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_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_magic_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_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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
36 v128_t vx_lo = wasm_v128_load(x);
37 v128_t vx_hi = wasm_v128_load(x + 4);
38 x += 8;
39
40 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
41 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
42
43 vx_lo = wasm_f32x4_add(vx_lo, vmagic_bias);
44 vx_hi = wasm_f32x4_add(vx_hi, vmagic_bias);
45
46 v128_t vacc_lo = wasm_i32x4_max(vx_lo, vmagic_min);
47 v128_t vacc_hi = wasm_i32x4_max(vx_hi, vmagic_min);
48
49 vacc_lo = wasm_i32x4_sub(vacc_lo, vmagic_bias_less_zero_point);
50 vacc_hi = wasm_i32x4_sub(vacc_hi, vmagic_bias_less_zero_point);
51
52 const v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
53
54 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
55 vy = wasm_u8x16_min(vy, voutput_max);
56 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
57 y += 8;
58 }
59 if XNN_UNLIKELY(n != 0) {
60 assert(n >= 1 * sizeof(float));
61 assert(n <= 7 * sizeof(float));
62 v128_t vx_lo = wasm_v128_load(x);
63 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
64 v128_t vx_hi = wasm_v128_load(x_hi);
65
66 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
67 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
68
69 vx_lo = wasm_f32x4_add(vx_lo, vmagic_bias);
70 vx_hi = wasm_f32x4_add(vx_hi, vmagic_bias);
71
72 v128_t vacc_lo = wasm_i32x4_max(vx_lo, vmagic_min);
73 v128_t vacc_hi = wasm_i32x4_max(vx_hi, vmagic_min);
74
75 vacc_lo = wasm_i32x4_sub(vacc_lo, vmagic_bias_less_zero_point);
76 vacc_hi = wasm_i32x4_sub(vacc_hi, vmagic_bias_less_zero_point);
77
78 const v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
79
80 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
81 vy = wasm_u8x16_min(vy, voutput_max);
82
83 if (n & (4 * sizeof(float))) {
84 *((float*) y) = wasm_f32x4_extract_lane(vy, 0);
85 y += 4;
86 vy = wasm_u64x2_shr(vy, 32);
87 }
88 uint32_t vy_lo = (uint32_t) wasm_i32x4_extract_lane(vy, 0);
89 if (n & (2 * sizeof(float))) {
90 *((uint16_t*) y) = (uint16_t) vy_lo;
91 y += 2;
92 vy_lo >>= 16;
93 }
94 if (n & (1 * sizeof(float))) {
95 *y = (uint8_t) vy_lo;
96 }
97 }
98 }
99