Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x32c4-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_1x32c4__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 <= 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 __attribute__((aligned(64))) int32_t res1[1 * 16];
48 __attribute__((aligned(64))) int8_t weight_buffer[16 * 64];
49
50 kc = round_up_po2(kc, 4 * sizeof(int8_t));
51 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
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 float* c0 = c;
89
90 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
91 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
92 // XNN_FORCE_REALIZATION(voutput_min);
93 // XNN_FORCE_REALIZATION(voutput_max);
94 const __m512i vmask = _mm512_set1_epi8(0xF0);
95 const __m512i vshl4 = _mm512_set1_epi64(0x01020408);
96 XNN_FORCE_REALIZATION(vmask);
97 XNN_FORCE_REALIZATION(vshl4);
98
99 do {
100 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32((const int32_t*) w + 0);
101 const __m512i vksumGHIJKLMNOPQRSTUV = _mm512_load_epi32((const int32_t*) w + 16);
102 w = (const int32_t*) w + 32;
103
104 // Zero tile accumulator
105 __asm__ volatile (
106 "tilezero %%tmm0\n"
107 "tilezero %%tmm1\n"
108 ::);
109
110 size_t k = kc;
111 while (k >= 64 * sizeof(int8_t)) {
112 _tile_loadd(4, a, a_stride);
113 a += 64;
114 const __m512i vb0x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 0);
115 const __m512i vb1x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 128);
116 const __m512i vb2x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 256);
117 const __m512i vb3x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 384);
118 const __m512i vb4x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 512);
119 const __m512i vb5x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 640);
120 const __m512i vb6x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 768);
121 const __m512i vb7x0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 896);
122 const __m512i vl0x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb0x0123456789ABCDEF, vshl4, 0);
123 const __m512i vh0x0123456789ABCDEF = _mm512_and_si512(vb0x0123456789ABCDEF, vmask);
124 const __m512i vl1x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb1x0123456789ABCDEF, vshl4, 0);
125 const __m512i vh1x0123456789ABCDEF = _mm512_and_si512(vb1x0123456789ABCDEF, vmask);
126 const __m512i vl2x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb2x0123456789ABCDEF, vshl4, 0);
127 const __m512i vh2x0123456789ABCDEF = _mm512_and_si512(vb2x0123456789ABCDEF, vmask);
128 const __m512i vl3x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb3x0123456789ABCDEF, vshl4, 0);
129 const __m512i vh3x0123456789ABCDEF = _mm512_and_si512(vb3x0123456789ABCDEF, vmask);
130 const __m512i vl4x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb4x0123456789ABCDEF, vshl4, 0);
131 const __m512i vh4x0123456789ABCDEF = _mm512_and_si512(vb4x0123456789ABCDEF, vmask);
132 const __m512i vl5x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb5x0123456789ABCDEF, vshl4, 0);
