Debian Codegraph @

src/x32-packw/gen/x32-packw-x16-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_x16__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 == 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
74 // KC remainder loop
75 for (; k > 0; --k) {
76 const __m256 v0 = _mm256_loadu_ps(w + 0);
77 const __m256 v8 = _mm256_loadu_ps(w + 8);
78 _mm256_store_ps(packed_w + 0, v0);
79 _mm256_store_ps(packed_w + 8, v8);
80 w += k_stride;
81 packed_w += 16;
82 }
83 w = w - kc * k_stride + 16; // Advance to next column of 16 floats
84 }
85
86 // NC remainder (1..15)
87 if XNN_UNLIKELY(n != 0) {
88 assert(n >= 1);
89 assert(n <= 15);
90 const __m256i vmask0 = _mm256_loadu_si256((const __m256i*) &mask_table[16 - n]);
91 const __m256i vmask8 = _mm256_loadu_si256((const __m256i*) &mask_table[24 - n]);
92
93 if XNN_LIKELY(b != NULL) {
94 const __m256 vb0 = _mm256_maskload_ps(b + 0, vmask0);
95 const __m256 vb8 = _mm256_maskload_ps(b + 8, vmask8);
96 _mm256_store_ps(packed_w + 0, vb0);
97 _mm256_store_ps(packed_w + 8, vb8);
98 b += n;
99 } else {
100 _mm256_store_ps(packed_w + 0, vzero);
101 _mm256_store_ps(packed_w + 8, vzero);
102 }
103 packed_w += 16;
104
105 // KC main loop
106 for (size_t k = kc; k > 0; --k) {
107 const __m256 v0 = _mm256_maskload_ps(w + 0, vmask0);
108 const __m256 v8 = _mm256_maskload_ps(w + 8, vmask8);
109 _mm256_store_ps(packed_w + 0, v0);
110 _mm256_store_ps(packed_w + 8, v8);
111 w += k_stride;
112 packed_w += 16;
113 }
114 }
115 weights += nc * kc;
116 } while (--g != 0);
117 }