Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x4c2s4-minmax-wasmsimd-dot16x2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c2s4-wasmsimd-dot16x2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 <wasm_simd128.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_4x4c2s4__wasmsimd_dot16x2_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 <= 4);
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 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr != 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61
62 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
63 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
64 XNN_FORCE_REALIZATION(vmin);
65 XNN_FORCE_REALIZATION(vmax);
66
67 kc = round_up_po2(kc, 8 * sizeof(int8_t));
68 do {
69 const v128_t vksum0123 = wasm_v128_load(w);
70 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
71 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
72 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
73 const v128_t vinput_zero_point3 = wasm_v128_load32_splat(&quantization_params[3].zero_point);
74 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
75 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
76 v128_t vacc2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
77 v128_t vacc3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point3);
78 w = (const int32_t*) w + 4;
79
80 size_t k = kc;
81 do {
82 v128_t vxa0 = wasm_i16x8_load8x8((const v128_t*) a0);
83 a0 += 8;
84 v128_t vxa1 = wasm_i16x8_load8x8((const v128_t*) a1);
85 a1 += 8;
86 v128_t vxa2 = wasm_i16x8_load8x8((const v128_t*) a2);
87 a2 += 8;
88 v128_t vxa3 = wasm_i16x8_load8x8((const v128_t*) a3);
89 a3 += 8;
90
91 const v128_t vb01 = wasm_v128_load(w);
92 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
93 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
94
95 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
96 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
97 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
98 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
99 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
100 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
101 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb0));
102 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
103
104 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
105 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
106 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
107 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
108 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
109 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
110 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb1));
111 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
112 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
113 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
114 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
115
116 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
117 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
118 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
119 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
120 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
121 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
122 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb2));
123 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
124
125 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
126 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
127 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
128 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb3));
129
130 w = (const int8_t*) w + 32;
131 k -= 8 * sizeof(int8_t);
132 } while (k != 0);
133
134 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
135 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
136 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
137 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
138
139 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
140 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
141 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
142 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
143
144 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
145 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
146 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
147 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
148
149 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
150 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
151 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
152 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
153 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
154 w = (const float*) w + 4;
155
156 const v128_t vbias0123 = wasm_v128_load(w);
157 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
158 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
159 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
160 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
161 w = (const float*) w + 4;
162
163 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
164 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
165 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
166 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
167
168 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
169 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
170 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
171 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
172
173 if XNN_LIKELY(nc >= 4) {
174 wasm_v128_store(c0, vacc0x0123);
175 wasm_v128_store(c1, vacc1x0123);
176 wasm_v128_store(c2, vacc2x0123);
177 wasm_v128_store(c3, vacc3x0123);
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 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
183
184 c0 = (float*) ((uintptr_t) c0 + cn_stride);
185 c1 = (float*) ((uintptr_t) c1 + cn_stride);
186 c2 = (float*) ((uintptr_t) c2 + cn_stride);
187 c3 = (float*) ((uintptr_t) c3 + cn_stride);
188
189 nc -= 4;
190 } else {
191 if (nc & 2) {
192 wasm_v128_store64_lane(c0, vacc0x0123, 0);
193 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
194 c0 += 2;
195 wasm_v128_store64_lane(c1, vacc1x0123, 0);
196 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
197 c1 += 2;
198 wasm_v128_store64_lane(c2, vacc2x0123, 0);
199 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
200 c2 += 2;
201 wasm_v128_store64_lane(c3, vacc3x0123, 0);
202 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
203 c3 += 2;
204 }
205 if (nc & 1) {
206 wasm_v128_store32_lane(c0, vacc0x0123, 0);
207 wasm_v128_store32_lane(c1, vacc1x0123, 0);
208 wasm_v128_store32_lane(c2, vacc2x0123, 0);
209 wasm_v128_store32_lane(c3, vacc3x0123, 0);
210 }
211 nc = 0;
212 }
213 } while (nc != 0);
214 }