133 const __m512i vh5x0123456789ABCDEF = _mm512_and_si512(vb5x0123456789ABCDEF, vmask);
134 const __m512i vl6x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb6x0123456789ABCDEF, vshl4, 0);
135 const __m512i vh6x0123456789ABCDEF = _mm512_and_si512(vb6x0123456789ABCDEF, vmask);
136 const __m512i vl7x0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb7x0123456789ABCDEF, vshl4, 0);
137 const __m512i vh7x0123456789ABCDEF = _mm512_and_si512(vb7x0123456789ABCDEF, vmask);
138 _mm512_store_epi32(weight_buffer + 0, vl0x0123456789ABCDEF);
139 _mm512_store_epi32(weight_buffer + 64, vh0x0123456789ABCDEF);
140 _mm512_store_epi32(weight_buffer + 128, vl1x0123456789ABCDEF);
141 _mm512_store_epi32(weight_buffer + 192, vh1x0123456789ABCDEF);
142 _mm512_store_epi32(weight_buffer + 256, vl2x0123456789ABCDEF);
143 _mm512_store_epi32(weight_buffer + 320, vh2x0123456789ABCDEF);
144 _mm512_store_epi32(weight_buffer + 384, vl3x0123456789ABCDEF);
145 _mm512_store_epi32(weight_buffer + 448, vh3x0123456789ABCDEF);
146 _mm512_store_epi32(weight_buffer + 512, vl4x0123456789ABCDEF);
147 _mm512_store_epi32(weight_buffer + 576, vh4x0123456789ABCDEF);
148 _mm512_store_epi32(weight_buffer + 640, vl5x0123456789ABCDEF);
149 _mm512_store_epi32(weight_buffer + 704, vh5x0123456789ABCDEF);
150 _mm512_store_epi32(weight_buffer + 768, vl6x0123456789ABCDEF);
151 _mm512_store_epi32(weight_buffer + 832, vh6x0123456789ABCDEF);
152 _mm512_store_epi32(weight_buffer + 896, vl7x0123456789ABCDEF);
153 _mm512_store_epi32(weight_buffer + 960, vh7x0123456789ABCDEF);
154 _tile_loadd(5, weight_buffer, 64);
155 _tile_dpbssd(0, 4, 5);
156 const __m512i vb0xGHIJKLMNOPQRSTUV = _mm512_load_epi32((const int8_t*) w + 64);
157 const __m512i vb1xGHIJKLMNOPQRSTUV = _mm512_load_epi32((const int8_t*) w + 192);
158 const __m512i vb2xGHIJKLMNOPQRSTUV = _mm512_load_epi32((const int8_t*) w + 320);
159 const __m512i vb3xGHIJKLMNOPQRSTUV = _mm512_load_epi32((const int8_t*) w + 448);
160 const __m512i vb4xGHIJKLMNOPQRSTUV = _mm512_load_epi32((const int8_t*) w + 576);
161 const __m512i vb5xGHIJKLMNOPQRSTUV = _mm512_load_epi32((const int8_t*) w + 704);
162 const __m512i vb6xGHIJKLMNOPQRSTUV = _mm512_load_epi32((const int8_t*) w + 832);
163 const __m512i vb7xGHIJKLMNOPQRSTUV = _mm512_load_epi32((const int8_t*) w + 960);
164 const __m512i vl0xGHIJKLMNOPQRSTUV = _mm512_gf2p8affine_epi64_epi8(vb0xGHIJKLMNOPQRSTUV, vshl4, 0);
165 const __m512i vh0xGHIJKLMNOPQRSTUV = _mm512_and_si512(vb0xGHIJKLMNOPQRSTUV, vmask);
166 const __m512i vl1xGHIJKLMNOPQRSTUV = _mm512_gf2p8affine_epi64_epi8(vb1xGHIJKLMNOPQRSTUV, vshl4, 0);
167 const __m512i vh1xGHIJKLMNOPQRSTUV = _mm512_and_si512(vb1xGHIJKLMNOPQRSTUV, vmask);
168 const __m512i vl2xGHIJKLMNOPQRSTUV = _mm512_gf2p8affine_epi64_epi8(vb2xGHIJKLMNOPQRSTUV, vshl4, 0);
169 const __m512i vh2xGHIJKLMNOPQRSTUV = _mm512_and_si512(vb2xGHIJKLMNOPQRSTUV, vmask);
170 const __m512i vl3xGHIJKLMNOPQRSTUV = _mm512_gf2p8affine_epi64_epi8(vb3xGHIJKLMNOPQRSTUV, vshl4, 0);
