Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-1x8c8-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_1x8c8__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 <= 1);
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
51 const __m128i vmask = _mm_set1_epi8(0xF0);
52 XNN_FORCE_REALIZATION(vmask);
53 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
54 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
55 XNN_FORCE_REALIZATION(vmin);
56 XNN_FORCE_REALIZATION(vmax);
57
58 do {
59 const __m128 vinit0 = _mm_load_ss(&((const float*) w)[0]);
60 const __m128 vinit1 = _mm_load_ss(&((const float*) w)[1]);
61 const __m256 vinit01 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit0), vinit1, 1);
62 const __m128 vinit2 = _mm_load_ss(&((const float*) w)[2]);
63 const __m128 vinit3 = _mm_load_ss(&((const float*) w)[3]);
64 const __m256 vinit23 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit2), vinit3, 1);
65 const __m128 vinit4 = _mm_load_ss(&((const float*) w)[4]);
66 const __m128 vinit5 = _mm_load_ss(&((const float*) w)[5]);
67 const __m256 vinit45 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit4), vinit5, 1);
68 const __m128 vinit6 = _mm_load_ss(&((const float*) w)[6]);
69 const __m128 vinit7 = _mm_load_ss(&((const float*) w)[7]);
70 const __m256 vinit67 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit6), vinit7, 1);
71 const __m256 vinput_zero_point0 = _mm256_set1_ps((float) quantization_params[0].zero_point);
72 __m256 vout0x01 = _mm256_mul_ps(vinit01, vinput_zero_point0);
73 __m256 vout0x23 = _mm256_mul_ps(vinit23, vinput_zero_point0);
74 __m256 vout0x45 = _mm256_mul_ps(vinit45, vinput_zero_point0);
75 __m256 vout0x67 = _mm256_mul_ps(vinit67, vinput_zero_point0);
76 w = (const int32_t*) w + 8;
77
78 for (size_t kb=0; kb < kc; kb += bl) {
79 __m256i vacc0x01 = _mm256_setzero_si256();
80 __m256i vacc0x23 = _mm256_setzero_si256();
81 __m256i vacc0x45 = _mm256_setzero_si256();
82 __m256i vacc0x67 = _mm256_setzero_si256();
83
84 size_t k = bl;
85 while (k >= 16 * sizeof(int8_t)) {
86 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
87 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
88 a0 += 8;
89
90 __m128i vb01 = _mm_load_si128((const __m128i*) w);
91 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
92 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
93 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
94
95 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
96 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
97 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
98 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
99 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
100
101 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
102 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
103 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
104 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
105 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
106
107 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
108 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
109 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
110 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
111 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
112
113 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
114
115 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
116 vxa0 = _mm256_cvtepi8_epi16(va0);
117 a0 += 8;
118
119 vbm01 = _mm_and_si128(vb01, vmask);
120 vxb01 = _mm256_cvtepi8_epi16(vbm01);
121
122 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
123 vbm23 = _mm_and_si128(vb23, vmask);
124 vxb23 = _mm256_cvtepi8_epi16(vbm23);
125
126 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
127 vbm45 = _mm_and_si128(vb45, vmask);
128 vxb45 = _mm256_cvtepi8_epi16(vbm45);
129
130 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
131 vbm67 = _mm_and_si128(vb67, vmask);
132 vxb67 = _mm256_cvtepi8_epi16(vbm67);
133
134 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
135
136 w = (const int8_t*) w + 64;
137 k -= 16 * sizeof(int8_t);
138 }
139
140 while (k >= 8 * sizeof(int8_t)) {
141 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
142 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
143 a0 += 8;
144
145 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
146 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
147 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
148 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
149
150 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
151 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
152 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
153 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
154 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
155
156 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
157 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
158 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
159 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
160 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
161
162 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
163 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
164 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
165 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
166 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
167
168 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
169
170 w = (const int8_t*) w + 64;
171 k -= 8 * sizeof(int8_t);
172 }
173
174 const __m128 vfilter_output_scale0 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[0] << 16));
175 const __m128 vfilter_output_scale1 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[1] << 16));
176 const __m256 vfilter_output_scale01 = _mm256_insertf128_ps(
177 _mm256_castps128_ps256(vfilter_output_scale0), vfilter_output_scale1, 1);
178 vout0x01 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x01), vfilter_output_scale01, vout0x01);
179 const __m128 vfilter_output_scale2 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[2] << 16));
180 const __m128 vfilter_output_scale3 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[3] << 16));
181 const __m256 vfilter_output_scale23 = _mm256_insertf128_ps(
182 _mm256_castps128_ps256(vfilter_output_scale2), vfilter_output_scale3, 1);
183 vout0x23 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x23), vfilter_output_scale23, vout0x23);
184 const __m128 vfilter_output_scale4 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[4] << 16));
185 const __m128 vfilter_output_scale5 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[5] << 16));
186 const __m256 vfilter_output_scale45 = _mm256_insertf128_ps(
187 _mm256_castps128_ps256(vfilter_output_scale4), vfilter_output_scale5, 1);
188 vout0x45 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x45), vfilter_output_scale45, vout0x45);
189 const __m128 vfilter_output_scale6 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[6] << 16));
190 const __m128 vfilter_output_scale7 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[7] << 16));
191 const __m256 vfilter_output_scale67 = _mm256_insertf128_ps(
192 _mm256_castps128_ps256(vfilter_output_scale6), vfilter_output_scale7, 1);
193 vout0x67 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x67), vfilter_output_scale67, vout0x67);
194
195 w = (const uint16_t*) w + 8;
196 }
197
198 const __m256 vout0x0213 = _mm256_hadd_ps(vout0x01, vout0x23);
199 const __m256 vout0x4657 = _mm256_hadd_ps(vout0x45, vout0x67);
200
201 const __m256 vout0x02461357 = _mm256_hadd_ps(vout0x0213, vout0x4657);
202
203 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
204 __m256 vout0x01234567 = _mm256_permutevar8x32_ps(vout0x02461357, vpermute_mask);
205
206 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
207
208 const __m256 vbias01234567 = _mm256_loadu_ps((const float*) w);
209 w = (const float*) w + 8;
210 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vinput_scale0, vbias01234567);
211
212
213
214 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
215
216 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
217 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
218 if XNN_LIKELY(nc >= 8) {
219 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
220 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
221
222 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
223
224 nc -= 8;
225 } else {
226 if (nc & 4) {
227 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
228
229 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
230
231 c0 += 4;
232 }
233 if (nc & 2) {
234 _mm_storeu_si32(c0, vfp16out0x01234567);
235
236 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
237
238 c0 += 2;
239 }
240 if (nc & 1) {
241 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
242 }
243 nc = 0;
244 }
245 } while (nc != 0);
246 }