Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x4c8-minmax-avx-ld128.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_ld128(
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 vb01c01 = _mm_loadu_si128((const __m128i*) w);
131 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
132 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
133 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
134 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
135 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
136
137 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
138 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
139 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
140 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
141 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
142 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
143 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c0, vxb0c0));
144 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c0, vxb1c0));
145 const __m128i vb23c01 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
146 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
147 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
148 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
149 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
150 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
151
152 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
153 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
154 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
155 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
156 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
157 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
158 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c0, vxb2c0));
159 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c0, vxb3c0));
160
161 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
162 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
163 a0 += 8;
164 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
165 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
166 a1 += 8;
167 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
168 const __m128i vxa2c1 = _mm_cvtepi8_epi16(va2c1);
169 a2 += 8;
170 const __m128i va3c1 = _mm_loadl_epi64((const __m128i*) a3);
171 const __m128i vxa3c1 = _mm_cvtepi8_epi16(va3c1);
172 a3 += 8;
173
174 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
175 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
176 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
177 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
178
179 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
180 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
181 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
182 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
183 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
184 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
185 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c1, vxb0c1));
186 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c1, vxb1c1));
187 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
188 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
189 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
190 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
191
192 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
193 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
194 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
195 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
196 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
197 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
198 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c1, vxb2c1));
199 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c1, vxb3c1));
200
201
202 w = (const int8_t*) w + 32;
203 k -= 16 * sizeof(int8_t);
204 }
205
206 while (k >= 8 * sizeof(int8_t)) {
207 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
208 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
209 a0 += 8;
210 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
211 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
212 a1 += 8;
213 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
214 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
215 a2 += 8;
216 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
217 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
218 a3 += 8;
219
220 __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
221 vb01 = _mm_slli_epi32(vb01, 4);
222 vb01 = _mm_and_si128(vb01, vmask);
223
224 const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01);
225 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
226 const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8);
227
228 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
229 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
230 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
231 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
232 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
233 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
234 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3, vxb0));
235 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3, vxb1));
236 __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
237 vb23 = _mm_slli_epi32(vb23, 4);
238 vb23 = _mm_and_si128(vb23, vmask);
239
240 const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23);
241 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
242 const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8);
243
244 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
245 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
246 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
247 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
248 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
249 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
250 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3, vxb2));
251 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3, vxb3));
252
253 w = (const int8_t*) w + 32;
254 k -= 8 * sizeof(int8_t);
255 }
256 // accumulate float
257 const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) w)), 16));
258 w = (const uint16_t*) w + 4;
259
260 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
261 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
262 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
263 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
264 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
265 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
266 const __m128i vacc3x01 = _mm_hadd_epi32(vacc3x0, vacc3x1);
267 const __m128i vacc3x23 = _mm_hadd_epi32(vacc3x2, vacc3x3);
268
269 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
270 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
271 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
272 __m128i vacc3x0123 = _mm_hadd_epi32(vacc3x01, vacc3x23);
273
274 vout0x0123 = _mm_add_ps(vout0x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc0x0123), vfilter_output_scale0123));
275 vout1x0123 = _mm_add_ps(vout1x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc1x0123), vfilter_output_scale0123));
276 vout2x0123 = _mm_add_ps(vout2x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc2x0123), vfilter_output_scale0123));
277 vout3x0123 = _mm_add_ps(vout3x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc3x0123), vfilter_output_scale0123));
278 }
279
280 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
281 const __m128 vinput_scale1 = _mm_broadcast_ss(&quantization_params[1].inv_scale);
282 const __m128 vinput_scale2 = _mm_broadcast_ss(&quantization_params[2].inv_scale);
283 const __m128 vinput_scale3 = _mm_broadcast_ss(&quantization_params[3].inv_scale);
284
285 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
286 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
287 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
288 vout3x0123 = _mm_mul_ps(vout3x0123, vinput_scale3);
289
290
291 const __m128 vbias0123 = _mm_loadu_ps((const float*) w);
292 w = (const float*) w + 4;
293 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
294 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
295 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
296 vout3x0123 = _mm_add_ps(vout3x0123, vbias0123);
297
298 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
299 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
300 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
301 vout3x0123 = _mm_max_ps(vout3x0123, vmin);
302
303 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
304 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
305 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
306 vout3x0123 = _mm_min_ps(vout3x0123, vmax);
307
308 if XNN_LIKELY(nc >= 4) {
309 _mm_storeu_ps(c0, vout0x0123);
310 _mm_storeu_ps(c1, vout1x0123);
311 _mm_storeu_ps(c2, vout2x0123);
312 _mm_storeu_ps(c3, vout3x0123);
313
314 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
315 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
316 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
317 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
318
319 c0 = (float*) ((uintptr_t) c0 + cn_stride);
320 c1 = (float*) ((uintptr_t) c1 + cn_stride);
321 c2 = (float*) ((uintptr_t) c2 + cn_stride);
322 c3 = (float*) ((uintptr_t) c3 + cn_stride);
323
324 nc -= 4;
325 } else {
326 if (nc & 2) {
327 _mm_storel_pi((__m64*) c0, vout0x0123);
328 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
329 c0 += 2;
330 _mm_storel_pi((__m64*) c1, vout1x0123);
331 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
332 c1 += 2;
333 _mm_storel_pi((__m64*) c2, vout2x0123);
334 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
335 c2 += 2;
336 _mm_storel_pi((__m64*) c3, vout3x0123);
337 vout3x0123 = _mm_unpackhi_ps(vout3x0123, vout3x0123);
338 c3 += 2;
339 }
340 if (nc & 1) {
341 _mm_store_ss(c0, vout0x0123);
342 _mm_store_ss(c1, vout1x0123);
343 _mm_store_ss(c2, vout2x0123);
344 _mm_store_ss(c3, vout3x0123);
345 }
346 nc = 0;
347 }
348 } while (nc != 0);
349 }