Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-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_qc4w_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 struct xnn_f32_qc4w_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 __attribute__((aligned(64))) int8_t weight_buffer[16 * 64];
48
49 kc = round_up_po2(kc, 4 * sizeof(int8_t));
50 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
51
52 // Define tile config data structure
53 struct __tile_config {
54 uint8_t palette_id;
55 uint8_t start_row;
56 uint8_t reserved_0[14];
57 uint16_t colsb[8];
58 uint16_t reserved_1[8];
59 uint8_t rows[8];
60 uint8_t reserved_2[8];
61 };
62
63 // Load tile configuration
64 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
65 tile_data.palette_id = 1;
66 tile_data.rows[0] = mr; // tmm0 = res 0
67 tile_data.rows[1] = mr; // tmm1 = res 1
68 tile_data.rows[2] = mr; // tmm2 = res 2
69 tile_data.rows[3] = mr; // tmm3 = res 3
70 tile_data.rows[4] = mr; // tmm4 = input
71 tile_data.rows[5] = 16; // tmm5 = weights
72 tile_data.rows[6] = mr; // tmm6 = input remainder
73 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
74
75 tile_data.colsb[0] = 64; // tmm0 = res 0
76 tile_data.colsb[1] = 64; // tmm1 = res 1
77 tile_data.colsb[2] = 64; // tmm2 = res 1
78 tile_data.colsb[3] = 64; // tmm3 = res 1
79 tile_data.colsb[4] = 64; // tmm4 = input
80 tile_data.colsb[5] = 64; // tmm5 = weights
81 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
82 tile_data.colsb[7] = 64; // tmm7 = weights remainder
83
84 //_tile_loadconfig(&tile_data);
85 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
86
87 float* c0 = c;
88 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
89 if XNN_UNPREDICTABLE(mr < 2) {
90 c1 = c0;
91 }
92 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
93 if XNN_UNPREDICTABLE(mr <= 2) {
94 c2 = c1;
95 }
96 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
97 if XNN_UNPREDICTABLE(mr < 4) {
98 c3 = c2;
99 }
100 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
101 if XNN_UNPREDICTABLE(mr <= 4) {
102 c4 = c3;
103 }
104 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
105 if XNN_UNPREDICTABLE(mr < 6) {
106 c5 = c4;
107 }
108 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
109 if XNN_UNPREDICTABLE(mr <= 6) {
110 c6 = c5;
111 }
112
113 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
114 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
115 // XNN_FORCE_REALIZATION(voutput_min);
116 // XNN_FORCE_REALIZATION(voutput_max);
117 const __m512i vmask = _mm512_set1_epi8(0xF0);
118 const __m512i vshl4 = _mm512_set1_epi64(0x01020408);
119 XNN_FORCE_REALIZATION(vmask);
120 XNN_FORCE_REALIZATION(vshl4);
121
122 do {
123 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32((const int32_t*) w + 0);
124 w = (const int32_t*) w + 16;
125
126 // Zero tile accumulator
127 __asm__ volatile (
128 "tilezero %%tmm0\n"
129 ::);
130
131 size_t k = kc;
132 while (k >= 64 * sizeof(int8_t)) {
133 _tile_loadd(4, a, a_stride);
134 a += 64;
135 const __m512i vb0x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 0);
136 const __m512i vb1x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 64);
137 const __m512i vb2x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 128);
138 const __m512i vb3x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 192);
139 const __m512i vb4x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 256);
140 const __m512i vb5x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 320);
141 const __m512i vb6x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 384);
142 const __m512i vb7x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 448);
143 const __m512i vl0x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb0x0123456789ABCDEF, vshl4, 0);
144 const __m512i vh0x0123456789ABCDEF = _mm512_and_si512(vb0x0123456789ABCDEF, vmask);
145 const __m512i vl1x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb1x0123456789ABCDEF, vshl4, 0);
146 const __m512i vh1x0123456789ABCDEF = _mm512_and_si512(vb1x0123456789ABCDEF, vmask);
147 const __m512i vl2x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb2x0123456789ABCDEF, vshl4, 0);
