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_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_cvt_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_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 >= 24 * sizeof(float); n -= 24 * 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 x += 24;
42
43 vx0123 = wasm_f32x4_mul(vx0123, vscale);
44 vx4567 = wasm_f32x4_mul(vx4567, vscale);
45 vx89AB = wasm_f32x4_mul(vx89AB, vscale);
46 vxCDEF = wasm_f32x4_mul(vxCDEF, vscale);
47 vxGHIJ = wasm_f32x4_mul(vxGHIJ, vscale);
48 vxKLMN = wasm_f32x4_mul(vxKLMN, vscale);
49
50 vx0123 = wasm_f32x4_nearest(vx0123);
51 vx4567 = wasm_f32x4_nearest(vx4567);
52 vx89AB = wasm_f32x4_nearest(vx89AB);
53 vxCDEF = wasm_f32x4_nearest(vxCDEF);
54 vxGHIJ = wasm_f32x4_nearest(vxGHIJ);
55 vxKLMN = wasm_f32x4_nearest(vxKLMN);
56
57 v128_t vacc0123 = wasm_i32x4_trunc_sat_f32x4(vx0123);
58 v128_t vacc4567 = wasm_i32x4_trunc_sat_f32x4(vx4567);
59 v128_t vacc89AB = wasm_i32x4_trunc_sat_f32x4(vx89AB);
60 v128_t vaccCDEF = wasm_i32x4_trunc_sat_f32x4(vxCDEF);
61 v128_t vaccGHIJ = wasm_i32x4_trunc_sat_f32x4(vxGHIJ);
62 v128_t vaccKLMN = wasm_i32x4_trunc_sat_f32x4(vxKLMN);
63
64 v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
65 v128_t vacc89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
66 v128_t vaccGHIJKLMN = wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN);
67
68 vacc01234567 = wasm_i16x8_add_sat(vacc01234567, voutput_zero_point);
69 vacc89ABCDEF = wasm_i16x8_add_sat(vacc89ABCDEF, voutput_zero_point);
70 vaccGHIJKLMN = wasm_i16x8_add_sat(vaccGHIJKLMN, voutput_zero_point);
71
72 v128_t vy0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vacc01234567, vacc89ABCDEF);
73 v128_t vyGHIJKLMN = wasm_u8x16_narrow_i16x8(vaccGHIJKLMN, vaccGHIJKLMN);
74
75 vy0123456789ABCDEF = wasm_u8x16_max(vy0123456789ABCDEF, voutput_min);
76 vyGHIJKLMN = wasm_u8x16_max(vyGHIJKLMN, voutput_min);
77
78 vy0123456789ABCDEF = wasm_u8x16_min(vy0123456789ABCDEF, voutput_max);
79 vyGHIJKLMN = wasm_u8x16_min(vyGHIJKLMN, voutput_max);
80
81 wasm_v128_store(y, vy0123456789ABCDEF);
82 *((double*) (y + 16)) = wasm_f64x2_extract_lane(vyGHIJKLMN, 0);
83 y += 24;
84 }
85 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
86 v128_t vx_lo = wasm_v128_load(x);
87 v128_t vx_hi = wasm_v128_load(x + 4);
88 x += 8;
89
90 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
91 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
92
93 vx_lo = wasm_f32x4_nearest(vx_lo);
94 vx_hi = wasm_f32x4_nearest(vx_hi);
95
96 v128_t vacc_lo = wasm_i32x4_trunc_sat_f32x4(vx_lo);
97 v128_t vacc_hi = wasm_i32x4_trunc_sat_f32x4(vx_hi);
98
99 v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
100 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
101
102 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
103 vy = wasm_u8x16_max(vy, voutput_min);
104 vy = wasm_u8x16_min(vy, voutput_max);
105
106 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
107 y += 8;
108 }
109 if XNN_UNLIKELY(n != 0) {
110 assert(n >= 1 * sizeof(float));
111 assert(n <= 7 * sizeof(float));
112 v128_t vx_lo = wasm_v128_load(x);
113 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
114 v128_t vx_hi = wasm_v128_load(x_hi);
115
116 vx_lo = wasm_f32x4_mul(vx_lo, vscale);
117 vx_hi = wasm_f32x4_mul(vx_hi, vscale);
118
119 vx_lo = wasm_f32x4_nearest(vx_lo);
120 vx_hi = wasm_f32x4_nearest(vx_hi);
121
122 v128_t vacc_lo = wasm_i32x4_trunc_sat_f32x4(vx_lo);
123 v128_t vacc_hi = wasm_i32x4_trunc_sat_f32x4(vx_hi);
124
125 v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);
126 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
127
128 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
129 vy = wasm_u8x16_max(vy, voutput_min);
130 vy = wasm_u8x16_min(vy, voutput_max);
131
132 if (n & (4 * sizeof(float))) {
133 *((float*) y) = wasm_f32x4_extract_lane(vy, 0);
134 y += 4;
135 vy = wasm_u64x2_shr(vy, 32);
136 }
137 uint32_t vy_lo = (uint32_t) wasm_i32x4_extract_lane(vy, 0);
138 if (n & (2 * sizeof(float))) {
139 *((uint16_t*) y) = (uint16_t) vy_lo;
140 y += 2;
141 vy_lo >>= 16;
142 }
143 if (n & (1 * sizeof(float))) {
144 *y = (uint8_t) vy_lo;
145 }
146 }
147 }
148