Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x4c8-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_3x4c8__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 <= 3);
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
66 const __m128 vmin = _mm_set1_ps(params->scalar.min);
67 const __m128 vmax = _mm_set1_ps(params->scalar.max);
68 XNN_FORCE_REALIZATION(vmin);
69 XNN_FORCE_REALIZATION(vmax);
70
71 const __m128i vmask = _mm_set1_epi8(0xF0);
72 XNN_FORCE_REALIZATION(vmask);
73
74 do {
75 const __m128 vksum = _mm_load_ps((const float*) w);
76 const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
77 const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
78 const __m128i vinput_zero_point2 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[2].zero_point));
79
80 __m128 vinput_zero_point0_float = _mm_cvtepi32_ps(vinput_zero_point0);
81 __m128 vout0x0123 = _mm_mul_ps(vksum, vinput_zero_point0_float);
82 __m128 vinput_zero_point1_float = _mm_cvtepi32_ps(vinput_zero_point1);
83 __m128 vout1x0123 = _mm_mul_ps(vksum, vinput_zero_point1_float);
84 __m128 vinput_zero_point2_float = _mm_cvtepi32_ps(vinput_zero_point2);
85 __m128 vout2x0123 = _mm_mul_ps(vksum, vinput_zero_point2_float);
86 w = (const int32_t*) w + 4;
87
88 for (size_t kb=0; kb < kc; kb += bl) {
89 __m128i vacc0x0 = _mm_setzero_si128();
90 __m128i vacc0x1 = _mm_setzero_si128();
91 __m128i vacc0x2 = _mm_setzero_si128();
92 __m128i vacc0x3 = _mm_setzero_si128();
93 __m128i vacc1x0 = _mm_setzero_si128();
94 __m128i vacc1x1 = _mm_setzero_si128();
95 __m128i vacc1x2 = _mm_setzero_si128();
96 __m128i vacc1x3 = _mm_setzero_si128();
97 __m128i vacc2x0 = _mm_setzero_si128();
98 __m128i vacc2x1 = _mm_setzero_si128();
99 __m128i vacc2x2 = _mm_setzero_si128();
100 __m128i vacc2x3 = _mm_setzero_si128();
101 size_t k = bl;
102
103 while (k >= 16 * sizeof(int8_t)) {
104 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
105 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
106 a0 += 8;
107 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
108 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
109 a1 += 8;
110 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
111 const __m128i vxa2c0 = _mm_cvtepi8_epi16(va2c0);
112 a2 += 8;
113
114 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
115 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
116 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
117 const __m128i vxb0c0 = _mm_cvtepi8_epi16(vb0c0);
118
119 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
120 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
121 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
122 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
123 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
124 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
125 const __m128i vxb1c0 = _mm_cvtepi8_epi16(vb1c0);
126
127 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
128 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
129 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
130 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
131 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
132 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
133 const __m128i vxb2c0 = _mm_cvtepi8_epi16(vb2c0);
134
135 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
136 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
137 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
138 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
139 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
140 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
141 const __m128i vxb3c0 = _mm_cvtepi8_epi16(vb3c0);
142
143 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
144 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
145 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
146
147 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
148 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
149 a0 += 8;
150 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
151 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
152 a1 += 8;
153 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
154 const __m128i vxa2c1 = _mm_cvtepi8_epi16(va2c1);
155 a2 += 8;
156
157 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
158 const __m128i vxb0c1 = _mm_cvtepi8_epi16(vb0c1);
159
160 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
161 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
162 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
163 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
164 const __m128i vxb1c1 = _mm_cvtepi8_epi16(vb1c1);
165
166 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
167 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
168 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
169 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
170 const __m128i vxb2c1 = _mm_cvtepi8_epi16(vb2c1);
171
172 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
173 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
174 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
175 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
176 const __m128i vxb3c1 = _mm_cvtepi8_epi16(vb3c1);
177
178 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
179 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
180 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
181
182
183 w = (const int8_t*) w + 32;
184 k -= 16 * sizeof(int8_t);
185 }
186
187 while (k >= 8 * sizeof(int8_t)) {
188 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
189 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
190 a0 += 8;
191 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
192 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
193 a1 += 8;
194 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
195 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
196 a2 += 8;
197
198 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
199 vb0 = _mm_slli_epi32(vb0, 4);
200 vb0 = _mm_and_si128(vb0, vmask);
201
202 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
203
204 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
205 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
206 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
207 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
208 vb1 = _mm_slli_epi32(vb1, 4);
209 vb1 = _mm_and_si128(vb1, vmask);
210
211 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
212
213 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
214 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
215 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
216 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
217 vb2 = _mm_slli_epi32(vb2, 4);
218 vb2 = _mm_and_si128(vb2, vmask);
219
220 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
221
222 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
223 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
224 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
225 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
226 vb3 = _mm_slli_epi32(vb3, 4);
227 vb3 = _mm_and_si128(vb3, vmask);
228
229 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
230
231 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
232 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
233 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
234
235 w = (const int8_t*) w + 32;
236 k -= 8 * sizeof(int8_t);
237 }
238 // accumulate float
239 const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) w)), 16));
240 w = (const uint16_t*) w + 4;
241
242 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
243 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
244 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
245 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
246 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
247 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
248
249 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
250 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
251 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
252
253 vout0x0123 = _mm_add_ps(vout0x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc0x0123), vfilter_output_scale0123));
254 vout1x0123 = _mm_add_ps(vout1x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc1x0123), vfilter_output_scale0123));
255 vout2x0123 = _mm_add_ps(vout2x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc2x0123), vfilter_output_scale0123));
256 }
257
258 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
259 const __m128 vinput_scale1 = _mm_broadcast_ss(&quantization_params[1].inv_scale);
260 const __m128 vinput_scale2 = _mm_broadcast_ss(&quantization_params[2].inv_scale);
261
262 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
263 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
264 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
265
266
267 const __m128 vbias0123 = _mm_loadu_ps((const float*) w);
268 w = (const float*) w + 4;
269 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
270 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
271 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
272
273 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
274 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
275 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
276
277 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
278 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
279 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
280
281 if XNN_LIKELY(nc >= 4) {
282 _mm_storeu_ps(c0, vout0x0123);
283 _mm_storeu_ps(c1, vout1x0123);
284 _mm_storeu_ps(c2, vout2x0123);
285
286 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
287 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
288 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
289
290 c0 = (float*) ((uintptr_t) c0 + cn_stride);
291 c1 = (float*) ((uintptr_t) c1 + cn_stride);
292 c2 = (float*) ((uintptr_t) c2 + cn_stride);
293
294 nc -= 4;
295 } else {
296 if (nc & 2) {
297 _mm_storel_pi((__m64*) c0, vout0x0123);
298 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
299 c0 += 2;
300 _mm_storel_pi((__m64*) c1, vout1x0123);
301 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
302 c1 += 2;
303 _mm_storel_pi((__m64*) c2, vout2x0123);
304 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
305 c2 += 2;
306 }
307 if (nc & 1) {
308 _mm_store_ss(c0, vout0x0123);
309 _mm_store_ss(c1, vout1x0123);
310 _mm_store_ss(c2, vout2x0123);
311 }
312 nc = 0;
313 }
314 } while (nc != 0);
315 }