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_5x8__wasmsimd_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_5x8__wasmsimd_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 <= 5);
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 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
59 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 a4 = a3;
62 c4 = c3;
63 }
64
65 do {
66 v128_t vacc0x0123 = wasm_v128_load(w + 0);
67 v128_t vacc0x4567 = wasm_v128_load(w + 4);
68 v128_t vacc1x0123 = vacc0x0123;
69 v128_t vacc1x4567 = vacc0x4567;
70 v128_t vacc2x0123 = vacc0x0123;
71 v128_t vacc2x4567 = vacc0x4567;
72 v128_t vacc3x0123 = vacc0x0123;
73 v128_t vacc3x4567 = vacc0x4567;
74 v128_t vacc4x0123 = vacc0x0123;
75 v128_t vacc4x4567 = vacc0x4567;
76 w += 8;
77
78 size_t k = kc;
79 do {
80 const v128_t va0 = wasm_v128_load32_splat(a0);
81 a0 += 1;
82 const v128_t va1 = wasm_v128_load32_splat(a1);
83 a1 += 1;
84 const v128_t va2 = wasm_v128_load32_splat(a2);
85 a2 += 1;
86 const v128_t va3 = wasm_v128_load32_splat(a3);
87 a3 += 1;
88 const v128_t va4 = wasm_v128_load32_splat(a4);
89 a4 += 1;
90
91 const v128_t vb0123 = wasm_v128_load(w);
92 const v128_t vb4567 = wasm_v128_load(w + 4);
93 w += 8;
94
95 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
96 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
97 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
98 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
99 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
100 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
101 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
102 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
103 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
104 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
105
106 k -= sizeof(float);
107 } while (k != 0);
108
109 const v128_t vzero = wasm_i32x4_const_splat(0);
110 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
111 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
112 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
113 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
114 vacc4x0123 = wasm_i32x4_max(vacc4x0123, vzero);
115 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
116 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
117 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
118 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
119 vacc4x4567 = wasm_i32x4_max(vacc4x4567, vzero);
120
121 if XNN_LIKELY(nc >= 8) {
122 wasm_v128_store(c4, vacc4x0123);
123 wasm_v128_store(c4 + 4, vacc4x4567);
124 c4 = (float*) ((uintptr_t) c4 + cn_stride);
125 wasm_v128_store(c3, vacc3x0123);
126 wasm_v128_store(c3 + 4, vacc3x4567);
127 c3 = (float*) ((uintptr_t) c3 + cn_stride);
128 wasm_v128_store(c2, vacc2x0123);
129 wasm_v128_store(c2 + 4, vacc2x4567);
130 c2 = (float*) ((uintptr_t) c2 + cn_stride);
131 wasm_v128_store(c1, vacc1x0123);
132 wasm_v128_store(c1 + 4, vacc1x4567);
133 c1 = (float*) ((uintptr_t) c1 + cn_stride);
134 wasm_v128_store(c0, vacc0x0123);
135 wasm_v128_store(c0 + 4, vacc0x4567);
136 c0 = (float*) ((uintptr_t) c0 + cn_stride);
137
138 a4 = (const float*) ((uintptr_t) a4 - kc);
139 a3 = (const float*) ((uintptr_t) a3 - kc);
140 a2 = (const float*) ((uintptr_t) a2 - kc);
141 a1 = (const float*) ((uintptr_t) a1 - kc);
142 a0 = (const float*) ((uintptr_t) a0 - kc);
143
144 nc -= 8;
145 } else {
146 if (nc & 4) {
147 wasm_v128_store(c4, vacc4x0123);
148 wasm_v128_store(c3, vacc3x0123);
149 wasm_v128_store(c2, vacc2x0123);
150 wasm_v128_store(c1, vacc1x0123);
151 wasm_v128_store(c0, vacc0x0123);
152
153 vacc4x0123 = vacc4x4567;
154 vacc3x0123 = vacc3x4567;
155 vacc2x0123 = vacc2x4567;
156 vacc1x0123 = vacc1x4567;
157 vacc0x0123 = vacc0x4567;
158
159 c4 += 4;
160 c3 += 4;
161 c2 += 4;
162 c1 += 4;
163 c0 += 4;
164 }
165 if (nc & 2) {
166 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
167 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
168 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
169 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
170 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
171
172 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
173 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
174 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
175 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
176 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
177
178 c4 += 2;
179 c3 += 2;
180 c2 += 2;
181 c1 += 2;
182 c0 += 2;
183 }
184 if (nc & 1) {
185 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
186 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
187 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
188 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
189 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
190 }
191
192 nc = 0;
193 }
194 } while (nc != 0);
195 }
196