Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-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_qc4w_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_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 <= 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 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
62 const __m128 vmin = _mm_set1_ps(params->scalar.min);
63 const __m128 vmax = _mm_set1_ps(params->scalar.max);
64 XNN_FORCE_REALIZATION(vmin);
65 XNN_FORCE_REALIZATION(vmax);
66
67 const __m128i vmask = _mm_set1_epi8(0xF0);
68 XNN_FORCE_REALIZATION(vmask);
69
70 do {
71 const __m128i vksum = _mm_load_si128((const __m128i*) w);
72 const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
73 const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
74 const __m128i vinput_zero_point2 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[2].zero_point));
75 const __m128i vzero = _mm_setzero_si128();
76 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
77 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point1);
78 const __m128i vinit2 = _mm_mullo_epi32(vksum, vinput_zero_point2);
79 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
80 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
81 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
82 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
83 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
84 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
85 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
86 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
87 __m128i vacc2x0 = _mm_blend_epi16(vinit2, vzero, 0xFC);
88 __m128i vacc2x1 = _mm_blend_epi16(vinit2, vzero, 0xF3);
89 __m128i vacc2x2 = _mm_blend_epi16(vinit2, vzero, 0xCF);
90 __m128i vacc2x3 = _mm_blend_epi16(vinit2, vzero, 0x3F);
91 w = (const int32_t*) w + 4;
92
93 size_t k = kc;
94
95 while (k >= 16 * sizeof(int8_t)) {
96 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
97 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
98 a0 += 8;
99 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
100 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
101 a1 += 8;
102 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
103 const __m128i vxa2c0 = _mm_cvtepi8_epi16(va2c0);
104 a2 += 8;
105
106 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
107 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
108 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
109 const __m128i vxb0c0 = _mm_cvtepi8_epi16(vb0c0);
110
111 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
112 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
113 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
114 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
115 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
116 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
117 const __m128i vxb1c0 = _mm_cvtepi8_epi16(vb1c0);
118
119 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
120 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
121 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
122 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
123 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
124 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
125 const __m128i vxb2c0 = _mm_cvtepi8_epi16(vb2c0);
126
127 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
128 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
129 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
130 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
131 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
132 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
133 const __m128i vxb3c0 = _mm_cvtepi8_epi16(vb3c0);
134
135 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
136 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
137 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
138
139 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
140 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
141 a0 += 8;
142 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
143 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
144 a1 += 8;
145 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
146 const __m128i vxa2c1 = _mm_cvtepi8_epi16(va2c1);
147 a2 += 8;
148
149 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
150 const __m128i vxb0c1 = _mm_cvtepi8_epi16(vb0c1);
151
152 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
153 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
154 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
155 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
156 const __m128i vxb1c1 = _mm_cvtepi8_epi16(vb1c1);
157
158 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
159 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
160 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
161 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
162 const __m128i vxb2c1 = _mm_cvtepi8_epi16(vb2c1);
163
164 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
165 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
166 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
167 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
168 const __m128i vxb3c1 = _mm_cvtepi8_epi16(vb3c1);
169
170 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
171 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
172 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
173
174
175 w = (const int8_t*) w + 32;
176 k -= 16 * sizeof(int8_t);
177 }
178
179 while (k >= 8 * sizeof(int8_t)) {
180 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
181 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
182 a0 += 8;
183 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
184 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
185 a1 += 8;
186 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
187 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
188 a2 += 8;
189
190 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
191 vb0 = _mm_slli_epi32(vb0, 4);
192 vb0 = _mm_and_si128(vb0, vmask);
193
194 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
195
196 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
197 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
198 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
199 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
200 vb1 = _mm_slli_epi32(vb1, 4);
201 vb1 = _mm_and_si128(vb1, vmask);
202
203 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
204
205 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
206 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
207 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
208 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
209 vb2 = _mm_slli_epi32(vb2, 4);
210 vb2 = _mm_and_si128(vb2, vmask);
211
212 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
213
214 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
215 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
216 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
217 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
218 vb3 = _mm_slli_epi32(vb3, 4);
219 vb3 = _mm_and_si128(vb3, vmask);
220
221 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
222
223 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
224 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
225 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
226
227 w = (const int8_t*) w + 32;
228 k -= 8 * sizeof(int8_t);
229 }
230
231 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
232 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
233 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
234 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
235 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
236 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
237
238 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
239 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
240 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
241
242 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
243 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
244 vacc2x0123 = _mm_srai_epi32(vacc2x0123, 4);
245 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
246 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
247 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
248
249 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
250 const __m128 vinput_scale1 = _mm_broadcast_ss(&quantization_params[1].inv_scale);
251 const __m128 vinput_scale2 = _mm_broadcast_ss(&quantization_params[2].inv_scale);
252
253 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
254 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
255 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
256
257 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
258 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
259 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
260 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
261
262 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
263 w = (const float*) w + 8;
264 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
265 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
266 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
267
268 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
269 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
270 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
271
272 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
273 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
274 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
275
276 if XNN_LIKELY(nc >= 4) {
277 _mm_storeu_ps(c0, vout0x0123);
278 _mm_storeu_ps(c1, vout1x0123);
279 _mm_storeu_ps(c2, vout2x0123);
280
281 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
282 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
283 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
284
285 c0 = (float*) ((uintptr_t) c0 + cn_stride);
286 c1 = (float*) ((uintptr_t) c1 + cn_stride);
287 c2 = (float*) ((uintptr_t) c2 + cn_stride);
288
289 nc -= 4;
290 } else {
291 if (nc & 2) {
292 _mm_storel_pi((__m64*) c0, vout0x0123);
293 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
294 c0 += 2;
295 _mm_storel_pi((__m64*) c1, vout1x0123);
296 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
297 c1 += 2;
298 _mm_storel_pi((__m64*) c2, vout2x0123);
299 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
300 c2 += 2;
301 }
302 if (nc & 1) {
303 _mm_store_ss(c0, vout0x0123);
304 _mm_store_ss(c1, vout1x0123);
305 _mm_store_ss(c2, vout2x0123);
306 }
307 nc = 0;
308 }
309 } while (nc != 0);
310 }