Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-5x16c8-minmax-avx512skx.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
19
20 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_5x16c8__avx512skx(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 5);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63
64 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
65 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point);
66 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
67 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
68 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
69 // XNN_FORCE_REALIZATION(vinput_zero_point);
70 // XNN_FORCE_REALIZATION(vinput_inv_scale);
71 // XNN_FORCE_REALIZATION(voutput_min);
72 // XNN_FORCE_REALIZATION(voutput_max);
73 do {
74 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
75 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
76 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
77 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
78 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point);
79 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point);
80 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point);
81 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point);
82 __m512i vacc1x0123 = vacc0x0123;
83 __m512i vacc1x4567 = vacc0x4567;
84 __m512i vacc1x89AB = vacc0x89AB;
85 __m512i vacc1xCDEF = vacc0xCDEF;
86 __m512i vacc2x0123 = vacc0x0123;
87 __m512i vacc2x4567 = vacc0x4567;
88 __m512i vacc2x89AB = vacc0x89AB;
89 __m512i vacc2xCDEF = vacc0xCDEF;
90 __m512i vacc3x0123 = vacc0x0123;
91 __m512i vacc3x4567 = vacc0x4567;
92 __m512i vacc3x89AB = vacc0x89AB;
93 __m512i vacc3xCDEF = vacc0xCDEF;
94 __m512i vacc4x0123 = vacc0x0123;
95 __m512i vacc4x4567 = vacc0x4567;
96 __m512i vacc4x89AB = vacc0x89AB;
97 __m512i vacc4xCDEF = vacc0xCDEF;
98 w = (const int32_t*) w + 16;
99
100 size_t p = ks;
101 do {
102 const int8_t* restrict a0 = a[0];
103 if XNN_UNPREDICTABLE(a0 != zero) {
104 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
105 } else {
106 a0 = zero_data;
107 }
108 const int8_t* restrict a1 = a[1];
109 if XNN_UNPREDICTABLE(a1 != zero) {
110 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
111 } else {
112 a1 = zero_data;
113 }
114 const int8_t* restrict a2 = a[2];
115 if XNN_UNPREDICTABLE(a2 != zero) {
116 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
117 } else {
118 a2 = zero_data;
119 }
120 const int8_t* restrict a3 = a[3];
121 if XNN_UNPREDICTABLE(a3 != zero) {
122 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
123 } else {
124 a3 = zero_data;
125 }
126 const int8_t* restrict a4 = a[4];
127 if XNN_UNPREDICTABLE(a4 != zero) {
128 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
129 } else {
130 a4 = zero_data;
131 }
132 a += 5;
133
134 size_t k = kc;
135
136 while (k >= 8 * sizeof(int8_t)) {
137 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
138 a0 += 8;
139 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
140 a1 += 8;
141 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
142 a2 += 8;
143 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
144 a3 += 8;
145 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
146 a4 += 8;
147
148 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
149
150 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
151 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
152 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
153 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
154 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
155 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
156
157 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
158 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
159 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
160 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
161 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
162 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
163
164 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
165 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
166 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
167 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
168 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
169 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
170
171 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
172 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
173 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
174 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
175 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
176
177 w = (const int8_t*) w + 128;
178 k -= 8 * sizeof(int8_t);
179 }
180 p -= 5 * sizeof(void*);
181 } while (p != 0);
182
183 // Add 4 adjacent sums
184 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
185 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
186 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
187 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
188 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
189 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
190 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
191 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
192 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
193 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
194 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
195 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
196 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
197 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
198 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
199 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
200 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
201 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
202 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
203 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
204 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
205
206 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
207 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
208 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
209 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
210 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
211
212 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
213 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
214 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
215 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
216 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
217
218 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
219 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
220 w = (const float*) w + 32;
221
222 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
223 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
224 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
225 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
226 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
227
228 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
229 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
230 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
231 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
232 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
233
234 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
235 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
236 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
237 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
238 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
239
240 if XNN_LIKELY(nc >= 16) {
241 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
242 c4 = (float*) ((uintptr_t) c4 + cn_stride);
243 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
244 c3 = (float*) ((uintptr_t) c3 + cn_stride);
245 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
246 c2 = (float*) ((uintptr_t) c2 + cn_stride);
247 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
248 c1 = (float*) ((uintptr_t) c1 + cn_stride);
249 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
250 c0 = (float*) ((uintptr_t) c0 + cn_stride);
251 a = (const int8_t**restrict) ((uintptr_t) a - ks);
252 nc -= 16;
253 } else {
254 // Prepare mask for valid 32-bit elements (depends on nc).
255 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
256 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
257 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
258 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
259 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
260 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
261 nc = 0;
262 }
263 } while (nc != 0);
264 }