1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/unipass-wasmsimd.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/gavgpool.h>
15
16
xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__wasmsimd_c24(size_t rows,size_t channels,const uint8_t * input,size_t input_stride,const uint8_t * zero,uint8_t * output,const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__wasmsimd_c24(
18 size_t rows,
19 size_t channels,
20 const uint8_t* input,
21 size_t input_stride,
22 const uint8_t* zero,
23 uint8_t* output,
24 const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(rows != 0);
27 assert(rows <= 7);
28 assert(channels != 0);
29
30 const uint8_t* i0 = input;
31 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
32 if XNN_UNPREDICTABLE(rows < 2) {
33 i1 = zero;
34 }
35 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
36 if XNN_UNPREDICTABLE(rows <= 2) {
37 i2 = zero;
38 }
39 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
40 if XNN_UNPREDICTABLE(rows < 4) {
41 i3 = zero;
42 }
43 const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
44 if XNN_UNPREDICTABLE(rows <= 4) {
45 i4 = zero;
46 }
47 const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
48 if XNN_UNPREDICTABLE(rows < 6) {
49 i5 = zero;
50 }
51 const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
52 if XNN_UNPREDICTABLE(rows <= 6) {
53 i6 = zero;
54 }
55
56 const v128_t vinit_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.init_bias);
57 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
58 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
59 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
60 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
61 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
62 for (; channels >= 24; channels -= 24) {
63 const v128_t vxi0x01234567 = wasm_u16x8_load8x8(i0);
64 const v128_t vxi0x89ABCDEF = wasm_u16x8_load8x8(i0 + 8);
65 const v128_t vxi0xGHIJKLMN = wasm_u16x8_load8x8(i0 + 16);
66 i0 += 24;
67 const v128_t vxi1x01234567 = wasm_u16x8_load8x8(i1);
68 const v128_t vxi1x89ABCDEF = wasm_u16x8_load8x8(i1 + 8);
69 const v128_t vxi1xGHIJKLMN = wasm_u16x8_load8x8(i1 + 16);
70 i1 += 24;
71
72 v128_t vacc01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567);
73 const v128_t vxi2x01234567 = wasm_u16x8_load8x8(i2);
74 v128_t vacc89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF);
75 const v128_t vxi2x89ABCDEF = wasm_u16x8_load8x8(i2 + 8);
76 v128_t vaccGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN);
77 const v128_t vxi2xGHIJKLMN = wasm_u16x8_load8x8(i2 + 16);
78 i2 += 24;
79
80 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi2x01234567);
81 const v128_t vxi3x01234567 = wasm_u16x8_load8x8(i3);
82 vacc89ABCDEF = wasm_i16x8_add(vacc89ABCDEF, vxi2x89ABCDEF);
83 const v128_t vxi3x89ABCDEF = wasm_u16x8_load8x8(i3 + 8);
84 vaccGHIJKLMN = wasm_i16x8_add(vaccGHIJKLMN, vxi2xGHIJKLMN);
85 const v128_t vxi3xGHIJKLMN = wasm_u16x8_load8x8(i3 + 16);
86 i3 += 24;
87 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi3x01234567);
88 const v128_t vxi4x01234567 = wasm_u16x8_load8x8(i4);
89 vacc89ABCDEF = wasm_i16x8_add(vacc89ABCDEF, vxi3x89ABCDEF);
90 const v128_t vxi4x89ABCDEF = wasm_u16x8_load8x8(i4 + 8);
91 vaccGHIJKLMN = wasm_i16x8_add(vaccGHIJKLMN, vxi3xGHIJKLMN);
92 const v128_t vxi4xGHIJKLMN = wasm_u16x8_load8x8(i4 + 16);
93 i4 += 24;
94 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi4x01234567);
95 const v128_t vxi5x01234567 = wasm_u16x8_load8x8(i5);
96 vacc89ABCDEF = wasm_i16x8_add(vacc89ABCDEF, vxi4x89ABCDEF);
97 const v128_t vxi5x89ABCDEF = wasm_u16x8_load8x8(i5 + 8);
98 vaccGHIJKLMN = wasm_i16x8_add(vaccGHIJKLMN, vxi4xGHIJKLMN);
99 const v128_t vxi5xGHIJKLMN = wasm_u16x8_load8x8(i5 + 16);
100 i5 += 24;
101 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi5x01234567);
102 const v128_t vxi6x01234567 = wasm_u16x8_load8x8(i6);
103 vacc89ABCDEF = wasm_i16x8_add(vacc89ABCDEF, vxi5x89ABCDEF);
104 const v128_t vxi6x89ABCDEF = wasm_u16x8_load8x8(i6 + 8);
105 vaccGHIJKLMN = wasm_i16x8_add(vaccGHIJKLMN, vxi5xGHIJKLMN);
106 const v128_t vxi6xGHIJKLMN = wasm_u16x8_load8x8(i6 + 16);
107 i6 += 24;
108
109 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi6x01234567);
110 vacc89ABCDEF = wasm_i16x8_add(vacc89ABCDEF, vxi6x89ABCDEF);
111 vaccGHIJKLMN = wasm_i16x8_add(vaccGHIJKLMN, vxi6xGHIJKLMN);
112
113 v128_t vacc0123 = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_low_u16x8(vacc01234567));
114 v128_t vacc4567 = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_high_u16x8(vacc01234567));
115 v128_t vacc89AB = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_low_u16x8(vacc89ABCDEF));
116 v128_t vaccCDEF = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_high_u16x8(vacc89ABCDEF));
117 v128_t vaccGHIJ = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_low_u16x8(vaccGHIJKLMN));
118 v128_t vaccKLMN = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_high_u16x8(vaccGHIJKLMN));
119
120 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
