Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-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_qb4w_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_qb4w_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 size_t bl = params->scalar.blocksize;
47 assert(bl <= round_up_po2(kc, 2));
48 assert(bl != 0);
49 assert(bl % 32 == 0);
50 kc = round_up_po2(kc, 8 * sizeof(int8_t));
51 const int8_t* a0 = a;
52 float* c0 = c;
53 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
54 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 2) {
56 a1 = a0;
57 c1 = c0;
58 }
59 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
60 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 2) {
62 a2 = a1;
63 c2 = c1;
64 }
65 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
66 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
67 if XNN_UNPREDICTABLE(mr != 4) {
68 a3 = a2;
69 c3 = c2;
70 }
71
72 const __m128 vmin = _mm_set1_ps(params->scalar.min);
73 const __m128 vmax = _mm_set1_ps(params->scalar.max);
74 XNN_FORCE_REALIZATION(vmin);
75 XNN_FORCE_REALIZATION(vmax);
76
77 const __m128i vmask = _mm_set1_epi8(0xF0);
78 XNN_FORCE_REALIZATION(vmask);
79
80 do {
81 const __m128 vksum = _mm_load_ps((const float*) w);
82 const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
83 const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
84 const __m128i vinput_zero_point2 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[2].zero_point));
85 const __m128i vinput_zero_point3 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[3].zero_point));
86
87 __m128 vinput_zero_point0_float = _mm_cvtepi32_ps(vinput_zero_point0);
88 __m128 vout0x0123 = _mm_mul_ps(vksum, vinput_zero_point0_float);
89 __m128 vinput_zero_point1_float = _mm_cvtepi32_ps(vinput_zero_point1);
90 __m128 vout1x0123 = _mm_mul_ps(vksum, vinput_zero_point1_float);
91 __m128 vinput_zero_point2_float = _mm_cvtepi32_ps(vinput_zero_point2);
92 __m128 vout2x0123 = _mm_mul_ps(vksum, vinput_zero_point2_float);
93 __m128 vinput_zero_point3_float = _mm_cvtepi32_ps(vinput_zero_point3);
94 __m128 vout3x0123 = _mm_mul_ps(vksum, vinput_zero_point3_float);
95 w = (const int32_t*) w + 4;
96
97 for (size_t kb=0; kb < kc; kb += bl) {
98 __m128i vacc0x0 = _mm_setzero_si128();
99 __m128i vacc0x1 = _mm_setzero_si128();
100 __m128i vacc0x2 = _mm_setzero_si128();
101 __m128i vacc0x3 = _mm_setzero_si128();
102 __m128i vacc1x0 = _mm_setzero_si128();
103 __m128i vacc1x1 = _mm_setzero_si128();
104 __m128i vacc1x2 = _mm_setzero_si128();
105 __m128i vacc1x3 = _mm_setzero_si128();
106 __m128i vacc2x0 = _mm_setzero_si128();
107 __m128i vacc2x1 = _mm_setzero_si128();
108 __m128i vacc2x2 = _mm_setzero_si128();
109 __m128i vacc2x3 = _mm_setzero_si128();
110 __m128i vacc3x0 = _mm_setzero_si128();
111 __m128i vacc3x1 = _mm_setzero_si128();
112 __m128i vacc3x2 = _mm_setzero_si128();
113 __m128i vacc3x3 = _mm_setzero_si128();
114 size_t k = bl;
115
116 while (k >= 16 * sizeof(int8_t)) {
117 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
118 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
119 a0 += 8;
120 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
121 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
122 a1 += 8;
123 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
124 const __m128i vxa2c0 = _mm_cvtepi8_epi16(va2c0);
125 a2 += 8;
126 const __m128i va3c0 = _mm_loadl_epi64((const __m128i*) a3);
127 const __m128i vxa3c0 = _mm_cvtepi8_epi16(va3c0);
128 a3 += 8;
129
130 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
131 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
132 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
133 const __m128i vxb0c0 = _mm_cvtepi8_epi16(vb0c0);
134
135 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
136 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
137 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
138 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c0, vxb0c0));
139 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
140 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
141 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
142 const __m128i vxb1c0 = _mm_cvtepi8_epi16(vb1c0);
143
144 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
145 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
146 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
147 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c0, vxb1c0));
148 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
149 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
150 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
151 const __m128i vxb2c0 = _mm_cvtepi8_epi16(vb2c0);
152
153 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
154 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
155 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
156 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c0, vxb2c0));
157 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
158 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
159 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
160 const __m128i vxb3c0 = _mm_cvtepi8_epi16(vb3c0);
161
162 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
163 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
164 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
165 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c0, vxb3c0));
166
167 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
168 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
169 a0 += 8;
170 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
171 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
172 a1 += 8;
173 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
174 const __m128i vxa2c1 = _mm_cvtepi8_epi16(va2c1);
175 a2 += 8;
176 const __m128i va3c1 = _mm_loadl_epi64((const __m128i*) a3);
177 const __m128i vxa3c1 = _mm_cvtepi8_epi16(va3c1);
178 a3 += 8;
179
180 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
181 const __m128i vxb0c1 = _mm_cvtepi8_epi16(vb0c1);
182
183 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
184 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
185 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
186 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c1, vxb0c1));
187 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
188 const __m128i vxb1c1 = _mm_cvtepi8_epi16(vb1c1);
189
190 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
191 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
192 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
193 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c1, vxb1c1));
194 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
195 const __m128i vxb2c1 = _mm_cvtepi8_epi16(vb2c1);
196
197 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
198 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
199 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
