1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
15
16
xnn_f32_igemm_ukernel_5x8__wasmsimd_loadsplat(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_ukernel_5x8__wasmsimd_loadsplat(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 5);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (5 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr < 4) {
54 c3 = c2;
55 }
56 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 4) {
58 c4 = c3;
59 }
60
61 do {
62 v128_t vacc0x0123 = wasm_v128_load(w);
63 v128_t vacc0x4567 = wasm_v128_load(w + 4);
64 v128_t vacc1x0123 = vacc0x0123;
65 v128_t vacc1x4567 = vacc0x4567;
66 v128_t vacc2x0123 = vacc0x0123;
67 v128_t vacc2x4567 = vacc0x4567;
68 v128_t vacc3x0123 = vacc0x0123;
69 v128_t vacc3x4567 = vacc0x4567;
70 v128_t vacc4x0123 = vacc0x0123;
71 v128_t vacc4x4567 = vacc0x4567;
72 w += 8;
73
74 size_t p = ks;
75 do {
76 const float* restrict a0 = a[0];
77 assert(a0 != NULL);
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const float*) ((uintptr_t) a0 + a_offset);
80 }
81 const float* restrict a1 = a[1];
82 assert(a1 != NULL);
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const float*) ((uintptr_t) a1 + a_offset);
85 }
86 const float* restrict a2 = a[2];
87 assert(a2 != NULL);
88 if XNN_UNPREDICTABLE(a2 != zero) {
89 a2 = (const float*) ((uintptr_t) a2 + a_offset);
90 }
91 const float* restrict a3 = a[3];
92 assert(a3 != NULL);
93 if XNN_UNPREDICTABLE(a3 != zero) {
94 a3 = (const float*) ((uintptr_t) a3 + a_offset);
95 }
96 const float* restrict a4 = a[4];
97 assert(a4 != NULL);
98 if XNN_UNPREDICTABLE(a4 != zero) {
99 a4 = (const float*) ((uintptr_t) a4 + a_offset);
100 }
101 a += 5;
102
103 size_t k = kc;
104 do {
105 const v128_t vb0123 = wasm_v128_load(w);
106 const v128_t vb4567 = wasm_v128_load(w + 4);
107 w += 8;
108
109 const v128_t va0 = wasm_v128_load32_splat(a0);
110 a0 += 1;
111 const v128_t va1 = wasm_v128_load32_splat(a1);
112 a1 += 1;
113 const v128_t va2 = wasm_v128_load32_splat(a2);
114 a2 += 1;
115 const v128_t va3 = wasm_v128_load32_splat(a3);
116 a3 += 1;
117 const v128_t va4 = wasm_v128_load32_splat(a4);
118 a4 += 1;
119
120 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
121 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
122 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
123 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
124 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
125 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
126 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
127 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
128 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
129 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
130 k -= sizeof(float);
131 } while (k != 0);
132 p -= 5 * sizeof(void*);
133 } while (p != 0);
134
135
136 if XNN_LIKELY(nc >= 8) {
137 wasm_v128_store(c4, vacc4x0123);
138 wasm_v128_store(c4 + 4, vacc4x4567);
139 c4 = (float*) ((uintptr_t) c4 + cn_stride);
140 wasm_v128_store(c3, vacc3x0123);
141 wasm_v128_store(c3 + 4, vacc3x4567);
142 c3 = (float*) ((uintptr_t) c3 + cn_stride);
143 wasm_v128_store(c2, vacc2x0123);
144 wasm_v128_store(c2 + 4, vacc2x4567);
145 c2 = (float*) ((uintptr_t) c2 + cn_stride);
146 wasm_v128_store(c1, vacc1x0123);
147 wasm_v128_store(c1 + 4, vacc1x4567);
148 c1 = (float*) ((uintptr_t) c1 + cn_stride);
149 wasm_v128_store(c0, vacc0x0123);
150 wasm_v128_store(c0 + 4, vacc0x4567);
151 c0 = (float*) ((uintptr_t) c0 + cn_stride);
152
153 a = (const float**restrict) ((uintptr_t) a - ks);
154 nc -= 8;
155 } else {
156 if (nc & 4) {
157 wasm_v128_store(c4, vacc4x0123);
158 wasm_v128_store(c3, vacc3x0123);
159 wasm_v128_store(c2, vacc2x0123);
160 wasm_v128_store(c1, vacc1x0123);
161 wasm_v128_store(c0, vacc0x0123);
162
163 vacc4x0123 = vacc4x4567;
164 vacc3x0123 = vacc3x4567;
165 vacc2x0123 = vacc2x4567;
166 vacc1x0123 = vacc1x4567;
167 vacc0x0123 = vacc0x4567;
168
169 c4 += 4;
170 c3 += 4;
171 c2 += 4;
172 c1 += 4;
173 c0 += 4;
174 }
175 if (nc & 2) {
176 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
177 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
178 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
179 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
180 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
181
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 c4 += 2;
189 c3 += 2;
190 c2 += 2;
191 c1 += 2;
192 c0 += 2;
193 }
194 if (nc & 1) {
195 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
196 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
197 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
198 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
199 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
200 }
201
202 nc = 0;
203 }
204 } while (nc != 0);
205 }
206