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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
35 v128_t vx_lo = wasm_v128_load(x);
36 v128_t vx_hi = wasm_v128_load(x + 4);
37 x += 8;
38
39 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
40 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
41
42 vx_lo = wasm_f32x4_nearest(vx_lo);
43 vx_hi = wasm_f32x4_nearest(vx_hi);
44
45 v128_t vacc_lo = wasm_i32x4_trunc_sat_f32x4(vx_lo);
46 v128_t vacc_hi = wasm_i32x4_trunc_sat_f32x4(vx_hi);
47
48 v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
49 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
50
51 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
52 vy = wasm_u8x16_max(vy, voutput_min);
53 vy = wasm_u8x16_min(vy, voutput_max);
54
55 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
56 y += 8;
57 }
58 if XNN_UNLIKELY(n != 0) {
59 assert(n >= 1 * sizeof(float));
60 assert(n <= 7 * sizeof(float));
61 v128_t vx_lo = wasm_v128_load(x);
62 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
63 v128_t vx_hi = wasm_v128_load(x_hi);
64
65 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
66 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
67
68 vx_lo = wasm_f32x4_nearest(vx_lo);
69 vx_hi = wasm_f32x4_nearest(vx_hi);
70
71 v128_t vacc_lo = wasm_i32x4_trunc_sat_f32x4(vx_lo);
72 v128_t vacc_hi = wasm_i32x4_trunc_sat_f32x4(vx_hi);
73
74 v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
75 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
76
77 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
78 vy = wasm_u8x16_max(vy, voutput_min);
79 vy = wasm_u8x16_min(vy, voutput_max);
80
81 if (n & (4 * sizeof(float))) {
82 *((float*) y) = wasm_f32x4_extract_lane(vy, 0);
83 y += 4;
84 vy = wasm_u64x2_shr(vy, 32);
85 }
86 uint32_t vy_lo = (uint32_t) wasm_i32x4_extract_lane(vy, 0);
87 if (n & (2 * sizeof(float))) {
88 *((uint16_t*) y) = (uint16_t) vy_lo;
89 y += 2;
90 vy_lo >>= 16;
91 }
92 if (n & (1 * sizeof(float))) {
93 *y = (uint8_t) vy_lo;
94 }
95 }
96 }
97