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_4x8__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_4x8__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 <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (4 * 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
57 do {
58 v128_t vacc0x0123 = wasm_v128_load(w);
59 v128_t vacc0x4567 = wasm_v128_load(w + 4);
60 v128_t vacc1x0123 = vacc0x0123;
61 v128_t vacc1x4567 = vacc0x4567;
62 v128_t vacc2x0123 = vacc0x0123;
63 v128_t vacc2x4567 = vacc0x4567;
64 v128_t vacc3x0123 = vacc0x0123;
65 v128_t vacc3x4567 = vacc0x4567;
66 w += 8;
67
68 size_t p = ks;
69 do {
70 const float* restrict a0 = a[0];
71 assert(a0 != NULL);
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const float*) ((uintptr_t) a0 + a_offset);
74 }
75 const float* restrict a1 = a[1];
76 assert(a1 != NULL);
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const float*) ((uintptr_t) a1 + a_offset);
79 }
80 const float* restrict a2 = a[2];
81 assert(a2 != NULL);
82 if XNN_UNPREDICTABLE(a2 != zero) {
83 a2 = (const float*) ((uintptr_t) a2 + a_offset);
84 }
85 const float* restrict a3 = a[3];
86 assert(a3 != NULL);
87 if XNN_UNPREDICTABLE(a3 != zero) {
88 a3 = (const float*) ((uintptr_t) a3 + a_offset);
89 }
90 a += 4;
91
92 size_t k = kc;
93 do {
94 const v128_t vb0123 = wasm_v128_load(w);
95 const v128_t vb4567 = wasm_v128_load(w + 4);
96 w += 8;
97
98 const v128_t va0 = wasm_v128_load32_splat(a0);
99 a0 += 1;
100 const v128_t va1 = wasm_v128_load32_splat(a1);
101 a1 += 1;
102 const v128_t va2 = wasm_v128_load32_splat(a2);
103 a2 += 1;
104 const v128_t va3 = wasm_v128_load32_splat(a3);
105 a3 += 1;
106
107 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
108 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
109 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
110 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
111 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
112 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
113 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
114 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
115 k -= sizeof(float);
116 } while (k != 0);
117 p -= 4 * sizeof(void*);
118 } while (p != 0);
119
120
121 if XNN_LIKELY(nc >= 8) {
122 wasm_v128_store(c3, vacc3x0123);
123 wasm_v128_store(c3 + 4, vacc3x4567);
124 c3 = (float*) ((uintptr_t) c3 + cn_stride);
125 wasm_v128_store(c2, vacc2x0123);
126 wasm_v128_store(c2 + 4, vacc2x4567);
127 c2 = (float*) ((uintptr_t) c2 + cn_stride);
128 wasm_v128_store(c1, vacc1x0123);
129 wasm_v128_store(c1 + 4, vacc1x4567);
130 c1 = (float*) ((uintptr_t) c1 + cn_stride);
131 wasm_v128_store(c0, vacc0x0123);
132 wasm_v128_store(c0 + 4, vacc0x4567);
133 c0 = (float*) ((uintptr_t) c0 + cn_stride);
134
135 a = (const float**restrict) ((uintptr_t) a - ks);
136 nc -= 8;
137 } else {
138 if (nc & 4) {
139 wasm_v128_store(c3, vacc3x0123);
140 wasm_v128_store(c2, vacc2x0123);
141 wasm_v128_store(c1, vacc1x0123);
142 wasm_v128_store(c0, vacc0x0123);
143
144 vacc3x0123 = vacc3x4567;
145 vacc2x0123 = vacc2x4567;
146 vacc1x0123 = vacc1x4567;
147 vacc0x0123 = vacc0x4567;
148
149 c3 += 4;
150 c2 += 4;
151 c1 += 4;
152 c0 += 4;
153 }
154 if (nc & 2) {
155 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
156 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
157 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
158 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
159
160 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
161 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
162 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
163 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
164
165 c3 += 2;
166 c2 += 2;
167 c1 += 2;
168 c0 += 2;
169 }
170 if (nc & 1) {
171 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
172 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
173 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
174 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
175 }
176
177 nc = 0;
178 }
179 } while (nc != 0);
180 }
181