1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/wasmsimd-cvt.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_cvt_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_cvt_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_cvt.scale);
31 const v128_t voutput_zero_point = wasm_v128_load64_splat(params->wasmsimd_cvt.output_zero_point);
32 const v128_t voutput_min = wasm_v128_load64_splat(params->wasmsimd_cvt.output_min);
33 const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd_cvt.output_max);
34 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
35 v128_t vx0123 = wasm_v128_load(x);
36 v128_t vx4567 = wasm_v128_load(x + 4);
37 v128_t vx89AB = wasm_v128_load(x + 8);
38 v128_t vxCDEF = wasm_v128_load(x + 12);
39 v128_t vxGHIJ = wasm_v128_load(x + 16);
40 v128_t vxKLMN = wasm_v128_load(x + 20);
41 v128_t vxOPQR = wasm_v128_load(x + 24);
42 v128_t vxSTUV = wasm_v128_load(x + 28);
43 x += 32;
44
45 vx0123 = wasm_f32x4_mul(vx0123, vscale);
46 vx4567 = wasm_f32x4_mul(vx4567, vscale);
47 vx89AB = wasm_f32x4_mul(vx89AB, vscale);
48 vxCDEF = wasm_f32x4_mul(vxCDEF, vscale);
49 vxGHIJ = wasm_f32x4_mul(vxGHIJ, vscale);
50 vxKLMN = wasm_f32x4_mul(vxKLMN, vscale);
51 vxOPQR = wasm_f32x4_mul(vxOPQR, vscale);
52 vxSTUV = wasm_f32x4_mul(vxSTUV, vscale);
53
54 vx0123 = wasm_f32x4_nearest(vx0123);
55 vx4567 = wasm_f32x4_nearest(vx4567);
56 vx89AB = wasm_f32x4_nearest(vx89AB);
57 vxCDEF = wasm_f32x4_nearest(vxCDEF);
58 vxGHIJ = wasm_f32x4_nearest(vxGHIJ);
59 vxKLMN = wasm_f32x4_nearest(vxKLMN);
60 vxOPQR = wasm_f32x4_nearest(vxOPQR);
61 vxSTUV = wasm_f32x4_nearest(vxSTUV);
62
63 v128_t vacc0123 = wasm_i32x4_trunc_sat_f32x4(vx0123);
64 v128_t vacc4567 = wasm_i32x4_trunc_sat_f32x4(vx4567);
65 v128_t vacc89AB = wasm_i32x4_trunc_sat_f32x4(vx89AB);
66 v128_t vaccCDEF = wasm_i32x4_trunc_sat_f32x4(vxCDEF);
67 v128_t vaccGHIJ = wasm_i32x4_trunc_sat_f32x4(vxGHIJ);
68 v128_t vaccKLMN = wasm_i32x4_trunc_sat_f32x4(vxKLMN);
69 v128_t vaccOPQR = wasm_i32x4_trunc_sat_f32x4(vxOPQR);
70 v128_t vaccSTUV = wasm_i32x4_trunc_sat_f32x4(vxSTUV);
71
72 v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
73 v128_t vacc89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
74 v128_t vaccGHIJKLMN = wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN);
75 v128_t vaccOPQRSTUV = wasm_i16x8_narrow_i32x4(vaccOPQR, vaccSTUV);
76
77 vacc01234567 = wasm_i16x8_add_sat(vacc01234567, voutput_zero_point);
78 vacc89ABCDEF = wasm_i16x8_add_sat(vacc89ABCDEF, voutput_zero_point);
79 vaccGHIJKLMN = wasm_i16x8_add_sat(vaccGHIJKLMN, voutput_zero_point);
80 vaccOPQRSTUV = wasm_i16x8_add_sat(vaccOPQRSTUV, voutput_zero_point);
81
82 v128_t vy0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vacc01234567, vacc89ABCDEF);
83 v128_t vyGHIJKLMNOPQRSTUV = wasm_u8x16_narrow_i16x8(vaccGHIJKLMN, vaccOPQRSTUV);
84
85 vy0123456789ABCDEF = wasm_u8x16_max(vy0123456789ABCDEF, voutput_min);
86 vyGHIJKLMNOPQRSTUV = wasm_u8x16_max(vyGHIJKLMNOPQRSTUV, voutput_min);
87
88 vy0123456789ABCDEF = wasm_u8x16_min(vy0123456789ABCDEF, voutput_max);
89 vyGHIJKLMNOPQRSTUV = wasm_u8x16_min(vyGHIJKLMNOPQRSTUV, voutput_max);
90
91 wasm_v128_store(y, vy0123456789ABCDEF);
92 wasm_v128_store(y + 16, vyGHIJKLMNOPQRSTUV);
93 y += 32;
94 }
95 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
96 v128_t vx_lo = wasm_v128_load(x);
97 v128_t vx_hi = wasm_v128_load(x + 4);
98 x += 8;
99
100 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
101 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
102
103 vx_lo = wasm_f32x4_nearest(vx_lo);
104 vx_hi = wasm_f32x4_nearest(vx_hi);
105
106 v128_t vacc_lo = wasm_i32x4_trunc_sat_f32x4(vx_lo);
107 v128_t vacc_hi = wasm_i32x4_trunc_sat_f32x4(vx_hi);
108
109 v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
110 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
111
112 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
113 vy = wasm_u8x16_max(vy, voutput_min);
114 vy = wasm_u8x16_min(vy, voutput_max);
115
116 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
117 y += 8;
118 }
119 if XNN_UNLIKELY(n != 0) {
120 assert(n >= 1 * sizeof(float));
121 assert(n <= 7 * sizeof(float));
122 v128_t vx_lo = wasm_v128_load(x);
123 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
124 v128_t vx_hi = wasm_v128_load(x_hi);
125
126 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
127 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
128
129 vx_lo = wasm_f32x4_nearest(vx_lo);
130 vx_hi = wasm_f32x4_nearest(vx_hi);
131
132 v128_t vacc_lo = wasm_i32x4_trunc_sat_f32x4(vx_lo);
133 v128_t vacc_hi = wasm_i32x4_trunc_sat_f32x4(vx_hi);
134
135 v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
136 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
137
138 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
139 vy = wasm_u8x16_max(vy, voutput_min);
140 vy = wasm_u8x16_min(vy, voutput_max);
141
142 if (n & (4 * sizeof(float))) {
143 *((float*) y) = wasm_f32x4_extract_lane(vy, 0);
144 y += 4;
145 vy = wasm_u64x2_shr(vy, 32);
146 }
147 uint32_t vy_lo = (uint32_t) wasm_i32x4_extract_lane(vy, 0);
148 if (n & (2 * sizeof(float))) {
149 *((uint16_t*) y) = (uint16_t) vy_lo;
150 y += 2;
151 vy_lo >>= 16;
152 }
153 if (n & (1 * sizeof(float))) {
154 *y = (uint8_t) vy_lo;
155 }
156 }
157 }
158