Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x16c8-minmax-avx512skx-prfm.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 #include "xnnpack/prefetch.h"
18
19 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_1x16c8__avx512skx_prfm(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 8 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 float* c0 = c;
44
45 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
46 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point);
47 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
48 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
49 // XNN_FORCE_REALIZATION(voutput_min);
50 // XNN_FORCE_REALIZATION(voutput_max);
51 const __m256i vmask = _mm256_set1_epi8(0xF0);
52 XNN_FORCE_REALIZATION(vmask);
53
54 do {
55 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
56 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
57 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
58 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
59
60 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point0);
61 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point0);
62 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point0);
63 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point0);
64 w = (const int32_t*) w + 16;
65
66 size_t k = kc;
67 while (k >= 16 * sizeof(int8_t)) {
68 __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
69 a0 += 8;
70
71 __m256i vbb0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
72 __m256i vbs0123 = _mm256_slli_epi32(vbb0123, 4);
73 __m256i vbm0123 = _mm256_and_si256(vbs0123, vmask);
74 __m512i vb0123 = _mm512_cvtepi8_epi16(vbm0123);
75 xnn_prefetch_to_l1((const int8_t*) w + 896);
76
77 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
78 __m256i vbb4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
79 __m256i vbs4567 = _mm256_slli_epi32(vbb4567, 4);
80 __m256i vbm4567 = _mm256_and_si256(vbs4567, vmask);
81 __m512i vb4567 = _mm512_cvtepi8_epi16(vbm4567);
82
83 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
84 __m256i vbb89AB = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
85 __m256i vbs89AB = _mm256_slli_epi32(vbb89AB, 4);
86 __m256i vbm89AB = _mm256_and_si256(vbs89AB, vmask);
87 __m512i vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
88 xnn_prefetch_to_l1((const int8_t*) w + 960);
89
90 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
91 __m256i vbbCDEF = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
92 __m256i vbsCDEF = _mm256_slli_epi32(vbbCDEF, 4);
93 __m256i vbmCDEF = _mm256_and_si256(vbsCDEF, vmask);
94 __m512i vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
95
96 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
97
98 va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
99 a0 += 8;
100
101 vbm0123 = _mm256_and_si256(vbb0123, vmask);
102 vb0123 = _mm512_cvtepi8_epi16(vbm0123);
103
104 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
105 vbm4567 = _mm256_and_si256(vbb4567, vmask);
106 vb4567 = _mm512_cvtepi8_epi16(vbm4567);
107
108 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
109 vbm89AB = _mm256_and_si256(vbb89AB, vmask);
110 vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
111
112 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
113 vbmCDEF = _mm256_and_si256(vbbCDEF, vmask);
114 vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
115
116 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
117
118 w = (const int8_t*) w + 128;
119 k -= 16 * sizeof(int8_t);
120 }
121
122 while (k >= 8 * sizeof(int8_t)) {
123 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
124 a0 += 8;
125
126 const __m256i vbb0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
127 const __m256i vbs0123 = _mm256_slli_epi32(vbb0123, 4);
128 const __m256i vbm0123 = _mm256_and_si256(vbs0123, vmask);
129 const __m512i vb0123 = _mm512_cvtepi8_epi16(vbm0123);
130 xnn_prefetch_to_l1((const int8_t*) w + 896);
131
132 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
133 const __m256i vbb4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
134 const __m256i vbs4567 = _mm256_slli_epi32(vbb4567, 4);
135 const __m256i vbm4567 = _mm256_and_si256(vbs4567, vmask);
136 const __m512i vb4567 = _mm512_cvtepi8_epi16(vbm4567);
137
138 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
139 const __m256i vbb89AB = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
140 const __m256i vbs89AB = _mm256_slli_epi32(vbb89AB, 4);
141 const __m256i vbm89AB = _mm256_and_si256(vbs89AB, vmask);
142 const __m512i vb89AB = _mm512_cvtepi8_epi16(vbm89AB);
143 xnn_prefetch_to_l1((const int8_t*) w + 960);
144
145 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
146 const __m256i vbbCDEF = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
147 const __m256i vbsCDEF = _mm256_slli_epi32(vbbCDEF, 4);
148 const __m256i vbmCDEF = _mm256_and_si256(vbsCDEF, vmask);
149 const __m512i vbCDEF = _mm512_cvtepi8_epi16(vbmCDEF);
150
151 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
152
153 w = (const int8_t*) w + 128;
154 k -= 8 * sizeof(int8_t);
155 }
156
157 // Add 4 adjacent sums
158 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
159 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
160
161 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
162
163 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
164 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
165
166 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
167 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
168
169 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
170
171 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
172 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
173 w = (const float*) w + 32;
174
175 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
176
177 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
178
179 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
180
181 if XNN_LIKELY(nc >= 16) {
182 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
183
184 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
185
186 c0 = (float*) ((uintptr_t) c0 + cn_stride);
187
188 nc -= 16;
189 } else {
190 // Prepare mask for valid 32-bit elements (depends on nc).
191 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
192 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
193 nc = 0;
194 }
195 } while (nc != 0);
196 }