Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x4c2-minmax-fp32-wasmsimd-dot16x2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-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/igemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64(
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 <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * 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, 2 * 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
54
55 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
56 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
57 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
58 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
59 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
60 XNN_FORCE_REALIZATION(vmagic_bias);
61 XNN_FORCE_REALIZATION(vmagic_min);
62 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
63 XNN_FORCE_REALIZATION(voutput_max);
64
65
66 do {
67 v128_t vacc0x0123 = wasm_v128_load(w);
68 v128_t vacc1x0123 = vacc0x0123;
69 v128_t vacc2x0123 = vacc0x0123;
70 w = (const void*) ((const int32_t*) w + 4);
71
72 size_t p = ks;
73 do {
74 const int8_t* restrict a0 = a[0];
75 if XNN_UNPREDICTABLE(a0 != zero) {
76 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
77 }
78 const int8_t* restrict a1 = a[1];
79 if XNN_UNPREDICTABLE(a1 != zero) {
80 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
81 }
82 const int8_t* restrict a2 = a[2];
83 if XNN_UNPREDICTABLE(a2 != zero) {
84 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
85 }
86 a += 3;
87
88 size_t k = kc;
89 while (k >= 8 * sizeof(int8_t)) {
90 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
91 a0 += 8;
92 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
93 a1 += 8;
94 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
95 a2 += 8;
96
97 const v128_t vxb0 = wasm_i16x8_load8x8(w);
98
99 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
100 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
101 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
102 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
103 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
104 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
105 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
106
107 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
108 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
109 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
110 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
111 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
112 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
113 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
114
115 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
116 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
117 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
118 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
119 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
120 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
121 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
122
123 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
124 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
125 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
126 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
127 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
128 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
129
130 w = (const void*) ((const int8_t*) w + 32);
131 k -= 8 * sizeof(int8_t);
132 }
133 if (k != 0) {
134 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
135 a0 = (const int8_t*) ((uintptr_t) a0 + k);
136 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
137 a1 = (const int8_t*) ((uintptr_t) a1 + k);
138 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
139 a2 = (const int8_t*) ((uintptr_t) a2 + k);
140
141 const v128_t vxb0 = wasm_i16x8_load8x8(w);
142 w = (const int8_t*) w + 8;
143
144 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
145 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
146 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
147 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
148 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
149 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
150
151 if (k > 2 * sizeof(int8_t)) {
152 const v128_t vxb1 = wasm_i16x8_load8x8(w);
153 w = (const int8_t*) w + 8;
154
155 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
156 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
157 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
158 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
159 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
160 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
161
162 if (k > 4 * sizeof(int8_t)) {
163 const v128_t vxb2 = wasm_i16x8_load8x8(w);
164 w = (const int8_t*) w + 8;
165
166 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
167 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
168 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
169 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
170 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
171 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
172 }
173 }
174 }
175 p -= 3 * sizeof(void*);
176 } while (p != 0);
177
178 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
179 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
180 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
181
182 const v128_t vscale0123 = wasm_v128_load(w);
183 w = (const float*) w + 4;
184 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
185 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
186 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
187
188 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
189 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
190 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
191
192 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
193 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
194 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
195
196 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
197 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
198 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
199
200 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
201 v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);
202
203 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
204
205 vout = wasm_i8x16_min(vout, voutput_max);
206
207 if (nc >= 4) {
208 wasm_v128_store32_lane(c2, vout, 2);
209 wasm_v128_store32_lane(c1, vout, 1);
210 wasm_v128_store32_lane(c0, vout, 0);
211
212 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
213 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
214 c0 = (int8_t*) ((uintptr_t) c0 + 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_store16_lane(c2, vout, 4);
222 c2 += 2;
223 wasm_v128_store16_lane(c1, vout, 2);
224 c1 += 2;
225 wasm_v128_store16_lane(c0, vout, 0);
226 c0 += 2;
227
228 vout = wasm_u32x4_shr(vout, 16);
229 }
230 if (nc & 1) {
231 wasm_v128_store8_lane(c2, vout, 8);
232 wasm_v128_store8_lane(c1, vout, 4);
233 wasm_v128_store8_lane(c0, vout, 0);
234 }
235
236 nc = 0;
237 }
238 } while (nc != 0);
239 }