 xnn_prefetch_to_l1((const int8_t*) w + 960);
84 xnn_prefetch_to_l1((const int8_t*) w + 960);
85 xnn_prefetch_to_l1((const int8_t*) w + 960);
86 xnn_prefetch_to_l1((const int8_t*) w + 960);
87 _mm256_store_ps(packed_w + 0, v0_0);
88 _mm256_store_ps(packed_w + 8, v0_1);
89 _mm256_store_ps(packed_w + 16, v0_2);
90 _mm256_store_ps(packed_w + 24, v0_3);
91 _mm256_store_ps(packed_w + 32, v0_4);
92 _mm256_store_ps(packed_w + 40, v0_5);
93 _mm256_store_ps(packed_w + 48, v0_6);
94 _mm256_store_ps(packed_w + 56, v0_7);
95 w += k_stride * 8;
96 packed_w += 64;
97 }
98
99 // KC remainder loop
100 for (; k > 0; --k) {
101 const __m256 v0 = _mm256_loadu_ps(w + 0);
102 xnn_prefetch_to_l1((const int8_t*) w + 960);
103 _mm256_store_ps(packed_w + 0, v0);
104 w += k_stride;
105 packed_w += 8;
106 }
107 w = w - kc * k_stride + 8; // Advance to next column of 8 floats
108 }
109
110 // NC remainder (1..7)
111 if XNN_UNLIKELY(n != 0) {
112 assert(n >= 1);
113 assert(n <= 7);
114 const __m256i vmask0 = _mm256_loadu_si256((const __m256i*) &mask_table[8 - n]);
115
116 if XNN_LIKELY(b != NULL) {
117 const __m256 vb0 = _mm256_maskload_ps(b + 0, vmask0);
118 _mm256_store_ps(packed_w + 0, vb0);
119 b += n;
120 } else {
121 _mm256_store_ps(packed_w + 0, vzero);
122 }
123 packed_w += 8;
124
125 // KC main loop
126 for (size_t k = kc; k > 0; --k) {
127 const __m256 v0 = _mm256_maskload_ps(w + 0, vmask0);
128 _mm256_store_ps(packed_w + 0, v0);
129 w += k_stride;
130 packed_w += 8;
131 }
132 }
133 weights += nc * kc;
134 } while (--g != 0);
135 }