Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x16c8-minmax-fp32-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_qs8_qc8w_igemm_minmax_fp32_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 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 5);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 c3 = c2;
57 }
58 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59 if XNN_UNPREDICTABLE(mr <= 4) {
60 c4 = c3;
61 }
62
63 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
64 const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
65 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
66 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
67 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
68 // XNN_FORCE_REALIZATION(voutput_zero_point);
69 // XNN_FORCE_REALIZATION(voutput_min);
70 do {
71 __m512i vacc0x0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
72 __m512i vacc0x4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
73 __m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
74 __m512i vacc0xCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
75 __m512i vacc1x0123 = vacc0x0123;
76 __m512i vacc1x4567 = vacc0x4567;
77 __m512i vacc1x89AB = vacc0x89AB;
78 __m512i vacc1xCDEF = vacc0xCDEF;
79 __m512i vacc2x0123 = vacc0x0123;
80 __m512i vacc2x4567 = vacc0x4567;
81 __m512i vacc2x89AB = vacc0x89AB;
82 __m512i vacc2xCDEF = vacc0xCDEF;
83 __m512i vacc3x0123 = vacc0x0123;
84 __m512i vacc3x4567 = vacc0x4567;
85 __m512i vacc3x89AB = vacc0x89AB;
86 __m512i vacc3xCDEF = vacc0xCDEF;
87 __m512i vacc4x0123 = vacc0x0123;
88 __m512i vacc4x4567 = vacc0x4567;
89 __m512i vacc4x89AB = vacc0x89AB;
90 __m512i vacc4xCDEF = vacc0xCDEF;
91 w = (const int32_t*) w + 16;
92
93 size_t p = ks;
94 do {
95 const int8_t* restrict a0 = a[0];
96 if XNN_UNPREDICTABLE(a0 != zero) {
97 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
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 }
103 const int8_t* restrict a2 = a[2];
104 if XNN_UNPREDICTABLE(a2 != zero) {
105 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
106 }
107 const int8_t* restrict a3 = a[3];
108 if XNN_UNPREDICTABLE(a3 != zero) {
109 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
110 }
111 const int8_t* restrict a4 = a[4];
112 if XNN_UNPREDICTABLE(a4 != zero) {
113 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
114 }
115 a += 5;
116
117 size_t k = kc;
118
119 while (k >= 8 * sizeof(int8_t)) {
120 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
121 a0 += 8;
122 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
123 a1 += 8;
124 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
125 a2 += 8;
126 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
127 a3 += 8;
128 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
129 a4 += 8;
130
131 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
132 xnn_prefetch_to_l1((const int8_t*) w + 896);
133
134 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
135 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
136 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
137 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
138 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
139 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
140
141 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
142 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
143 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
144 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
145 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
146 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
147 xnn_prefetch_to_l1((const int8_t*) w + 960);
148
149 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
150 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
151 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
152 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
153 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
154 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
155
156 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
157 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
158 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
159 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
160 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
161
162 w = (const int8_t*) w + 128;
163 k -= 8 * sizeof(int8_t);
164 }
165 p -= 5 * sizeof(void*);
166 } while (p != 0);
167
168 // Add 4 adjacent sums
169 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
170 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
171 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
172 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
173 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
174 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
175 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
176 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
177 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
178 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
179 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
180 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
181 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
182 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
183 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
184 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
185 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
186 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
187 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
188 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
189 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
190
191 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
192 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
193 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
194 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
195 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
196
197 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
198 w = (const float*) w + 16;
199 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
200 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
201 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
202 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
203 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
204
205 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
206 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
207 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
208 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
209 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
210
211 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
212 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
213 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
214 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
215 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
216
217 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
218 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
219 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
220 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
221 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
222
223 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
224 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
225 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
226 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
227 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
228
229 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
230 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
231 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
232 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
233 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
234
235 if (nc >= 16) {
236 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
237 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
238 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
239 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
240 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
241 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
242 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
243 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
244 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
245 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
246 a = (const int8_t**restrict) ((uintptr_t) a - ks);
247 nc -= 16;
248 } else {
249 // Prepare mask for valid 8-bit elements (depends on nc).
250 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
251 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
252 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
253 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
254 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
255 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
256 nc = 0;
257 }
258 } while (nc != 0);
259 }