Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-7x16c4-minmax-avx512amx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_7x16c4__avx512amx(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)])
35 {
36 assert(mr != 0);
37 assert(mr <= 7);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 // TODO: amxintrin.h only provide intrinsics for __x86_64__
46 // Update if amxintrin changes
47 #if defined(__x86_64__)
48 __attribute__((aligned(64))) int32_t vintile[7 * 16];
49 __attribute__((aligned(64))) int32_t res0[7 * 16];
50
51 kc = round_up_po2(kc, 4 * sizeof(int8_t));
52 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
53 const __mmask16 kremainder_mask = _cvtu32_mask16((UINT32_C(1) << (kremainder >> 2)) - 1);
54
55 // Define tile config data structure
56 struct __tile_config {
57 uint8_t palette_id;
58 uint8_t start_row;
59 uint8_t reserved_0[14];
60 uint16_t colsb[8];
61 uint16_t reserved_1[8];
62 uint8_t rows[8];
63 uint8_t reserved_2[8];
64 };
65
66 // Load tile configuration
67 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
68 tile_data.palette_id = 1;
69 tile_data.rows[0] = mr; // tmm0 = res 0
70 tile_data.rows[1] = mr; // tmm1 = res 1
71 tile_data.rows[2] = mr; // tmm2 = res 2
72 tile_data.rows[3] = mr; // tmm3 = res 3
73 tile_data.rows[4] = mr; // tmm4 = input
74 tile_data.rows[5] = 16; // tmm5 = weights
75 tile_data.rows[6] = mr; // tmm6 = input remainder
76 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
77
78 tile_data.colsb[0] = 64; // tmm0 = res 0
79 tile_data.colsb[1] = 64; // tmm1 = res 1
80 tile_data.colsb[2] = 64; // tmm2 = res 1
81 tile_data.colsb[3] = 64; // tmm3 = res 1
82 tile_data.colsb[4] = 64; // tmm4 = input
83 tile_data.colsb[5] = 64; // tmm5 = weights
84 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
85 tile_data.colsb[7] = 64; // tmm7 = weights remainder
86
87 //_tile_loadconfig(&tile_data);
88 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
89
90 float* c0 = c;
91 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
92 if XNN_UNPREDICTABLE(mr < 2) {
93 c1 = c0;
94 }
95 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
96 if XNN_UNPREDICTABLE(mr <= 2) {
97 c2 = c1;
98 }
99 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
100 if XNN_UNPREDICTABLE(mr < 4) {
101 c3 = c2;
102 }
103 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
104 if XNN_UNPREDICTABLE(mr <= 4) {
105 c4 = c3;
106 }
107 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
108 if XNN_UNPREDICTABLE(mr < 6) {
109 c5 = c4;
110 }
111 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
112 if XNN_UNPREDICTABLE(mr <= 6) {
113 c6 = c5;
114 }
115
116 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
117 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
118 // XNN_FORCE_REALIZATION(voutput_min);
119 // XNN_FORCE_REALIZATION(voutput_max);
120
121 do {
122 const __m512i vksum0123456789ABCDEF = _mm512_loadu_epi32((const int32_t*) w + 0);
123 w = (const int32_t*) w + 16;
124
125 // Zero tile accumulator
126 __asm__ volatile (
127 "tilezero %%tmm0\n"
128 ::);
129
130 size_t p = ks;
131 do {
132 const int8_t* restrict a0 = a[0];
133 if XNN_UNPREDICTABLE(a0 != zero) {
134 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
135 } else {
136 a0 = zero_data;
137 }
138 const int8_t* restrict a1 = a[1];
139 if XNN_UNPREDICTABLE(a1 != zero) {
140 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
141 } else {
142 a1 = zero_data;
143 }
144 const int8_t* restrict a2 = a[2];
145 if XNN_UNPREDICTABLE(a2 != zero) {
146 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
147 } else {
148 a2 = zero_data;
149 }
150 const int8_t* restrict a3 = a[3];
151 if XNN_UNPREDICTABLE(a3 != zero) {
152 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
153 } else {
154 a3 = zero_data;
155 }
156 const int8_t* restrict a4 = a[4];
157 if XNN_UNPREDICTABLE(a4 != zero) {
158 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
159 } else {
160 a4 = zero_data;
161 }
162 const int8_t* restrict a5 = a[5];
163 if XNN_UNPREDICTABLE(a5 != zero) {
164 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
165 } else {
166 a5 = zero_data;
167 }
168 const int8_t* restrict a6 = a[6];
169 if XNN_UNPREDICTABLE(a6 != zero) {
170 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
171 } else {
172 a6 = zero_data;
173 }
174 a += 7;
175
176 size_t k = kc;
177 if (mr == 1)
178 {
179 while (k >= 64 * sizeof(int8_t)) {
180 _tile_loadd(4, a0, 64); // Directly load input for mr=1
181 a6 += 64;
182 _tile_loadd(5, (const int8_t*) w + 0, 64);
183 _tile_dpbssd(0, 4, 5);
184
185 w = (const int8_t*) w + 1024;
186 k -= 64 * sizeof(int8_t);
187 }
188 }
189 else {
190 while (k >= 64 * sizeof(int8_t)) {
191 const __m512i vin0 = _mm512_loadu_epi32(a0);
192 a0 += 64;
193 _mm512_store_epi32(vintile + 0, vin0);
194 const __m512i vin1 = _mm512_loadu_epi32(a1);
195 a1 += 64;
196 _mm512_store_epi32(vintile + 16, vin1);
197 const __m512i vin2 = _mm512_loadu_epi32(a2);
198 a2 += 64;
199 _mm512_store_epi32(vintile + 32, vin2);
200 const __m512i vin3 = _mm512_loadu_epi32(a3);
201 a3 += 64;
202 _mm512_store_epi32(vintile + 48, vin3);
203 const __m512i vin4 = _mm512_loadu_epi32(a4);
204 a4 += 64;
205 _mm512_store_epi32(vintile + 64, vin4);
206 const __m512i vin5 = _mm512_loadu_epi32(a5);
207 a5 += 64;
208 _mm512_store_epi32(vintile + 80, vin5);
209 const __m512i vin6 = _mm512_loadu_epi32(a6);
210 a6 += 64;
211 _mm512_store_epi32(vintile + 96, vin6);
212 _tile_loadd(4, vintile, 64);
213 _tile_loadd(5, (const int8_t*) w + 0, 64);
214 _tile_dpbssd(0, 4, 5);
215
216 w = (const int8_t*) w + 1024;
217 k -= 64 * sizeof(int8_t);
218 }
219 }
220
221 if XNN_UNLIKELY(k != 0) {
222 const __m512i vin0 = _mm512_maskz_loadu_epi32(kremainder_mask, a0);
223 a0 += kremainder;
224 _mm512_store_epi32(vintile + 0, vin0);
