Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-2x8c8-minmax-avx256skx.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_qc4w_gemm_minmax_ukernel_2x8c8__avx256skx(
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_qc4w_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 const int8_t* a0 = a;
45 uint16_t* c0 = (uint16_t*) c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr != 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52
53 const __m128i vmask = _mm_set1_epi8(0xF0);
54 XNN_FORCE_REALIZATION(vmask);
55 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
56 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
57 XNN_FORCE_REALIZATION(vmin);
58 XNN_FORCE_REALIZATION(vmax);
59
60 do {
61 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
62 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
63 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
64 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
65 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
66 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
67 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
68 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
69 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
70 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
71 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
72 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
73 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
74 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
75 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
76 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
77 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
78 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
79 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
80 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
81 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
82 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
83 w = (const int32_t*) w + 8;
84
85 size_t k = kc;
86 while (k >= 16 * sizeof(int8_t)) {
87 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
88 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
89 a0 += 8;
90 __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
91 __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
92 a1 += 8;
93
94 __m128i vb01 = _mm_load_si128((const __m128i*) w);
95 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
96 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
97 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
98
99 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
100 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
101 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
102 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
103 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
104 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
105
106 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
107 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
108 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
109 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
110 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
111 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
112
113 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
114 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
115 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
116 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
117 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
118 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
119
120 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
121 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
122
123 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
124 vxa0 = _mm256_cvtepi8_epi16(va0);
125 a0 += 8;
126 va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
127 vxa1 = _mm256_cvtepi8_epi16(va1);
128 a1 += 8;
129
130 vbm01 = _mm_and_si128(vb01, vmask);
131 vxb01 = _mm256_cvtepi8_epi16(vbm01);
132
133 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
134 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
135 vbm23 = _mm_and_si128(vb23, vmask);
136 vxb23 = _mm256_cvtepi8_epi16(vbm23);
137
138 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
139 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
140 vbm45 = _mm_and_si128(vb45, vmask);
141 vxb45 = _mm256_cvtepi8_epi16(vbm45);
142
143 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
144 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
145 vbm67 = _mm_and_si128(vb67, vmask);
146 vxb67 = _mm256_cvtepi8_epi16(vbm67);
147
148 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
149 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
150
151 w = (const int8_t*) w + 64;
152 k -= 16 * sizeof(int8_t);
153 }
154
155 while (k >= 8 * sizeof(int8_t)) {
156 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
157 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
158 a0 += 8;
159 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
160 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
161 a1 += 8;
162
163 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
164 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
165 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
166 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
167
168 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
169 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
170 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
171 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
172 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
173 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
174
175 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
176 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
177 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
178 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
179 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
180 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
181
182 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
183 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
184 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
185 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
186 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
187 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
188
189 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
190 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
191
192 w = (const int8_t*) w + 64;
193 k -= 8 * sizeof(int8_t);
194 }
195
196 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
197 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
198 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
199 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
200
201 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
202 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
203
204 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
205 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
206 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
207
208 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
209 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
210 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
211 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
212 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
213 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
214
215 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
216 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
217
218 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
219 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
220 w = (const float*) w + 16;
221 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
222 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
223
224 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
225 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
226
227 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
228 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
229 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
230 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
231 if XNN_LIKELY(nc >= 8) {
232 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
233 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
234 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
235 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
236
237 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
238 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
239
240 nc -= 8;
241 } else {
242 // Prepare mask for valid 16-bit elements (depends on nc).
243 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
244 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
245 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
246 nc = 0;
247 }
248 } while (nc != 0);
249 }