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_fp32_ukernel_7x__wasmsimd_c8(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__wasmsimd_c8(
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_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 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 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 >= 8; channels -= 8) {
63 const v128_t vxi0x01234567 = wasm_i16x8_load8x8(i0);
64 i0 += 8;
65 const v128_t vxi1x01234567 = wasm_i16x8_load8x8(i1);
66 i1 += 8;
67
68 v128_t vacc01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567);
69 const v128_t vxi2x01234567 = wasm_i16x8_load8x8(i2);
70 i2 += 8;
71
72 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi2x01234567);
73 const v128_t vxi3x01234567 = wasm_i16x8_load8x8(i3);
74 i3 += 8;
75 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi3x01234567);
76 const v128_t vxi4x01234567 = wasm_i16x8_load8x8(i4);
77 i4 += 8;
78 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi4x01234567);
79 const v128_t vxi5x01234567 = wasm_i16x8_load8x8(i5);
80 i5 += 8;
81 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi5x01234567);
82 const v128_t vxi6x01234567 = wasm_i16x8_load8x8(i6);
83 i6 += 8;
84
85 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi6x01234567);
86
87 v128_t vacc0123 = wasm_i32x4_add(vinit_bias, wasm_i32x4_extend_low_i16x8(vacc01234567));
88 v128_t vacc4567 = wasm_i32x4_add(vinit_bias, wasm_i32x4_extend_high_i16x8(vacc01234567));
89
90 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
91 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
92
93 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
94 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
95
96 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
97 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
98
99 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
100 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
101
102 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
103 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
104
105 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
106
107 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
108
109 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
110
111 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
112 output += 8;
113 }
114 if XNN_UNLIKELY(channels != 0) {
115 {
116 const v128_t vxi0x01234567 = wasm_i16x8_load8x8(i0);
117 i0 += 8;
118 const v128_t vxi1x01234567 = wasm_i16x8_load8x8(i1);
119 i1 += 8;
120
121 v128_t vacc01234567 = wasm_i16x8_add(vxi0x01234567, vxi1x01234567);
122 const v128_t vxi2x01234567 = wasm_i16x8_load8x8(i2);
123 i2 += 8;
124
125 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi2x01234567);
126 const v128_t vxi3x01234567 = wasm_i16x8_load8x8(i3);
127 i3 += 8;
128 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi3x01234567);
129 const v128_t vxi4x01234567 = wasm_i16x8_load8x8(i4);
130 i4 += 8;
131 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi4x01234567);
132 const v128_t vxi5x01234567 = wasm_i16x8_load8x8(i5);
133 i5 += 8;
134 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi5x01234567);
135 const v128_t vxi6x01234567 = wasm_i16x8_load8x8(i6);
136 i6 += 8;
137
138 vacc01234567 = wasm_i16x8_add(vacc01234567, vxi6x01234567);
139
140 v128_t vacc0123 = wasm_i32x4_add(vinit_bias, wasm_i32x4_extend_low_i16x8(vacc01234567));
141 v128_t vacc4567 = wasm_i32x4_add(vinit_bias, wasm_i32x4_extend_high_i16x8(vacc01234567));
142
143 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
144 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
145
146 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
147 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
148
149 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
150 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
151
152 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
153 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
154
155 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
156 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
157
158 const v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
159 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
160 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
161
162 if (channels & 4) {
163 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
164 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
165 output += 4;
166 }
167 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
168 if (channels & 2) {
169 *((uint16_t*) output) = (uint16_t) vout0123;
170 vout0123 >>= 16;
171 output += 2;
172 }
173 if (channels & 1) {
174 *output = (int8_t) vout0123;
175 }
176 }
177 }
178 }
179