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_3x8__wasmsimd_arm_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_3x8__wasmsimd_arm_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 <= 3);
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 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min);
56 const v128_t vmax = wasm_v32x4_load_splat(¶ms->scalar.max);
57 do {
58 v128_t vacc0x0123 = wasm_v128_load(acc + 0);
59 v128_t vacc0x4567 = wasm_v128_load(acc + 4);
60 v128_t vacc1x0123 = wasm_v128_load(acc + 8);
61 v128_t vacc1x4567 = wasm_v128_load(acc + 12);
62 v128_t vacc2x0123 = wasm_v128_load(acc + 16);
63 v128_t vacc2x4567 = wasm_v128_load(acc + 20);
64 acc += 24;
65
66 size_t k = kc;
67 do {
68 const v128_t va0 = wasm_v32x4_load_splat(a0);
69 a0 += 1;
70 const v128_t va1 = wasm_v32x4_load_splat(a1);
71 a1 += 1;
72 const v128_t va2 = wasm_v32x4_load_splat(a2);
73 a2 += 1;
74
75 const v128_t vb0123 = wasm_v128_load(w);
76 const v128_t vb4567 = wasm_v128_load(w + 4);
77 w += 8;
78
79 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
80 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
81 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
82 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
83 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
84 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
85
86 k -= sizeof(float);
87 } while (k != 0);
88
89 vacc0x0123 = wasm_f32x4_max(vacc0x0123, vmin);
90 vacc1x0123 = wasm_f32x4_max(vacc1x0123, vmin);
91 vacc2x0123 = wasm_f32x4_max(vacc2x0123, vmin);
92 vacc0x4567 = wasm_f32x4_max(vacc0x4567, vmin);
93 vacc1x4567 = wasm_f32x4_max(vacc1x4567, vmin);
94 vacc2x4567 = wasm_f32x4_max(vacc2x4567, vmin);
95
96 vacc0x0123 = wasm_f32x4_min(vacc0x0123, vmax);
97 vacc1x0123 = wasm_f32x4_min(vacc1x0123, vmax);
98 vacc2x0123 = wasm_f32x4_min(vacc2x0123, vmax);
99 vacc0x4567 = wasm_f32x4_min(vacc0x4567, vmax);
100 vacc1x4567 = wasm_f32x4_min(vacc1x4567, vmax);
101 vacc2x4567 = wasm_f32x4_min(vacc2x4567, vmax);
102
103 if XNN_LIKELY(nc >= 8) {
104 wasm_v128_store(c2, vacc2x0123);
105 wasm_v128_store(c2 + 4, vacc2x4567);
106 c2 = (float*) ((uintptr_t) c2 + cn_stride);
107 wasm_v128_store(c1, vacc1x0123);
108 wasm_v128_store(c1 + 4, vacc1x4567);
109 c1 = (float*) ((uintptr_t) c1 + cn_stride);
110 wasm_v128_store(c0, vacc0x0123);
111 wasm_v128_store(c0 + 4, vacc0x4567);
112 c0 = (float*) ((uintptr_t) c0 + cn_stride);
113
114 a2 = (const float*) ((uintptr_t) a2 - kc);
115 a1 = (const float*) ((uintptr_t) a1 - kc);
116 a0 = (const float*) ((uintptr_t) a0 - kc);
117
118 nc -= 8;
119 } else {
120 if (nc & 4) {
121 wasm_v128_store(c2, vacc2x0123);
122 wasm_v128_store(c1, vacc1x0123);
123 wasm_v128_store(c0, vacc0x0123);
124
125 vacc2x0123 = vacc2x4567;
126 vacc1x0123 = vacc1x4567;
127 vacc0x0123 = vacc0x4567;
128
129 c2 += 4;
130 c1 += 4;
131 c0 += 4;
132 }
133 if (nc & 2) {
134 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
135 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
136 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
137
138 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
139 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
140 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
141
142 c2 += 2;
143 c1 += 2;
144 c0 += 2;
145 }
146 if (nc & 1) {
147 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
148 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
149 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
150 }
151
152 nc = 0;
153 }
154 } while (nc != 0);
155 }
156