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_ukernel_6x8__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_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_ukernel_6x8__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_default_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 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123);
106 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123);
107 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123);
108 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123);
109 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123);
110 vacc5x0123 = __builtin_wasm_fma_f32x4(vacc5x0123, va5, vb0123);
111 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567);
112 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, va1, vb4567);
113 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, va2, vb4567);
114 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, va3, vb4567);
115 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567);
116 vacc5x4567 = __builtin_wasm_fma_f32x4(vacc5x4567, va5, vb4567);
117
118 k -= sizeof(float);
119 } while (k != 0);
120
121
122 if XNN_LIKELY(nc >= 8) {
123 wasm_v128_store(c5, vacc5x0123);
124 wasm_v128_store(c5 + 4, vacc5x4567);
125 c5 = (float*) ((uintptr_t) c5 + cn_stride);
126 wasm_v128_store(c4, vacc4x0123);
127 wasm_v128_store(c4 + 4, vacc4x4567);
128 c4 = (float*) ((uintptr_t) c4 + cn_stride);
129 wasm_v128_store(c3, vacc3x0123);
130 wasm_v128_store(c3 + 4, vacc3x4567);
131 c3 = (float*) ((uintptr_t) c3 + cn_stride);
132 wasm_v128_store(c2, vacc2x0123);
133 wasm_v128_store(c2 + 4, vacc2x4567);
134 c2 = (float*) ((uintptr_t) c2 + cn_stride);
135 wasm_v128_store(c1, vacc1x0123);
136 wasm_v128_store(c1 + 4, vacc1x4567);
137 c1 = (float*) ((uintptr_t) c1 + cn_stride);
138 wasm_v128_store(c0, vacc0x0123);
139 wasm_v128_store(c0 + 4, vacc0x4567);
140 c0 = (float*) ((uintptr_t) c0 + cn_stride);
141
142 a5 = (const float*) ((uintptr_t) a5 - kc);
143 a4 = (const float*) ((uintptr_t) a4 - kc);
144 a3 = (const float*) ((uintptr_t) a3 - kc);
145 a2 = (const float*) ((uintptr_t) a2 - kc);
146 a1 = (const float*) ((uintptr_t) a1 - kc);
147 a0 = (const float*) ((uintptr_t) a0 - kc);
148
149 nc -= 8;
150 } else {
151 if (nc & 4) {
152 wasm_v128_store(c5, vacc5x0123);
153 wasm_v128_store(c4, vacc4x0123);
154 wasm_v128_store(c3, vacc3x0123);
155 wasm_v128_store(c2, vacc2x0123);
156 wasm_v128_store(c1, vacc1x0123);
157 wasm_v128_store(c0, vacc0x0123);
158
159 vacc5x0123 = vacc5x4567;
160 vacc4x0123 = vacc4x4567;
161 vacc3x0123 = vacc3x4567;
162 vacc2x0123 = vacc2x4567;
163 vacc1x0123 = vacc1x4567;
164 vacc0x0123 = vacc0x4567;
165
166 c5 += 4;
167 c4 += 4;
168 c3 += 4;
169 c2 += 4;
170 c1 += 4;
171 c0 += 4;
172 }
173 if (nc & 2) {
174 *((double*) c5) = wasm_f64x2_extract_lane(vacc5x0123, 0);
175 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
176 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
177 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
178 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
179 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
180
181 vacc5x0123 = wasm_v32x4_shuffle(vacc5x0123, vacc5x0123, 2, 3, 2, 3);
182 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
183 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
184 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
185 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
186 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
187
188 c5 += 2;
189 c4 += 2;
190 c3 += 2;
191 c2 += 2;
192 c1 += 2;
193 c0 += 2;
194 }
195 if (nc & 1) {
196 *c5 = wasm_f32x4_extract_lane(vacc5x0123, 0);
197 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
198 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
199 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
200 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
201 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
202 }
203
204 nc = 0;
205 }
206 } while (nc != 0);
207 }
208