Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-1x8c8-minmax-avx256skx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c8-avx2.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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19
20
21 void xnn_qd8_f16_qc4w_gemm_minmax_ukernel_1x8c8__avx256skx(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f16_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 uint16_t* c0 = (uint16_t*) c;
46
47 const __m128i vmask = _mm_set1_epi8(0xF0);
48 XNN_FORCE_REALIZATION(vmask);
49 const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
50 const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
51 XNN_FORCE_REALIZATION(vmin);
52 XNN_FORCE_REALIZATION(vmax);
53
54 do {
55 const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
56 const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
57 const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
58 const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
59 const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
60 const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
61 const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
62 const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
63 const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
64 const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
65 const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
66 const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
67 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
68 __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0);
69 __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0);
70 __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0);
71 __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0);
72 w = (const int32_t*) w + 8;
73
74 size_t k = kc;
75 while (k >= 16 * sizeof(int8_t)) {
76 __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
77 __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
78 a0 += 8;
79
80 __m128i vb01 = _mm_load_si128((const __m128i*) w);
81 __m128i vbs01 = _mm_slli_epi32(vb01, 4);
82 __m128i vbm01 = _mm_and_si128(vbs01, vmask);
83 __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
84
85 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
86 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
87 __m128i vbs23 = _mm_slli_epi32(vb23, 4);
88 __m128i vbm23 = _mm_and_si128(vbs23, vmask);
89 __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
90
91 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
92 __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
93 __m128i vbs45 = _mm_slli_epi32(vb45, 4);
94 __m128i vbm45 = _mm_and_si128(vbs45, vmask);
95 __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
96
97 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
98 __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
99 __m128i vbs67 = _mm_slli_epi32(vb67, 4);
100 __m128i vbm67 = _mm_and_si128(vbs67, vmask);
101 __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
102
103 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
104
105 va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
106 vxa0 = _mm256_cvtepi8_epi16(va0);
107 a0 += 8;
108
109 vbm01 = _mm_and_si128(vb01, vmask);
110 vxb01 = _mm256_cvtepi8_epi16(vbm01);
111
112 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
113 vbm23 = _mm_and_si128(vb23, vmask);
114 vxb23 = _mm256_cvtepi8_epi16(vbm23);
115
116 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
117 vbm45 = _mm_and_si128(vb45, vmask);
118 vxb45 = _mm256_cvtepi8_epi16(vbm45);
119
120 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
121 vbm67 = _mm_and_si128(vb67, vmask);
122 vxb67 = _mm256_cvtepi8_epi16(vbm67);
123
124 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
125
126 w = (const int8_t*) w + 64;
127 k -= 16 * sizeof(int8_t);
128 }
129
130 while (k >= 8 * sizeof(int8_t)) {
131 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
132 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
133 a0 += 8;
134
135 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
136 const __m128i vbs01 = _mm_slli_epi32(vb01, 4);
137 const __m128i vbm01 = _mm_and_si128(vbs01, vmask);
138 const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01);
139
140 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
141 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
142 const __m128i vbs23 = _mm_slli_epi32(vb23, 4);
143 const __m128i vbm23 = _mm_and_si128(vbs23, vmask);
144 const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23);
145
146 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
147 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
148 const __m128i vbs45 = _mm_slli_epi32(vb45, 4);
149 const __m128i vbm45 = _mm_and_si128(vbs45, vmask);
150 const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45);
151
152 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
153 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
154 const __m128i vbs67 = _mm_slli_epi32(vb67, 4);
155 const __m128i vbm67 = _mm_and_si128(vbs67, vmask);
156 const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67);
157
158 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
159
160 w = (const int8_t*) w + 64;
161 k -= 8 * sizeof(int8_t);
162 }
163
164 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
165 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
166
167 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
168
169 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
170 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
171
172 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
173 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
174 const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale);
175
176 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0);
177
178 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
179 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
180 w = (const float*) w + 16;
181 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
182
183 vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);
184
185 vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);
186 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
187 if XNN_LIKELY(nc >= 8) {
188 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
189 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
190
191 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
192
193 nc -= 8;
194 } else {
195 // Prepare mask for valid 16-bit elements (depends on nc).
196 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
197 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
198 nc = 0;
199 }
200 } while (nc != 0);
201 }