Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x4c2s4-minmax-fp32-wasmsimd-dot16x2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 4);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (4 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
59
60 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
61 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
62 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
63 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
64 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
65 XNN_FORCE_REALIZATION(vmagic_bias);
66 XNN_FORCE_REALIZATION(vmagic_min);
67 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
68 XNN_FORCE_REALIZATION(voutput_max);
69
70 do {
71 v128_t vacc0x0123 = wasm_v128_load(w);
72 v128_t vacc1x0123 = vacc0x0123;
73 v128_t vacc2x0123 = vacc0x0123;
74 v128_t vacc3x0123 = vacc0x0123;
75 w = (const void*) ((const int32_t*) w + 4);
76
77 size_t p = ks;
78 do {
79 const int8_t* restrict a0 = a[0];
80 if XNN_UNPREDICTABLE(a0 != zero) {
81 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
82 }
83 const int8_t* restrict a1 = a[1];
84 if XNN_UNPREDICTABLE(a1 != zero) {
85 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
86 }
87 const int8_t* restrict a2 = a[2];
88 if XNN_UNPREDICTABLE(a2 != zero) {
89 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
90 }
91 const int8_t* restrict a3 = a[3];
92 if XNN_UNPREDICTABLE(a3 != zero) {
93 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
94 }
95 a += 4;
96
97 size_t k = kc;
98 do {
99 v128_t vxa0 = wasm_i16x8_load8x8(a0);
100 a0 += 8;
101 v128_t vxa1 = wasm_i16x8_load8x8(a1);
102 a1 += 8;
103 v128_t vxa2 = wasm_i16x8_load8x8(a2);
104 a2 += 8;
105 v128_t vxa3 = wasm_i16x8_load8x8(a3);
106 a3 += 8;
107
108 const v128_t vb01 = wasm_v128_load(w);
109 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
110 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
111
112 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
113 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
114 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
115 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
116 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
117 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
118 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb0));
119 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
120
121 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
122 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
123 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
124 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
125 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
126 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
127 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb1));
128 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
129 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
130 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
131 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
132
133 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
134 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
135 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
136 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
137 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
138 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
139 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb2));
140 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
141
142 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
143 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
144 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
145 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb3));
146
147 w = (const int8_t*) w + 32;
148 k -= 8 * sizeof(int8_t);
149 } while (k != 0);
150 p -= 4 * sizeof(void*);
151 } while (p != 0);
152
153 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
154 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
155 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
156 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
157
158 const v128_t vscale0123 = wasm_v128_load(w);
159 w = (const float*) w + 4;
160 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
161 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
162 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
163 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
164
165 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
166 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
167 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
168 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
169
170 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
171 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
172 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
173 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
174
175 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
176 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
177 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
178 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
179
180 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
181 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
182
183 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
184
185 vout = wasm_i8x16_min(vout, voutput_max);
186
187 if (nc >= 4) {
188 wasm_v128_store32_lane(c3, vout, 3);
189 wasm_v128_store32_lane(c2, vout, 2);
190 wasm_v128_store32_lane(c1, vout, 1);
191 wasm_v128_store32_lane(c0, vout, 0);
192
193 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
194 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
195 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
196 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
197
198 a = (const int8_t**restrict) ((uintptr_t) a - ks);
199
200 nc -= 4;
201 } else {
202 if (nc & 2) {
203 wasm_v128_store16_lane(c3, vout, 6);
204 c3 += 2;
205 wasm_v128_store16_lane(c2, vout, 4);
206 c2 += 2;
207 wasm_v128_store16_lane(c1, vout, 2);
208 c1 += 2;
209 wasm_v128_store16_lane(c0, vout, 0);
210 c0 += 2;
211
212 vout = wasm_u32x4_shr(vout, 16);
213 }
214 if (nc & 1) {
215 wasm_v128_store8_lane(c3, vout, 12);
216 wasm_v128_store8_lane(c2, vout, 8);
217 wasm_v128_store8_lane(c1, vout, 4);
218 wasm_v128_store8_lane(c0, vout, 0);
219 }
220
221 nc = 0;
222 }
223 } while (nc != 0);
224 }