Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x4c8-minmax-avx-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-sse.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 #ifdef _MSC_VER
13 #include <intrin.h>
14 #else
15 #include <x86intrin.h>
16 #endif
17
18 #include "xnnpack/common.h"
19 #include "xnnpack/gemm.h"
20 #include "xnnpack/math.h"
21
22
23 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_3x4c8__avx_ld64(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 const int8_t* restrict a,
28 size_t a_stride,
29 const void* restrict w,
30 float* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 3);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 const int8_t* a0 = a;
47 float* c0 = c;
48 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
49 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 a1 = a0;
52 c1 = c0;
53 }
54 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
55 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 a2 = a1;
58 c2 = c1;
59 }
60
61 const __m128 vmin = _mm_set1_ps(params->scalar.min);
62 const __m128 vmax = _mm_set1_ps(params->scalar.max);
63 XNN_FORCE_REALIZATION(vmin);
64 XNN_FORCE_REALIZATION(vmax);
65
66
67 do {
68 const __m128i vksum = _mm_load_si128((const __m128i*) w);
69 const __m128i vinput_zero_point0 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[0].zero_point));
70 const __m128i vinput_zero_point1 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[1].zero_point));
71 const __m128i vinput_zero_point2 = _mm_castps_si128(_mm_broadcast_ss((const float*) &quantization_params[2].zero_point));
72 const __m128i vzero = _mm_setzero_si128();
73 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
74 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point1);
75 const __m128i vinit2 = _mm_mullo_epi32(vksum, vinput_zero_point2);
76 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
77 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
78 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
79 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
80 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
81 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
82 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
83 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
84 __m128i vacc2x0 = _mm_blend_epi16(vinit2, vzero, 0xFC);
85 __m128i vacc2x1 = _mm_blend_epi16(vinit2, vzero, 0xF3);
86 __m128i vacc2x2 = _mm_blend_epi16(vinit2, vzero, 0xCF);
87 __m128i vacc2x3 = _mm_blend_epi16(vinit2, vzero, 0x3F);
88 w = (const int32_t*) w + 4;
89
90 size_t k = kc;
91
92
93 while (k >= 8 * sizeof(int8_t)) {
94 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
95 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
96 a0 += 8;
97 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
98 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
99 a1 += 8;
100 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
101 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
102 a2 += 8;
103
104 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
105
106 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
107
108 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
109 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
110 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
111 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
112
113 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
114
115 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
116 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
117 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
118 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
119
120 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
121
122 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
123 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
124 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
125 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
126
127 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
128
129 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
130 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
131 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
132
133 w = (const int8_t*) w + 32;
134 k -= 8 * sizeof(int8_t);
135 }
136
137 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
138 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
139 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
140 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
141 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
142 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
143
144 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
145 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
146 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
147
148 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
149 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
150 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
151
152 const __m128 vinput_scale0 = _mm_broadcast_ss(&quantization_params[0].inv_scale);
153 const __m128 vinput_scale1 = _mm_broadcast_ss(&quantization_params[1].inv_scale);
154 const __m128 vinput_scale2 = _mm_broadcast_ss(&quantization_params[2].inv_scale);
155
156 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
157 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
158 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
159
160 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
161 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
162 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
163 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
164
165 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
166 w = (const float*) w + 8;
167 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
168 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
169 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
170
171 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
172 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
173 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
174
175 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
176 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
177 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
178
179 if XNN_LIKELY(nc >= 4) {
180 _mm_storeu_ps(c0, vout0x0123);
181 _mm_storeu_ps(c1, vout1x0123);
182 _mm_storeu_ps(c2, vout2x0123);
183
184 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
185 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
186 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
187
188 c0 = (float*) ((uintptr_t) c0 + cn_stride);
189 c1 = (float*) ((uintptr_t) c1 + cn_stride);
190 c2 = (float*) ((uintptr_t) c2 + cn_stride);
191
192 nc -= 4;
193 } else {
194 if (nc & 2) {
195 _mm_storel_pi((__m64*) c0, vout0x0123);
196 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
197 c0 += 2;
198 _mm_storel_pi((__m64*) c1, vout1x0123);
199 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
200 c1 += 2;
201 _mm_storel_pi((__m64*) c2, vout2x0123);
202 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
203 c2 += 2;
204 }
205 if (nc & 1) {
206 _mm_store_ss(c0, vout0x0123);
207 _mm_store_ss(c1, vout1x0123);
208 _mm_store_ss(c2, vout2x0123);
209 }
210 nc = 0;
211 }
212 } while (nc != 0);
213 }