1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/qs8-f32-vcvt/wasmsimd.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2021 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker #include <wasm_simd128.h>
13*4bdc9457SAndroid Build Coastguard Worker
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/intrinsics-polyfill.h>
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vcvt.h>
17*4bdc9457SAndroid Build Coastguard Worker
18*4bdc9457SAndroid Build Coastguard Worker
xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32(size_t n,const uint8_t * x,float * y,const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19*4bdc9457SAndroid Build Coastguard Worker void xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32(
20*4bdc9457SAndroid Build Coastguard Worker size_t n,
21*4bdc9457SAndroid Build Coastguard Worker const uint8_t* x,
22*4bdc9457SAndroid Build Coastguard Worker float* y,
23*4bdc9457SAndroid Build Coastguard Worker const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24*4bdc9457SAndroid Build Coastguard Worker {
25*4bdc9457SAndroid Build Coastguard Worker assert(n != 0);
26*4bdc9457SAndroid Build Coastguard Worker assert(n % sizeof(uint8_t) == 0);
27*4bdc9457SAndroid Build Coastguard Worker assert(x != NULL);
28*4bdc9457SAndroid Build Coastguard Worker assert(y != NULL);
29*4bdc9457SAndroid Build Coastguard Worker
30*4bdc9457SAndroid Build Coastguard Worker const v128_t vminus_zero_point = wasm_v128_load64_splat(params->wasmsimd.minus_zero_point);
31*4bdc9457SAndroid Build Coastguard Worker const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd.scale);
32*4bdc9457SAndroid Build Coastguard Worker for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
33*4bdc9457SAndroid Build Coastguard Worker v128_t vx01234567 = wasm_u16x8_load8x8(x);
34*4bdc9457SAndroid Build Coastguard Worker v128_t vx89ABCDEF = wasm_u16x8_load8x8(x + 8);
35*4bdc9457SAndroid Build Coastguard Worker v128_t vxGHIJKLMN = wasm_u16x8_load8x8(x + 16);
36*4bdc9457SAndroid Build Coastguard Worker v128_t vxOPQRSTUV = wasm_u16x8_load8x8(x + 24);
37*4bdc9457SAndroid Build Coastguard Worker x += 32;
38*4bdc9457SAndroid Build Coastguard Worker
39*4bdc9457SAndroid Build Coastguard Worker vx01234567 = wasm_i16x8_add(vx01234567, vminus_zero_point);
40*4bdc9457SAndroid Build Coastguard Worker vx89ABCDEF = wasm_i16x8_add(vx89ABCDEF, vminus_zero_point);
41*4bdc9457SAndroid Build Coastguard Worker vxGHIJKLMN = wasm_i16x8_add(vxGHIJKLMN, vminus_zero_point);
42*4bdc9457SAndroid Build Coastguard Worker vxOPQRSTUV = wasm_i16x8_add(vxOPQRSTUV, vminus_zero_point);
43*4bdc9457SAndroid Build Coastguard Worker
44*4bdc9457SAndroid Build Coastguard Worker v128_t vy0123 = wasm_i32x4_extend_low_i16x8(vx01234567);
45*4bdc9457SAndroid Build Coastguard Worker v128_t vy4567 = wasm_i32x4_extend_high_i16x8(vx01234567);
46*4bdc9457SAndroid Build Coastguard Worker v128_t vy89AB = wasm_i32x4_extend_low_i16x8(vx89ABCDEF);
47*4bdc9457SAndroid Build Coastguard Worker v128_t vyCDEF = wasm_i32x4_extend_high_i16x8(vx89ABCDEF);
48*4bdc9457SAndroid Build Coastguard Worker v128_t vyGHIJ = wasm_i32x4_extend_low_i16x8(vxGHIJKLMN);
49*4bdc9457SAndroid Build Coastguard Worker v128_t vyKLMN = wasm_i32x4_extend_high_i16x8(vxGHIJKLMN);
50*4bdc9457SAndroid Build Coastguard Worker v128_t vyOPQR = wasm_i32x4_extend_low_i16x8(vxOPQRSTUV);
51*4bdc9457SAndroid Build Coastguard Worker v128_t vySTUV = wasm_i32x4_extend_high_i16x8(vxOPQRSTUV);
52*4bdc9457SAndroid Build Coastguard Worker
53*4bdc9457SAndroid Build Coastguard Worker vy0123 = wasm_f32x4_convert_i32x4(vy0123);
54*4bdc9457SAndroid Build Coastguard Worker vy4567 = wasm_f32x4_convert_i32x4(vy4567);
55*4bdc9457SAndroid Build Coastguard Worker vy89AB = wasm_f32x4_convert_i32x4(vy89AB);
56*4bdc9457SAndroid Build Coastguard Worker vyCDEF = wasm_f32x4_convert_i32x4(vyCDEF);
57*4bdc9457SAndroid Build Coastguard Worker vyGHIJ = wasm_f32x4_convert_i32x4(vyGHIJ);
58*4bdc9457SAndroid Build Coastguard Worker vyKLMN = wasm_f32x4_convert_i32x4(vyKLMN);
59*4bdc9457SAndroid Build Coastguard Worker vyOPQR = wasm_f32x4_convert_i32x4(vyOPQR);
60*4bdc9457SAndroid Build Coastguard Worker vySTUV = wasm_f32x4_convert_i32x4(vySTUV);
61*4bdc9457SAndroid Build Coastguard Worker
62*4bdc9457SAndroid Build Coastguard Worker vy0123 = wasm_f32x4_mul(vy0123, vscale);
63*4bdc9457SAndroid Build Coastguard Worker vy4567 = wasm_f32x4_mul(vy4567, vscale);
