1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/wasmsimd-loadsplat.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <wasm_simd128.h>
13
14 #include <xnnpack/gemm.h>
15
16
xnn_f32_gemm_ukernel_1x8__wasmrelaxedsimd_fma_loadsplat(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_ukernel_1x8__wasmrelaxedsimd_fma_loadsplat(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float*restrict a,
22 size_t a_stride,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 const float* a0 = a;
39 float* c0 = c;
40
41 do {
42 v128_t vacc0x0123 = wasm_v128_load(w + 0);
43 v128_t vacc0x4567 = wasm_v128_load(w + 4);
44 w += 8;
45
46 size_t k = kc;
47 do {
48 const v128_t va0 = wasm_v128_load32_splat(a0);
49 a0 += 1;
50
51 const v128_t vb0123 = wasm_v128_load(w);
52 const v128_t vb4567 = wasm_v128_load(w + 4);
53 w += 8;
54
55 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123);
56 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567);
57
58 k -= sizeof(float);
59 } while (k != 0);
60
61
62 if XNN_LIKELY(nc >= 8) {
63 wasm_v128_store(c0, vacc0x0123);
64 wasm_v128_store(c0 + 4, vacc0x4567);
65 c0 = (float*) ((uintptr_t) c0 + cn_stride);
66
67 a0 = (const float*) ((uintptr_t) a0 - kc);
68
69 nc -= 8;
70 } else {
71 if (nc & 4) {
72 wasm_v128_store(c0, vacc0x0123);
73
74 vacc0x0123 = vacc0x4567;
75
76 c0 += 4;
77 }
78 if (nc & 2) {
79 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
80
81 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
82
83 c0 += 2;
84 }
85 if (nc & 1) {
86 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
87 }
88
89 nc = 0;
90 }
91 } while (nc != 0);
92 }
93