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_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2(size_t rows,size_t channels,const int8_t * input,size_t input_stride,const int8_t * zero,int8_t * output,const union xnn_qs8_avgpool_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2(
18 size_t rows,
19 size_t channels,
20 const int8_t* input,
21 size_t input_stride,
22 const int8_t* zero,
23 int8_t* output,
24 const union xnn_qs8_avgpool_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
25 {
26 assert(rows != 0);
27 assert(rows <= 7);
28 assert(channels != 0);
29
30 const int8_t* i0 = input;
31 const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride);
32 if XNN_UNPREDICTABLE(rows < 2) {
33 i1 = zero;
34 }
35 const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride);
36 if XNN_UNPREDICTABLE(rows <= 2) {
37 i2 = zero;
38 }
39 const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride);
40 if XNN_UNPREDICTABLE(rows < 4) {
41 i3 = zero;
42 }
43 const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride);
44 if XNN_UNPREDICTABLE(rows <= 4) {
45 i4 = zero;
46 }
47 const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride);
48 if XNN_UNPREDICTABLE(rows < 6) {
49 i5 = zero;
50 }
51 const int8_t* i6 = (const int8_t*) ((uintptr_t) i5 + input_stride);
52 if XNN_UNPREDICTABLE(rows <= 6) {
53 i6 = zero;
54 }
55
56 const v128_t vbias = wasm_v128_load(params->wasmsimd.bias);
57 const v128_t vmultiplier = wasm_v128_load(params->wasmsimd.multiplier);
58 const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding);
59 const int32_t vshift = params->wasmsimd.shift;
60 const v128_t vzero = wasm_f64x2_splat(0.0);
61 while (channels >= 16) {
62 const v128_t vxi0x01234567 = wasm_i16x8_load_8x8(i0);
63 const v128_t vxi0x89ABCDEF = wasm_i16x8_load_8x8(i0 + 8);
64 i0 += 16;
65 const v128_t vxi1x01234567 = wasm_i16x8_load_8x8(i1);
66 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8);
67 i1 += 16;
68 const v128_t vxi2x01234567 = wasm_i16x8_load_8x8(i2);
69 const v128_t vxi2x89ABCDEF = wasm_i16x8_load_8x8(i2 + 8);
70 i2 += 16;
71 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3);
72 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8);
73 i3 += 16;
74 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4);
75 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8);
76 i4 += 16;
77 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5);
78 const v128_t vxi5x89ABCDEF = wasm_i16x8_load_8x8(i5 + 8);
79 i5 += 16;
80 const v128_t vxi6x01234567 = wasm_i16x8_load_8x8(i6);
81 const v128_t vxi6x89ABCDEF = wasm_i16x8_load_8x8(i6 + 8);
82 i6 += 16;
83
84 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567);
85 v128_t vacc0x89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF);
86 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567);
87 v128_t vacc1x89ABCDEF = wasm_i16x8_add(vxi2x89ABCDEF, vxi3x89ABCDEF);
88
89 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567);
90 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vxi4x89ABCDEF);
91 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567);
92 vacc1x89ABCDEF = wasm_i16x8_add(vacc1x89ABCDEF, vxi5x89ABCDEF);
93 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi6x01234567);
94 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vxi6x89ABCDEF);
95
96 // Add up all accumulators to vacc0x0123456789ABCDEF
97 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vacc1x01234567);
98 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vacc1x89ABCDEF);
99
100 const v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_widen_low_i16x8(vacc0x01234567));
101 const v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_widen_high_i16x8(vacc0x01234567));
102 const v128_t vacc89AB = wasm_i32x4_add(vbias, wasm_i32x4_widen_low_i16x8(vacc0x89ABCDEF));