200 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c1, vxb2c1));
201 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
202 const __m128i vxb3c1 = _mm_cvtepi8_epi16(vb3c1);
203
204 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
205 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
206 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
207 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c1, vxb3c1));
208
209
210 w = (const int8_t*) w + 32;
211 k -= 16 * sizeof(int8_t);
212 }
213
214 while (k >= 8 * sizeof(int8_t)) {
215 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
216 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
217 a0 += 8;
218 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
219 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
220 a1 += 8;
221 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
222 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
223 a2 += 8;
224 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
225 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
226 a3 += 8;
227
228 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
229 vb0 = _mm_slli_epi32(vb0, 4);
230 vb0 = _mm_and_si128(vb0, vmask);
231
232 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
233
234 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
235 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
236 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
237 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3, vxb0));
238 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
239 vb1 = _mm_slli_epi32(vb1, 4);
240 vb1 = _mm_and_si128(vb1, vmask);
241
242 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
243
244 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
245 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
246 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
247 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3, vxb1));
248 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
249 vb2 = _mm_slli_epi32(vb2, 4);
250 vb2 = _mm_and_si128(vb2, vmask);
251
252 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
253
254 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
255 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
256 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
257 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3, vxb2));
258 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
259 vb3 = _mm_slli_epi32(vb3, 4);
260 vb3 = _mm_and_si128(vb3, vmask);
261
262 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
263
264 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
265 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
266 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
267 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3, vxb3));
268
269 w = (const int8_t*) w + 32;
270 k -= 8 * sizeof(int8_t);
271 }
272 // accumulate float
273 const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) w)), 16));
274 w = (const uint16_t*) w + 4;
275
276 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
277 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
278 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
279 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
280 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
281 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
282 const __m128i vacc3x01 = _mm_hadd_epi32(vacc3x0, vacc3x1);
283 const __m128i vacc3x23 = _mm_hadd_epi32(vacc3x2, vacc3x3);
284
285 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
286 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
287 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
288 __m128i vacc3x0123 = _mm_hadd_epi32(vacc3x01, vacc3x23);
289
290 vout0x0123 = _mm_add_ps(vout0x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc0x0123), vfilter_output_scale0123));
291 vout1x0123 = _mm_add_ps(vout1x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc1x0123), vfilter_output_scale0123));
292 vout2x0123 = _mm_add_ps(vout2x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc2x0123), vfilter_output_scale0123));
293 vout3x0123 = _mm_add_ps(vout3x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc3x0123), vfilter_output_scale0123));
294 }
295
296 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
297 const __m128 vinput_scale1 = _mm_broadcast_ss(&quantization_params[1].inv_scale);
298 const __m128 vinput_scale2 = _mm_broadcast_ss(&quantization_params[2].inv_scale);
299 const __m128 vinput_scale3 = _mm_broadcast_ss(&quantization_params[3].inv_scale);
300
301 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
302 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
303 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
304 vout3x0123 = _mm_mul_ps(vout3x0123, vinput_scale3);
305
306
307 const __m128 vbias0123 = _mm_loadu_ps((const float*) w);
308 w = (const float*) w + 4;
309 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
310 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
311 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
312 vout3x0123 = _mm_add_ps(vout3x0123, vbias0123);
313
314 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
315 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
316 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
317 vout3x0123 = _mm_max_ps(vout3x0123, vmin);
318
319 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
320 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
321 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
322 vout3x0123 = _mm_min_ps(vout3x0123, vmax);
323
324 if XNN_LIKELY(nc >= 4) {
325 _mm_storeu_ps(c0, vout0x0123);
326 _mm_storeu_ps(c1, vout1x0123);
327 _mm_storeu_ps(c2, vout2x0123);
328 _mm_storeu_ps(c3, vout3x0123);
329
330 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
331 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
332 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
333 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
334
335 c0 = (float*) ((uintptr_t) c0 + cn_stride);
336 c1 = (float*) ((uintptr_t) c1 + cn_stride);
337 c2 = (float*) ((uintptr_t) c2 + cn_stride);
338 c3 = (float*) ((uintptr_t) c3 + cn_stride);
339
340 nc -= 4;
341 } else {
342 if (nc & 2) {
343 _mm_storel_pi((__m64*) c0, vout0x0123);
344 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
345 c0 += 2;
346 _mm_storel_pi((__m64*) c1, vout1x0123);
347 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
348 c1 += 2;
349 _mm_storel_pi((__m64*) c2, vout2x0123);
350 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
351 c2 += 2;
352 _mm_storel_pi((__m64*) c3, vout3x0123);
353 vout3x0123 = _mm_unpackhi_ps(vout3x0123, vout3x0123);
354 c3 += 2;
355 }
356 if (nc & 1) {
357 _mm_store_ss(c0, vout0x0123);
358 _mm_store_ss(c1, vout1x0123);
359 _mm_store_ss(c2, vout2x0123);
360 _mm_store_ss(c3, vout3x0123);
361 }
362 nc = 0;
363 }
364 } while (nc != 0);
365 }