1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/wasmsimd-magic.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 <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vcvt.h>
17
18
xnn_f32_qu8_vcvt_ukernel__wasmsimd_magic_x32(size_t n,const float * x,uint8_t * y,const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_qu8_vcvt_ukernel__wasmsimd_magic_x32(
20 size_t n,
21 const float* x,
22 uint8_t* y,
23 const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd_magic.scale);
31 const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_magic.magic_bias);
32 const v128_t vmagic_min = wasm_v128_load64_splat(params->wasmsimd_magic.magic_min);
33 const v128_t vmagic_bias_less_zero_point = wasm_v128_load64_splat(params->wasmsimd_magic.magic_bias_less_zero_point);
34 const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd_magic.output_max);
35 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
36 v128_t vx0123 = wasm_v128_load(x);
37 v128_t vx4567 = wasm_v128_load(x + 4);
38 v128_t vx89AB = wasm_v128_load(x + 8);
39 v128_t vxCDEF = wasm_v128_load(x + 12);
40 v128_t vxGHIJ = wasm_v128_load(x + 16);
41 v128_t vxKLMN = wasm_v128_load(x + 20);
42 v128_t vxOPQR = wasm_v128_load(x + 24);
43 v128_t vxSTUV = wasm_v128_load(x + 28);
44 x += 32;
45
46 vx0123 = wasm_f32x4_mul(vx0123, vscale);
47 vx4567 = wasm_f32x4_mul(vx4567, vscale);
48 vx89AB = wasm_f32x4_mul(vx89AB, vscale);
49 vxCDEF = wasm_f32x4_mul(vxCDEF, vscale);
50 vxGHIJ = wasm_f32x4_mul(vxGHIJ, vscale);
51 vxKLMN = wasm_f32x4_mul(vxKLMN, vscale);
52 vxOPQR = wasm_f32x4_mul(vxOPQR, vscale);
53 vxSTUV = wasm_f32x4_mul(vxSTUV, vscale);
54
55 vx0123 = wasm_f32x4_add(vx0123, vmagic_bias);
56 vx4567 = wasm_f32x4_add(vx4567, vmagic_bias);
57 vx89AB = wasm_f32x4_add(vx89AB, vmagic_bias);
58 vxCDEF = wasm_f32x4_add(vxCDEF, vmagic_bias);
59 vxGHIJ = wasm_f32x4_add(vxGHIJ, vmagic_bias);
60 vxKLMN = wasm_f32x4_add(vxKLMN, vmagic_bias);
61 vxOPQR = wasm_f32x4_add(vxOPQR, vmagic_bias);
62 vxSTUV = wasm_f32x4_add(vxSTUV, vmagic_bias);
63
64 v128_t vacc0123 = wasm_i32x4_max(vx0123, vmagic_min);
65 v128_t vacc4567 = wasm_i32x4_max(vx4567, vmagic_min);
66 v128_t vacc89AB = wasm_i32x4_max(vx89AB, vmagic_min);
67 v128_t vaccCDEF = wasm_i32x4_max(vxCDEF, vmagic_min);
68 v128_t vaccGHIJ = wasm_i32x4_max(vxGHIJ, vmagic_min);
69 v128_t vaccKLMN = wasm_i32x4_max(vxKLMN, vmagic_min);
70 v128_t vaccOPQR = wasm_i32x4_max(vxOPQR, vmagic_min);
71 v128_t vaccSTUV = wasm_i32x4_max(vxSTUV, vmagic_min);
72
73 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_zero_point);
74 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_zero_point);
75 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_zero_point);
76 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_zero_point);
77 vaccGHIJ = wasm_i32x4_sub(vaccGHIJ, vmagic_bias_less_zero_point);
78 vaccKLMN = wasm_i32x4_sub(vaccKLMN, vmagic_bias_less_zero_point);
79 vaccOPQR = wasm_i32x4_sub(vaccOPQR, vmagic_bias_less_zero_point);
80 vaccSTUV = wasm_i32x4_sub(vaccSTUV, vmagic_bias_less_zero_point);
81
82 const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
83 const v128_t vacc89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
84 const v128_t vaccGHIJKLMN = wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN);
85 const v128_t vaccOPQRSTUV = wasm_i16x8_narrow_i32x4(vaccOPQR, vaccSTUV);
86
87 v128_t vy0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vacc01234567, vacc89ABCDEF);
88 v128_t vyGHIJKLMNOPQRSTUV = wasm_u8x16_narrow_i16x8(vaccGHIJKLMN, vaccOPQRSTUV);
89
90 vy0123456789ABCDEF = wasm_u8x16_min(vy0123456789ABCDEF, voutput_max);
91 vyGHIJKLMNOPQRSTUV = wasm_u8x16_min(vyGHIJKLMNOPQRSTUV, voutput_max);
92
93 wasm_v128_store(y, vy0123456789ABCDEF);
94 wasm_v128_store(y + 16, vyGHIJKLMNOPQRSTUV);
95 y += 32;
96 }
97 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
98 v128_t vx_lo = wasm_v128_load(x);
99 v128_t vx_hi = wasm_v128_load(x + 4);
100 x += 8;
101
102 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
103 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
104
105 vx_lo = wasm_f32x4_add(vx_lo, vmagic_bias);
106 vx_hi = wasm_f32x4_add(vx_hi, vmagic_bias);
107
108 v128_t vacc_lo = wasm_i32x4_max(vx_lo, vmagic_min);
109 v128_t vacc_hi = wasm_i32x4_max(vx_hi, vmagic_min);
110
111 vacc_lo = wasm_i32x4_sub(vacc_lo, vmagic_bias_less_zero_point);
112 vacc_hi = wasm_i32x4_sub(vacc_hi, vmagic_bias_less_zero_point);
113
114 const v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
115
116 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
117 vy = wasm_u8x16_min(vy, voutput_max);
118 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
119 y += 8;
120 }
121 if XNN_UNLIKELY(n != 0) {
122 assert(n >= 1 * sizeof(float));
123 assert(n <= 7 * sizeof(float));
124 v128_t vx_lo = wasm_v128_load(x);
125 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
126 v128_t vx_hi = wasm_v128_load(x_hi);
127
128 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
129 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
130
131 vx_lo = wasm_f32x4_add(vx_lo, vmagic_bias);
132 vx_hi = wasm_f32x4_add(vx_hi, vmagic_bias);
133
134 v128_t vacc_lo = wasm_i32x4_max(vx_lo, vmagic_min);
135 v128_t vacc_hi = wasm_i32x4_max(vx_hi, vmagic_min);
136
137 vacc_lo = wasm_i32x4_sub(vacc_lo, vmagic_bias_less_zero_point);
138 vacc_hi = wasm_i32x4_sub(vacc_hi, vmagic_bias_less_zero_point);
139
140 const v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
141
142 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
143 vy = wasm_u8x16_min(vy, voutput_max);
144
145 if (n & (4 * sizeof(float))) {
146 *((float*) y) = wasm_f32x4_extract_lane(vy, 0);
147 y += 4;
148 vy = wasm_u64x2_shr(vy, 32);
149 }
150 uint32_t vy_lo = (uint32_t) wasm_i32x4_extract_lane(vy, 0);
151 if (n & (2 * sizeof(float))) {
152 *((uint16_t*) y) = (uint16_t) vy_lo;
153 y += 2;
154 vy_lo >>= 16;
155 }
156 if (n & (1 * sizeof(float))) {
157 *y = (uint8_t) vy_lo;
158 }
159 }
160 }
161