Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-4x8c8-minmax-avx2-madd.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
19
20 void xnn_qd8_f16_qc4w_gemm_minmax_ukernel_4x8c8__avx2_madd(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 xnn_float16* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f16_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 4);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 uint16_t* c0 = (uint16_t*) c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
52 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
58 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
59 if XNN_UNPREDICTABLE(mr != 4) {
60 a3 = a2;
61 c3 = c2;
62 }
63
64 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
65 XNN_FORCE_REALIZATION(vsign_mask);
66 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
67 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
68 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
69 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
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(voutput_min);
73 // XNN_FORCE_REALIZATION(voutput_max);
74 const __m256i vmask = _mm256_set1_epi8(0x0F);
75 XNN_FORCE_REALIZATION(vmask);
76 do {
77 const __m256i vksum01234567 = _mm256_load_si256(w);
78 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
79 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
80 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
81 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
82 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
83 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
84 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
85 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
86 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
87 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
88 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
89 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
90 w = (const int32_t*) w + 8;
91
92 size_t k = kc;
93 while (k >= 16 * sizeof(int8_t)) {
94 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
95 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
96 a0 += 16;
97 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
98 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
99 a1 += 16;
100 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
101 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
102 a2 += 16;
103 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
104 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
105 a3 += 16;
106
107 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
108 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
109 const __m256i vbs01234567x0123 = _mm256_srli_epi32(vbb01234567x01234567, 4);
110 const __m256i vbs89ABCDEFx0123 = _mm256_srli_epi32(vbb89ABCDEFx01234567, 4);
111 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
112 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
113 const __m256i vb01234567x4567 = _mm256_and_si256(vbs01234567x0123, vmask);
114 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
115
116 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
117 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
118 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
119 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
120 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
121 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
122 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
123 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
124 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
125 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
126 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
127 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
128 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
129 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
130 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
131 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
132
133 w = (const int8_t*) w + 64;
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 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
145 a3 += 8;
146
147 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
148 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
149 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
150 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
151
152 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
153 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
154 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
155 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
156 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
157 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
158 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
159 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
160
161 w = (const int8_t*) w + 64;
162 k -= 8 * sizeof(int8_t);
163 }
164
165 // Add adjacent pairs
166 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
167 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
168 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
169 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
170 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
171 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
172 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
173 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
174
175 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
176 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
177 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
178 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
179
180 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
181 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
182 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
183 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
184
185 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
186 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
187 w = (const float*) w + 16;
188
189 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
190 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
191 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
192 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
193
194 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
195 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
196 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
197 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
198
199 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
200 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
201 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
202 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
203
204 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
205 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
206 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
207 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
208 if XNN_LIKELY(nc >= 8) {
209 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
210 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
211 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
212 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
213 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
214 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
215 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
216 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
217 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
218 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
219 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
220 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
221 nc -= 8;
222 } else {
223 if (nc & 4) {
224 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
225 c0 += 4;
226 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
227 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
228 c1 += 4;
229 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
230 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
231 c2 += 4;
232 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
233 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
234 c3 += 4;
235 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
236 }
237 if (nc & 2) {
238 _mm_storeu_si32(c0, vfp16out0x01234567);
239 c0 += 2;
240 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
241 _mm_storeu_si32(c1, vfp16out1x01234567);
242 c1 += 2;
243 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
244 _mm_storeu_si32(c2, vfp16out2x01234567);
245 c2 += 2;
246 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
247 _mm_storeu_si32(c3, vfp16out3x01234567);
248 c3 += 2;
249 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
250 }
251 if (nc & 1) {
252 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
253 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
254 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
255 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
256 }
257 nc = 0;
258 }
259 } while (nc != 0);
260 }