148 const __m512i vh2x0123456789ABCDEF = _mm512_and_si512(vb2x0123456789ABCDEF, vmask);
149 const __m512i vl3x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb3x0123456789ABCDEF, vshl4, 0);
150 const __m512i vh3x0123456789ABCDEF = _mm512_and_si512(vb3x0123456789ABCDEF, vmask);
151 const __m512i vl4x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb4x0123456789ABCDEF, vshl4, 0);
152 const __m512i vh4x0123456789ABCDEF = _mm512_and_si512(vb4x0123456789ABCDEF, vmask);
153 const __m512i vl5x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb5x0123456789ABCDEF, vshl4, 0);
154 const __m512i vh5x0123456789ABCDEF = _mm512_and_si512(vb5x0123456789ABCDEF, vmask);
155 const __m512i vl6x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb6x0123456789ABCDEF, vshl4, 0);
156 const __m512i vh6x0123456789ABCDEF = _mm512_and_si512(vb6x0123456789ABCDEF, vmask);
157 const __m512i vl7x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb7x0123456789ABCDEF, vshl4, 0);
158 const __m512i vh7x0123456789ABCDEF = _mm512_and_si512(vb7x0123456789ABCDEF, vmask);
159 _mm512_store_epi32(weight_buffer + 0, vl0x0123456789ABCDEF);
160 _mm512_store_epi32(weight_buffer + 64, vh0x0123456789ABCDEF);
161 _mm512_store_epi32(weight_buffer + 128, vl1x0123456789ABCDEF);
162 _mm512_store_epi32(weight_buffer + 192, vh1x0123456789ABCDEF);
163 _mm512_store_epi32(weight_buffer + 256, vl2x0123456789ABCDEF);
164 _mm512_store_epi32(weight_buffer + 320, vh2x0123456789ABCDEF);
165 _mm512_store_epi32(weight_buffer + 384, vl3x0123456789ABCDEF);
166 _mm512_store_epi32(weight_buffer + 448, vh3x0123456789ABCDEF);
167 _mm512_store_epi32(weight_buffer + 512, vl4x0123456789ABCDEF);
168 _mm512_store_epi32(weight_buffer + 576, vh4x0123456789ABCDEF);
169 _mm512_store_epi32(weight_buffer + 640, vl5x0123456789ABCDEF);
170 _mm512_store_epi32(weight_buffer + 704, vh5x0123456789ABCDEF);
171 _mm512_store_epi32(weight_buffer + 768, vl6x0123456789ABCDEF);
172 _mm512_store_epi32(weight_buffer + 832, vh6x0123456789ABCDEF);
173 _mm512_store_epi32(weight_buffer + 896, vl7x0123456789ABCDEF);
174 _mm512_store_epi32(weight_buffer + 960, vh7x0123456789ABCDEF);
175 _tile_loadd(5, weight_buffer, 64);
176 _tile_dpbssd(0, 4, 5);
177
178 w = (const int8_t*) w + 512;
179 k -= 64 * sizeof(int8_t);
180 }
181
182 if XNN_UNLIKELY(k != 0) {
183 _tile_loadd(6, a, a_stride);
184 a += kremainder;
185 for (size_t k = 0; k < ((kremainder + 7) >> 3); ++k) {
186 const __m512i vb0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 0 + 64 * k);
187 const __m512i vl0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb0123456789ABCDEF, vshl4, 0);
188 const __m512i vh0123456789ABCDEF = _mm512_and_si512(vb0123456789ABCDEF, vmask);
189 _mm512_store_epi32(weight_buffer + 128 * k + 0, vl0123456789ABCDEF);
190 _mm512_store_epi32(weight_buffer + 128 * k + 64, vh0123456789ABCDEF);
191 }
192 _tile_loadd(7, weight_buffer, 64);
193 _tile_dpbssd(0, 6, 7);
194
195 w = (const int8_t*) w + ((kremainder + 7) >> 3) * 64;
196 k -= kremainder * sizeof(int8_t);
197 }
198
199 // Add tile to bias
200 _tile_stored(0, res0, 64);
201
202 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[0].zero_point));
203 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[1].zero_point));
204 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[2].zero_point));
205 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[3].zero_point));
206 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[4].zero_point));
207 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[5].zero_point));
208 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[6].zero_point));
209 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
210 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
211 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
212 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
213 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
214 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
215 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
216
217 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
218 vacc1x0123456789ABCDEF = _mm512_srai_epi32(vacc1x0123456789ABCDEF, 4);
