Debian Codegraph @

src/x32-packw/gen/x32-packw-x32-gemm-gio-avx-u1.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
19
20 void xnn_x32_packw_gemm_gio_ukernel_x32__avx_u1(
21 size_t g,
22 size_t nc,
23 size_t kc,
24 size_t nr,
25 size_t kr,
26 size_t sr,
27 size_t k_stride,
28 const uint32_t* weights,
29 const uint32_t* bias,
30 const void* scale,
31 uint32_t* packed_weights,
32 size_t extra_bytes,
33 const void* params)
34 {
35 assert(g != 0);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(nr == 32); // This kernel is for NR=32
39 assert(kr == 1);
40 assert(sr == 1);
41 assert(k_stride != 0);
42 assert(weights != NULL);
43 assert(packed_weights != NULL);
44 static const int32_t mask_table[64] = {
45 -1, -1, -1, -1, -1, -1, -1, -1,
46 -1, -1, -1, -1, -1, -1, -1, -1,
47 -1, -1, -1, -1, -1, -1, -1, -1,
48 -1, -1, -1, -1, -1, -1, -1, -1,
49 0, 0, 0, 0, 0, 0, 0, 0,
50 0, 0, 0, 0, 0, 0, 0, 0,
51 0, 0, 0, 0, 0, 0, 0, 0,
52 0, 0, 0, 0, 0, 0, 0, 0,
53 };
54
55 const __m256 vzero = _mm256_setzero_ps();
56 const float* b = (const float*) bias;
57 float* packed_w = (float*) packed_weights;
58 do {
59 // NC main loop multiple of 32
60 const float* w = (const float*) weights;
61 size_t n = nc;
62
63 for (; n >= 32; n -= 32) {
64 if XNN_LIKELY(b != NULL) {
65 const __m256 vb0 = _mm256_loadu_ps(b + 0);
66 const __m256 vb8 = _mm256_loadu_ps(b + 8);
67 const __m256 vb16 = _mm256_loadu_ps(b + 16);
68 const __m256 vb24 = _mm256_loadu_ps(b + 24);
69 _mm256_store_ps(packed_w + 0, vb0);
70 _mm256_store_ps(packed_w + 8, vb8);
71 _mm256_store_ps(packed_w + 16, vb16);
72 _mm256_store_ps(packed_w + 24, vb24);
73 b += 32;
74 } else {
75 _mm256_store_ps(packed_w + 0, vzero);
76 _mm256_store_ps(packed_w + 8, vzero);
77 _mm256_store_ps(packed_w + 16, vzero);
78 _mm256_store_ps(packed_w + 24, vzero);
79 }
80 packed_w += 32;
81
82 size_t k = kc;
83
84 // KC remainder loop
85 for (; k > 0; --k) {
86 const __m256 v0 = _mm256_loadu_ps(w + 0);
87 const __m256 v8 = _mm256_loadu_ps(w + 8);
88 const __m256 v16 = _mm256_loadu_ps(w + 16);
89 const __m256 v24 = _mm256_loadu_ps(w + 24);
90 _mm256_store_ps(packed_w + 0, v0);
91 _mm256_store_ps(packed_w + 8, v8);
92 _mm256_store_ps(packed_w + 16, v16);
93 _mm256_store_ps(packed_w + 24, v24);
94 w += k_stride;
95 packed_w += 32;
96 }
97 w = w - kc * k_stride + 32; // Advance to next column of 32 floats
98 }
99
100 // NC remainder (1..31)
101 if XNN_UNLIKELY(n != 0) {
102 assert(n >= 1);
103 assert(n <= 31);
104 const __m256i vmask0 = _mm256_loadu_si256((const __m256i*) &mask_table[32 - n]);
105 const __m256i vmask8 = _mm256_loadu_si256((const __m256i*) &mask_table[40 - n]);
106 const __m256i vmask16 = _mm256_loadu_si256((const __m256i*) &mask_table[48 - n]);
107 const __m256i vmask24 = _mm256_loadu_si256((const __m256i*) &mask_table[56 - n]);
108
109 if XNN_LIKELY(b != NULL) {
110 const __m256 vb0 = _mm256_maskload_ps(b + 0, vmask0);
111 const __m256 vb8 = _mm256_maskload_ps(b + 8, vmask8);
112 const __m256 vb16 = _mm256_maskload_ps(b + 16, vmask16);
113 const __m256 vb24 = _mm256_maskload_ps(b + 24, vmask24);
114 _mm256_store_ps(packed_w + 0, vb0);
115 _mm256_store_ps(packed_w + 8, vb8);
116 _mm256_store_ps(packed_w + 16, vb16);
117 _mm256_store_ps(packed_w + 24, vb24);
118 b += n;
119 } else {
120 _mm256_store_ps(packed_w + 0, vzero);
121 _mm256_store_ps(packed_w + 8, vzero);
122 _mm256_store_ps(packed_w + 16, vzero);
123 _mm256_store_ps(packed_w + 24, vzero);
124 }
125 packed_w += 32;
126
127 // KC main loop
128 for (size_t k = kc; k > 0; --k) {
129 const __m256 v0 = _mm256_maskload_ps(w + 0, vmask0);
130 const __m256 v8 = _mm256_maskload_ps(w + 8, vmask8);
131 const __m256 v16 = _mm256_maskload_ps(w + 16, vmask16);
132 const __m256 v24 = _mm256_maskload_ps(w + 24, vmask24);
133 _mm256_store_ps(packed_w + 0, v0);
134 _mm256_store_ps(packed_w + 8, v8);
135 _mm256_store_ps(packed_w + 16, v16);
136 _mm256_store_ps(packed_w + 24, v24);
137 w += k_stride;
138 packed_w += 32;
139 }
140 }
141 weights += nc * kc;
142 } while (--g != 0);
143 }