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_x24(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_x24(
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 >= 24 * sizeof(float); n -= 24 * 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 x += 24;
43
44 vx0123 = wasm_f32x4_mul(vx0123, vscale);
45 vx4567 = wasm_f32x4_mul(vx4567, vscale);
46 vx89AB = wasm_f32x4_mul(vx89AB, vscale);
47 vxCDEF = wasm_f32x4_mul(vxCDEF, vscale);
48 vxGHIJ = wasm_f32x4_mul(vxGHIJ, vscale);
49 vxKLMN = wasm_f32x4_mul(vxKLMN, vscale);
50
51 vx0123 = wasm_f32x4_add(vx0123, vmagic_bias);
52 vx4567 = wasm_f32x4_add(vx4567, vmagic_bias);
53 vx89AB = wasm_f32x4_add(vx89AB, vmagic_bias);
54 vxCDEF = wasm_f32x4_add(vxCDEF, vmagic_bias);
55 vxGHIJ = wasm_f32x4_add(vxGHIJ, vmagic_bias);
56 vxKLMN = wasm_f32x4_add(vxKLMN, vmagic_bias);
57
58 v128_t vacc0123 = wasm_i32x4_max(vx0123, vmagic_min);
59 v128_t vacc4567 = wasm_i32x4_max(vx4567, vmagic_min);
60 v128_t vacc89AB = wasm_i32x4_max(vx89AB, vmagic_min);
61 v128_t vaccCDEF = wasm_i32x4_max(vxCDEF, vmagic_min);
62 v128_t vaccGHIJ = wasm_i32x4_max(vxGHIJ, vmagic_min);
63 v128_t vaccKLMN = wasm_i32x4_max(vxKLMN, vmagic_min);
64
65 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_zero_point);
66 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_zero_point);
67 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_zero_point);
68 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_zero_point);
69 vaccGHIJ = wasm_i32x4_sub(vaccGHIJ, vmagic_bias_less_zero_point);
70 vaccKLMN = wasm_i32x4_sub(vaccKLMN, vmagic_bias_less_zero_point);
71
72 const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
73 const v128_t vacc89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
74 const v128_t vaccGHIJKLMN = wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN);
75
76 v128_t vy0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vacc01234567, vacc89ABCDEF);
77 v128_t vyGHIJKLMN = wasm_u8x16_narrow_i16x8(vaccGHIJKLMN, vaccGHIJKLMN);
78
79 vy0123456789ABCDEF = wasm_u8x16_min(vy0123456789ABCDEF, voutput_max);
80 vyGHIJKLMN = wasm_u8x16_min(vyGHIJKLMN, voutput_max);
81
82 wasm_v128_store(y, vy0123456789ABCDEF);
83 *((double*) (y + 16)) = wasm_f64x2_extract_lane(vyGHIJKLMN, 0);
84 y += 24;
85 }
86 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
87 v128_t vx_lo = wasm_v128_load(x);
88 v128_t vx_hi = wasm_v128_load(x + 4);
89 x += 8;
90
91 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
92 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
93
94 vx_lo = wasm_f32x4_add(vx_lo, vmagic_bias);
95 vx_hi = wasm_f32x4_add(vx_hi, vmagic_bias);
96
97 v128_t vacc_lo = wasm_i32x4_max(vx_lo, vmagic_min);
98 v128_t vacc_hi = wasm_i32x4_max(vx_hi, vmagic_min);
99
100 vacc_lo = wasm_i32x4_sub(vacc_lo, vmagic_bias_less_zero_point);
101 vacc_hi = wasm_i32x4_sub(vacc_hi, vmagic_bias_less_zero_point);
102
103 const v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
104
105 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
106 vy = wasm_u8x16_min(vy, voutput_max);
107 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
108 y += 8;
109 }
110 if XNN_UNLIKELY(n != 0) {
111 assert(n >= 1 * sizeof(float));
112 assert(n <= 7 * sizeof(float));
113 v128_t vx_lo = wasm_v128_load(x);
114 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
115 v128_t vx_hi = wasm_v128_load(x_hi);
116
117 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
118 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
119
120 vx_lo = wasm_f32x4_add(vx_lo, vmagic_bias);
121 vx_hi = wasm_f32x4_add(vx_hi, vmagic_bias);
122
123 v128_t vacc_lo = wasm_i32x4_max(vx_lo, vmagic_min);
124 v128_t vacc_hi = wasm_i32x4_max(vx_hi, vmagic_min);
125
126 vacc_lo = wasm_i32x4_sub(vacc_lo, vmagic_bias_less_zero_point);
127 vacc_hi = wasm_i32x4_sub(vacc_hi, vmagic_bias_less_zero_point);
128
129 const v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
130
131 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
132 vy = wasm_u8x16_min(vy, voutput_max);
133
134 if (n & (4 * sizeof(float))) {
135 *((float*) y) = wasm_f32x4_extract_lane(vy, 0);
136 y += 4;
137 vy = wasm_u64x2_shr(vy, 32);
138 }
139 uint32_t vy_lo = (uint32_t) wasm_i32x4_extract_lane(vy, 0);
140 if (n & (2 * sizeof(float))) {
141 *((uint16_t*) y) = (uint16_t) vy_lo;
142 y += 2;
143 vy_lo >>= 16;
144 }
145 if (n & (1 * sizeof(float))) {
146 *y = (uint8_t) vy_lo;
147 }
148 }
149 }
150