219 vacc2x0123456789ABCDEF = _mm512_srai_epi32(vacc2x0123456789ABCDEF, 4);
220 vacc3x0123456789ABCDEF = _mm512_srai_epi32(vacc3x0123456789ABCDEF, 4);
221 vacc4x0123456789ABCDEF = _mm512_srai_epi32(vacc4x0123456789ABCDEF, 4);
222 vacc5x0123456789ABCDEF = _mm512_srai_epi32(vacc5x0123456789ABCDEF, 4);
223 vacc6x0123456789ABCDEF = _mm512_srai_epi32(vacc6x0123456789ABCDEF, 4);
224 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
225 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
226 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
227 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
228 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
229 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
230 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
231
232 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
233 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
234 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
235 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
236 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
237 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
238 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
239
240 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
241 w = (const int32_t*) w + 16;
242 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
243 w = (const int32_t*) w + 16;
244
245 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
246 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
247 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
248 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
249 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
250 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
251 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
252
253 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
254 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
255 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
256 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
257 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
258 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
259 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
260
261 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
262 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
263 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
264 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
265 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
266 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
267 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
268
269 if XNN_LIKELY(nc >= 16) {
270 _mm512_storeu_ps(c6 + 0, vscaled6x0123456789ABCDEF);
271 c6 = (float*) ((uintptr_t) c6 + cn_stride);
272 _mm512_storeu_ps(c5 + 0, vscaled5x0123456789ABCDEF);
273 c5 = (float*) ((uintptr_t) c5 + cn_stride);
274 _mm512_storeu_ps(c4 + 0, vscaled4x0123456789ABCDEF);
275 c4 = (float*) ((uintptr_t) c4 + cn_stride);
276 _mm512_storeu_ps(c3 + 0, vscaled3x0123456789ABCDEF);
277 c3 = (float*) ((uintptr_t) c3 + cn_stride);
278 _mm512_storeu_ps(c2 + 0, vscaled2x0123456789ABCDEF);
279 c2 = (float*) ((uintptr_t) c2 + cn_stride);
280 _mm512_storeu_ps(c1 + 0, vscaled1x0123456789ABCDEF);
281 c1 = (float*) ((uintptr_t) c1 + cn_stride);
282 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
283 c0 = (float*) ((uintptr_t) c0 + cn_stride);
284
285 a -= kc;
286 nc -= 16;
287 } else {
288 // Prepare mask for valid 32-bit elements (depends on nc).
289 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
290 _mm512_mask_storeu_ps(c6 + 0, vmask0, vscaled6x0123456789ABCDEF);
291 _mm512_mask_storeu_ps(c5 + 0, vmask0, vscaled5x0123456789ABCDEF);
292 _mm512_mask_storeu_ps(c4 + 0, vmask0, vscaled4x0123456789ABCDEF);
293 _mm512_mask_storeu_ps(c3 + 0, vmask0, vscaled3x0123456789ABCDEF);
294 _mm512_mask_storeu_ps(c2 + 0, vmask0, vscaled2x0123456789ABCDEF);
295 _mm512_mask_storeu_ps(c1 + 0, vmask0, vscaled1x0123456789ABCDEF);
296 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
297 nc = 0;
298 }
299 } while (nc != 0);
300 // Release tile config
301 // _tile_release();
302 __asm__ volatile ("tilerelease" ::);
303 #endif // defined(__x86_64__)
304 }