Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-7x16c4-minmax-fp32-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_qs8_qc8w_igemm_minmax_fp32_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 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const union xnn_qs8_qc8w_conv_minmax_params 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 vintile[7 * 16];
47 __attribute__((aligned(64))) int32_t res0[7 * 16];
48
49 kc = round_up_po2(kc, 4 * sizeof(int8_t));
50 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
51 const __mmask16 kremainder_mask = _cvtu32_mask16((UINT32_C(1) << (kremainder >> 2)) - 1);
52
53 // Define tile config data structure
54 struct __tile_config {
55 uint8_t palette_id;
56 uint8_t start_row;
57 uint8_t reserved_0[14];
58 uint16_t colsb[8];
59 uint16_t reserved_1[8];
60 uint8_t rows[8];
61 uint8_t reserved_2[8];
62 };
63
64 // Load tile configuration
65 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
66 tile_data.palette_id = 1;
67 tile_data.rows[0] = mr; // tmm0 = res 0
68 tile_data.rows[1] = mr; // tmm1 = res 1
69 tile_data.rows[2] = mr; // tmm2 = res 2
70 tile_data.rows[3] = mr; // tmm3 = res 3
71 tile_data.rows[4] = mr; // tmm4 = input
72 tile_data.rows[5] = 16; // tmm5 = weights
73 tile_data.rows[6] = mr; // tmm6 = input remainder
74 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
75
76 tile_data.colsb[0] = 64; // tmm0 = res 0
77 tile_data.colsb[1] = 64; // tmm1 = res 1
78 tile_data.colsb[2] = 64; // tmm2 = res 1
79 tile_data.colsb[3] = 64; // tmm3 = res 1
80 tile_data.colsb[4] = 64; // tmm4 = input
81 tile_data.colsb[5] = 64; // tmm5 = weights
82 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
83 tile_data.colsb[7] = 64; // tmm7 = weights remainder
84
85 //_tile_loadconfig(&tile_data);
86 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
87
88 int8_t* c0 = c;
89 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
90 if XNN_UNPREDICTABLE(mr < 2) {
91 c1 = c0;
92 }
93 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
94 if XNN_UNPREDICTABLE(mr <= 2) {
95 c2 = c1;
96 }
97 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
98 if XNN_UNPREDICTABLE(mr < 4) {
99 c3 = c2;
100 }
101 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
102 if XNN_UNPREDICTABLE(mr <= 4) {
103 c4 = c3;
104 }
105 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
106 if XNN_UNPREDICTABLE(mr < 6) {
107 c5 = c4;
108 }
109 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
110 if XNN_UNPREDICTABLE(mr <= 6) {
111 c6 = c5;
112 }
113
114 const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
115 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
116 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
117 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
118 // XNN_FORCE_REALIZATION(voutput_zero_point);
119 // XNN_FORCE_REALIZATION(voutput_min);
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 }
136 const int8_t* restrict a1 = a[1];
137 if XNN_UNPREDICTABLE(a1 != zero) {
138 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
139 }
140 const int8_t* restrict a2 = a[2];
141 if XNN_UNPREDICTABLE(a2 != zero) {
142 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
143 }
144 const int8_t* restrict a3 = a[3];
145 if XNN_UNPREDICTABLE(a3 != zero) {
146 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
147 }
148 const int8_t* restrict a4 = a[4];
149 if XNN_UNPREDICTABLE(a4 != zero) {
150 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
151 }
152 const int8_t* restrict a5 = a[5];
153 if XNN_UNPREDICTABLE(a5 != zero) {
154 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
155 }
156 const int8_t* restrict a6 = a[6];
157 if XNN_UNPREDICTABLE(a6 != zero) {
158 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
159 }
160 a += 7;
161
162 size_t k = kc;
163 if (mr == 1)
164 {
165 while (k >= 64 * sizeof(int8_t)) {
166 _tile_loadd(4, a0, 64); // Directly load input for mr=1
167 a6 += 64;
168 _tile_loadd(5, (const int8_t*) w + 0, 64);
169 _tile_dpbssd(0, 4, 5);
170
171 w = (const int8_t*) w + 1024;
172 k -= 64 * sizeof(int8_t);
173 }
174 }
175 else {
176 while (k >= 64 * sizeof(int8_t)) {
177 const __m512i vin0 = _mm512_loadu_epi32(a0);
178 a0 += 64;
179 _mm512_store_epi32(vintile + 0, vin0);
180 const __m512i vin1 = _mm512_loadu_epi32(a1);
181 a1 += 64;
182 _mm512_store_epi32(vintile + 16, vin1);
183 const __m512i vin2 = _mm512_loadu_epi32(a2);
184 a2 += 64;
185 _mm512_store_epi32(vintile + 32, vin2);
186 const __m512i vin3 = _mm512_loadu_epi32(a3);
187 a3 += 64;
188 _mm512_store_epi32(vintile + 48, vin3);
189 const __m512i vin4 = _mm512_loadu_epi32(a4);
190 a4 += 64;
191 _mm512_store_epi32(vintile + 64, vin4);
192 const __m512i vin5 = _mm512_loadu_epi32(a5);
193 a5 += 64;
194 _mm512_store_epi32(vintile + 80, vin5);
195 const __m512i vin6 = _mm512_loadu_epi32(a6);
196 a6 += 64;
197 _mm512_store_epi32(vintile + 96, vin6);
198 _tile_loadd(4, vintile, 64);
199 _tile_loadd(5, (const int8_t*) w + 0, 64);
200 _tile_dpbssd(0, 4, 5);
201
202 w = (const int8_t*) w + 1024;
203 k -= 64 * sizeof(int8_t);
204 }
205 }
206
207 if XNN_UNLIKELY(k != 0) {
208 const __m512i vin0 = _mm512_maskz_loadu_epi32(kremainder_mask, a0);
209 a0 += kremainder;
210 _mm512_store_epi32(vintile + 0, vin0);
211 const __m512i vin1 = _mm512_maskz_loadu_epi32(kremainder_mask, a1);
212 a1 += kremainder;
213 _mm512_store_epi32(vintile + 16, vin1);
214 const __m512i vin2 = _mm512_maskz_loadu_epi32(kremainder_mask, a2);
215 a2 += kremainder;
216 _mm512_store_epi32(vintile + 32, vin2);
217 const __m512i vin3 = _mm512_maskz_loadu_epi32(kremainder_mask, a3);
218 a3 += kremainder;
219 _mm512_store_epi32(vintile + 48, vin3);
220 const __m512i vin4 = _mm512_maskz_loadu_epi32(kremainder_mask, a4);
221 a4 += kremainder;
222 _mm512_store_epi32(vintile + 64, vin4);
223 const __m512i vin5 = _mm512_maskz_loadu_epi32(kremainder_mask, a5);
224 a5 += kremainder;
