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_c16(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_c16(
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 >= 16; channels -= 16) {
63 const v128_t vxi0x01234567 = wasm_u16x8_load8x8(i0);
64 const v128_t vxi0x89ABCDEF = wasm_u16x8_load8x8(i0 + 8);
65 i0 += 16;
66 const v128_t vxi1x01234567 = wasm_u16x8_load8x8(i1);
67 const v128_t vxi1x89ABCDEF = wasm_u16x8_load8x8(i1 + 8);
68 i1 += 16;
69
70 v128_t vacc01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567);
71 const v128_t vxi2x01234567 = wasm_u16x8_load8x8(i2);
72 v128_t vacc89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF);
73 const v128_t vxi2x89ABCDEF = wasm_u16x8_load8x8(i2 + 8);
74 i2 += 16;
75
76 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi2x01234567);
77 const v128_t vxi3x01234567 = wasm_u16x8_load8x8(i3);
78 vacc89ABCDEF = wasm_i16x8_add(vacc89ABCDEF, vxi2x89ABCDEF);
79 const v128_t vxi3x89ABCDEF = wasm_u16x8_load8x8(i3 + 8);
80 i3 += 16;
81 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi3x01234567);
82 const v128_t vxi4x01234567 = wasm_u16x8_load8x8(i4);
83 vacc89ABCDEF = wasm_i16x8_add(vacc89ABCDEF, vxi3x89ABCDEF);
84 const v128_t vxi4x89ABCDEF = wasm_u16x8_load8x8(i4 + 8);
85 i4 += 16;
86 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi4x01234567);
87 const v128_t vxi5x01234567 = wasm_u16x8_load8x8(i5);
88 vacc89ABCDEF = wasm_i16x8_add(vacc89ABCDEF, vxi4x89ABCDEF);
89 const v128_t vxi5x89ABCDEF = wasm_u16x8_load8x8(i5 + 8);
90 i5 += 16;
91 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi5x01234567);
92 const v128_t vxi6x01234567 = wasm_u16x8_load8x8(i6);
93 vacc89ABCDEF = wasm_i16x8_add(vacc89ABCDEF, vxi5x89ABCDEF);
94 const v128_t vxi6x89ABCDEF = wasm_u16x8_load8x8(i6 + 8);
95 i6 += 16;
96
97 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi6x01234567);
98 vacc89ABCDEF = wasm_i16x8_add(vacc89ABCDEF, vxi6x89ABCDEF);
99
100 v128_t vacc0123 = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_low_u16x8(vacc01234567));
101 v128_t vacc4567 = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_high_u16x8(vacc01234567));
102 v128_t vacc89AB = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_low_u16x8(vacc89ABCDEF));
103 v128_t vaccCDEF = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_high_u16x8(vacc89ABCDEF));
104
105 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
106 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
107 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
108 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
109
110 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
111 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
112 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
113 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
114
115 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
116 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
117 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
118 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
119
120 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
121 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
122 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
123 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
124
125 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
126 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
127 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
128 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
129
130 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
131 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
132
133 v128_t vout0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
134
135 vout0123456789ABCDEF = wasm_u8x16_min(vout0123456789ABCDEF, voutput_max);
136
137 wasm_v128_store(output, vout0123456789ABCDEF);
138 output += 16;
139 }
140 if XNN_UNLIKELY(channels != 0) {
141 do {
142 const v128_t vxi0x01234567 = wasm_u16x8_load8x8(i0);
143 i0 += 8;
144 const v128_t vxi1x01234567 = wasm_u16x8_load8x8(i1);
145 i1 += 8;
146
147 v128_t vacc01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567);
148 const v128_t vxi2x01234567 = wasm_u16x8_load8x8(i2);
149 i2 += 8;
150
151 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi2x01234567);
152 const v128_t vxi3x01234567 = wasm_u16x8_load8x8(i3);
153 i3 += 8;
154 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi3x01234567);
155 const v128_t vxi4x01234567 = wasm_u16x8_load8x8(i4);
156 i4 += 8;
157 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi4x01234567);
158 const v128_t vxi5x01234567 = wasm_u16x8_load8x8(i5);
159 i5 += 8;
160 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi5x01234567);
161 const v128_t vxi6x01234567 = wasm_u16x8_load8x8(i6);
162 i6 += 8;
163
164 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi6x01234567);
165
166 v128_t vacc0123 = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_low_u16x8(vacc01234567));
167 v128_t vacc4567 = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_high_u16x8(vacc01234567));
168
169 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
170 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
171
172 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
173 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
174
175 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
176 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
177
178 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
179 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
180
181 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
182 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
183
184 const v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
185 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
186 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
187
188 if XNN_LIKELY(channels >= 8) {
189 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
190 output += 8;
191 channels -= 8;
192 } else {
193 if (channels & 4) {
194 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
195 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
196 output += 4;
197 }
198 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
199 if (channels & 2) {
200 *((uint16_t*) output) = (uint16_t) vout0123;
201 vout0123 >>= 16;
202 output += 2;
203 }
204 if (channels & 1) {
205 *output = (uint8_t) vout0123;
206 output += 1;
207 }
208 channels = 0;
209 }
210 } while (channels != 0);
211 }
212 }
213