Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x4c16-minmax-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c16-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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/gemm.h"
15 #include "xnnpack/math.h"
16
17 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x4c16__wasmusdot(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 float* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
28 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 16 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 float* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60
61 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
62 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
63 XNN_FORCE_REALIZATION(vmin);
64 XNN_FORCE_REALIZATION(vmax);
65
66 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
67 do {
68 v128_t vksum0 = wasm_v128_load32_zero(w);
69 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
70 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
71 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
72 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
73 const v128_t vinput_zero_point1 = wasm_i32x4_splat((int32_t) quantization_params[1].zero_point + 128);
74 const v128_t vinput_zero_point2 = wasm_i32x4_splat((int32_t) quantization_params[2].zero_point + 128);
75 const v128_t vinput_zero_point3 = wasm_i32x4_splat((int32_t) quantization_params[3].zero_point + 128);
76 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point0);
77 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point0);
78 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point0);
79 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point0);
80 v128_t vacc1x0 = wasm_i32x4_mul(vksum0, vinput_zero_point1);
81 v128_t vacc1x1 = wasm_i32x4_mul(vksum1, vinput_zero_point1);
82 v128_t vacc1x2 = wasm_i32x4_mul(vksum2, vinput_zero_point1);
83 v128_t vacc1x3 = wasm_i32x4_mul(vksum3, vinput_zero_point1);
84 v128_t vacc2x0 = wasm_i32x4_mul(vksum0, vinput_zero_point2);
85 v128_t vacc2x1 = wasm_i32x4_mul(vksum1, vinput_zero_point2);
86 v128_t vacc2x2 = wasm_i32x4_mul(vksum2, vinput_zero_point2);
87 v128_t vacc2x3 = wasm_i32x4_mul(vksum3, vinput_zero_point2);
88 v128_t vacc3x0 = wasm_i32x4_mul(vksum0, vinput_zero_point3);
89 v128_t vacc3x1 = wasm_i32x4_mul(vksum1, vinput_zero_point3);
90 v128_t vacc3x2 = wasm_i32x4_mul(vksum2, vinput_zero_point3);
91 v128_t vacc3x3 = wasm_i32x4_mul(vksum3, vinput_zero_point3);
92 w = (const int32_t*) w + 4;
93
94 size_t k = kc;
95 do {
96 const v128_t va0 = wasm_v128_xor(wasm_v128_load(a0), vsign_mask);
97 a0 += 16;
98 const v128_t va1 = wasm_v128_xor(wasm_v128_load(a1), vsign_mask);
99 a1 += 16;
100 const v128_t va2 = wasm_v128_xor(wasm_v128_load(a2), vsign_mask);
101 a2 += 16;
102 const v128_t va3 = wasm_v128_xor(wasm_v128_load(a3), vsign_mask);
103 a3 += 16;
104
105 const v128_t vb0 = wasm_v128_load(w);
106
107 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
108 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
109 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
110 vacc3x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va3, vacc3x0);
111 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
112
113 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
114 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
115 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
116 vacc3x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va3, vacc3x1);
117 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
118
119 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
120 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
121 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
122 vacc3x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va3, vacc3x2);
123 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
124
125 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
126 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
127 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
128 vacc3x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va3, vacc3x3);
129
130 w = (const int8_t*) w + 64;
131 k -= 16 * sizeof(int8_t);
132 } while (k != 0);
133
134 const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7));
135 const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7));
136 const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7));
137 const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7));
138 const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
139 const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));
140 const v128_t vacc3x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x0, vacc3x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x0, vacc3x2, 2, 6, 3, 7));
141 const v128_t vacc3x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x1, vacc3x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x1, vacc3x3, 2, 6, 3, 7));
142
143 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
144 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
145 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
146 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
147
148 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
149 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
150 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
151 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
152
153 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
154 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
155 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
156 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
157
158 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
159 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
160 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
161 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
162
163 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
164 w = (const float*) w + 4;
165 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
166 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
167 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
168 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
169
170 const v128_t vbias0123 = wasm_v128_load(w);
171 w = (const float*) w + 4;
172 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
173 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
174 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
175 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
176
177 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
178 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
179 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
180 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
181
182 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
183 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
184 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
185 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
186
187 if XNN_LIKELY(nc >= 4) {
188 wasm_v128_store(c0, vacc0x0123);
189 wasm_v128_store(c1, vacc1x0123);
190 wasm_v128_store(c2, vacc2x0123);
191 wasm_v128_store(c3, vacc3x0123);
192
193 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
194 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
195 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
196 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
197
198 c0 = (float*) ((uintptr_t) c0 + cn_stride);
199 c1 = (float*) ((uintptr_t) c1 + cn_stride);
200 c2 = (float*) ((uintptr_t) c2 + cn_stride);
201 c3 = (float*) ((uintptr_t) c3 + cn_stride);
202
203 nc -= 4;
204 } else {
205 if (nc & 2) {
206 wasm_v128_store64_lane(c0, vacc0x0123, 0);
207 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
208 c0 += 2;
209 wasm_v128_store64_lane(c1, vacc1x0123, 0);
210 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
211 c1 += 2;
212 wasm_v128_store64_lane(c2, vacc2x0123, 0);
213 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
214 c2 += 2;
215 wasm_v128_store64_lane(c3, vacc3x0123, 0);
216 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
217 c3 += 2;
218 }
219 if (nc & 1) {
220 wasm_v128_store32_lane(c0, vacc0x0123, 0);
221 wasm_v128_store32_lane(c1, vacc1x0123, 0);
222 wasm_v128_store32_lane(c2, vacc2x0123, 0);
223 wasm_v128_store32_lane(c3, vacc3x0123, 0);
224 }
225 nc = 0;
226 }
227 } while (nc != 0);
228 }