103 const v128_t vaccCDEF = wasm_i32x4_add(vbias, wasm_i32x4_widen_high_i16x8(vacc0x89ABCDEF));
104
105 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123);
106 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567);
107 const v128_t vabsacc89AB = wasm_i32x4_abs(vacc89AB);
108 const v128_t vabsaccCDEF = wasm_i32x4_abs(vaccCDEF);
109
110 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123);
111 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567);
112 const v128_t vsgnacc89AB = wasm_i32x4_gt(vabsacc89AB, vacc89AB);
113 const v128_t vsgnaccCDEF = wasm_i32x4_gt(vabsaccCDEF, vaccCDEF);
114
115 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5);
116 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7);
117 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5);
118 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7);
119 const v128_t vabsacc89 = wasm_v32x4_shuffle(vabsacc89AB, vzero, 0, 4, 1, 5);
120 const v128_t vabsaccAB = wasm_v32x4_shuffle(vabsacc89AB, vzero, 2, 6, 3, 7);
121 const v128_t vabsaccCD = wasm_v32x4_shuffle(vabsaccCDEF, vzero, 0, 4, 1, 5);
122 const v128_t vabsaccEF = wasm_v32x4_shuffle(vabsaccCDEF, vzero, 2, 6, 3, 7);
123
124 const v128_t vabsprod01 = wasm_i64x2_mul(vabsacc01, vmultiplier);
125 const v128_t vabsprod23 = wasm_i64x2_mul(vabsacc23, vmultiplier);
126 const v128_t vabsprod45 = wasm_i64x2_mul(vabsacc45, vmultiplier);
127 const v128_t vabsprod67 = wasm_i64x2_mul(vabsacc67, vmultiplier);
128 const v128_t vabsprod89 = wasm_i64x2_mul(vabsacc89, vmultiplier);
129 const v128_t vabsprodAB = wasm_i64x2_mul(vabsaccAB, vmultiplier);
130 const v128_t vabsprodCD = wasm_i64x2_mul(vabsaccCD, vmultiplier);
131 const v128_t vabsprodEF = wasm_i64x2_mul(vabsaccEF, vmultiplier);
132
133 const v128_t vabsout01 = wasm_u64x2_shr(wasm_i64x2_add(vabsprod01, vrounding), vshift);
134 const v128_t vabsout23 = wasm_u64x2_shr(wasm_i64x2_add(vabsprod23, vrounding), vshift);
135 const v128_t vabsout45 = wasm_u64x2_shr(wasm_i64x2_add(vabsprod45, vrounding), vshift);
136 const v128_t vabsout67 = wasm_u64x2_shr(wasm_i64x2_add(vabsprod67, vrounding), vshift);
137 const v128_t vabsout89 = wasm_u64x2_shr(wasm_i64x2_add(vabsprod89, vrounding), vshift);
138 const v128_t vabsoutAB = wasm_u64x2_shr(wasm_i64x2_add(vabsprodAB, vrounding), vshift);
139 const v128_t vabsoutCD = wasm_u64x2_shr(wasm_i64x2_add(vabsprodCD, vrounding), vshift);
140 const v128_t vabsoutEF = wasm_u64x2_shr(wasm_i64x2_add(vabsprodEF, vrounding), vshift);
141
142 const v128_t vabsout0123 = wasm_v32x4_shuffle(vabsout01, vabsout23, 0, 2, 4, 6);
143 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6);
144 const v128_t vabsout89AB = wasm_v32x4_shuffle(vabsout89, vabsoutAB, 0, 2, 4, 6);
145 const v128_t vabsoutCDEF = wasm_v32x4_shuffle(vabsoutCD, vabsoutEF, 0, 2, 4, 6);
146
147 const v128_t vout0123 = wasm_i32x4_sub(wasm_v128_xor(vabsout0123, vsgnacc0123), vsgnacc0123);
148 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567);
149 const v128_t vout89AB = wasm_i32x4_sub(wasm_v128_xor(vabsout89AB, vsgnacc89AB), vsgnacc89AB);
150 const v128_t voutCDEF = wasm_i32x4_sub(wasm_v128_xor(vabsoutCDEF, vsgnaccCDEF), vsgnaccCDEF);
151
152 const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
153 const v128_t vout01234567 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vout0123, vout4567), voutput_zero_point);
154 const v128_t vout89ABCDEF = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vout89AB, voutCDEF), voutput_zero_point);
155
156 const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
157 const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
158 const v128_t vout0123456789ABCDEF = wasm_i8x16_min(wasm_i8x16_max(wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF), voutput_min), voutput_max);
159
160 wasm_v128_store(output, vout0123456789ABCDEF);
161 output += 16;
162
163 channels -= 16;
164 }
165 if XNN_UNLIKELY(channels != 0) {
