Debian Codegraph @

src/x32-packw/gen/x32-packw-x16-gemm-gio-avx-u8.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_x16__avx_u8(
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 == 16); // This kernel is for NR=16
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[32] = {
45 -1, -1, -1, -1, -1, -1, -1, -1,
46 -1, -1, -1, -1, -1, -1, -1, -1,
47 0, 0, 0, 0, 0, 0, 0, 0,
48 0, 0, 0, 0, 0, 0, 0, 0,
49 };
50
51 const __m256 vzero = _mm256_setzero_ps();
52 const float* b = (const float*) bias;
53 float* packed_w = (float*) packed_weights;
54 do {
55 // NC main loop multiple of 16
56 const float* w = (const float*) weights;
57 size_t n = nc;
58
59 for (; n >= 16; n -= 16) {
60 if XNN_LIKELY(b != NULL) {
61 const __m256 vb0 = _mm256_loadu_ps(b + 0);
62 const __m256 vb8 = _mm256_loadu_ps(b + 8);
63 _mm256_store_ps(packed_w + 0, vb0);
64 _mm256_store_ps(packed_w + 8, vb8);
65 b += 16;
66 } else {
67 _mm256_store_ps(packed_w + 0, vzero);
68 _mm256_store_ps(packed_w + 8, vzero);
69 }
70 packed_w += 16;
71
72 size_t k = kc;
73 // KC main loop 8x16
74 for (; k >= 8; k -= 8) {
75 const __m256 v0_0 = _mm256_loadu_ps(w + 0 + 0 * k_stride);
76 const __m256 v8_0 = _mm256_loadu_ps(w + 8 + 0 * k_stride);
77 const __m256 v0_1 = _mm256_loadu_ps(w + 0 + 1 * k_stride);
78 const __m256 v8_1 = _mm256_loadu_ps(w + 8 + 1 * k_stride);
79 const __m256 v0_2 = _mm256_loadu_ps(w + 0 + 2 * k_stride);
80 const __m256 v8_2 = _mm256_loadu_ps(w + 8 + 2 * k_stride);
81 const __m256 v0_3 = _mm256_loadu_ps(w + 0 + 3 * k_stride);
82 const __m256 v8_3 = _mm256_loadu_ps(w + 8 + 3 * k_stride);
83 const __m256 v0_4 = _mm256_loadu_ps(w + 0 + 4 * k_stride);
84 const __m256 v8_4 = _mm256_loadu_ps(w + 8 + 4 * k_stride);
85 const __m256 v0_5 = _mm256_loadu_ps(w + 0 + 5 * k_stride);
86 const __m256 v8_5 = _mm256_loadu_ps(w + 8 + 5 * k_stride);
87 const __m256 v0_6 = _mm256_loadu_ps(w + 0 + 6 * k_stride);
88 const __m256 v8_6 = _mm256_loadu_ps(w + 8 + 6 * k_stride);
89 const __m256 v0_7 = _mm256_loadu_ps(w + 0 + 7 * k_stride);
90 const __m256 v8_7 = _mm256_loadu_ps(w + 8 + 7 * k_stride);
91 _mm256_store_ps(packed_w + 0, v0_0);
92 _mm256_store_ps(packed_w + 8, v8_0);
93 _mm256_store_ps(packed_w + 16, v0_1);
94 _mm256_store_ps(packed_w + 24, v8_1);
95 _mm256_store_ps(packed_w + 32, v0_2);
96 _mm256_store_ps(packed_w + 40, v8_2);
97 _mm256_store_ps(packed_w + 48, v0_3);
98 _mm256_store_ps(packed_w + 56, v8_3);
99 _mm256_store_ps(packed_w + 64, v0_4);
100 _mm256_store_ps(packed_w + 72, v8_4);
101 _mm256_store_ps(packed_w + 80, v0_5);
102 _mm256_store_ps(packed_w + 88, v8_5);
103 _mm256_store_ps(packed_w + 96, v0_6);
104 _mm256_store_ps(packed_w + 104, v8_6);
105 _mm256_store_ps(packed_w + 112, v0_7);
106 _mm256_store_ps(packed_w + 120, v8_7);
107 w += k_stride * 8;
108 packed_w += 128;
109 }
110
111 // KC remainder loop
112 for (; k > 0; --k) {
113 const __m256 v0 = _mm256_loadu_ps(w + 0);
114 const __m256 v8 = _mm256_loadu_ps(w + 8);
115 _mm256_store_ps(packed_w + 0, v0);
116 _mm256_store_ps(packed_w + 8, v8);
117 w += k_stride;
118 packed_w += 16;
119 }
120 w = w - kc * k_stride + 16; // Advance to next column of 16 floats
121 }
122
123 // NC remainder (1..15)
124 if XNN_UNLIKELY(n != 0) {
125 assert(n >= 1);
126 assert(n <= 15);
127 const __m256i vmask0 = _mm256_loadu_si256((const __m256i*) &mask_table[16 - n]);
128 const __m256i vmask8 = _mm256_loadu_si256((const __m256i*) &mask_table[24 - n]);
129
130 if XNN_LIKELY(b != NULL) {
131 const __m256 vb0 = _mm256_maskload_ps(b + 0, vmask0);
132 const __m256 vb8 = _mm256_maskload_ps(b + 8, vmask8);
133 _mm256_store_ps(packed_w + 0, vb0);
134 _mm256_store_ps(packed_w + 8, vb8);
135 b += n;
136 } else {
137 _mm256_store_ps(packed_w + 0, vzero);
138 _mm256_store_ps(packed_w + 8, vzero);
139 }
140 packed_w += 16;
141
142 // KC main loop
143 for (size_t k = kc; k > 0; --k) {
144 const __m256 v0 = _mm256_maskload_ps(w + 0, vmask0);
145 const __m256 v8 = _mm256_maskload_ps(w + 8, vmask8);
146 _mm256_store_ps(packed_w + 0, v0);
147 _mm256_store_ps(packed_w + 8, v8);
148 w += k_stride;
149 packed_w += 16;
150 }
151 }
152 weights += nc * kc;
153 } while (--g != 0);
154 }