121 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
122 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
123 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
124 vaccGHIJ = wasm_f32x4_convert_i32x4(vaccGHIJ);
125 vaccKLMN = wasm_f32x4_convert_i32x4(vaccKLMN);
126
127 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
128 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
129 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
130 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
131 vaccGHIJ = wasm_f32x4_mul(vaccGHIJ, vscale);
132 vaccKLMN = wasm_f32x4_mul(vaccKLMN, vscale);
133
134 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
135 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
136 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
137 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
138 vaccGHIJ = wasm_f32x4_add(vaccGHIJ, vmagic_bias);
139 vaccKLMN = wasm_f32x4_add(vaccKLMN, vmagic_bias);
140
141 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
142 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
143 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
144 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
145 vaccGHIJ = wasm_i32x4_max(vaccGHIJ, vmagic_min);
146 vaccKLMN = wasm_i32x4_max(vaccKLMN, vmagic_min);
147
148 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
149 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
150 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
151 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
152 vaccGHIJ = wasm_i32x4_sub(vaccGHIJ, vmagic_bias_less_output_zero_point);
153 vaccKLMN = wasm_i32x4_sub(vaccKLMN, vmagic_bias_less_output_zero_point);
154
155 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
156 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
157 v128_t voutGHIJKLMN = wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN);
158
159 v128_t vout0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
160 v128_t voutGHIJKLMNGHIJKLMN = wasm_u8x16_narrow_i16x8(voutGHIJKLMN, voutGHIJKLMN);
161
162 vout0123456789ABCDEF = wasm_u8x16_min(vout0123456789ABCDEF, voutput_max);
163 voutGHIJKLMNGHIJKLMN = wasm_u8x16_min(voutGHIJKLMNGHIJKLMN, voutput_max);
164
165 wasm_v128_store(output, vout0123456789ABCDEF);
166 *((double*) (output + 16)) = wasm_f64x2_extract_lane(voutGHIJKLMNGHIJKLMN, 0);
167 output += 24;
168 }
169 if XNN_UNLIKELY(channels != 0) {
170 do {
171 const v128_t vxi0x01234567 = wasm_u16x8_load8x8(i0);
172 i0 += 8;
173 const v128_t vxi1x01234567 = wasm_u16x8_load8x8(i1);
174 i1 += 8;
175
176 v128_t vacc01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567);
177 const v128_t vxi2x01234567 = wasm_u16x8_load8x8(i2);
178 i2 += 8;
179
180 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi2x01234567);
181 const v128_t vxi3x01234567 = wasm_u16x8_load8x8(i3);
182 i3 += 8;
183 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi3x01234567);
184 const v128_t vxi4x01234567 = wasm_u16x8_load8x8(i4);
185 i4 += 8;
186 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi4x01234567);
187 const v128_t vxi5x01234567 = wasm_u16x8_load8x8(i5);
188 i5 += 8;
189 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi5x01234567);
190 const v128_t vxi6x01234567 = wasm_u16x8_load8x8(i6);
191 i6 += 8;
192
193 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi6x01234567);
194
195 v128_t vacc0123 = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_low_u16x8(vacc01234567));
196 v128_t vacc4567 = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_high_u16x8(vacc01234567));
197
198 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
199 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
200
201 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
202 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
203
204 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
205 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
206
207 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
208 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
209
210 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
211 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
212
213 const v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
214 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
215 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
216
217 if XNN_LIKELY(channels >= 8) {
218 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
219 output += 8;
220 channels -= 8;
221 } else {
222 if (channels & 4) {
223 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
224 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
225 output += 4;
226 }
227 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
228 if (channels & 2) {
229 *((uint16_t*) output) = (uint16_t) vout0123;
230 vout0123 >>= 16;
231 output += 2;
232 }
233 if (channels & 1) {
234 *output = (uint8_t) vout0123;
235 output += 1;
236 }
237 channels = 0;
238 }
239 } while (channels != 0);
240 }
241 }
242