225 _mm512_store_epi32(vintile + 80, vin5);
226 const __m512i vin6 = _mm512_maskz_loadu_epi32(kremainder_mask, a6);
227 a6 += kremainder;
228 _mm512_store_epi32(vintile + 96, vin6);
229 _tile_loadd(6, vintile, 64);
230 _tile_loadd(7, (const int8_t*) w + 0, 64);
231 _tile_dpbssd(0, 6, 7);
232
233 w = (const int8_t*) w + kremainder * 16;
234 k -= kremainder * sizeof(int8_t);
235 }
236
237 p -= 7 * sizeof(void*);
238 } while (p != 0);
239
240 // Add tile to bias
241 _tile_stored(0, res0, 64);
242
243 __m512i vacc0x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
244 __m512i vacc1x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
245 __m512i vacc2x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
246 __m512i vacc3x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
247 __m512i vacc4x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
248 __m512i vacc5x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
249 __m512i vacc6x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
250
251 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
252 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
253 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
254 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
255 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
256 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
257 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
258
259 const __m512 vscale0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
260 w = (const int32_t*) w + 16;
261
262 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale0123456789ABCDEF);
263 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale0123456789ABCDEF);
264 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale0123456789ABCDEF);
265 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale0123456789ABCDEF);
266 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale0123456789ABCDEF);
267 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale0123456789ABCDEF);
268 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale0123456789ABCDEF);
269
270 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
271 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
272 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
273 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
274 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
275 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
276 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
277
278 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
279 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
280 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
281 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
282 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
283 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
284 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
285
286 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
287 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
288 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
289 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
290 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
291 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
292 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
293
294 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
295 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
296 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
297 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
298 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
299 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
300 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
301
302 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
303 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
304 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
305 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
306 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
307 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
308 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
309
310 if XNN_LIKELY(nc >= 16) {
311 _mm_storeu_si128((__m128i*) (c6 + 0), vout6x0123456789ABCDEF);
312 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
313 _mm_storeu_si128((__m128i*) (c5 + 0), vout5x0123456789ABCDEF);
314 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
315 _mm_storeu_si128((__m128i*) (c4 + 0), vout4x0123456789ABCDEF);
316 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
317 _mm_storeu_si128((__m128i*) (c3 + 0), vout3x0123456789ABCDEF);
318 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
319 _mm_storeu_si128((__m128i*) (c2 + 0), vout2x0123456789ABCDEF);
320 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
321 _mm_storeu_si128((__m128i*) (c1 + 0), vout1x0123456789ABCDEF);
322 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
323 _mm_storeu_si128((__m128i*) (c0 + 0), vout0x0123456789ABCDEF);
324 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
325 a = (const int8_t**restrict) ((uintptr_t) a - ks);
326 nc -= 16;
327 } else {
328 // Prepare mask for valid 8-bit elements (depends on nc).
329 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
330
331 _mm_mask_storeu_epi8(c6 + 0, vmask0, vout6x0123456789ABCDEF);
332 _mm_mask_storeu_epi8(c5 + 0, vmask0, vout5x0123456789ABCDEF);
333 _mm_mask_storeu_epi8(c4 + 0, vmask0, vout4x0123456789ABCDEF);
334 _mm_mask_storeu_epi8(c3 + 0, vmask0, vout3x0123456789ABCDEF);
335 _mm_mask_storeu_epi8(c2 + 0, vmask0, vout2x0123456789ABCDEF);
336 _mm_mask_storeu_epi8(c1 + 0, vmask0, vout1x0123456789ABCDEF);
337 _mm_mask_storeu_epi8(c0 + 0, vmask0, vout0x0123456789ABCDEF);
338 nc = 0;
339 }
340 } while (nc != 0);
341
342 // Release tile config
343 // _tile_release();
344 __asm__ volatile ("tilerelease" ::);
345 #endif // defined(__x86_64__)
346 }