Debian Codegraph @

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

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_5x16c8__avx512skx(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 5);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
62 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67
68 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
69 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point);
70 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point);
71 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point);
72 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point);
73 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point);
74 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
75 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
76 // XNN_FORCE_REALIZATION(voutput_min);
77 // XNN_FORCE_REALIZATION(voutput_max);
78
79 do {
80 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
81 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
82 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
83 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
84
85 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point0);
86 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point0);
87 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point0);
88 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point0);
89 __m512i vacc1x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point1);
90 __m512i vacc1x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point1);
91 __m512i vacc1x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point1);
92 __m512i vacc1xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point1);
93 __m512i vacc2x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point2);
94 __m512i vacc2x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point2);
95 __m512i vacc2x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point2);
96 __m512i vacc2xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point2);
97 __m512i vacc3x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point3);
98 __m512i vacc3x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point3);
99 __m512i vacc3x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point3);
100 __m512i vacc3xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point3);
101 __m512i vacc4x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point4);
102 __m512i vacc4x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point4);
103 __m512i vacc4x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point4);
104 __m512i vacc4xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point4);
105 w = (const int32_t*) w + 16;
106
107 size_t k = kc;
108
109 while (k >= 8 * sizeof(int8_t)) {
110 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
111 a0 += 8;
112 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
113 a1 += 8;
114 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
115 a2 += 8;
116 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
117 a3 += 8;
118 const __m512i va4 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a4)));
119 a4 += 8;
120
121 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
122
123 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
124 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
125 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
126 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
127 vacc4x0123 = _mm512_add_epi32(vacc4x0123, _mm512_madd_epi16(va4, vb0123));
128 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
129
130 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
131 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
132 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
133 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
134 vacc4x4567 = _mm512_add_epi32(vacc4x4567, _mm512_madd_epi16(va4, vb4567));
135 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
136
137 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
138 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
139 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
140 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
141 vacc4x89AB = _mm512_add_epi32(vacc4x89AB, _mm512_madd_epi16(va4, vb89AB));
142 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
143
144 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
145 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
146 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
147 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
148 vacc4xCDEF = _mm512_add_epi32(vacc4xCDEF, _mm512_madd_epi16(va4, vbCDEF));
149
150 w = (const int8_t*) w + 128;
151 k -= 8 * sizeof(int8_t);
152 }
153
154 // Add 4 adjacent sums
155 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
156 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
157 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
158 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
159 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
160 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
161 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
162 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
163 const __m512i vacc4x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x0123, vacc4x4567), _mm512_unpackhi_epi32(vacc4x0123, vacc4x4567));
164 const __m512i vacc4x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x89AB, vacc4xCDEF), _mm512_unpackhi_epi32(vacc4x89AB, vacc4xCDEF));
165
166 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
167 const __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
168 const __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
169 const __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
170 const __m512i vacc4x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc4x04152637, vacc4x8C9DAEBF), _mm512_unpackhi_epi32(vacc4x04152637, vacc4x8C9DAEBF));
171
172 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
173 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
174 __m512i vacc1x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc1x084C195D2A6E3B7F);
175 __m512i vacc2x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc2x084C195D2A6E3B7F);
176 __m512i vacc3x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc3x084C195D2A6E3B7F);
177 __m512i vacc4x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc4x084C195D2A6E3B7F);
178
179 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
180 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
181 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
182 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
183 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
184
185 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
186 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
187 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
188 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
189 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
190
191 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
192 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
193 w = (const float*) w + 32;
194
195 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
196 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
197 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
198 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
199 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
200
201 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
202 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
203 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
204 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
205 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
206
207 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
208 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
209 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
210 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
211 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
212
213 if XNN_LIKELY(nc >= 16) {
214 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
215 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
216 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
217 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
218 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
219
220 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
221 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
222 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
223 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
224 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
225
226 c0 = (float*) ((uintptr_t) c0 + cn_stride);
227 c1 = (float*) ((uintptr_t) c1 + cn_stride);
228 c2 = (float*) ((uintptr_t) c2 + cn_stride);
229 c3 = (float*) ((uintptr_t) c3 + cn_stride);
230 c4 = (float*) ((uintptr_t) c4 + cn_stride);
231
232 nc -= 16;
233 } else {
234 // Prepare mask for valid 32-bit elements (depends on nc).
235 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
236 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
237 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
238 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
239 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
240 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
241 nc = 0;
242 }
243 } while (nc != 0);
244 }