171 const __m512i vh3xGHIJKLMNOPQRSTUV = _mm512_and_si512(vb3xGHIJKLMNOPQRSTUV, vmask);
172 const __m512i vl4xGHIJKLMNOPQRSTUV = _mm512_gf2p8affine_epi64_epi8(vb4xGHIJKLMNOPQRSTUV, vshl4, 0);
173 const __m512i vh4xGHIJKLMNOPQRSTUV = _mm512_and_si512(vb4xGHIJKLMNOPQRSTUV, vmask);
174 const __m512i vl5xGHIJKLMNOPQRSTUV = _mm512_gf2p8affine_epi64_epi8(vb5xGHIJKLMNOPQRSTUV, vshl4, 0);
175 const __m512i vh5xGHIJKLMNOPQRSTUV = _mm512_and_si512(vb5xGHIJKLMNOPQRSTUV, vmask);
176 const __m512i vl6xGHIJKLMNOPQRSTUV = _mm512_gf2p8affine_epi64_epi8(vb6xGHIJKLMNOPQRSTUV, vshl4, 0);
177 const __m512i vh6xGHIJKLMNOPQRSTUV = _mm512_and_si512(vb6xGHIJKLMNOPQRSTUV, vmask);
178 const __m512i vl7xGHIJKLMNOPQRSTUV = _mm512_gf2p8affine_epi64_epi8(vb7xGHIJKLMNOPQRSTUV, vshl4, 0);
179 const __m512i vh7xGHIJKLMNOPQRSTUV = _mm512_and_si512(vb7xGHIJKLMNOPQRSTUV, vmask);
180 _mm512_store_epi32(weight_buffer + 0, vl0xGHIJKLMNOPQRSTUV);
181 _mm512_store_epi32(weight_buffer + 64, vh0xGHIJKLMNOPQRSTUV);
182 _mm512_store_epi32(weight_buffer + 128, vl1xGHIJKLMNOPQRSTUV);
183 _mm512_store_epi32(weight_buffer + 192, vh1xGHIJKLMNOPQRSTUV);
184 _mm512_store_epi32(weight_buffer + 256, vl2xGHIJKLMNOPQRSTUV);
185 _mm512_store_epi32(weight_buffer + 320, vh2xGHIJKLMNOPQRSTUV);
186 _mm512_store_epi32(weight_buffer + 384, vl3xGHIJKLMNOPQRSTUV);
187 _mm512_store_epi32(weight_buffer + 448, vh3xGHIJKLMNOPQRSTUV);
188 _mm512_store_epi32(weight_buffer + 512, vl4xGHIJKLMNOPQRSTUV);
189 _mm512_store_epi32(weight_buffer + 576, vh4xGHIJKLMNOPQRSTUV);
190 _mm512_store_epi32(weight_buffer + 640, vl5xGHIJKLMNOPQRSTUV);
191 _mm512_store_epi32(weight_buffer + 704, vh5xGHIJKLMNOPQRSTUV);
192 _mm512_store_epi32(weight_buffer + 768, vl6xGHIJKLMNOPQRSTUV);
193 _mm512_store_epi32(weight_buffer + 832, vh6xGHIJKLMNOPQRSTUV);
194 _mm512_store_epi32(weight_buffer + 896, vl7xGHIJKLMNOPQRSTUV);
195 _mm512_store_epi32(weight_buffer + 960, vh7xGHIJKLMNOPQRSTUV);
196 _tile_loadd(5, weight_buffer, 64);
197 _tile_dpbssd(1, 4, 5);
198
199 w = (const int8_t*) w + 1024;
200 k -= 64 * sizeof(int8_t);
201 }
202
203 if XNN_UNLIKELY(k != 0) {
204 _tile_loadd(6, a, a_stride);
205 a += kremainder;
206 for (size_t k = 0; k < ((kremainder + 7) >> 3); ++k) {
207 const __m512i vb0123456789ABCDEF = _mm512_load_epi32((const int8_t*) w + 0 + 128 * k);
208 const __m512i vl0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vb0123456789ABCDEF, vshl4, 0);
209 const __m512i vh0123456789ABCDEF = _mm512_and_si512(vb0123456789ABCDEF, vmask);
210 _mm512_store_epi32(weight_buffer + 128 * k + 0, vl0123456789ABCDEF);
211 _mm512_store_epi32(weight_buffer + 128 * k + 64, vh0123456789ABCDEF);
212 }
213 _tile_loadd(7, weight_buffer, 64);
214 _tile_dpbssd(0, 6, 7);
215 for (size_t k = 0; k < ((kremainder + 7) >> 3); ++k) {
216 const __m512i vbGHIJKLMNOPQRSTUV = _mm512_load_epi32((const int8_t*) w + 64 + 128 * k);
217 const __m512i vlGHIJKLMNOPQRSTUV = _mm512_gf2p8affine_epi64_epi8(vbGHIJKLMNOPQRSTUV, vshl4, 0);
218 const __m512i vhGHIJKLMNOPQRSTUV = _mm512_and_si512(vbGHIJKLMNOPQRSTUV, vmask);
