Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-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_qc8w_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 union xnn_f32_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
52 do {
53 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
54 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
55 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
56 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
57
58 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point0);
59 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point0);
60 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point0);
61 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point0);
62 w = (const int32_t*) w + 16;
63
64 size_t k = kc;
65
66 while (k >= 8 * sizeof(int8_t)) {
67 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
68 a0 += 8;
69
70 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
71 xnn_prefetch_to_l1((const int8_t*) w + 896);
72
73 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
74 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
75
76 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
77 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
78 xnn_prefetch_to_l1((const int8_t*) w + 960);
79
80 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
81 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
82
83 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
84
85 w = (const int8_t*) w + 128;
86 k -= 8 * sizeof(int8_t);
87 }
88
89 // Add 4 adjacent sums
90 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
91 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
92
93 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
94
95 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
96 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
97
98 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
99
100 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
101
102 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
103 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
104 w = (const float*) w + 32;
105
106 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
107
108 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
109
110 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
111
112 if XNN_LIKELY(nc >= 16) {
113 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
114
115 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
116
117 c0 = (float*) ((uintptr_t) c0 + cn_stride);
118
119 nc -= 16;
120 } else {
121 // Prepare mask for valid 32-bit elements (depends on nc).
122 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
123 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
124 nc = 0;
125 }
126 } while (nc != 0);
127 }