1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/wasmsimd-splat.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/igemm.h>
15
16
xnn_f32_igemm_ukernel_1x8__wasmrelaxedsimd_fma_splat(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_ukernel_1x8__wasmrelaxedsimd_fma_splat(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (1 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44
45 do {
46 v128_t vacc0x0123 = wasm_v128_load(w);
47 v128_t vacc0x4567 = wasm_v128_load(w + 4);
48 w += 8;
49
50 size_t p = ks;
51 do {
52 const float* restrict a0 = a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const float*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 while (k >= 4 * sizeof(float)) {
61 const v128_t va0 = wasm_v128_load(a0);
62 a0 += 4;
63
64 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
65
66 const v128_t vb0123c0 = wasm_v128_load(w + 0);
67 const v128_t vb4567c0 = wasm_v128_load(w + 4);
68
69 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0c0, vb0123c0);
70 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0c0, vb4567c0);
71 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
72
73 const v128_t vb0123c1 = wasm_v128_load(w + 8);
74 const v128_t vb4567c1 = wasm_v128_load(w + 12);
75
76 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0c1, vb0123c1);
77 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0c1, vb4567c1);
78 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
79
80 const v128_t vb0123c2 = wasm_v128_load(w + 16);
81 const v128_t vb4567c2 = wasm_v128_load(w + 20);
82
83 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0c2, vb0123c2);
84 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0c2, vb4567c2);
85 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
86
87 const v128_t vb0123c3 = wasm_v128_load(w + 24);
88 const v128_t vb4567c3 = wasm_v128_load(w + 28);
89
90 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0c3, vb0123c3);
91 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0c3, vb4567c3);
92
93 w += 32;
94 k -= 4 * sizeof(float);
95 }
96 if XNN_UNLIKELY(k != 0) {
97 do {
98 const v128_t vb0123 = wasm_v128_load(w);
99 const v128_t vb4567 = wasm_v128_load(w + 4);
100 w += 8;
101
102 const v128_t va0 = wasm_v128_load32_splat(a0);
103 a0 += 1;
104
105 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123);
106 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567);
107 k -= sizeof(float);
108 } while (k != 0);
109 }
110 p -= 1 * sizeof(void*);
111 } while (p != 0);
112
113
114 if XNN_LIKELY(nc >= 8) {
115 wasm_v128_store(c0, vacc0x0123);
116 wasm_v128_store(c0 + 4, vacc0x4567);
117 c0 = (float*) ((uintptr_t) c0 + cn_stride);
118
119 a = (const float**restrict) ((uintptr_t) a - ks);
120 nc -= 8;
121 } else {
122 if (nc & 4) {
123 wasm_v128_store(c0, vacc0x0123);
124
125 vacc0x0123 = vacc0x4567;
126
127 c0 += 4;
128 }
129 if (nc & 2) {
130 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
131
132 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
133
134 c0 += 2;
135 }
136 if (nc & 1) {
137 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
138 }
139
140 nc = 0;
141 }
142 } while (nc != 0);
143 }
144