Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x4c16-minmax-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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
18 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_4x4c16__wasmsdot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (4 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 16 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
61 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
62 XNN_FORCE_REALIZATION(vmin);
63 XNN_FORCE_REALIZATION(vmax);
64
65 do {
66 v128_t vksum0 = wasm_v128_load32_zero(w);
67 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
68 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
69 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
70 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
71 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point);
72 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point);
73 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point);
74 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point);
75 v128_t vacc1x0 = wasm_i32x4_mul(vksum0, vinput_zero_point);
76 v128_t vacc1x1 = wasm_i32x4_mul(vksum1, vinput_zero_point);
77 v128_t vacc1x2 = wasm_i32x4_mul(vksum2, vinput_zero_point);
78 v128_t vacc1x3 = wasm_i32x4_mul(vksum3, vinput_zero_point);
79 v128_t vacc2x0 = wasm_i32x4_mul(vksum0, vinput_zero_point);
80 v128_t vacc2x1 = wasm_i32x4_mul(vksum1, vinput_zero_point);
81 v128_t vacc2x2 = wasm_i32x4_mul(vksum2, vinput_zero_point);
82 v128_t vacc2x3 = wasm_i32x4_mul(vksum3, vinput_zero_point);
83 v128_t vacc3x0 = wasm_i32x4_mul(vksum0, vinput_zero_point);
84 v128_t vacc3x1 = wasm_i32x4_mul(vksum1, vinput_zero_point);
85 v128_t vacc3x2 = wasm_i32x4_mul(vksum2, vinput_zero_point);
86 v128_t vacc3x3 = wasm_i32x4_mul(vksum3, vinput_zero_point);
87 w = (const int32_t*) w + 4;
88
89 size_t p = ks;
90 do {
91 const int8_t* restrict a0 = a[0];
92 if XNN_UNPREDICTABLE(a0 != zero) {
93 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
94 } else {
95 a0 = zero_data;
96 }
97 const int8_t* restrict a1 = a[1];
98 if XNN_UNPREDICTABLE(a1 != zero) {
99 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
100 } else {
101 a1 = zero_data;
102 }
103 const int8_t* restrict a2 = a[2];
104 if XNN_UNPREDICTABLE(a2 != zero) {
105 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
106 } else {
107 a2 = zero_data;
108 }
109 const int8_t* restrict a3 = a[3];
110 if XNN_UNPREDICTABLE(a3 != zero) {
111 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
112 } else {
113 a3 = zero_data;
114 }
115 a += 4;
116
117 size_t k = kc;
118 do {
119 const v128_t va0 = wasm_v128_load(a0);
120 a0 += 16;
121 const v128_t va1 = wasm_v128_load(a1);
122 a1 += 16;
123 const v128_t va2 = wasm_v128_load(a2);
124 a2 += 16;
125 const v128_t va3 = wasm_v128_load(a3);
126 a3 += 16;
127
128 const v128_t vb0 = wasm_v128_load(w);
129 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
130 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
131 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
132 vacc3x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va3, vacc3x0);
133 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
134 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
135 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
136 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
137 vacc3x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va3, vacc3x1);
138 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
139 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
140 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
141 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
142 vacc3x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va3, vacc3x2);
143 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
144 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
145 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
146 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
147 vacc3x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va3, vacc3x3);
148
149 w = (const int8_t*) w + 64;
150 k -= 16 * sizeof(int8_t);
151 } while (k != 0);
152 p -= 4 * sizeof(void*);
153 } while (p != 0);
154
155 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));
156 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));
157 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));
158 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));
159 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));
160 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));
161 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));
162 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));
163
164 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
165 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
166 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
167 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
168
169 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
170 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
171 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
172 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
173
174 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
175
176 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
177 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
178 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale);
179 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale);
180
181 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
182 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
183 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
184 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
185 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
186 w = (const float*) w + 4;
187
188 const v128_t vbias0123 = wasm_v128_load(w);
189 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
190 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
191 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
192 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
193 w = (const float*) w + 4;
194
195 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
196 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
197 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
198 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
199
200 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
201 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
202 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
203 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
204
205 if XNN_LIKELY(nc >= 4) {
206 wasm_v128_store(c3, vacc3x0123);
207 wasm_v128_store(c2, vacc2x0123);
208 wasm_v128_store(c1, vacc1x0123);
209 wasm_v128_store(c0, vacc0x0123);
210
211 c0 = (float*) ((uintptr_t) c0 + cn_stride);
212 c1 = (float*) ((uintptr_t) c1 + cn_stride);
213 c2 = (float*) ((uintptr_t) c2 + cn_stride);
214 c3 = (float*) ((uintptr_t) c3 + cn_stride);
215
216 a = (const int8_t**restrict) ((uintptr_t) a - ks);
217
218 nc -= 4;
219 } else {
220 if (nc & 2) {
221 wasm_v128_store64_lane(c3, vacc3x0123, 0);
222 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
223 c3 += 2;
224 wasm_v128_store64_lane(c2, vacc2x0123, 0);
225 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
226 c2 += 2;
227 wasm_v128_store64_lane(c1, vacc1x0123, 0);
228 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
229 c1 += 2;
230 wasm_v128_store64_lane(c0, vacc0x0123, 0);
231 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
232 c0 += 2;
233 }
234 if (nc & 1) {
235 wasm_v128_store32_lane(c3, vacc3x0123, 0);
236 wasm_v128_store32_lane(c2, vacc2x0123, 0);
237 wasm_v128_store32_lane(c1, vacc1x0123, 0);
238 wasm_v128_store32_lane(c0, vacc0x0123, 0);
239 }
240 nc = 0;
241 }
242 } while (nc != 0);
243 }