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_gemminc_minmax_ukernel_1x8__wasmsimd_x86_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 float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_minmax_ukernel_1x8__wasmsimd_x86_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 float*restrict acc,
28     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(float) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38   assert(acc != NULL);
39 
40   const float* a0 = a;
41   float* c0 = c;
42 
43   do {
44     v128_t vacc0x0123 = wasm_v128_load(acc + 0);
45     v128_t vacc0x4567 = wasm_v128_load(acc + 4);
46     acc += 8;
47 
48     size_t k = kc;
49     do {
50       const v128_t va0 = wasm_v32x4_load_splat(a0);
51       a0 += 1;
52 
53       const v128_t vb0123 = wasm_v128_load(w);
54       const v128_t vb4567 = wasm_v128_load(w + 4);
55       w += 8;
56 
57       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
58       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
59 
60       k -= sizeof(float);
61     } while (k != 0);
62 
63     const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
64     vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin));
65     vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin));
66 
67     const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
68     vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax));
69     vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax));
70 
71     if XNN_LIKELY(nc >= 8) {
72       wasm_v128_store(c0, vacc0x0123);
73       wasm_v128_store(c0 + 4, vacc0x4567);
74       c0 = (float*) ((uintptr_t) c0 + cn_stride);
75 
76       a0 = (const float*) ((uintptr_t) a0 - kc);
77 
78       nc -= 8;
79     } else {
80       if (nc & 4) {
81         wasm_v128_store(c0, vacc0x0123);
82 
83         vacc0x0123 = vacc0x4567;
84 
85         c0 += 4;
86       }
87       if (nc & 2) {
88         *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
89 
90         vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
91 
92         c0 += 2;
93       }
94       if (nc & 1) {
95         *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
96       }
97 
98       nc = 0;
99     }
100   } while (nc != 0);
101 }
102