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_relu_ukernel_4x8__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_relu_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_relu_ukernel_4x8__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_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 4);
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 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
41 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr < 2) {
43 a1 = a0;
44 c1 = c0;
45 }
46 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
47 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 a2 = a1;
50 c2 = c1;
51 }
52 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 a3 = a2;
56 c3 = c2;
57 }
58
59 do {
60 v128_t vacc0x0123 = wasm_v128_load(w + 0);
61 v128_t vacc0x4567 = wasm_v128_load(w + 4);
62 v128_t vacc1x0123 = vacc0x0123;
63 v128_t vacc1x4567 = vacc0x4567;
64 v128_t vacc2x0123 = vacc0x0123;
65 v128_t vacc2x4567 = vacc0x4567;
66 v128_t vacc3x0123 = vacc0x0123;
67 v128_t vacc3x4567 = vacc0x4567;
68 w += 8;
69
70 size_t k = kc;
71 do {
72 const v128_t va0 = wasm_v128_load32_splat(a0);
73 a0 += 1;
74 const v128_t va1 = wasm_v128_load32_splat(a1);
75 a1 += 1;
76 const v128_t va2 = wasm_v128_load32_splat(a2);
77 a2 += 1;
78 const v128_t va3 = wasm_v128_load32_splat(a3);
79 a3 += 1;
80
81 const v128_t vb0123 = wasm_v128_load(w);
82 const v128_t vb4567 = wasm_v128_load(w + 4);
83 w += 8;
84
85 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123);
86 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123);
87 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123);
88 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123);
89 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567);
90 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, va1, vb4567);
91 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, va2, vb4567);
92 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, va3, vb4567);
93
94 k -= sizeof(float);
95 } while (k != 0);
96
97 const v128_t vzero = wasm_i32x4_const_splat(0);
98 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
99 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
100 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
101 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
102 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
103 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
104 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
105 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
106
107 if XNN_LIKELY(nc >= 8) {
108 wasm_v128_store(c3, vacc3x0123);
109 wasm_v128_store(c3 + 4, vacc3x4567);
110 c3 = (float*) ((uintptr_t) c3 + cn_stride);
111 wasm_v128_store(c2, vacc2x0123);
112 wasm_v128_store(c2 + 4, vacc2x4567);
113 c2 = (float*) ((uintptr_t) c2 + cn_stride);
114 wasm_v128_store(c1, vacc1x0123);
115 wasm_v128_store(c1 + 4, vacc1x4567);
116 c1 = (float*) ((uintptr_t) c1 + cn_stride);
117 wasm_v128_store(c0, vacc0x0123);
118 wasm_v128_store(c0 + 4, vacc0x4567);
119 c0 = (float*) ((uintptr_t) c0 + cn_stride);
120
121 a3 = (const float*) ((uintptr_t) a3 - kc);
122 a2 = (const float*) ((uintptr_t) a2 - kc);
123 a1 = (const float*) ((uintptr_t) a1 - kc);
124 a0 = (const float*) ((uintptr_t) a0 - kc);
125
126 nc -= 8;
127 } else {
128 if (nc & 4) {
129 wasm_v128_store(c3, vacc3x0123);
130 wasm_v128_store(c2, vacc2x0123);
131 wasm_v128_store(c1, vacc1x0123);
132 wasm_v128_store(c0, vacc0x0123);
133
134 vacc3x0123 = vacc3x4567;
135 vacc2x0123 = vacc2x4567;
136 vacc1x0123 = vacc1x4567;
137 vacc0x0123 = vacc0x4567;
138
139 c3 += 4;
140 c2 += 4;
141 c1 += 4;
142 c0 += 4;
143 }
144 if (nc & 2) {
145 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
146 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
147 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
148 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
149
150 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
151 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
152 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
153 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
154
155 c3 += 2;
156 c2 += 2;
157 c1 += 2;
158 c0 += 2;
159 }
160 if (nc & 1) {
161 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
162 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
163 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
164 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
165 }
166
167 nc = 0;
168 }
169 } while (nc != 0);
170 }
171