Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-2x8c8-minmax-avx2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c8-avx2.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 #include <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19
20
21 void xnn_qd8_f16_qb4w_gemm_minmax_ukernel_2x8c8__avx2(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f16_qb4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 2);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 size_t bl = params->scalar.blocksize;
45 assert(bl <= round_up_po2(kc, 16));
46 assert(bl != 0);
47 assert(bl % 32 == 0);
48 const int8_t* a0 = a;
49 uint16_t* c0 = (uint16_t*) c;
50 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
51 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
52 if XNN_UNPREDICTABLE(mr != 2) {
53 a1 = a0;
54 c1 = c0;
55 }
56
57 const __m128i vmask = _mm_set1_epi8(0xF0);
58 XNN_FORCE_REALIZATION(vmask);
59 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
60 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
61 XNN_FORCE_REALIZATION(vmin);
62 XNN_FORCE_REALIZATION(vmax);
63
64 do {
65 const __m128 vinit0 = _mm_load_ss(&((const float*) w)[0]);
66 const __m128 vinit1 = _mm_load_ss(&((const float*) w)[1]);
67 const __m256 vinit01 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit0), vinit1, 1);
68 const __m128 vinit2 = _mm_load_ss(&((const float*) w)[2]);
69 const __m128 vinit3 = _mm_load_ss(&((const float*) w)[3]);
70 const __m256 vinit23 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit2), vinit3, 1);
71 const __m128 vinit4 = _mm_load_ss(&((const float*) w)[4]);
72 const __m128 vinit5 = _mm_load_ss(&((const float*) w)[5]);
73 const __m256 vinit45 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit4), vinit5, 1);
74 const __m128 vinit6 = _mm_load_ss(&((const float*) w)[6]);
75 const __m128 vinit7 = _mm_load_ss(&((const float*) w)[7]);
76 const __m256 vinit67 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit6), vinit7, 1);
77 const __m256 vinput_zero_point0 = _mm256_set1_ps((float) quantization_params[0].zero_point);
78 __m256 vout0x01 = _mm256_mul_ps(vinit01, vinput_zero_point0);
79 __m256 vout0x23 = _mm256_mul_ps(vinit23, vinput_zero_point0);
80 __m256 vout0x45 = _mm256_mul_ps(vinit45, vinput_zero_point0);
81 __m256 vout0x67 = _mm256_mul_ps(vinit67, vinput_zero_point0);
82 const __m256 vinput_zero_point1 = _mm256_set1_ps((float) quantization_params[1].zero_point);
83 __m256 vout1x01 = _mm256_mul_ps(vinit01, vinput_zero_point1);
84 __m256 vout1x23 = _mm256_mul_ps(vinit23, vinput_zero_point1);
85 __m256 vout1x45 = _mm256_mul_ps(vinit45, vinput_zero_point1);
86 __m256 vout1x67 = _mm256_mul_ps(vinit67, vinput_zero_point1);
87 w = (const int32_t*) w + 8;
88
89 for (size_t kb=0; kb < kc; kb += bl) {
90 __m256i vacc0x01 = _mm256_setzero_si256();
91 __m256i vacc0x23 = _mm256_setzero_si256();
92 __m256i vacc0x45 = _mm256_setzero_si256();
93 __m256i vacc0x67 = _mm256_setzero_si256();
94 __m256i vacc1x01 = _mm256_setzero_si256();
95 __m256i vacc1x23 = _mm256_setzero_si256();
96 __m256i vacc1x45 = _mm256_setzero_si256();
97 __m256i vacc1x67 = _mm256_setzero_si256();
98
99 size_t k = bl;
100 while (k >= 16 * sizeof(int8_t)) {
101 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
102 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
103 a0 += 8;
104 __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
105 __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
106 a1 += 8;
107
108 __m128i vb01 = _mm_load_si128((const __m128i*) w);
109 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
110 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
111 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
112
113 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
114 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
115 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
116 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
117 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
118 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
119
120 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
121 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
122 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
123 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
124 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
125 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
126
127 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
128 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
129 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
130 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
131 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
132 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
133
134 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
135 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
136
137 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
138 vxa0 = _mm256_cvtepi8_epi16(va0);
139 a0 += 8;
140 va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
141 vxa1 = _mm256_cvtepi8_epi16(va1);
142 a1 += 8;
143
144 vbm01 = _mm_and_si128(vb01, vmask);
145 vxb01 = _mm256_cvtepi8_epi16(vbm01);
146
147 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
148 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
149 vbm23 = _mm_and_si128(vb23, vmask);
150 vxb23 = _mm256_cvtepi8_epi16(vbm23);
151
152 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
153 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
154 vbm45 = _mm_and_si128(vb45, vmask);
155 vxb45 = _mm256_cvtepi8_epi16(vbm45);
156
157 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
158 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
159 vbm67 = _mm_and_si128(vb67, vmask);
160 vxb67 = _mm256_cvtepi8_epi16(vbm67);
161
162 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
163 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
164
165 w = (const int8_t*) w + 64;
166 k -= 16 * sizeof(int8_t);
167 }
168
169 while (k >= 8 * sizeof(int8_t)) {
170 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
171 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
172 a0 += 8;
173 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
174 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
175 a1 += 8;
176
177 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
