Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-5x8c8-minmax-avx256vnni-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_5x8c8__avx256vnni_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 <= 5);
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 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr < 4) {
59 c3 = c2;
60 }
61 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 c4 = c3;
64 }
65
66 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
67 XNN_FORCE_REALIZATION(vsign_mask);
68 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
69 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
70 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
71 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
72 // XNN_FORCE_REALIZATION(vinput_zero_point);
73 // XNN_FORCE_REALIZATION(vinput_inv_scale);
74 // XNN_FORCE_REALIZATION(voutput_min);
75 // XNN_FORCE_REALIZATION(voutput_max);
76 do {
77 const __m256i vksum01234567 = _mm256_load_si256(w);
78 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
79 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
80 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
81 __m256i vacc1x0123 = vacc0x0123;
82 __m256i vacc1x4567 = vacc0x4567;
83 __m256i vacc2x0123 = vacc0x0123;
84 __m256i vacc2x4567 = vacc0x4567;
85 __m256i vacc3x0123 = vacc0x0123;
86 __m256i vacc3x4567 = vacc0x4567;
87 __m256i vacc4x0123 = vacc0x0123;
88 __m256i vacc4x4567 = vacc0x4567;
89 w = (const int32_t*) w + 8;
90
91 size_t p = ks;
92 do {
93 const int8_t* restrict a0 = a[0];
94 if XNN_UNPREDICTABLE(a0 != zero) {
95 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
96 } else {
97 a0 = zero_data;
98 }
99 const int8_t* restrict a1 = a[1];
100 if XNN_UNPREDICTABLE(a1 != zero) {
101 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
102 } else {
103 a1 = zero_data;
104 }
105 const int8_t* restrict a2 = a[2];
106 if XNN_UNPREDICTABLE(a2 != zero) {
107 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
108 } else {
109 a2 = zero_data;
110 }
111 const int8_t* restrict a3 = a[3];
112 if XNN_UNPREDICTABLE(a3 != zero) {
113 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
114 } else {
115 a3 = zero_data;
116 }
117 const int8_t* restrict a4 = a[4];
118 if XNN_UNPREDICTABLE(a4 != zero) {
119 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
120 } else {
121 a4 = zero_data;
122 }
123 a += 5;
124
125 size_t k = kc;
126 while (k >= 16 * sizeof(int8_t)) {
127 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
128 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
129 a0 += 16;
130 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
131 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
132 a1 += 16;
133 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
134 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
135 a2 += 16;
136 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
137 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
138 a3 += 16;
139 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
140 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
141 a4 += 16;
142
143 const __m256i vb01234567x0123 = _mm256_load_si256(w);
144 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
145 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
146 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
147 xnn_prefetch_to_l1((const int8_t*) w + 896);
148
149 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
150 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
151 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
152 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
153 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
154 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
155 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
156 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
157 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
158 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
159 xnn_prefetch_to_l1((const int8_t*) w + 960);
160 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
161 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
162 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
163 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
164 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
165 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
166 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
167 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
168 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
169 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
170
171 w = (const int8_t*) w + 128;
172 k -= 16 * sizeof(int8_t);
173 }
174
175 if (k != 0) {
176 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
177 a0 += 8;
178 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
179 a1 += 8;
180 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
181 a2 += 8;
182 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
183 a3 += 8;
184 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
185 a4 += 8;
186
187 const __m256i vb01234567x0123 = _mm256_load_si256(w);
188 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
189
190 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
191 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
192 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
193 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
194 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
195 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
196 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
197 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
198 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
199 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
200 xnn_prefetch_to_l1((const int8_t*) w + 960);
201
202 w = (const int8_t*) w + 64;
203 k -= 8 * sizeof(int8_t);
204 }
205
206 p -= 5 * sizeof(void*);
207 } while (p != 0);
208
209
210 // Add adjacent pairs
211 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
212 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
213 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
214 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
215 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
216 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
217 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
218 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
219 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
220 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
221
222 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
223 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
224 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
225 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
226 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
227
228 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
229 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
230 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
231 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
232 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
233
234 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
235 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
236 w = (const float*) w + 16;
237 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
238 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
239 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
240 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
241 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
242
243 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
244 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
245 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
246 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
247 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
248
249 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
250 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
251 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
252 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
253 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
254
255 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
256 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
257 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
258 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
259 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
260 if XNN_LIKELY(nc >= 8) {
261 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
262 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
263 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
264 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
265 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
266 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
267 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
268 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
269 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
270 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
271
272 a = (const int8_t**restrict) ((uintptr_t) a - ks);
273 nc -= 8;
274 } else {
275 // Prepare mask for valid 16-bit elements (depends on nc).
276 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
277 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
278 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
279 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
280 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
281 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
282 nc = 0;
283 }
284 } while (nc != 0);
285 }