Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-7x16c4-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_7x16c4__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 <= 7);
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[7 * 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 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
88 if XNN_UNPREDICTABLE(mr < 2) {
89 c1 = c0;
90 }
91 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
92 if XNN_UNPREDICTABLE(mr <= 2) {
93 c2 = c1;
94 }
95 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
96 if XNN_UNPREDICTABLE(mr < 4) {
97 c3 = c2;
98 }
99 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
100 if XNN_UNPREDICTABLE(mr <= 4) {
101 c4 = c3;
102 }
103 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
104 if XNN_UNPREDICTABLE(mr < 6) {
105 c5 = c4;
106 }
107 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
108 if XNN_UNPREDICTABLE(mr <= 6) {
109 c6 = c5;
110 }
111
112 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
113 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
114 // XNN_FORCE_REALIZATION(voutput_min);
115 // XNN_FORCE_REALIZATION(voutput_max);
116
117 do {
118 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32((const int32_t*) w + 0);
119 w = (const int32_t*) w + 16;
120
121 // Zero tile accumulator
122 __asm__ volatile (
123 "tilezero %%tmm0\n"
124 ::);
125
126 size_t k = kc;
127 while (k >= 64 * sizeof(int8_t)) {
128 _tile_loadd(4, a, a_stride);
129 a += 64;
130 _tile_loadd(5, (const int8_t*) w + 0, 64);
131 _tile_dpbssd(0, 4, 5);
132
133 w = (const int8_t*) w + 1024;
134 k -= 64 * sizeof(int8_t);
135 }
136
137 if XNN_UNLIKELY(k != 0) {
138 _tile_loadd(6, a, a_stride);
139 a += kremainder;
140 _tile_loadd(7, (const int8_t*) w + 0, 64);
141 _tile_dpbssd(0, 6, 7);
142
143 w = (const int8_t*) w + kremainder * 16;
144 k -= kremainder * sizeof(int8_t);
145 }
146
147 // Add tile to bias
148 _tile_stored(0, res0, 64);
149
150 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[0].zero_point));
151 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[1].zero_point));
152 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[2].zero_point));
153 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[3].zero_point));
154 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[4].zero_point));
155 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[5].zero_point));
156 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[6].zero_point));
157 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
158 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
159 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
160 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
161 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
162 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
163 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
164
165 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
166 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
167 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
168 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
169 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
170 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
171 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
172
173 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
174 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
175 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
176 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
177 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
178 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
179 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
180
181 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
182 w = (const int32_t*) w + 16;
183 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
184 w = (const int32_t*) w + 16;
185
186 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
187 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
188 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
189 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
190 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
191 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
192 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
193
194 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
195 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
196 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
197 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
198 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
199 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
200 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
201
202 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
203 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
204 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
205 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
206 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
207 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
208 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
209
210 if XNN_LIKELY(nc >= 16) {
211 _mm512_storeu_ps(c6 + 0, vscaled6x0123456789ABCDEF);
212 c6 = (float*) ((uintptr_t) c6 + cn_stride);
213 _mm512_storeu_ps(c5 + 0, vscaled5x0123456789ABCDEF);
214 c5 = (float*) ((uintptr_t) c5 + cn_stride);
215 _mm512_storeu_ps(c4 + 0, vscaled4x0123456789ABCDEF);
216 c4 = (float*) ((uintptr_t) c4 + cn_stride);
217 _mm512_storeu_ps(c3 + 0, vscaled3x0123456789ABCDEF);
218 c3 = (float*) ((uintptr_t) c3 + cn_stride);
219 _mm512_storeu_ps(c2 + 0, vscaled2x0123456789ABCDEF);
220 c2 = (float*) ((uintptr_t) c2 + cn_stride);
221 _mm512_storeu_ps(c1 + 0, vscaled1x0123456789ABCDEF);
222 c1 = (float*) ((uintptr_t) c1 + cn_stride);
223 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
224 c0 = (float*) ((uintptr_t) c0 + cn_stride);
225
226 a -= kc;
227 nc -= 16;
228 } else {
229 // Prepare mask for valid 32-bit elements (depends on nc).
230 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
231 _mm512_mask_storeu_ps(c6 + 0, vmask0, vscaled6x0123456789ABCDEF);
232 _mm512_mask_storeu_ps(c5 + 0, vmask0, vscaled5x0123456789ABCDEF);
233 _mm512_mask_storeu_ps(c4 + 0, vmask0, vscaled4x0123456789ABCDEF);
234 _mm512_mask_storeu_ps(c3 + 0, vmask0, vscaled3x0123456789ABCDEF);
235 _mm512_mask_storeu_ps(c2 + 0, vmask0, vscaled2x0123456789ABCDEF);
236 _mm512_mask_storeu_ps(c1 + 0, vmask0, vscaled1x0123456789ABCDEF);
237 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
238 nc = 0;
239 }
240 } while (nc != 0);
241 // Release tile config
242 // _tile_release();
243 __asm__ volatile ("tilerelease" ::);
244 #endif // defined(__x86_64__)
245 }