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_x16(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_x16(
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 >= 16 * sizeof(float); n -= 16 * 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 x += 16;
40
41 vx0123 = wasm_f32x4_mul(vx0123, vscale);
42 vx4567 = wasm_f32x4_mul(vx4567, vscale);
43 vx89AB = wasm_f32x4_mul(vx89AB, vscale);
44 vxCDEF = wasm_f32x4_mul(vxCDEF, vscale);
45
46 vx0123 = wasm_f32x4_nearest(vx0123);
47 vx4567 = wasm_f32x4_nearest(vx4567);
48 vx89AB = wasm_f32x4_nearest(vx89AB);
49 vxCDEF = wasm_f32x4_nearest(vxCDEF);
50
51 v128_t vacc0123 = wasm_i32x4_trunc_sat_f32x4(vx0123);
52 v128_t vacc4567 = wasm_i32x4_trunc_sat_f32x4(vx4567);
53 v128_t vacc89AB = wasm_i32x4_trunc_sat_f32x4(vx89AB);
54 v128_t vaccCDEF = wasm_i32x4_trunc_sat_f32x4(vxCDEF);
55
56 v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
57 v128_t vacc89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
58
59 vacc01234567 = wasm_i16x8_add_sat(vacc01234567, voutput_zero_point);
60 vacc89ABCDEF = wasm_i16x8_add_sat(vacc89ABCDEF, voutput_zero_point);
61
62 v128_t vy0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vacc01234567, vacc89ABCDEF);
63
64 vy0123456789ABCDEF = wasm_u8x16_max(vy0123456789ABCDEF, voutput_min);
65
66 vy0123456789ABCDEF = wasm_u8x16_min(vy0123456789ABCDEF, voutput_max);
67
68 wasm_v128_store(y, vy0123456789ABCDEF);
69 y += 16;
70 }
71 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
72 v128_t vx_lo = wasm_v128_load(x);
73 v128_t vx_hi = wasm_v128_load(x + 4);
74 x += 8;
75
76 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
77 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
78
79 vx_lo = wasm_f32x4_nearest(vx_lo);
80 vx_hi = wasm_f32x4_nearest(vx_hi);
81
82 v128_t vacc_lo = wasm_i32x4_trunc_sat_f32x4(vx_lo);
83 v128_t vacc_hi = wasm_i32x4_trunc_sat_f32x4(vx_hi);
84
85 v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
86 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
87
88 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
89 vy = wasm_u8x16_max(vy, voutput_min);
90 vy = wasm_u8x16_min(vy, voutput_max);
91
92 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
93 y += 8;
94 }
95 if XNN_UNLIKELY(n != 0) {
96 assert(n >= 1 * sizeof(float));
97 assert(n <= 7 * sizeof(float));
98 v128_t vx_lo = wasm_v128_load(x);
99 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
100 v128_t vx_hi = wasm_v128_load(x_hi);
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_nearest(vx_lo);
106 vx_hi = wasm_f32x4_nearest(vx_hi);
107
108 v128_t vacc_lo = wasm_i32x4_trunc_sat_f32x4(vx_lo);
109 v128_t vacc_hi = wasm_i32x4_trunc_sat_f32x4(vx_hi);
110
111 v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
112 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
113
114 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
115 vy = wasm_u8x16_max(vy, voutput_min);
116 vy = wasm_u8x16_min(vy, voutput_max);
117
118 if (n & (4 * sizeof(float))) {
119 *((float*) y) = wasm_f32x4_extract_lane(vy, 0);
120 y += 4;
121 vy = wasm_u64x2_shr(vy, 32);
122 }
123 uint32_t vy_lo = (uint32_t) wasm_i32x4_extract_lane(vy, 0);
124 if (n & (2 * sizeof(float))) {
125 *((uint16_t*) y) = (uint16_t) vy_lo;
126 y += 2;
127 vy_lo >>= 16;
128 }
129 if (n & (1 * sizeof(float))) {
130 *y = (uint8_t) vy_lo;
131 }
132 }
133 }
134