Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-3x8c8-minmax-avxvnni.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
20
21 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_3x8c8__avxvnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 3);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 uint16_t* c0 = (uint16_t*) c;
48 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56
57 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
58 XNN_FORCE_REALIZATION(vsign_mask);
59 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
60 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
61 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
62 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
63 // XNN_FORCE_REALIZATION(vinput_zero_point);
64 // XNN_FORCE_REALIZATION(vinput_inv_scale);
65 // XNN_FORCE_REALIZATION(voutput_min);
66 // XNN_FORCE_REALIZATION(voutput_max);
67 do {
68 const __m256i vksum01234567 = _mm256_load_si256(w);
69 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
70 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
71 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
72 __m256i vacc1x0123 = vacc0x0123;
73 __m256i vacc1x4567 = vacc0x4567;
74 __m256i vacc2x0123 = vacc0x0123;
75 __m256i vacc2x4567 = vacc0x4567;
76 w = (const int32_t*) w + 8;
77
78 size_t p = ks;
79 do {
80 const int8_t* restrict a0 = a[0];
81 if XNN_UNPREDICTABLE(a0 != zero) {
82 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
83 } else {
84 a0 = zero_data;
85 }
86 const int8_t* restrict a1 = a[1];
87 if XNN_UNPREDICTABLE(a1 != zero) {
88 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
89 } else {
90 a1 = zero_data;
91 }
92 const int8_t* restrict a2 = a[2];
93 if XNN_UNPREDICTABLE(a2 != zero) {
94 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
95 } else {
96 a2 = zero_data;
97 }
98 a += 3;
99
100 size_t k = kc;
101 while (k >= 16 * sizeof(int8_t)) {
102 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
103 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
104 a0 += 16;
105 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
106 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
107 a1 += 16;
108 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
109 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
110 a2 += 16;
111
112 const __m256i vb01234567x0123 = _mm256_load_si256(w);
113 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
114 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
115 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
116
117 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
118 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
119 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
120 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
121 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
122 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
123 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
124 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
125 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
126 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
127 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
128 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
129
130 w = (const int8_t*) w + 128;
131 k -= 16 * sizeof(int8_t);
132 }
133
134 if (k != 0) {
135 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
136 a0 += 8;
137 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
138 a1 += 8;
139 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
140 a2 += 8;
141
142 const __m256i vb01234567x0123 = _mm256_load_si256(w);
143 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
144
145 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
146 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
147 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
148 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
149 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
150 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
151
152 w = (const int8_t*) w + 64;
153 k -= 8 * sizeof(int8_t);
154 }
155
156 p -= 3 * sizeof(void*);
157 } while (p != 0);
158
159
160 // Add adjacent pairs
161 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
162 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
163 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
164 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
165 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
166 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
167
168 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
169 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
170 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
171
172 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
173 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
174 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
175
176 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
177 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
178 w = (const float*) w + 16;
179 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
180 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
181 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
182
183 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
184 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
185 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
186
187 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
188 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
189 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
190
191 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
192 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
193 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
194 if XNN_LIKELY(nc >= 8) {
195 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
196 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
197 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
198 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
199 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
200 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
201
202 a = (const int8_t**restrict) ((uintptr_t) a - ks);
203 nc -= 8;
204 } else {
205 if (nc & 4) {
206 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
207 c2 += 4;
208 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
209 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
210 c1 += 4;
211 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
212 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
213 c0 += 4;
214 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
215 }
216 if (nc & 2) {
217 _mm_storeu_si32(c2, vfp16out2x01234567);
218 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
219 c2 += 2;
220 _mm_storeu_si32(c1, vfp16out1x01234567);
221 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
222 c1 += 2;
223 _mm_storeu_si32(c0, vfp16out0x01234567);
224 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
225 c0 += 2;
226 }
227 if (nc & 1) {
228 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
229 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
230 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
231 }
232 nc = 0;
233 }
234 } while (nc != 0);
235 }