225 const __m512i vin1 = _mm512_maskz_loadu_epi32(kremainder_mask, a1);
226 a1 += kremainder;
227 _mm512_store_epi32(vintile + 16, vin1);
228 const __m512i vin2 = _mm512_maskz_loadu_epi32(kremainder_mask, a2);
229 a2 += kremainder;
230 _mm512_store_epi32(vintile + 32, vin2);
231 const __m512i vin3 = _mm512_maskz_loadu_epi32(kremainder_mask, a3);
232 a3 += kremainder;
233 _mm512_store_epi32(vintile + 48, vin3);
234 const __m512i vin4 = _mm512_maskz_loadu_epi32(kremainder_mask, a4);
235 a4 += kremainder;
236 _mm512_store_epi32(vintile + 64, vin4);
237 const __m512i vin5 = _mm512_maskz_loadu_epi32(kremainder_mask, a5);
238 a5 += kremainder;
239 _mm512_store_epi32(vintile + 80, vin5);
240 const __m512i vin6 = _mm512_maskz_loadu_epi32(kremainder_mask, a6);
241 a6 += kremainder;
242 _mm512_store_epi32(vintile + 96, vin6);
243 _tile_loadd(6, vintile, 64);
244 _tile_loadd(7, (const int8_t*) w + 0, 64);
245 _tile_dpbssd(0, 6, 7);
246
247 w = (const int8_t*) w + kremainder * 16;
248 k -= kremainder * sizeof(int8_t);
249 }
250
251 p -= 7 * sizeof(void*);
252 } while (p != 0);
253
254 // Add tile to bias
255 _tile_stored(0, res0, 64);
256
257 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
258 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
259 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
260 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
261 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
262 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
263 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
264 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
265 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
266 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
267 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
268 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
269 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
270 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
271
272 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
273 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
274 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
275 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
276 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
277 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
278 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
279
280 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
281 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
282 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
283 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
284 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
285 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
286 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
287
288 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
289 w = (const int32_t*) w + 16;
290 const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
291 w = (const int32_t*) w + 16;
292
293 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
294 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
295 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
296 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
297 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
298 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
299 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
300
301 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
302 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
303 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
304 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
305 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
306 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
307 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
308
309 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
310 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
311 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
312 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
313 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
314 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
315 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
316
317 if XNN_LIKELY(nc >= 16) {
318 _mm512_storeu_ps(c6 + 0, vscaled6x0123456789ABCDEF);
319 c6 = (float*) ((uintptr_t) c6 + cn_stride);
320 _mm512_storeu_ps(c5 + 0, vscaled5x0123456789ABCDEF);
321 c5 = (float*) ((uintptr_t) c5 + cn_stride);
322 _mm512_storeu_ps(c4 + 0, vscaled4x0123456789ABCDEF);
323 c4 = (float*) ((uintptr_t) c4 + cn_stride);
324 _mm512_storeu_ps(c3 + 0, vscaled3x0123456789ABCDEF);
325 c3 = (float*) ((uintptr_t) c3 + cn_stride);
326 _mm512_storeu_ps(c2 + 0, vscaled2x0123456789ABCDEF);
327 c2 = (float*) ((uintptr_t) c2 + cn_stride);
328 _mm512_storeu_ps(c1 + 0, vscaled1x0123456789ABCDEF);
329 c1 = (float*) ((uintptr_t) c1 + cn_stride);
330 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
331 c0 = (float*) ((uintptr_t) c0 + cn_stride);
332
333 a = (const int8_t**restrict) ((uintptr_t) a - ks);
334 nc -= 16;
335 } else {
336 // Prepare mask for valid 32-bit elements (depends on nc).
337 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
338 _mm512_mask_storeu_ps(c6 + 0, vmask0, vscaled6x0123456789ABCDEF);
339 _mm512_mask_storeu_ps(c5 + 0, vmask0, vscaled5x0123456789ABCDEF);
340 _mm512_mask_storeu_ps(c4 + 0, vmask0, vscaled4x0123456789ABCDEF);
341 _mm512_mask_storeu_ps(c3 + 0, vmask0, vscaled3x0123456789ABCDEF);
342 _mm512_mask_storeu_ps(c2 + 0, vmask0, vscaled2x0123456789ABCDEF);
343 _mm512_mask_storeu_ps(c1 + 0, vmask0, vscaled1x0123456789ABCDEF);
344 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
345 nc = 0;
346 }
347 } while (nc != 0);
348
349 // Release tile config
350 // _tile_release();
351 __asm__ volatile ("tilerelease" ::);
352 #endif // defined(__x86_64__)
353 }