Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-3x8c8-minmax-avxvnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx8c8-avxvnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_3x8c8__avxvnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 xnn_float16* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const int8_t* zero_data,
35 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
36 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 3);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(kc % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 kc = round_up_po2(kc, 8 * sizeof(int8_t));
48 uint16_t* c0 = (uint16_t*) c;
49 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 c1 = c0;
52 }
53 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 c2 = c1;
56 }
57
58 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
59 XNN_FORCE_REALIZATION(vsign_mask);
60 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
61 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
62 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
63 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
64 // XNN_FORCE_REALIZATION(vinput_zero_point);
65 // XNN_FORCE_REALIZATION(vinput_inv_scale);
66 // XNN_FORCE_REALIZATION(voutput_min);
67 // XNN_FORCE_REALIZATION(voutput_max);
68 do {
69 const __m256i vksum01234567 = _mm256_load_si256(w);
70 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
71 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
72 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
73 __m256i vacc1x0123 = vacc0x0123;
74 __m256i vacc1x4567 = vacc0x4567;
75 __m256i vacc2x0123 = vacc0x0123;
76 __m256i vacc2x4567 = vacc0x4567;
77 w = (const int32_t*) w + 8;
78
79 size_t p = ks;
80 do {
81 const int8_t* restrict a0 = a[0];
82 if XNN_UNPREDICTABLE(a0 != zero) {
83 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
84 } else {
85 a0 = zero_data;
86 }
87 const int8_t* restrict a1 = a[1];
88 if XNN_UNPREDICTABLE(a1 != zero) {
89 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
90 } else {
91 a1 = zero_data;
92 }
93 const int8_t* restrict a2 = a[2];
94 if XNN_UNPREDICTABLE(a2 != zero) {
95 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
96 } else {
97 a2 = zero_data;
98 }
99 a += 3;
100
101 size_t k = kc;
102 while (k >= 16 * sizeof(int8_t)) {
103 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
104 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
105 a0 += 16;
106 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
107 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
108 a1 += 16;
109 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
110 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
111 a2 += 16;
112
113 const __m256i vb01234567x0123 = _mm256_load_si256(w);
114 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
115 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
116 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
117 xnn_prefetch_to_l1((const int8_t*) w + 896);
118
119 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
120 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
121 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
122 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
123 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
124 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
125 xnn_prefetch_to_l1((const int8_t*) w + 960);
126 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
127 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
128 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
129 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
130 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
131 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
132
133 w = (const int8_t*) w + 128;
134 k -= 16 * sizeof(int8_t);
135 }
136
137 if (k != 0) {
138 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
139 a0 += 8;
140 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
141 a1 += 8;
142 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
143 a2 += 8;
144
145 const __m256i vb01234567x0123 = _mm256_load_si256(w);
146 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
147
148 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
149 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
150 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
151 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
152 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
153 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
154 xnn_prefetch_to_l1((const int8_t*) w + 960);
155
156 w = (const int8_t*) w + 64;
157 k -= 8 * sizeof(int8_t);
158 }
159
160 p -= 3 * sizeof(void*);
161 } while (p != 0);
162
163
164 // Add adjacent pairs
165 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
166 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
167 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
168 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
169 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
170 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
171
172 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
173 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
174 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
175
176 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
177 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
178 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
179
180 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
181 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
182 w = (const float*) w + 16;
183 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
184 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
185 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
186
187 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
188 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
189 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
190
191 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
192 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
193 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
194
195 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
196 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
197 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
198 if XNN_LIKELY(nc >= 8) {
199 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
200 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
201 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
202 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
203 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
204 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
205
206 a = (const int8_t**restrict) ((uintptr_t) a - ks);
207 nc -= 8;
208 } else {
209 if (nc & 4) {
210 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
211 c2 += 4;
212 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
213 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
214 c1 += 4;
215 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
216 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
217 c0 += 4;
218 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
219 }
220 if (nc & 2) {
221 _mm_storeu_si32(c2, vfp16out2x01234567);
222 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
223 c2 += 2;
224 _mm_storeu_si32(c1, vfp16out1x01234567);
225 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
226 c1 += 2;
227 _mm_storeu_si32(c0, vfp16out0x01234567);
228 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
229 c0 += 2;
230 }
231 if (nc & 1) {
232 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
233 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
234 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
235 }
236 nc = 0;
237 }
238 } while (nc != 0);
239 }