64*4bdc9457SAndroid Build Coastguard Worker vy89AB = wasm_f32x4_mul(vy89AB, vscale);
65*4bdc9457SAndroid Build Coastguard Worker vyCDEF = wasm_f32x4_mul(vyCDEF, vscale);
66*4bdc9457SAndroid Build Coastguard Worker vyGHIJ = wasm_f32x4_mul(vyGHIJ, vscale);
67*4bdc9457SAndroid Build Coastguard Worker vyKLMN = wasm_f32x4_mul(vyKLMN, vscale);
68*4bdc9457SAndroid Build Coastguard Worker vyOPQR = wasm_f32x4_mul(vyOPQR, vscale);
69*4bdc9457SAndroid Build Coastguard Worker vySTUV = wasm_f32x4_mul(vySTUV, vscale);
70*4bdc9457SAndroid Build Coastguard Worker
71*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y, vy0123);
72*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y + 4, vy4567);
73*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y + 8, vy89AB);
74*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y + 12, vyCDEF);
75*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y + 16, vyGHIJ);
76*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y + 20, vyKLMN);
77*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y + 24, vyOPQR);
78*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y + 28, vySTUV);
79*4bdc9457SAndroid Build Coastguard Worker y += 32;
80*4bdc9457SAndroid Build Coastguard Worker }
81*4bdc9457SAndroid Build Coastguard Worker for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
82*4bdc9457SAndroid Build Coastguard Worker v128_t vx = wasm_u16x8_load8x8(x);
83*4bdc9457SAndroid Build Coastguard Worker vx = wasm_i16x8_add(vx, vminus_zero_point);
84*4bdc9457SAndroid Build Coastguard Worker x += 8;
85*4bdc9457SAndroid Build Coastguard Worker
86*4bdc9457SAndroid Build Coastguard Worker v128_t vy_lo = wasm_i32x4_extend_low_i16x8(vx);
87*4bdc9457SAndroid Build Coastguard Worker v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx);
88*4bdc9457SAndroid Build Coastguard Worker
89*4bdc9457SAndroid Build Coastguard Worker vy_lo = wasm_f32x4_convert_i32x4(vy_lo);
90*4bdc9457SAndroid Build Coastguard Worker vy_hi = wasm_f32x4_convert_i32x4(vy_hi);
91*4bdc9457SAndroid Build Coastguard Worker
92*4bdc9457SAndroid Build Coastguard Worker vy_lo = wasm_f32x4_mul(vy_lo, vscale);
93*4bdc9457SAndroid Build Coastguard Worker vy_hi = wasm_f32x4_mul(vy_hi, vscale);
94*4bdc9457SAndroid Build Coastguard Worker
95*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y, vy_lo);
96*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y + 4, vy_hi);
97*4bdc9457SAndroid Build Coastguard Worker y += 8;
98*4bdc9457SAndroid Build Coastguard Worker }
99*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(n != 0) {
100*4bdc9457SAndroid Build Coastguard Worker assert(n >= 1 * sizeof(uint8_t));
101*4bdc9457SAndroid Build Coastguard Worker assert(n <= 7 * sizeof(uint8_t));
102*4bdc9457SAndroid Build Coastguard Worker
103*4bdc9457SAndroid Build Coastguard Worker v128_t vx = wasm_u16x8_load8x8(x);
104*4bdc9457SAndroid Build Coastguard Worker vx = wasm_i16x8_add(vx, vminus_zero_point);
105*4bdc9457SAndroid Build Coastguard Worker x += 8;
106*4bdc9457SAndroid Build Coastguard Worker
107*4bdc9457SAndroid Build Coastguard Worker v128_t vy = wasm_i32x4_extend_low_i16x8(vx);
108*4bdc9457SAndroid Build Coastguard Worker vy = wasm_f32x4_convert_i32x4(vy);
109*4bdc9457SAndroid Build Coastguard Worker vy = wasm_f32x4_mul(vy, vscale);
110*4bdc9457SAndroid Build Coastguard Worker
111*4bdc9457SAndroid Build Coastguard Worker if (n & (4 * sizeof(uint8_t))) {
112*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(y, vy); y += 4;
113*4bdc9457SAndroid Build Coastguard Worker vy = wasm_i32x4_extend_high_i16x8(vx);
114*4bdc9457SAndroid Build Coastguard Worker vy = wasm_f32x4_convert_i32x4(vy);
115*4bdc9457SAndroid Build Coastguard Worker vy = wasm_f32x4_mul(vy, vscale);
116*4bdc9457SAndroid Build Coastguard Worker }
117*4bdc9457SAndroid Build Coastguard Worker if (n & (2 * sizeof(uint8_t))) {
118*4bdc9457SAndroid Build Coastguard Worker *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
119*4bdc9457SAndroid Build Coastguard Worker vy = wasm_v64x2_shuffle(vy, vy, 1, 1);
120*4bdc9457SAndroid Build Coastguard Worker y += 2;
121*4bdc9457SAndroid Build Coastguard Worker }
122*4bdc9457SAndroid Build Coastguard Worker if (n & (1 * sizeof(uint8_t))) {
123*4bdc9457SAndroid Build Coastguard Worker *y = wasm_f32x4_extract_lane(vy, 0);
124*4bdc9457SAndroid Build Coastguard Worker }
125*4bdc9457SAndroid Build Coastguard Worker }
126*4bdc9457SAndroid Build Coastguard Worker }
127