Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c8-minmax-fp32-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-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_qs8_qc8w_igemm_minmax_fp32_ukernel_4x8c8__wasmusdot(
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 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_qc8w_conv_minmax_params 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(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
59 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
60 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
61 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
62 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
63 XNN_FORCE_REALIZATION(vmagic_bias);
64 XNN_FORCE_REALIZATION(vmagic_min);
65 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
66 XNN_FORCE_REALIZATION(voutput_max);
67
68 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
69 XNN_FORCE_REALIZATION(vsign_mask);
70 do {
71 v128_t vacc0x01 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
72 v128_t vacc0x23 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
73 v128_t vacc0x45 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
74 v128_t vacc0x67 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
75 v128_t vacc1x01 = vacc0x01;
76 v128_t vacc2x01 = vacc0x01;
77 v128_t vacc3x01 = vacc0x01;
78 v128_t vacc1x23 = vacc0x23;
79 v128_t vacc2x23 = vacc0x23;
80 v128_t vacc3x23 = vacc0x23;
81 v128_t vacc1x45 = vacc0x45;
82 v128_t vacc2x45 = vacc0x45;
83 v128_t vacc3x45 = vacc0x45;
84 v128_t vacc1x67 = vacc0x67;
85 v128_t vacc2x67 = vacc0x67;
86 v128_t vacc3x67 = vacc0x67;
87
88 size_t p = ks;
89 do {
90 const int8_t* restrict a0 = a[0];
91 if XNN_UNPREDICTABLE(a0 != zero) {
92 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
93 }
94 const int8_t* restrict a1 = a[1];
95 if XNN_UNPREDICTABLE(a1 != zero) {
96 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
97 }
98 const int8_t* restrict a2 = a[2];
99 if XNN_UNPREDICTABLE(a2 != zero) {
100 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
101 }
102 const int8_t* restrict a3 = a[3];
103 if XNN_UNPREDICTABLE(a3 != zero) {
104 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
105 }
106 a += 4;
107
108 size_t k = kc;
109 while (k != 0) {
110 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
111 a0 += 8;
112 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
113 a1 += 8;
114 const v128_t va2x01 = wasm_v128_xor(wasm_v128_load64_splat(a2), vsign_mask);
115 a2 += 8;
116 const v128_t va3x01 = wasm_v128_xor(wasm_v128_load64_splat(a3), vsign_mask);
117 a3 += 8;
118
119 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
120 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
121 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
122 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
123
124 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
125 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
126 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
127 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
128 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
129 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
130 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
131 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
132 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
133 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
134 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
135 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
136 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va3x01, vacc3x01);
137 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va3x01, vacc3x23);
138 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va3x01, vacc3x45);
139 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va3x01, vacc3x67);
140
141 k -= 8 * sizeof(int8_t);
142 }
143
144 p -= 4 * sizeof(void*);
145 } while (p != 0);
146
147 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
148 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
149 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
150 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
151 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
152 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
153 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x01, vacc3x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc3x01, vacc3x23, 1, 3, 5, 7));
154 v128_t vacc3x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x45, vacc3x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc3x45, vacc3x67, 1, 3, 5, 7));
155
156 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
157 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
158 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
159 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
160 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
161 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
162 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
163 vacc3x4567 = wasm_f32x4_convert_i32x4(vacc3x4567);
164
165 const v128_t vscale0123 = wasm_v128_load(w);
166 w = (const float*) w + 4;
167 const v128_t vscale4567 = wasm_v128_load(w);
168 w = (const float*) w + 4;
169 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
170 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
171 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
172 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
173 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
174 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vscale4567);
175 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
176 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vscale4567);
177
178 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
179 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
180 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
181 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
182 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
183 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vmagic_bias);
184 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
185 vacc3x4567 = wasm_f32x4_add(vacc3x4567, vmagic_bias);
186
187 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
188 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
189 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
190 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
191 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
192 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vmagic_min);
193 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
194 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vmagic_min);
195
196 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
197 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
198 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
199 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
200 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
201 vacc2x4567 = wasm_i32x4_sub(vacc2x4567, vmagic_bias_less_output_zero_point);
202 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
203 vacc3x4567 = wasm_i32x4_sub(vacc3x4567, vmagic_bias_less_output_zero_point);
204
205 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
206 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
207 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x4567);
208 v128_t vacc3x01234567 = wasm_i16x8_narrow_i32x4(vacc3x0123, vacc3x4567);
209
210 v128_t vout0x01234567_1x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc1x01234567);
211 v128_t vout2x01234567_3x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc3x01234567);
212
213 vout0x01234567_1x01234567 = wasm_i8x16_min(vout0x01234567_1x01234567, voutput_max);
214 vout2x01234567_3x01234567 = wasm_i8x16_min(vout2x01234567_3x01234567, voutput_max);
215
216 if (nc >= 8) {
217 wasm_v128_store64_lane(c3, vout2x01234567_3x01234567, 1);
218 wasm_v128_store64_lane(c2, vout2x01234567_3x01234567, 0);
219 wasm_v128_store64_lane(c1, vout0x01234567_1x01234567, 1);
220 wasm_v128_store64_lane(c0, vout0x01234567_1x01234567, 0);
221
222 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
223 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
224 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
225 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
226
227 a = (const int8_t**restrict) ((uintptr_t) a - ks);
228
229 nc -= 8;
230 } else {
231 if (nc & 4) {
232 wasm_v128_store32_lane(c3, vout2x01234567_3x01234567, 2);
233 c3 += 4;
234 wasm_v128_store32_lane(c2, vout2x01234567_3x01234567, 0);
235 c2 += 4;
236 wasm_v128_store32_lane(c1, vout0x01234567_1x01234567, 2);
237 c1 += 4;
238 wasm_v128_store32_lane(c0, vout0x01234567_1x01234567, 0);
239 c0 += 4;
240
241 vout0x01234567_1x01234567 = wasm_u64x2_shr(vout0x01234567_1x01234567, 32);
242 vout2x01234567_3x01234567 = wasm_u64x2_shr(vout2x01234567_3x01234567, 32);
243 }
244 if (nc & 2) {
245 wasm_v128_store16_lane(c3, vout2x01234567_3x01234567, 4);
246 c3 += 2;
247 wasm_v128_store16_lane(c2, vout2x01234567_3x01234567, 0);
248 c2 += 2;
249 wasm_v128_store16_lane(c1, vout0x01234567_1x01234567, 4);
250 c1 += 2;
251 wasm_v128_store16_lane(c0, vout0x01234567_1x01234567, 0);
252 c0 += 2;
253
254 vout0x01234567_1x01234567 = wasm_u32x4_shr(vout0x01234567_1x01234567, 16);
255 vout2x01234567_3x01234567 = wasm_u32x4_shr(vout2x01234567_3x01234567, 16);
256 }
257 if (nc & 1) {
258 wasm_v128_store8_lane(c3, vout2x01234567_3x01234567, 8);
259 wasm_v128_store8_lane(c2, vout2x01234567_3x01234567, 0);
260 wasm_v128_store8_lane(c1, vout0x01234567_1x01234567, 8);
261 wasm_v128_store8_lane(c0, vout0x01234567_1x01234567, 0);
262 }
263
264 nc = 0;
265 }
266 } while (nc != 0);
267 }