• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_5x8__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_5x8__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 <= 5);
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 
65   do {
66     v128_t vacc0x0123 = wasm_v128_load(w + 0);
67     v128_t vacc0x4567 = wasm_v128_load(w + 4);
68     v128_t vacc1x0123 = vacc0x0123;
69     v128_t vacc1x4567 = vacc0x4567;
70     v128_t vacc2x0123 = vacc0x0123;
71     v128_t vacc2x4567 = vacc0x4567;
72     v128_t vacc3x0123 = vacc0x0123;
73     v128_t vacc3x4567 = vacc0x4567;
74     v128_t vacc4x0123 = vacc0x0123;
75     v128_t vacc4x4567 = vacc0x4567;
76     w += 8;
77 
78     size_t k = kc;
79     do {
80       const v128_t va0 = wasm_v128_load32_splat(a0);
81       a0 += 1;
82       const v128_t va1 = wasm_v128_load32_splat(a1);
83       a1 += 1;
84       const v128_t va2 = wasm_v128_load32_splat(a2);
85       a2 += 1;
86       const v128_t va3 = wasm_v128_load32_splat(a3);
87       a3 += 1;
88       const v128_t va4 = wasm_v128_load32_splat(a4);
89       a4 += 1;
90 
91       const v128_t vb0123 = wasm_v128_load(w);
92       const v128_t vb4567 = wasm_v128_load(w + 4);
93       w += 8;
94 
95       vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123);
96       vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123);
97       vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123);
98       vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123);
99       vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123);
100       vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567);
101       vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, va1, vb4567);
102       vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, va2, vb4567);
103       vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, va3, vb4567);
104       vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567);
105 
106       k -= sizeof(float);
107     } while (k != 0);
108 
109 
110     if XNN_LIKELY(nc >= 8) {
111       wasm_v128_store(c4, vacc4x0123);
112       wasm_v128_store(c4 + 4, vacc4x4567);
113       c4 = (float*) ((uintptr_t) c4 + cn_stride);
114       wasm_v128_store(c3, vacc3x0123);
115       wasm_v128_store(c3 + 4, vacc3x4567);
116       c3 = (float*) ((uintptr_t) c3 + cn_stride);
117       wasm_v128_store(c2, vacc2x0123);
118       wasm_v128_store(c2 + 4, vacc2x4567);
119       c2 = (float*) ((uintptr_t) c2 + cn_stride);
120       wasm_v128_store(c1, vacc1x0123);
121       wasm_v128_store(c1 + 4, vacc1x4567);
122       c1 = (float*) ((uintptr_t) c1 + cn_stride);
123       wasm_v128_store(c0, vacc0x0123);
124       wasm_v128_store(c0 + 4, vacc0x4567);
125       c0 = (float*) ((uintptr_t) c0 + cn_stride);
126 
127       a4 = (const float*) ((uintptr_t) a4 - kc);
128       a3 = (const float*) ((uintptr_t) a3 - kc);
129       a2 = (const float*) ((uintptr_t) a2 - kc);
130       a1 = (const float*) ((uintptr_t) a1 - kc);
131       a0 = (const float*) ((uintptr_t) a0 - kc);
132 
133       nc -= 8;
134     } else {
135       if (nc & 4) {
136         wasm_v128_store(c4, vacc4x0123);
137         wasm_v128_store(c3, vacc3x0123);
138         wasm_v128_store(c2, vacc2x0123);
139         wasm_v128_store(c1, vacc1x0123);
140         wasm_v128_store(c0, vacc0x0123);
141 
142         vacc4x0123 = vacc4x4567;
143         vacc3x0123 = vacc3x4567;
144         vacc2x0123 = vacc2x4567;
145         vacc1x0123 = vacc1x4567;
146         vacc0x0123 = vacc0x4567;
147 
148         c4 += 4;
149         c3 += 4;
150         c2 += 4;
151         c1 += 4;
152         c0 += 4;
153       }
154       if (nc & 2) {
155         *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
156         *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
157         *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
158         *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
159         *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
160 
161         vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
162         vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
163         vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
164         vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
165         vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
166 
167         c4 += 2;
168         c3 += 2;
169         c2 += 2;
170         c1 += 2;
171         c0 += 2;
172       }
173       if (nc & 1) {
174         *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
175         *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
176         *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
177         *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
178         *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
179       }
180 
181       nc = 0;
182     }
183   } while (nc != 0);
184 }
185