178 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
179 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
180 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
181
182 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
183 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
184 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
185 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
186 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
187 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
188
189 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
190 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
191 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
192 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
193 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
194 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
195
196 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
197 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
198 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
199 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
200 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
201 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
202
203 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
204 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
205
206 w = (const int8_t*) w + 64;
207 k -= 8 * sizeof(int8_t);
208 }
209
210 const __m128 vfilter_output_scale0 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[0] << 16));
211 const __m128 vfilter_output_scale1 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[1] << 16));
212 const __m256 vfilter_output_scale01 = _mm256_insertf128_ps(
213 _mm256_castps128_ps256(vfilter_output_scale0), vfilter_output_scale1, 1);
214 vout0x01 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x01), vfilter_output_scale01, vout0x01);
215 vout1x01 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x01), vfilter_output_scale01, vout1x01);
216 const __m128 vfilter_output_scale2 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[2] << 16));
217 const __m128 vfilter_output_scale3 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[3] << 16));
218 const __m256 vfilter_output_scale23 = _mm256_insertf128_ps(
219 _mm256_castps128_ps256(vfilter_output_scale2), vfilter_output_scale3, 1);
220 vout0x23 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x23), vfilter_output_scale23, vout0x23);
221 vout1x23 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x23), vfilter_output_scale23, vout1x23);
222 const __m128 vfilter_output_scale4 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[4] << 16));
223 const __m128 vfilter_output_scale5 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[5] << 16));
224 const __m256 vfilter_output_scale45 = _mm256_insertf128_ps(
225 _mm256_castps128_ps256(vfilter_output_scale4), vfilter_output_scale5, 1);
226 vout0x45 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x45), vfilter_output_scale45, vout0x45);
227 vout1x45 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x45), vfilter_output_scale45, vout1x45);
228 const __m128 vfilter_output_scale6 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[6] << 16));
229 const __m128 vfilter_output_scale7 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[7] << 16));
230 const __m256 vfilter_output_scale67 = _mm256_insertf128_ps(
231 _mm256_castps128_ps256(vfilter_output_scale6), vfilter_output_scale7, 1);
232 vout0x67 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x67), vfilter_output_scale67, vout0x67);
233 vout1x67 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x67), vfilter_output_scale67, vout1x67);
234
235 w = (const uint16_t*) w + 8;
236 }
237
238 const __m256 vout0x0213 = _mm256_hadd_ps(vout0x01, vout0x23);
239 const __m256 vout0x4657 = _mm256_hadd_ps(vout0x45, vout0x67);
240 const __m256 vout1x0213 = _mm256_hadd_ps(vout1x01, vout1x23);
241 const __m256 vout1x4657 = _mm256_hadd_ps(vout1x45, vout1x67);
242
243 const __m256 vout0x02461357 = _mm256_hadd_ps(vout0x0213, vout0x4657);
244 const __m256 vout1x02461357 = _mm256_hadd_ps(vout1x0213, vout1x4657);
245
246 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
247 __m256 vout0x01234567 = _mm256_permutevar8x32_ps(vout0x02461357, vpermute_mask);
248 __m256 vout1x01234567 = _mm256_permutevar8x32_ps(vout1x02461357, vpermute_mask);
249
250 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
251 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
252
253 const __m256 vbias01234567 = _mm256_loadu_ps((const float*) w);
254 w = (const float*) w + 8;
255 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vinput_scale0, vbias01234567);
256 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vinput_scale1, vbias01234567);
257
258
259
260 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
261 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
262
263 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
264 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
265 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
266 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
267 if XNN_LIKELY(nc >= 8) {
268 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
269 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
270 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
271 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
272
273 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
274 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
275
276 nc -= 8;
277 } else {
278 if (nc & 4) {
279 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
280 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
281
282 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
283 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
284
285 c0 += 4;
286 c1 += 4;
287 }
288 if (nc & 2) {
289 _mm_storeu_si32(c0, vfp16out0x01234567);
290 _mm_storeu_si32(c1, vfp16out1x01234567);
291
292 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
293 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
294
295 c0 += 2;
296 c1 += 2;
297 }
298 if (nc & 1) {
299 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
300 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
301 }
302 nc = 0;
303 }
304 } while (nc != 0);
305 }