1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemm_ukernel_3x8__wasmrelaxedsimd_fma_splat(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_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_ukernel_3x8__wasmrelaxedsimd_fma_splat(
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_default_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 3);
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
53 do {
54 v128_t vacc0x0123 = wasm_v128_load(w + 0);
55 v128_t vacc0x4567 = wasm_v128_load(w + 4);
56 v128_t vacc1x0123 = vacc0x0123;
57 v128_t vacc1x4567 = vacc0x4567;
58 v128_t vacc2x0123 = vacc0x0123;
59 v128_t vacc2x4567 = vacc0x4567;
60 w += 8;
61
62 size_t k = kc;
63 while (k >= 4 * sizeof(float)) {
64 const v128_t va0 = wasm_v128_load(a0);
65 a0 += 4;
66 const v128_t va1 = wasm_v128_load(a1);
67 a1 += 4;
68 const v128_t va2 = wasm_v128_load(a2);
69 a2 += 4;
70
71 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
72 const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
73 const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
74
75 const v128_t vb0123c0 = wasm_v128_load(w + 0);
76 const v128_t vb4567c0 = wasm_v128_load(w + 4);
77
78 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
79 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
80 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
81 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
82 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
83 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
84 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
85 const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
86 const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
87
88 const v128_t vb0123c1 = wasm_v128_load(w + 8);
89 const v128_t vb4567c1 = wasm_v128_load(w + 12);
90
91 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
92 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
93 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
94 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
95 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
96 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
97 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
98 const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
99 const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
100
101 const v128_t vb0123c2 = wasm_v128_load(w + 16);
102 const v128_t vb4567c2 = wasm_v128_load(w + 20);
103
104 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
105 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
106 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
107 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
108 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
109 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
110 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
111 const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
112 const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
113
114 const v128_t vb0123c3 = wasm_v128_load(w + 24);
115 const v128_t vb4567c3 = wasm_v128_load(w + 28);
116
117 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
118 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
119 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
120 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
121 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
122 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
123
124 w += 32;
125 k -= 4 * sizeof(float);
126 }
127 if XNN_UNLIKELY(k != 0) {
128 do {
129 const v128_t va0 = wasm_v128_load32_splat(a0);
130 a0 += 1;
131 const v128_t va1 = wasm_v128_load32_splat(a1);
132 a1 += 1;
133 const v128_t va2 = wasm_v128_load32_splat(a2);
134 a2 += 1;
135
136 const v128_t vb0123 = wasm_v128_load(w);
137 const v128_t vb4567 = wasm_v128_load(w + 4);
138 w += 8;
139
140 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123);
141 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123);
142 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123);
143 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567);
144 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, va1, vb4567);
145 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, va2, vb4567);
146
147 k -= sizeof(float);
148 } while (k != 0);
149 }
150
151
152 if XNN_LIKELY(nc >= 8) {
153 wasm_v128_store(c2, vacc2x0123);
154 wasm_v128_store(c2 + 4, vacc2x4567);
155 c2 = (float*) ((uintptr_t) c2 + cn_stride);
156 wasm_v128_store(c1, vacc1x0123);
157 wasm_v128_store(c1 + 4, vacc1x4567);
158 c1 = (float*) ((uintptr_t) c1 + cn_stride);
159 wasm_v128_store(c0, vacc0x0123);
160 wasm_v128_store(c0 + 4, vacc0x4567);
161 c0 = (float*) ((uintptr_t) c0 + cn_stride);
162
163 a2 = (const float*) ((uintptr_t) a2 - kc);
164 a1 = (const float*) ((uintptr_t) a1 - kc);
165 a0 = (const float*) ((uintptr_t) a0 - kc);
166
167 nc -= 8;
168 } else {
169 if (nc & 4) {
170 wasm_v128_store(c2, vacc2x0123);
171 wasm_v128_store(c1, vacc1x0123);
172 wasm_v128_store(c0, vacc0x0123);
173
174 vacc2x0123 = vacc2x4567;
175 vacc1x0123 = vacc1x4567;
176 vacc0x0123 = vacc0x4567;
177
178 c2 += 4;
179 c1 += 4;
180 c0 += 4;
181 }
182 if (nc & 2) {
183 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
184 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
185 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
186
187 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
188 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
189 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
190
191 c2 += 2;
192 c1 += 2;
193 c0 += 2;
194 }
195 if (nc & 1) {
196 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
197 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
198 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
199 }
200
201 nc = 0;
202 }
203 } while (nc != 0);
204 }
205