Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-5x8c8-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_f32_qc8w_gemm_minmax_ukernel_5x8c8__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 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const union xnn_f32_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 <= 5);
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 float* c0 = c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr < 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
65 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 4) {
67 a4 = a3;
68 c4 = c3;
69 }
70
71 const __m256 vmin = _mm256_set1_ps(params->scalar.min);
72 const __m256 vmax = _mm256_set1_ps(params->scalar.max);
73 XNN_FORCE_REALIZATION(vmin);
74 XNN_FORCE_REALIZATION(vmax);
75
76 do {
77 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
78 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
79 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
80 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
81 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
82 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
83 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
84 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
85 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
86 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
87 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
88 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
89 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
90 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
91 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
92 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
93 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
94 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
95 __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1);
96 __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1);
97 __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1);
98 __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1);
99 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
100 __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2);
101 __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2);
102 __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2);
103 __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2);
104 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point);
105 __m256i vacc3x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point3);
106 __m256i vacc3x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point3);
107 __m256i vacc3x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point3);
108 __m256i vacc3x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point3);
109 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point);
110 __m256i vacc4x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point4);
111 __m256i vacc4x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point4);
112 __m256i vacc4x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point4);
113 __m256i vacc4x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point4);
114 w = (const int32_t*) w + 8;
115
116 size_t k = kc;
117
118 while (k >= 8 * sizeof(int8_t)) {
119 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
120 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
121 a0 += 8;
122 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
123 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
124 a1 += 8;
125 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
126 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
127 a2 += 8;
128 const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3));
129 const __m256i vxa3 = _mm256_cvtepi8_epi16(va3);
130 a3 += 8;
131 const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4));
132 const __m256i vxa4 = _mm256_cvtepi8_epi16(va4);
133 a4 += 8;
134
135 const __m256i vxb01 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) w));
136
137 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
138 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
139 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
140 vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01));
141 vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01));
142 const __m256i vxb23 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 16)));
143
144 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
145 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
146 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
147 vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23));
148 vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23));
149 const __m256i vxb45 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 32)));
150
151 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
152 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
153 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
154 vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45));
155 vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45));
156 const __m256i vxb67 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*) ((const int8_t*) w + 48)));
157
158 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
159 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
160 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
161 vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67));
162 vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67));
163
164 w = (const int8_t*) w + 64;
165 k -= 8 * sizeof(int8_t);
166 }
167
168 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
169 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
170 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
171 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
172 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
173 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
174 const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23);
175 const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67);
176 const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23);
177 const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67);
178
179 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
180 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
181 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
182 const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657);
183 const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657);
184
185 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
186 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
187 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
188 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
189 __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask);
190 __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask);
191
192 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
193 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
194 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
195 const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale);
196 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
197 const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale);
198 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
199 const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale);
200 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
201 const __m256 vinput_scale4 = _mm256_broadcast_ss(&quantization_params[4].inv_scale);
202
203 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
204 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1);
205 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2);
206 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale3);
207 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale4);
208
209 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
210 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
211 w = (const float*) w + 16;
212 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
213 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
214 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
215 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
216 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
217
218 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
219 vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin);
220 vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin);
221 vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin);
222 vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin);
223
224 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
225 vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax);
226 vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax);
227 vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax);
228 vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax);
229
230 if XNN_LIKELY(nc >= 8) {
231 _mm256_storeu_ps(c0, vout0x01234567);
232 c0 = (float*) ((uintptr_t) c0 + cn_stride);
233 _mm256_storeu_ps(c1, vout1x01234567);
234 c1 = (float*) ((uintptr_t) c1 + cn_stride);
235 _mm256_storeu_ps(c2, vout2x01234567);
236 c2 = (float*) ((uintptr_t) c2 + cn_stride);
237 _mm256_storeu_ps(c3, vout3x01234567);
238 c3 = (float*) ((uintptr_t) c3 + cn_stride);
239 _mm256_storeu_ps(c4, vout4x01234567);
240 c4 = (float*) ((uintptr_t) c4 + cn_stride);
241
242 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
243 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
244 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
245 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
246 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
247
248 nc -= 8;
249 } else {
250 // Prepare mask for valid 32-bit elements (depends on nc).
251 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
252 _mm256_mask_storeu_ps(c0, vmask, vout0x01234567);
253 _mm256_mask_storeu_ps(c1, vmask, vout1x01234567);
254 _mm256_mask_storeu_ps(c2, vmask, vout2x01234567);
255 _mm256_mask_storeu_ps(c3, vmask, vout3x01234567);
256 _mm256_mask_storeu_ps(c4, vmask, vout4x01234567);
257 nc = 0;
258 }
259 } while (nc != 0);
260 }