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_6x8__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_6x8__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 <= 6);
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 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
65 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr != 6) {
67 a5 = a4;
68 c5 = c4;
69 }
70
71 do {
72 v128_t vacc0x0123 = wasm_v128_load(w + 0);
73 v128_t vacc0x4567 = wasm_v128_load(w + 4);
74 v128_t vacc1x0123 = vacc0x0123;
75 v128_t vacc1x4567 = vacc0x4567;
76 v128_t vacc2x0123 = vacc0x0123;
77 v128_t vacc2x4567 = vacc0x4567;
78 v128_t vacc3x0123 = vacc0x0123;
79 v128_t vacc3x4567 = vacc0x4567;
80 v128_t vacc4x0123 = vacc0x0123;
81 v128_t vacc4x4567 = vacc0x4567;
82 v128_t vacc5x0123 = vacc0x0123;
83 v128_t vacc5x4567 = vacc0x4567;
84 w += 8;
85
86 size_t k = kc;
87 do {
88 const v128_t va0 = wasm_v128_load32_splat(a0);
89 a0 += 1;
90 const v128_t va1 = wasm_v128_load32_splat(a1);
91 a1 += 1;
92 const v128_t va2 = wasm_v128_load32_splat(a2);
93 a2 += 1;
94 const v128_t va3 = wasm_v128_load32_splat(a3);
95 a3 += 1;
96 const v128_t va4 = wasm_v128_load32_splat(a4);
97 a4 += 1;
98 const v128_t va5 = wasm_v128_load32_splat(a5);
99 a5 += 1;
100
101 const v128_t vb0123 = wasm_v128_load(w);
102 const v128_t vb4567 = wasm_v128_load(w + 4);
103 w += 8;
104
105 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
106 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
107 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
108 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
109 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
110 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123));
111 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
112 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
113 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
114 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
115 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
116 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567));
117
118 k -= sizeof(float);
119 } while (k != 0);
120
121 const v128_t vzero = wasm_i32x4_const_splat(0);
122 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
123 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
124 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
125 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
126 vacc4x0123 = wasm_i32x4_max(vacc4x0123, vzero);
127 vacc5x0123 = wasm_i32x4_max(vacc5x0123, vzero);
128 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
129 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
130 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
131 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
132 vacc4x4567 = wasm_i32x4_max(vacc4x4567, vzero);
133 vacc5x4567 = wasm_i32x4_max(vacc5x4567, vzero);
134
135 if XNN_LIKELY(nc >= 8) {
136 wasm_v128_store(c5, vacc5x0123);
137 wasm_v128_store(c5 + 4, vacc5x4567);
138 c5 = (float*) ((uintptr_t) c5 + cn_stride);
139 wasm_v128_store(c4, vacc4x0123);
140 wasm_v128_store(c4 + 4, vacc4x4567);
141 c4 = (float*) ((uintptr_t) c4 + cn_stride);
142 wasm_v128_store(c3, vacc3x0123);
143 wasm_v128_store(c3 + 4, vacc3x4567);
144 c3 = (float*) ((uintptr_t) c3 + cn_stride);
145 wasm_v128_store(c2, vacc2x0123);
146 wasm_v128_store(c2 + 4, vacc2x4567);
147 c2 = (float*) ((uintptr_t) c2 + cn_stride);
148 wasm_v128_store(c1, vacc1x0123);
149 wasm_v128_store(c1 + 4, vacc1x4567);
150 c1 = (float*) ((uintptr_t) c1 + cn_stride);
151 wasm_v128_store(c0, vacc0x0123);
152 wasm_v128_store(c0 + 4, vacc0x4567);
153 c0 = (float*) ((uintptr_t) c0 + cn_stride);
154
155 a5 = (const float*) ((uintptr_t) a5 - kc);
156 a4 = (const float*) ((uintptr_t) a4 - kc);
157 a3 = (const float*) ((uintptr_t) a3 - kc);
158 a2 = (const float*) ((uintptr_t) a2 - kc);
159 a1 = (const float*) ((uintptr_t) a1 - kc);
160 a0 = (const float*) ((uintptr_t) a0 - kc);
161
162 nc -= 8;
163 } else {
164 if (nc & 4) {
165 wasm_v128_store(c5, vacc5x0123);
166 wasm_v128_store(c4, vacc4x0123);
167 wasm_v128_store(c3, vacc3x0123);
168 wasm_v128_store(c2, vacc2x0123);
169 wasm_v128_store(c1, vacc1x0123);
170 wasm_v128_store(c0, vacc0x0123);
171
172 vacc5x0123 = vacc5x4567;
173 vacc4x0123 = vacc4x4567;
174 vacc3x0123 = vacc3x4567;
175 vacc2x0123 = vacc2x4567;
176 vacc1x0123 = vacc1x4567;
177 vacc0x0123 = vacc0x4567;
178
179 c5 += 4;
180 c4 += 4;
181 c3 += 4;
182 c2 += 4;
183 c1 += 4;
184 c0 += 4;
185 }
186 if (nc & 2) {
187 *((double*) c5) = wasm_f64x2_extract_lane(vacc5x0123, 0);
188 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
189 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
190 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
191 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
192 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
193
194 vacc5x0123 = wasm_v32x4_shuffle(vacc5x0123, vacc5x0123, 2, 3, 2, 3);
195 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
196 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
197 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
198 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
199 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
200
201 c5 += 2;
202 c4 += 2;
203 c3 += 2;
204 c2 += 2;
205 c1 += 2;
206 c0 += 2;
207 }
208 if (nc & 1) {
209 *c5 = wasm_f32x4_extract_lane(vacc5x0123, 0);
210 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
211 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
212 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
213 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
214 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
215 }
216
217 nc = 0;
218 }
219 } while (nc != 0);
220 }
221