Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-5x16c8-minmax-avx512skx-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c8-avx512skx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/igemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_5x16c8__avx512skx_prfm(
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 float* 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_f32_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 <= 5);
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 float* c0 = c;
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 c3 = c2;
59 }
60 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 c4 = c3;
63 }
64
65 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
66 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point);
67 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
68 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
69 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
70 // XNN_FORCE_REALIZATION(vinput_zero_point);
71 // XNN_FORCE_REALIZATION(vinput_inv_scale);
72 // XNN_FORCE_REALIZATION(voutput_min);
73 // XNN_FORCE_REALIZATION(voutput_max);
74 do {
75 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
76 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
77 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
78 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
79 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point);
80 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point);
81 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point);
82 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point);
83 __m512i vacc1x0123 = vacc0x0123;
84 __m512i vacc1x4567 = vacc0x4567;
85 __m512i vacc1x89AB = vacc0x89AB;
86 __m512i vacc1xCDEF = vacc0xCDEF;
87 __m512i vacc2x0123 = vacc0x0123;
88 __m512i vacc2x4567 = vacc0x4567;
89 __m512i vacc2x89AB = vacc0x89AB;
90 __m512i vacc2xCDEF = vacc0xCDEF;
91 __m512i vacc3x0123 = vacc0x0123;
92 __m512i vacc3x4567 = vacc0x4567;
93 __m512i vacc3x89AB = vacc0x89AB;
94 __m512i vacc3xCDEF = vacc0xCDEF;
95 __m512i vacc4x0123 = vacc0x0123;
96 __m512i vacc4x4567 = vacc0x4567;
97 __m512i vacc4x89AB = vacc0x89AB;
98 __m512i vacc4xCDEF = vacc0xCDEF;
99 w = (const int32_t*) w + 16;
100
101 size_t p = ks;
102 do {
103 const int8_t* restrict a0 = a[0];
104 if XNN_UNPREDICTABLE(a0 != zero) {
105 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
106 } else {
107 a0 = zero_data;
108 }
109 const int8_t* restrict a1 = a[1];
110 if XNN_UNPREDICTABLE(a1 != zero) {
111 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
112 } else {
113 a1 = zero_data;
114 }
115 const int8_t* restrict a2 = a[2];
116 if XNN_UNPREDICTABLE(a2 != zero) {
117 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
118 } else {
119 a2 = zero_data;
120 }
121 const int8_t* restrict a3 = a[3];
122 if XNN_UNPREDICTABLE(a3 != zero) {
123 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
124 } else {
125 a3 = zero_data;
126 }
127 const int8_t* restrict a4 = a[4];
128 if XNN_UNPREDICTABLE(a4 != zero) {
129 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
130 } else {
131 a4 = zero_data;
132 }
133 a += 5;
134
135 size_t k = kc;
136
137 while (k >= 8 * sizeof(int8_t)) {
138 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
139 a0 += 8;
140 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
141 a1 += 8;
142 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
143 a2 += 8;
144 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
145 a3 += 8;
146 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
147 a4 += 8;
148
149 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
150 xnn_prefetch_to_l1((const int8_t*) w + 896);
151
152 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
153 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
154 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
155 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
156 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
157 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
158
159 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
160 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
161 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
162 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
163 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
164 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
165 xnn_prefetch_to_l1((const int8_t*) w + 960);
166
167 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
168 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
169 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
170 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
171 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
172 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
173
174 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
175 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
176 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
177 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
178 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
179
180 w = (const int8_t*) w + 128;
181 k -= 8 * sizeof(int8_t);
182 }
183 p -= 5 * sizeof(void*);
184 } while (p != 0);
185
186 // Add 4 adjacent sums
187 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
188 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
189 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
190 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
191 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
192 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
193 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
194 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
195 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
196 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
197 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
198 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
199 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
200 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
201 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
202 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
203 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
204 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
205 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
206 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
207 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
208
209 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
210 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
211 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
212 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
213 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
214
215 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
216 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
217 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
218 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
219 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
220
221 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
222 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
223 w = (const float*) w + 32;
224
225 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
226 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
227 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
228 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
229 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
230
231 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
232 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
233 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
234 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
235 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
236
237 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
238 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
239 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
240 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
241 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
242
243 if XNN_LIKELY(nc >= 16) {
244 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
245 c4 = (float*) ((uintptr_t) c4 + cn_stride);
246 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
247 c3 = (float*) ((uintptr_t) c3 + cn_stride);
248 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
249 c2 = (float*) ((uintptr_t) c2 + cn_stride);
250 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
251 c1 = (float*) ((uintptr_t) c1 + cn_stride);
252 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
253 c0 = (float*) ((uintptr_t) c0 + cn_stride);
254 a = (const int8_t**restrict) ((uintptr_t) a - ks);
255 nc -= 16;
256 } else {
257 // Prepare mask for valid 32-bit elements (depends on nc).
258 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
259 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
260 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
261 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
262 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
263 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
264 nc = 0;
265 }
266 } while (nc != 0);
267 }