xref: /aosp_15_r20/external/XNNPACK/src/f32-vbinary/gen/vmulc-wasmsimd-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-vbinary/vopc-wasmsimd.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 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/vbinary.h>
16*4bdc9457SAndroid Build Coastguard Worker 
17*4bdc9457SAndroid Build Coastguard Worker 
xnn_f32_vmulc_ukernel__wasmsimd_x16(size_t n,const float * a,const float * b,float * y,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])18*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_vmulc_ukernel__wasmsimd_x16(
19*4bdc9457SAndroid Build Coastguard Worker     size_t n,
20*4bdc9457SAndroid Build Coastguard Worker     const float* a,
21*4bdc9457SAndroid Build Coastguard Worker     const float* b,
22*4bdc9457SAndroid Build Coastguard Worker     float* y,
23*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_default_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(float) == 0);
27*4bdc9457SAndroid Build Coastguard Worker   assert(a != NULL);
28*4bdc9457SAndroid Build Coastguard Worker   assert(b != NULL);
29*4bdc9457SAndroid Build Coastguard Worker   assert(y != NULL);
30*4bdc9457SAndroid Build Coastguard Worker 
31*4bdc9457SAndroid Build Coastguard Worker   const v128_t vb = wasm_v128_load32_splat(b);
32*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
33*4bdc9457SAndroid Build Coastguard Worker     const v128_t va0123 = wasm_v128_load(a);
34*4bdc9457SAndroid Build Coastguard Worker     const v128_t va4567 = wasm_v128_load(a + 4);
35*4bdc9457SAndroid Build Coastguard Worker     const v128_t va89AB = wasm_v128_load(a + 8);
36*4bdc9457SAndroid Build Coastguard Worker     const v128_t vaCDEF = wasm_v128_load(a + 12);
37*4bdc9457SAndroid Build Coastguard Worker     a += 16;
38*4bdc9457SAndroid Build Coastguard Worker 
39*4bdc9457SAndroid Build Coastguard Worker     v128_t vy0123 = wasm_f32x4_mul(va0123, vb);
40*4bdc9457SAndroid Build Coastguard Worker     v128_t vy4567 = wasm_f32x4_mul(va4567, vb);
41*4bdc9457SAndroid Build Coastguard Worker     v128_t vy89AB = wasm_f32x4_mul(va89AB, vb);
42*4bdc9457SAndroid Build Coastguard Worker     v128_t vyCDEF = wasm_f32x4_mul(vaCDEF, vb);
43*4bdc9457SAndroid Build Coastguard Worker 
44*4bdc9457SAndroid Build Coastguard Worker 
45*4bdc9457SAndroid Build Coastguard Worker 
46*4bdc9457SAndroid Build Coastguard Worker     wasm_v128_store(y, vy0123);
47*4bdc9457SAndroid Build Coastguard Worker     wasm_v128_store(y + 4, vy4567);
48*4bdc9457SAndroid Build Coastguard Worker     wasm_v128_store(y + 8, vy89AB);
49*4bdc9457SAndroid Build Coastguard Worker     wasm_v128_store(y + 12, vyCDEF);
50*4bdc9457SAndroid Build Coastguard Worker     y += 16;
51*4bdc9457SAndroid Build Coastguard Worker   }
52*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
53*4bdc9457SAndroid Build Coastguard Worker     const v128_t va = wasm_v128_load(a);
54*4bdc9457SAndroid Build Coastguard Worker     a += 4;
55*4bdc9457SAndroid Build Coastguard Worker 
56*4bdc9457SAndroid Build Coastguard Worker     v128_t vy = wasm_f32x4_mul(va, vb);
57*4bdc9457SAndroid Build Coastguard Worker 
58*4bdc9457SAndroid Build Coastguard Worker 
59*4bdc9457SAndroid Build Coastguard Worker     wasm_v128_store(y, vy);
60*4bdc9457SAndroid Build Coastguard Worker     y += 4;
61*4bdc9457SAndroid Build Coastguard Worker   }
62*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(n != 0) {
63*4bdc9457SAndroid Build Coastguard Worker     const v128_t va = wasm_v128_load(a);
64*4bdc9457SAndroid Build Coastguard Worker 
65*4bdc9457SAndroid Build Coastguard Worker     v128_t vy = wasm_f32x4_mul(va, vb);
66*4bdc9457SAndroid Build Coastguard Worker 
67*4bdc9457SAndroid Build Coastguard Worker 
68*4bdc9457SAndroid Build Coastguard Worker     if (n & (2 * sizeof(float))) {
69*4bdc9457SAndroid Build Coastguard Worker       *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
70*4bdc9457SAndroid Build Coastguard Worker       vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
71*4bdc9457SAndroid Build Coastguard Worker       y += 2;
72*4bdc9457SAndroid Build Coastguard Worker     }
73*4bdc9457SAndroid Build Coastguard Worker     if (n & (1 * sizeof(float))) {
74*4bdc9457SAndroid Build Coastguard Worker       *y = wasm_f32x4_extract_lane(vy, 0);
75*4bdc9457SAndroid Build Coastguard Worker     }
76*4bdc9457SAndroid Build Coastguard Worker   }
77*4bdc9457SAndroid Build Coastguard Worker }
78