Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x16c8-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_1x16c8__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 <= 1);
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
44 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
45 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point);
46 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
47 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
48 // XNN_FORCE_REALIZATION(voutput_min);
49 // XNN_FORCE_REALIZATION(voutput_max);
50
51 do {
52 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
53 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
54 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
55 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
56
57 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point0);
58 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point0);
59 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point0);
60 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point0);
61 w = (const int32_t*) w + 16;
62
63 size_t k = kc;
64
65 while (k >= 8 * sizeof(int8_t)) {
66 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
67 a0 += 8;
68
69 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
70
71 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
72 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
73
74 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
75 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
76
77 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
78 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
79
80 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
81
82 w = (const int8_t*) w + 128;
83 k -= 8 * sizeof(int8_t);
84 }
85
86 // Add 4 adjacent sums
87 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
88 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
89
90 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
91
92 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
93 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
94
95 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
96
97 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
98
99 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
100 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
101 w = (const float*) w + 32;
102
103 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
104
105 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
106
107 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
108
109 if XNN_LIKELY(nc >= 16) {
110 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
111
112 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
113
114 c0 = (float*) ((uintptr_t) c0 + cn_stride);
115
116 nc -= 16;
117 } else {
118 // Prepare mask for valid 32-bit elements (depends on nc).
119 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
120 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
121 nc = 0;
122 }
123 } while (nc != 0);
124 }