Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x4c8-minmax-fp32-wasmsimd-dot16x2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c8-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/gemm.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_4x4c8__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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 c3 = c2;
57 }
58
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 do {
70 v128_t vacc0x0 = wasm_v128_load32_zero(w);
71 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
72 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
73 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
74 v128_t vacc1x0 = vacc0x0;
75 v128_t vacc1x1 = vacc0x1;
76 v128_t vacc1x2 = vacc0x2;
77 v128_t vacc1x3 = vacc0x3;
78 v128_t vacc2x0 = vacc0x0;
79 v128_t vacc2x1 = vacc0x1;
80 v128_t vacc2x2 = vacc0x2;
81 v128_t vacc2x3 = vacc0x3;
82 v128_t vacc3x0 = vacc0x0;
83 v128_t vacc3x1 = vacc0x1;
84 v128_t vacc3x2 = vacc0x2;
85 v128_t vacc3x3 = vacc0x3;
86 w = (const void*) ((const int32_t*) w + 4);
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 do {
110 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
111 a0 += 8;
112 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
113 a1 += 8;
114 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
115 a2 += 8;
116 const v128_t vxa3 = wasm_i16x8_load8x8(a3);
117 a3 += 8;
118
119 const v128_t vb01 = wasm_v128_load(w);
120 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
121 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
122
123 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
124 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
125 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
126 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
127 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
128 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
129 vacc3x0 = wasm_i32x4_add(vacc3x0, wasm_i32x4_dot_i16x8(vxa3, vxb0));
130 vacc3x1 = wasm_i32x4_add(vacc3x1, wasm_i32x4_dot_i16x8(vxa3, vxb1));
131 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
132 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
133 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
134
135 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
136 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
137 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
138 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
139 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
140 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
141 vacc3x2 = wasm_i32x4_add(vacc3x2, wasm_i32x4_dot_i16x8(vxa3, vxb2));
142 vacc3x3 = wasm_i32x4_add(vacc3x3, wasm_i32x4_dot_i16x8(vxa3, vxb3));
143
144 w = (const void*) ((const int8_t*) w + 32);
145 k -= 8 * sizeof(int8_t);
146 } while (k != 0);
147 p -= 4 * sizeof(void*);
148 } while (p != 0);
149
150 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));
151 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));
152 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));
153 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));
154 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));
155 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));
156 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));
157 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));
158
159 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
160 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
161 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
162 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
163
164 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
165 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
166 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
167 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
168
169
170 const v128_t vscale0123 = wasm_v128_load(w);
171 w = (const float*) w + 4;
172 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
173 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
174 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
175 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
176
177 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
178 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
179 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
180 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
181
182 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
183 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
184 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
185 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
186
187 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
188 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
189 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
190 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
191
192 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
193 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
194
195 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
196
197 vout = wasm_i8x16_min(vout, voutput_max);
198
199 if (nc >= 4) {
200 wasm_v128_store32_lane(c3, vout, 3);
201 wasm_v128_store32_lane(c2, vout, 2);
202 wasm_v128_store32_lane(c1, vout, 1);
203 wasm_v128_store32_lane(c0, vout, 0);
204
205 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
206 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
207 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
208 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
209
210 a = (const int8_t**restrict) ((uintptr_t) a - ks);
211
212 nc -= 4;
213 } else {
214 if (nc & 2) {
215 wasm_v128_store16_lane(c3, vout, 6);
216 c3 += 2;
217 wasm_v128_store16_lane(c2, vout, 4);
218 c2 += 2;
219 wasm_v128_store16_lane(c1, vout, 2);
220 c1 += 2;
221 wasm_v128_store16_lane(c0, vout, 0);
222 c0 += 2;
223
224 vout = wasm_u32x4_shr(vout, 16);
225 }
226 if (nc & 1) {
227 wasm_v128_store8_lane(c3, vout, 12);
228 wasm_v128_store8_lane(c2, vout, 8);
229 wasm_v128_store8_lane(c1, vout, 4);
230 wasm_v128_store8_lane(c0, vout, 0);
231 }
232
233 nc = 0;
234 }
235 } while (nc != 0);
236 }