166 do {
167 const v128_t vxi0x01234567 = wasm_i16x8_load_8x8(i0);
168 i0 += 8;
169 const v128_t vxi1x01234567 = wasm_i16x8_load_8x8(i1);
170 i1 += 8;
171 const v128_t vxi2x01234567 = wasm_i16x8_load_8x8(i2);
172 i2 += 8;
173 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3);
174 i3 += 8;
175 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4);
176 i4 += 8;
177 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5);
178 i5 += 8;
179 const v128_t vxi6x01234567 = wasm_i16x8_load_8x8(i6);
180 i6 += 8;
181
182 v128_t vacc0x01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567);
183 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567);
184
185 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567);
186 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567);
187 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi6x01234567);
188
189 // Add up all accumulators to vacc0x01234567
190 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vacc1x01234567);
191
192 const v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_widen_low_i16x8(vacc0x01234567));
193 const v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_widen_high_i16x8(vacc0x01234567));
194
195 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123);
196 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567);
197
198 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123);
199 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567);
200
201 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5);
202 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7);
203 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5);
204 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7);
205
206 const v128_t vabsprod01 = wasm_i64x2_mul(vabsacc01, vmultiplier);
207 const v128_t vabsprod23 = wasm_i64x2_mul(vabsacc23, vmultiplier);
208 const v128_t vabsprod45 = wasm_i64x2_mul(vabsacc45, vmultiplier);
209 const v128_t vabsprod67 = wasm_i64x2_mul(vabsacc67, vmultiplier);
210
211 const v128_t vabsout01 = wasm_u64x2_shr(wasm_i64x2_add(vabsprod01, vrounding), vshift);
212 const v128_t vabsout23 = wasm_u64x2_shr(wasm_i64x2_add(vabsprod23, vrounding), vshift);
213 const v128_t vabsout45 = wasm_u64x2_shr(wasm_i64x2_add(vabsprod45, vrounding), vshift);
214 const v128_t vabsout67 = wasm_u64x2_shr(wasm_i64x2_add(vabsprod67, vrounding), vshift);
215
216 const v128_t vabsout0123 = wasm_v32x4_shuffle(vabsout01, vabsout23, 0, 2, 4, 6);
217 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6);
218
219 const v128_t vout0123 = wasm_i32x4_sub(wasm_v128_xor(vabsout0123, vsgnacc0123), vsgnacc0123);
220 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567);
221
222 const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
223 const v128_t vout01234567 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vout0123, vout4567), voutput_zero_point);
224
225 const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
226 const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
227 v128_t vout0123456701234567 = wasm_i8x16_min(wasm_i8x16_max(wasm_i8x16_narrow_i16x8(vout01234567, vout01234567), voutput_min), voutput_max);
228
229 if XNN_LIKELY(channels >= 8) {
230 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
231 output += 8;
232 channels -= 8;
233 } else {
234 if (channels & 4) {
235 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
236 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
237 output += 4;
238 }
239 if (channels & 2) {
240 *((uint16_t*) output) = (uint16_t) wasm_i16x8_extract_lane(vout0123456701234567, 0);
241 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
242 output += 2;
243 }
244 if (channels & 1) {
245 *output = (int8_t) wasm_i8x16_extract_lane(vout0123456701234567, 0);
246 output += 1;
247 }
248 channels = 0;
249 }
250 } while (channels != 0);
251 }
252 }
253