Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x4c8-minmax-sse41-ld128.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 #include <smmintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_3x4c8__sse41_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 8 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 float* c0 = c;
44 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56
57 const __m128 vmin = _mm_set1_ps(params->scalar.min);
58 const __m128 vmax = _mm_set1_ps(params->scalar.max);
59 XNN_FORCE_REALIZATION(vmin);
60 XNN_FORCE_REALIZATION(vmax);
61
62
63 do {
64 const __m128i vksum = _mm_load_si128((const __m128i*) w);
65 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
66 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
67 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
68 __m128i vinput_zero_point2 = _mm_cvtsi32_si128(*((const int*) &quantization_params[2].zero_point));
69 vinput_zero_point2 = _mm_shuffle_epi32(vinput_zero_point2, _MM_SHUFFLE(0, 0, 0, 0));
70 const __m128i vzero = _mm_setzero_si128();
71 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
72 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point1);
73 const __m128i vinit2 = _mm_mullo_epi32(vksum, vinput_zero_point2);
74 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
75 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
76 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
77 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
78 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
79 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
80 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
81 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
82 __m128i vacc2x0 = _mm_blend_epi16(vinit2, vzero, 0xFC);
83 __m128i vacc2x1 = _mm_blend_epi16(vinit2, vzero, 0xF3);
84 __m128i vacc2x2 = _mm_blend_epi16(vinit2, vzero, 0xCF);
85 __m128i vacc2x3 = _mm_blend_epi16(vinit2, vzero, 0x3F);
86 w = (const int32_t*) w + 4;
87
88 size_t k = kc;
89
90
91 while (k >= 8 * sizeof(int8_t)) {
92 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
93 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
94 a0 += 8;
95 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
96 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
97 a1 += 8;
98 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
99 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
100 a2 += 8;
101
102 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
103
104 const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01);
105 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
106 const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8);
107
108 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
109 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
110 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
111 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
112 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
113 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
114 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
115
116 const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23);
117 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
118 const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8);
119
120 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
121 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
122 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
123 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
124 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
125 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
126
127 w = (const int8_t*) w + 32;
128 k -= 8 * sizeof(int8_t);
129 }
130
131 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
132 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
133 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
134 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
135 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
136 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
137
138 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
139 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
140 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
141
142 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
143 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
144 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
145
146 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
147 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
148 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
149 const __m128 vinput_scale2 = _mm_load1_ps(&quantization_params[2].inv_scale);
150
151 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
152 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
153 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
154
155 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
156 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
157 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
158 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
159
160 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
161 w = (const float*) w + 8;
162 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
163 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
164 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
165
166 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
167 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
168 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
169
170 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
171 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
172 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
173
174 if XNN_LIKELY(nc >= 4) {
175 _mm_storeu_ps(c0, vout0x0123);
176 _mm_storeu_ps(c1, vout1x0123);
177 _mm_storeu_ps(c2, vout2x0123);
178
179 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
180 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
181 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
182
183 c0 = (float*) ((uintptr_t) c0 + cn_stride);
184 c1 = (float*) ((uintptr_t) c1 + cn_stride);
185 c2 = (float*) ((uintptr_t) c2 + cn_stride);
186
187 nc -= 4;
188 } else {
189 if (nc & 2) {
190 _mm_storel_pi((__m64*) c0, vout0x0123);
191 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
192 c0 += 2;
193 _mm_storel_pi((__m64*) c1, vout1x0123);
194 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
195 c1 += 2;
196 _mm_storel_pi((__m64*) c2, vout2x0123);
197 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
198 c2 += 2;
199 }
200 if (nc & 1) {
201 _mm_store_ss(c0, vout0x0123);
202 _mm_store_ss(c1, vout1x0123);
203 _mm_store_ss(c2, vout2x0123);
204 }
205 nc = 0;
206 }
207 } while (nc != 0);
208 }