Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-2x8c8-minmax-avx2-madd.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_f16_qc4w_gemm_minmax_ukernel_2x8c8__avx2_madd(
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 xnn_float16* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f16_qc4w_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 <= 2);
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 uint16_t* c0 = (uint16_t*) c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr != 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51
52 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
53 XNN_FORCE_REALIZATION(vsign_mask);
54 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
55 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
56 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
57 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
58 // XNN_FORCE_REALIZATION(voutput_min);
59 // XNN_FORCE_REALIZATION(voutput_max);
60 const __m256i vmask = _mm256_set1_epi8(0x0F);
61 XNN_FORCE_REALIZATION(vmask);
62 do {
63 const __m256i vksum01234567 = _mm256_load_si256(w);
64 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
65 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
66 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
67 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
68 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
69 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
70 __m256i vacc1x0x0123 = _mm256_setzero_si256();
71 __m256i vacc1x0x4567 = _mm256_setzero_si256();
72 __m256i vacc1x1x0123 = _mm256_setzero_si256();
73 __m256i vacc1x1x4567 = _mm256_setzero_si256();
74 w = (const int32_t*) w + 8;
75
76 size_t k = kc;
77 while (k >= 16 * sizeof(int8_t)) {
78 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
79 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
80 a0 += 16;
81 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
82 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
83 a1 += 16;
84
85 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
86 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
87 const __m256i vbs01234567x0123 = _mm256_srli_epi32(vbb01234567x01234567, 4);
88 const __m256i vbs89ABCDEFx0123 = _mm256_srli_epi32(vbb89ABCDEFx01234567, 4);
89 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
90 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
91 const __m256i vb01234567x4567 = _mm256_and_si256(vbs01234567x0123, vmask);
92 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
93
94 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
95 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
96 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
97 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
98 vacc1x0x0123 = _mm256_dpbusd_epi32_madd(vacc1x0x0123, va0x89ABCDEF, vb01234567x4567);
99 vacc1x0x4567 = _mm256_dpbusd_epi32_madd(vacc1x0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
100 vacc1x1x0123 = _mm256_dpbusd_epi32_madd(vacc1x1x0123, va1x89ABCDEF, vb01234567x4567);
101 vacc1x1x4567 = _mm256_dpbusd_epi32_madd(vacc1x1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
102
103 w = (const int8_t*) w + 64;
104 k -= 16 * sizeof(int8_t);
105 }
106
107 if (k != 0) {
108 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
109 a0 += 8;
110 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
111 a1 += 8;
112
113 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
114 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
115 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
116 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
117
118 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
119 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
120 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
121 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
122
123 w = (const int8_t*) w + 64;
124 k -= 8 * sizeof(int8_t);
125 }
126 vacc0x0123 = _mm256_add_epi32(vacc0x0123, vacc1x0x0123);
127 vacc0x4567 = _mm256_add_epi32(vacc0x4567, vacc1x0x4567);
128 vacc1x0123 = _mm256_add_epi32(vacc1x0123, vacc1x1x0123);
129 vacc1x4567 = _mm256_add_epi32(vacc1x4567, vacc1x1x4567);
130
131 // Add adjacent pairs
132 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
133 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
134 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
135 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
136
137 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
138 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
139
140 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
141 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
142
143 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
144 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
145 w = (const float*) w + 16;
146
147 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
148 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
149
150 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
151 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
152
153 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
154 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
155
156 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
157 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
158 if XNN_LIKELY(nc >= 8) {
159 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
160 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
161 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
162 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
163 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
164 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
165 nc -= 8;
166 } else {
167 if (nc & 4) {
168 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
169 c0 += 4;
170 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
171 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
172 c1 += 4;
173 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
174 }
175 if (nc & 2) {
176 _mm_storeu_si32(c0, vfp16out0x01234567);
177 c0 += 2;
178 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
179 _mm_storeu_si32(c1, vfp16out1x01234567);
180 c1 += 2;
181 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
182 }
183 if (nc & 1) {
184 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
185 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
186 }
187 nc = 0;
188 }
189 } while (nc != 0);
190 }