Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-5x8c8-minmax-avx256vnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qd8_f16_qc4w_gemm_minmax_ukernel_5x8c8__avx256vnni_prfm(
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 <= 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 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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 uint16_t* c2 = (uint16_t*) ((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 uint16_t* c3 = (uint16_t*) ((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 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 4) {
67 a4 = a3;
68 c4 = c3;
69 }
70
71 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
72 XNN_FORCE_REALIZATION(vsign_mask);
73 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
74 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
75 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
76 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
77 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
78 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
79 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
80 // XNN_FORCE_REALIZATION(voutput_min);
81 // XNN_FORCE_REALIZATION(voutput_max);
82 const __m256i vmask = _mm256_set1_epi8(0xF0);
83 XNN_FORCE_REALIZATION(vmask);
84 do {
85 const __m256i vksum01234567 = _mm256_load_si256(w);
86 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
87 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
88 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
89 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
90 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
91 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
92 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
93 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
94 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
95 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
96 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
97 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
98 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
99 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
100 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
101 w = (const int32_t*) w + 8;
102
103 size_t k = kc;
104 while (k >= 16 * sizeof(int8_t)) {
105 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
106 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
107 a0 += 16;
108 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
109 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
110 a1 += 16;
111 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
112 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
113 a2 += 16;
114 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
115 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
116 a3 += 16;
117 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
118 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
119 a4 += 16;
120
121 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
122 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
123 const __m256i vbs01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
124 const __m256i vbs89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
125 const __m256i vb01234567x4567 = _mm256_and_si256(vbb01234567x01234567, vmask);
126 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
127 const __m256i vb01234567x0123 = _mm256_and_si256(vbs01234567x0123, vmask);
128 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
129
130 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
131 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
132 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
133 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
134 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
135 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
136 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
137 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
138 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
139 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
140 xnn_prefetch_to_l1((const int8_t*) w + 960);
141 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
142 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
143 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
144 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
145 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
146 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
147 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
148 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
149 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
150 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
151
152 w = (const int8_t*) w + 64;
153 k -= 16 * sizeof(int8_t);
154 }
155
156 if (k != 0) {
157 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
158 a0 += 8;
159 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
160 a1 += 8;
161 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
162 a2 += 8;
163 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
164 a3 += 8;
165 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
166 a4 += 8;
167
168 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
169 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
170 const __m256i vb01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
171 const __m256i vb89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
172
173 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
174 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
175 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
176 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
177 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
178 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
179 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
180 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
181 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
182 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
183 xnn_prefetch_to_l1((const int8_t*) w + 960);
184
185 w = (const int8_t*) w + 64;
186 k -= 8 * sizeof(int8_t);
187 }
188
189 // Add adjacent pairs
190 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
191 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
192 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
193 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
194 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
195 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
196 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
197 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
198 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
199 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
200
201 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
202 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
203 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
204 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
205 vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4);
206 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
207 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
208 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
209 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
210 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
211
212 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
213 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
214 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
215 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
216 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
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
222 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
223 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
224 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
225 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
226 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
227
228 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
229 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
230 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
231 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
232 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
233
234 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
235 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
236 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
237 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
238 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
239
240 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
241 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
242 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
243 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
244 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
245 if XNN_LIKELY(nc >= 8) {
246 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
247 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
248 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
249 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
250 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
251 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
252 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
253 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
254 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
255 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
256 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
257 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
258 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
259 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
260 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
261 nc -= 8;
262 } else {
263 // Prepare mask for valid 16-bit elements (depends on nc).
264 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
265 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
266 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
267 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
268 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
269 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
270 nc = 0;
271 }
272 } while (nc != 0);
273 }