Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-4x4c8-minmax-avx-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 #ifdef _MSC_VER
13 #include <intrin.h>
14 #else
15 #include <x86intrin.h>
16 #endif
17
18 #include "xnnpack/common.h"
19 #include "xnnpack/gemm.h"
20 #include "xnnpack/math.h"
21 #include "xnnpack/unaligned.h"
22
23
24 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_4x4c8__avx_ld64(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 const int8_t* restrict a,
29 size_t a_stride,
30 const void* restrict w,
31 float* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 4);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 const int8_t* a0 = a;
48 float* c0 = c;
49 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
50 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
56 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 2) {
58 a2 = a1;
59 c2 = c1;
60 }
61 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
62 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
63 if XNN_UNPREDICTABLE(mr != 4) {
64 a3 = a2;
65 c3 = c2;
66 }
67
68 const __m128 vmin = _mm_set1_ps(params->scalar.min);
69 const __m128 vmax = _mm_set1_ps(params->scalar.max);
70 XNN_FORCE_REALIZATION(vmin);
71 XNN_FORCE_REALIZATION(vmax);
72
73 const __m128i vmask = _mm_set1_epi8(0xF0);
74 XNN_FORCE_REALIZATION(vmask);
75
76 do {
77 const __m128i vksum = _mm_load_si128((const __m128i*) w);
78 const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
79 const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
80 const __m128i vinput_zero_point2 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[2].zero_point));
81 const __m128i vinput_zero_point3 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[3].zero_point));
82 const __m128i vzero = _mm_setzero_si128();
83 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
84 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point1);
85 const __m128i vinit2 = _mm_mullo_epi32(vksum, vinput_zero_point2);
86 const __m128i vinit3 = _mm_mullo_epi32(vksum, vinput_zero_point3);
87 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
88 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
89 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
90 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
91 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
92 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
93 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
94 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
95 __m128i vacc2x0 = _mm_blend_epi16(vinit2, vzero, 0xFC);
96 __m128i vacc2x1 = _mm_blend_epi16(vinit2, vzero, 0xF3);
97 __m128i vacc2x2 = _mm_blend_epi16(vinit2, vzero, 0xCF);
98 __m128i vacc2x3 = _mm_blend_epi16(vinit2, vzero, 0x3F);
99 __m128i vacc3x0 = _mm_blend_epi16(vinit3, vzero, 0xFC);
100 __m128i vacc3x1 = _mm_blend_epi16(vinit3, vzero, 0xF3);
101 __m128i vacc3x2 = _mm_blend_epi16(vinit3, vzero, 0xCF);
102 __m128i vacc3x3 = _mm_blend_epi16(vinit3, vzero, 0x3F);
103 w = (const int32_t*) w + 4;
104
105 size_t k = kc;
106
107 while (k >= 16 * sizeof(int8_t)) {
108 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
109 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
110 a0 += 8;
111 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
112 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
113 a1 += 8;
114 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
115 const __m128i vxa2c0 = _mm_cvtepi8_epi16(va2c0);
116 a2 += 8;
117 const __m128i va3c0 = _mm_loadl_epi64((const __m128i*) a3);
118 const __m128i vxa3c0 = _mm_cvtepi8_epi16(va3c0);
119 a3 += 8;
120
121 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
122 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
123 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
124 const __m128i vxb0c0 = _mm_cvtepi8_epi16(vb0c0);
125
126 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
127 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
128 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
129 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c0, vxb0c0));
130 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
131 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
132 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
133 const __m128i vxb1c0 = _mm_cvtepi8_epi16(vb1c0);
134
135 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
136 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
137 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
138 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c0, vxb1c0));
139 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
140 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
141 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
142 const __m128i vxb2c0 = _mm_cvtepi8_epi16(vb2c0);
143
144 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
145 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
146 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
147 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c0, vxb2c0));
148 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
149 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
150 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
151 const __m128i vxb3c0 = _mm_cvtepi8_epi16(vb3c0);
152
153 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
154 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
155 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
156 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c0, vxb3c0));
157
158 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
159 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
160 a0 += 8;
161 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
162 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
163 a1 += 8;
164 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
165 const __m128i vxa2c1 = _mm_cvtepi8_epi16(va2c1);
166 a2 += 8;
167 const __m128i va3c1 = _mm_loadl_epi64((const __m128i*) a3);
168 const __m128i vxa3c1 = _mm_cvtepi8_epi16(va3c1);
169 a3 += 8;
170
171 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
172 const __m128i vxb0c1 = _mm_cvtepi8_epi16(vb0c1);
173
174 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
175 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
176 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
177 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c1, vxb0c1));
178 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
179 const __m128i vxb1c1 = _mm_cvtepi8_epi16(vb1c1);
180
181 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
182 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
183 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
184 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c1, vxb1c1));
185 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
186 const __m128i vxb2c1 = _mm_cvtepi8_epi16(vb2c1);
187
188 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
189 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
190 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
