Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x16c4-minmax-avx512amx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-avx512amx.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
12 #include <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19
20
21 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_1x16c4__avx512amx(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)])
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 // TODO: amxintrin.h only provide intrinsics for __x86_64__
44 // Update if amxintrin changes
45 #if defined(__x86_64__)
46 __attribute__((aligned(64))) int32_t res0[1 * 16];
47
48 kc = round_up_po2(kc, 4 * sizeof(int8_t));
49 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
50
51 // Define tile config data structure
52 struct __tile_config {
53 uint8_t palette_id;
54 uint8_t start_row;
55 uint8_t reserved_0[14];
56 uint16_t colsb[8];
57 uint16_t reserved_1[8];
58 uint8_t rows[8];
59 uint8_t reserved_2[8];
60 };
61
62 // Load tile configuration
63 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
64 tile_data.palette_id = 1;
65 tile_data.rows[0] = mr; // tmm0 = res 0
66 tile_data.rows[1] = mr; // tmm1 = res 1
67 tile_data.rows[2] = mr; // tmm2 = res 2
68 tile_data.rows[3] = mr; // tmm3 = res 3
69 tile_data.rows[4] = mr; // tmm4 = input
70 tile_data.rows[5] = 16; // tmm5 = weights
71 tile_data.rows[6] = mr; // tmm6 = input remainder
72 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
73
74 tile_data.colsb[0] = 64; // tmm0 = res 0
75 tile_data.colsb[1] = 64; // tmm1 = res 1
76 tile_data.colsb[2] = 64; // tmm2 = res 1
77 tile_data.colsb[3] = 64; // tmm3 = res 1
78 tile_data.colsb[4] = 64; // tmm4 = input
79 tile_data.colsb[5] = 64; // tmm5 = weights
80 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
81 tile_data.colsb[7] = 64; // tmm7 = weights remainder
82
83 //_tile_loadconfig(&tile_data);
84 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
85
86 float* c0 = c;
87
88 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
89 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
90 // XNN_FORCE_REALIZATION(voutput_min);
91 // XNN_FORCE_REALIZATION(voutput_max);
92
93 do {
94 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32((const int32_t*) w + 0);
95 w = (const int32_t*) w + 16;
96
97 // Zero tile accumulator
98 __asm__ volatile (
99 "tilezero %%tmm0\n"
100 ::);
101
102 size_t k = kc;
103 while (k >= 64 * sizeof(int8_t)) {
104 _tile_loadd(4, a, a_stride);
105 a += 64;
106 _tile_loadd(5, (const int8_t*) w + 0, 64);
107 _tile_dpbssd(0, 4, 5);
108
109 w = (const int8_t*) w + 1024;
110 k -= 64 * sizeof(int8_t);
111 }
112
113 if XNN_UNLIKELY(k != 0) {
114 _tile_loadd(6, a, a_stride);
115 a += kremainder;
116 _tile_loadd(7, (const int8_t*) w + 0, 64);
117 _tile_dpbssd(0, 6, 7);
118
119 w = (const int8_t*) w + kremainder * 16;
120 k -= kremainder * sizeof(int8_t);
121 }
122
123 // Add tile to bias
124 _tile_stored(0, res0, 64);
125
126 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[0].zero_point));
127 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
128
129 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
130
131 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
132
133 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
134 w = (const int32_t*) w + 16;
135 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
136 w = (const int32_t*) w + 16;
137
138 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
139
140 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
141
142 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
143
144 if XNN_LIKELY(nc >= 16) {
145 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
146 c0 = (float*) ((uintptr_t) c0 + cn_stride);
147
148 a -= kc;
149 nc -= 16;
150 } else {
151 // Prepare mask for valid 32-bit elements (depends on nc).
152 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
153 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
154 nc = 0;
155 }
156 } while (nc != 0);
157 // Release tile config
158 // _tile_release();
159 __asm__ volatile ("tilerelease" ::);
160 #endif // defined(__x86_64__)
161 }