Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x4c2s4-minmax-wasmsimd-dot16x2-ld64.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_ld64(
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 vxb0 = wasm_i16x8_load8x8(w);
92
93 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
94 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
95 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
96 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
97 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
98 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
99 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb0));
100 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
101 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
102
103 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
104 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
105 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
106 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
107 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
108 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
109 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb1));
110 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
111 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
112
113 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
114 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
115 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
116 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
117 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
118 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
119 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb2));
120 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
121 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
122
123 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
124 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
125 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
126 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb3));
127
128 w = (const int8_t*) w + 32;
129 k -= 8 * sizeof(int8_t);
130 } while (k != 0);
131
132 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
133 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
134 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
135 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
136
137 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
138 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
139 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
140 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
141
142 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
143 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
144 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
145 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
146
147 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
148 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
149 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
150 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
151 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
152 w = (const float*) w + 4;
153
154 const v128_t vbias0123 = wasm_v128_load(w);
155 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
156 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
157 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
158 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
159 w = (const float*) w + 4;
160
161 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
162 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
163 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
164 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
165
166 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
167 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
168 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
169 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
170
171 if XNN_LIKELY(nc >= 4) {
172 wasm_v128_store(c0, vacc0x0123);
173 wasm_v128_store(c1, vacc1x0123);
174 wasm_v128_store(c2, vacc2x0123);
175 wasm_v128_store(c3, vacc3x0123);
176
177 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
178 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
179 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
180 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
181
182 c0 = (float*) ((uintptr_t) c0 + cn_stride);
183 c1 = (float*) ((uintptr_t) c1 + cn_stride);
184 c2 = (float*) ((uintptr_t) c2 + cn_stride);
185 c3 = (float*) ((uintptr_t) c3 + cn_stride);
186
187 nc -= 4;
188 } else {
189 if (nc & 2) {
190 wasm_v128_store64_lane(c0, vacc0x0123, 0);
191 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
192 c0 += 2;
193 wasm_v128_store64_lane(c1, vacc1x0123, 0);
194 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
195 c1 += 2;
196 wasm_v128_store64_lane(c2, vacc2x0123, 0);
197 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
198 c2 += 2;
199 wasm_v128_store64_lane(c3, vacc3x0123, 0);
200 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
201 c3 += 2;
202 }
203 if (nc & 1) {
204 wasm_v128_store32_lane(c0, vacc0x0123, 0);
205 wasm_v128_store32_lane(c1, vacc1x0123, 0);
206 wasm_v128_store32_lane(c2, vacc2x0123, 0);
207 wasm_v128_store32_lane(c3, vacc3x0123, 0);
208 }
209 nc = 0;
210 }
211 } while (nc != 0);
212 }