191 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c1, vxb2c1));
192 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
193 const __m128i vxb3c1 = _mm_cvtepi8_epi16(vb3c1);
194
195 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
196 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
197 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
198 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c1, vxb3c1));
199
200
201 w = (const int8_t*) w + 32;
202 k -= 16 * sizeof(int8_t);
203 }
204
205 while (k >= 8 * sizeof(int8_t)) {
206 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
207 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
208 a0 += 8;
209 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
210 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
211 a1 += 8;
212 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
213 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
214 a2 += 8;
215 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
216 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
217 a3 += 8;
218
219 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
220 vb0 = _mm_slli_epi32(vb0, 4);
221 vb0 = _mm_and_si128(vb0, vmask);
222
223 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
224
225 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
226 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
227 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
228 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3, vxb0));
229 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
230 vb1 = _mm_slli_epi32(vb1, 4);
231 vb1 = _mm_and_si128(vb1, vmask);
232
233 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
234
235 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
236 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
237 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
238 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3, vxb1));
239 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
240 vb2 = _mm_slli_epi32(vb2, 4);
241 vb2 = _mm_and_si128(vb2, vmask);
242
243 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
244
245 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
246 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
247 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
248 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3, vxb2));
249 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
250 vb3 = _mm_slli_epi32(vb3, 4);
251 vb3 = _mm_and_si128(vb3, vmask);
252
253 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
254
255 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
256 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
257 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
258 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3, vxb3));
259
260 w = (const int8_t*) w + 32;
261 k -= 8 * sizeof(int8_t);
262 }
263
264 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
265 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
266 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
267 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
268 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
269 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
270 const __m128i vacc3x01 = _mm_hadd_epi32(vacc3x0, vacc3x1);
271 const __m128i vacc3x23 = _mm_hadd_epi32(vacc3x2, vacc3x3);
272
273 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
274 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
275 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
276 __m128i vacc3x0123 = _mm_hadd_epi32(vacc3x01, vacc3x23);
277
278 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
279 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
280 vacc2x0123 = _mm_srai_epi32(vacc2x0123, 4);
281 vacc3x0123 = _mm_srai_epi32(vacc3x0123, 4);
282 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
283 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
284 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
285 __m128 vout3x0123 = _mm_cvtepi32_ps(vacc3x0123);
286
287 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
288 const __m128 vinput_scale1 = _mm_broadcast_ss(&quantization_params[1].inv_scale);
289 const __m128 vinput_scale2 = _mm_broadcast_ss(&quantization_params[2].inv_scale);
290 const __m128 vinput_scale3 = _mm_broadcast_ss(&quantization_params[3].inv_scale);
291
292 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
293 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
294 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
295 vout3x0123 = _mm_mul_ps(vout3x0123, vinput_scale3);
296
297 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
298 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
299 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
300 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
301 vout3x0123 = _mm_mul_ps(vout3x0123, vfilter_output_scale0123);
302
303 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
304 w = (const float*) w + 8;
305 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
306 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
307 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
308 vout3x0123 = _mm_add_ps(vout3x0123, vbias0123);
309
310 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
311 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
312 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
313 vout3x0123 = _mm_max_ps(vout3x0123, vmin);
314
315 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
316 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
317 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
318 vout3x0123 = _mm_min_ps(vout3x0123, vmax);
319
320 if XNN_LIKELY(nc >= 4) {
321 _mm_storeu_ps(c0, vout0x0123);
322 _mm_storeu_ps(c1, vout1x0123);
323 _mm_storeu_ps(c2, vout2x0123);
324 _mm_storeu_ps(c3, vout3x0123);
325
326 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
327 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
328 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
329 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
330
331 c0 = (float*) ((uintptr_t) c0 + cn_stride);
332 c1 = (float*) ((uintptr_t) c1 + cn_stride);
333 c2 = (float*) ((uintptr_t) c2 + cn_stride);
334 c3 = (float*) ((uintptr_t) c3 + cn_stride);
335
336 nc -= 4;
337 } else {
338 if (nc & 2) {
339 _mm_storel_pi((__m64*) c0, vout0x0123);
340 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
341 c0 += 2;
342 _mm_storel_pi((__m64*) c1, vout1x0123);
343 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
344 c1 += 2;
345 _mm_storel_pi((__m64*) c2, vout2x0123);
346 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
347 c2 += 2;
348 _mm_storel_pi((__m64*) c3, vout3x0123);
349 vout3x0123 = _mm_unpackhi_ps(vout3x0123, vout3x0123);
350 c3 += 2;
351 }
352 if (nc & 1) {
353 _mm_store_ss(c0, vout0x0123);
354 _mm_store_ss(c1, vout1x0123);
355 _mm_store_ss(c2, vout2x0123);
356 _mm_store_ss(c3, vout3x0123);
357 }
358 nc = 0;
359 }
360 } while (nc != 0);
361 }