Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x8c8-minmax-avx256vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c8-avxvnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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/unaligned.h"
18
19
20 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_1x8c8__avx256vnni(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45
46 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
47 XNN_FORCE_REALIZATION(vsign_mask);
48 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
49 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
50 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
51 // XNN_FORCE_REALIZATION(voutput_min);
52 // XNN_FORCE_REALIZATION(voutput_max);
53 do {
54 const __m256i vksum01234567 = _mm256_load_si256(w);
55 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
56 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
57 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
58 __m256i vacc1x0x0123 = _mm256_setzero_si256();
59 __m256i vacc1x0x4567 = _mm256_setzero_si256();
60 w = (const int32_t*) w + 8;
61
62 size_t k = kc;
63 while (k >= 16 * sizeof(int8_t)) {
64 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
65 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
66 a0 += 16;
67
68 const __m256i vb01234567x0123 = _mm256_load_si256(w);
69 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
70 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
71 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
72
73 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
74 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
75 vacc1x0x0123 = _mm256_dpbusd_epi32(vacc1x0x0123, va0x89ABCDEF, vb01234567x4567);
76 vacc1x0x4567 = _mm256_dpbusd_epi32(vacc1x0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
77
78 w = (const int8_t*) w + 128;
79 k -= 16 * sizeof(int8_t);
80 }
81
82 if (k != 0) {
83 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
84 a0 += 8;
85
86 const __m256i vb01234567x0123 = _mm256_load_si256(w);
87 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
88
89 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
90 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
91
92 w = (const int8_t*) w + 64;
93 k -= 8 * sizeof(int8_t);
94 }
95 vacc0x0123 = _mm256_add_epi32(vacc0x0123, vacc1x0x0123);
96 vacc0x4567 = _mm256_add_epi32(vacc0x4567, vacc1x0x4567);
97
98 // Add adjacent pairs
99 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
100 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
101
102 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
103
104 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
105
106 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
107 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
108 w = (const float*) w + 16;
109
110 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
111
112 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
113
114 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
115
116 if XNN_LIKELY(nc >= 8) {
117 _mm256_storeu_ps(c0, vout0x01234567);
118 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
119 c0 = (float*) ((uintptr_t) c0 + cn_stride);
120 nc -= 8;
121 } else {
122 // Prepare mask for valid 32-bit elements (depends on nc).
123 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
124 _mm256_mask_storeu_ps(c0, vmask, vout0x01234567);
125 nc = 0;
126 }
127 } while (nc != 0);
128 }