219 _mm512_store_epi32(weight_buffer + 128 * k + 0, vlGHIJKLMNOPQRSTUV);
220 _mm512_store_epi32(weight_buffer + 128 * k + 64, vhGHIJKLMNOPQRSTUV);
221 }
222 _tile_loadd(7, weight_buffer, 64);
223 _tile_dpbssd(1, 6, 7);
224
225 w = (const int8_t*) w + ((kremainder + 7) >> 3) * 128;
226 k -= kremainder * sizeof(int8_t);
227 }
228
229 // Add tile to bias
230 _tile_stored(0, res0, 64);
231 _tile_stored(1, res1, 64);
232
233 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params[0].zero_point));
234 __m512i vacc0xGHIJKLMNOPQRSTUV = _mm512_mullo_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_set1_epi32((int) quantization_params[0].zero_point));
235 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
236 vacc0xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc0xGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 0));
237
238 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
239 vacc0xGHIJKLMNOPQRSTUV = _mm512_srai_epi32(vacc0xGHIJKLMNOPQRSTUV, 4);
240 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
241 __m512 vscaled0xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc0xGHIJKLMNOPQRSTUV);
242
243 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
244 vscaled0xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled0xGHIJKLMNOPQRSTUV, _mm512_set1_ps(quantization_params[0].inv_scale));
245
246 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
247 const __m512 vfilter_output_scaleGHIJKLMNOPQRSTUV = _mm512_load_ps((const float*) w + 16);
248 w = (const int32_t*) w + 32;
249 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 0);
250 const __m512 vbiasGHIJKLMNOPQRSTUV = _mm512_load_ps((const float*) w + 16);
251 w = (const int32_t*) w + 32;
252
253 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
254 vscaled0xGHIJKLMNOPQRSTUV = _mm512_fmadd_ps(vscaled0xGHIJKLMNOPQRSTUV, vfilter_output_scaleGHIJKLMNOPQRSTUV, vbiasGHIJKLMNOPQRSTUV);
255
256 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
257 vscaled0xGHIJKLMNOPQRSTUV = _mm512_max_ps(vscaled0xGHIJKLMNOPQRSTUV, voutput_min);
258
259 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
260 vscaled0xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled0xGHIJKLMNOPQRSTUV, voutput_max);
261
262 if XNN_LIKELY(nc >= 32) {
263 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
264 _mm512_storeu_ps(c0 + 16, vscaled0xGHIJKLMNOPQRSTUV);
265 c0 = (float*) ((uintptr_t) c0 + cn_stride);
266
267 a -= kc;
268 nc -= 32;
269 } else {
270 // Prepare mask for valid 32-bit elements (depends on nc).
271 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
272 const __mmask16 vmask1 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 16) & 0xFFFF));
273 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
274 _mm512_mask_storeu_ps(c0 + 16, vmask1, vscaled0xGHIJKLMNOPQRSTUV);
275 nc = 0;
276 }
277 } while (nc != 0);
278 // Release tile config
279 // _tile_release();
280 __asm__ volatile ("tilerelease" ::);
281 #endif // defined(__x86_64__)
282 }