• 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_minmax_ukernel_4x8__wasmsimd_x86_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_x86_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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(mr != 0);
30   assert(mr <= 4);
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 
59   do {
60     v128_t vacc0x0123 = wasm_v128_load(w + 0);
61     v128_t vacc0x4567 = wasm_v128_load(w + 4);
62     v128_t vacc1x0123 = vacc0x0123;
63     v128_t vacc1x4567 = vacc0x4567;
64     v128_t vacc2x0123 = vacc0x0123;
65     v128_t vacc2x4567 = vacc0x4567;
66     v128_t vacc3x0123 = vacc0x0123;
67     v128_t vacc3x4567 = vacc0x4567;
68     w += 8;
69 
70     size_t k = kc;
71     do {
72       const v128_t va0 = wasm_v32x4_load_splat(a0);
73       a0 += 1;
74       const v128_t va1 = wasm_v32x4_load_splat(a1);
75       a1 += 1;
76       const v128_t va2 = wasm_v32x4_load_splat(a2);
77       a2 += 1;
78       const v128_t va3 = wasm_v32x4_load_splat(a3);
79       a3 += 1;
80 
81       const v128_t vb0123 = wasm_v128_load(w);
82       const v128_t vb4567 = wasm_v128_load(w + 4);
83       w += 8;
84 
85       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
86       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
87       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
88       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
89       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
90       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
91       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
92       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
93 
94       k -= sizeof(float);
95     } while (k != 0);
96 
97     const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
98     vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin));
99     vacc1x0123 = wasm_v128_bitselect(vmin, vacc1x0123, wasm_f32x4_lt(vacc1x0123, vmin));
100     vacc2x0123 = wasm_v128_bitselect(vmin, vacc2x0123, wasm_f32x4_lt(vacc2x0123, vmin));
101     vacc3x0123 = wasm_v128_bitselect(vmin, vacc3x0123, wasm_f32x4_lt(vacc3x0123, vmin));
102     vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin));
103     vacc1x4567 = wasm_v128_bitselect(vmin, vacc1x4567, wasm_f32x4_lt(vacc1x4567, vmin));
104     vacc2x4567 = wasm_v128_bitselect(vmin, vacc2x4567, wasm_f32x4_lt(vacc2x4567, vmin));
105     vacc3x4567 = wasm_v128_bitselect(vmin, vacc3x4567, wasm_f32x4_lt(vacc3x4567, vmin));
106 
107     const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
108     vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax));
109     vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax));
110     vacc2x0123 = wasm_v128_bitselect(vacc2x0123, vmax, wasm_f32x4_le(vacc2x0123, vmax));
111     vacc3x0123 = wasm_v128_bitselect(vacc3x0123, vmax, wasm_f32x4_le(vacc3x0123, vmax));
112     vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax));
113     vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax));
114     vacc2x4567 = wasm_v128_bitselect(vacc2x4567, vmax, wasm_f32x4_le(vacc2x4567, vmax));
115     vacc3x4567 = wasm_v128_bitselect(vacc3x4567, vmax, wasm_f32x4_le(vacc3x4567, vmax));
116 
117     if XNN_LIKELY(nc >= 8) {
118       wasm_v128_store(c3, vacc3x0123);
119       wasm_v128_store(c3 + 4, vacc3x4567);
120       c3 = (float*) ((uintptr_t) c3 + cn_stride);
121       wasm_v128_store(c2, vacc2x0123);
122       wasm_v128_store(c2 + 4, vacc2x4567);
123       c2 = (float*) ((uintptr_t) c2 + cn_stride);
124       wasm_v128_store(c1, vacc1x0123);
125       wasm_v128_store(c1 + 4, vacc1x4567);
126       c1 = (float*) ((uintptr_t) c1 + cn_stride);
127       wasm_v128_store(c0, vacc0x0123);
128       wasm_v128_store(c0 + 4, vacc0x4567);
129       c0 = (float*) ((uintptr_t) c0 + cn_stride);
130 
131       a3 = (const float*) ((uintptr_t) a3 - kc);
132       a2 = (const float*) ((uintptr_t) a2 - kc);
133       a1 = (const float*) ((uintptr_t) a1 - kc);
134       a0 = (const float*) ((uintptr_t) a0 - kc);
135 
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