1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/MRx2c4-wasmsimd.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_relu_ukernel_4x2c4__wasmsimd(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_relu_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_relu_ukernel_4x2c4__wasmsimd(
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_relu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
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 vacc0x0c4 = wasm_f32x4_replace_lane(wasm_f32x4_splat(0.0f), 0, w[0]);
59 v128_t vacc0x1c4 = wasm_f32x4_replace_lane(vacc0x0c4, 0, w[1]);
60 v128_t vacc1x0c4 = vacc0x0c4;
61 v128_t vacc1x1c4 = vacc0x1c4;
62 v128_t vacc2x0c4 = vacc0x0c4;
63 v128_t vacc2x1c4 = vacc0x1c4;
64 v128_t vacc3x0c4 = vacc0x0c4;
65 v128_t vacc3x1c4 = vacc0x1c4;
66 w += 2;
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 for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
94 const v128_t va0 = wasm_v128_load(a0);
95 a0 += 4;
96 const v128_t va1 = wasm_v128_load(a1);
97 a1 += 4;
98 const v128_t va2 = wasm_v128_load(a2);
99 a2 += 4;
100 const v128_t va3 = wasm_v128_load(a3);
101 a3 += 4;
102
103 const v128_t vb0 = wasm_v128_load(w);
104 const v128_t vb1 = wasm_v128_load(w + 4);
105 w += 8;
106
107 vacc0x0c4 = wasm_f32x4_add(vacc0x0c4, wasm_f32x4_mul(va0, vb0));
108 vacc0x1c4 = wasm_f32x4_add(vacc0x1c4, wasm_f32x4_mul(va0, vb1));
109 vacc1x0c4 = wasm_f32x4_add(vacc1x0c4, wasm_f32x4_mul(va1, vb0));
110 vacc1x1c4 = wasm_f32x4_add(vacc1x1c4, wasm_f32x4_mul(va1, vb1));
111 vacc2x0c4 = wasm_f32x4_add(vacc2x0c4, wasm_f32x4_mul(va2, vb0));
112 vacc2x1c4 = wasm_f32x4_add(vacc2x1c4, wasm_f32x4_mul(va2, vb1));
113 vacc3x0c4 = wasm_f32x4_add(vacc3x0c4, wasm_f32x4_mul(va3, vb0));
114 vacc3x1c4 = wasm_f32x4_add(vacc3x1c4, wasm_f32x4_mul(va3, vb1));
115 }
116 if XNN_UNLIKELY(k != 0) {
117 const v128_t va0 = wasm_v128_load(a0);
118 const v128_t va1 = wasm_v128_load(a1);
119 const v128_t va2 = wasm_v128_load(a2);
120 const v128_t va3 = wasm_v128_load(a3);
121
122 const v128_t vb0 = wasm_v128_load(w);
123 const v128_t vb1 = wasm_v128_load(w + 4);
124 w += 8;
125
126 const v128_t vzero = wasm_f32x4_splat(0.0f);
127 const v128_t vmask0 = wasm_f32x4_eq(vb0, vzero);
128 const v128_t vmask1 = wasm_f32x4_eq(vb1, vzero);
129
130 vacc0x0c4 = wasm_f32x4_add(vacc0x0c4, wasm_f32x4_mul(wasm_v128_andnot(va0, vmask0), vb0));
131 vacc0x1c4 = wasm_f32x4_add(vacc0x1c4, wasm_f32x4_mul(wasm_v128_andnot(va0, vmask1), vb1));
132 vacc1x0c4 = wasm_f32x4_add(vacc1x0c4, wasm_f32x4_mul(wasm_v128_andnot(va1, vmask0), vb0));
133 vacc1x1c4 = wasm_f32x4_add(vacc1x1c4, wasm_f32x4_mul(wasm_v128_andnot(va1, vmask1), vb1));
134 vacc2x0c4 = wasm_f32x4_add(vacc2x0c4, wasm_f32x4_mul(wasm_v128_andnot(va2, vmask0), vb0));
135 vacc2x1c4 = wasm_f32x4_add(vacc2x1c4, wasm_f32x4_mul(wasm_v128_andnot(va2, vmask1), vb1));
136 vacc3x0c4 = wasm_f32x4_add(vacc3x0c4, wasm_f32x4_mul(wasm_v128_andnot(va3, vmask0), vb0));
137 vacc3x1c4 = wasm_f32x4_add(vacc3x1c4, wasm_f32x4_mul(wasm_v128_andnot(va3, vmask1), vb1));
138 }
139 p -= 4 * sizeof(void*);
140 } while (p != 0);
141
142 const v128_t vacc0x01c2 = wasm_f32x4_add(
143 wasm_v32x4_shuffle(vacc0x0c4, vacc0x1c4, 0, 4, 1, 5),
144 wasm_v32x4_shuffle(vacc0x0c4, vacc0x1c4, 2, 6, 3, 7));
145 const v128_t vacc1x01c2 = wasm_f32x4_add(
146 wasm_v32x4_shuffle(vacc1x0c4, vacc1x1c4, 0, 4, 1, 5),
147 wasm_v32x4_shuffle(vacc1x0c4, vacc1x1c4, 2, 6, 3, 7));
148 const v128_t vacc2x01c2 = wasm_f32x4_add(
149 wasm_v32x4_shuffle(vacc2x0c4, vacc2x1c4, 0, 4, 1, 5),
150 wasm_v32x4_shuffle(vacc2x0c4, vacc2x1c4, 2, 6, 3, 7));
151 const v128_t vacc3x01c2 = wasm_f32x4_add(
152 wasm_v32x4_shuffle(vacc3x0c4, vacc3x1c4, 0, 4, 1, 5),
153 wasm_v32x4_shuffle(vacc3x0c4, vacc3x1c4, 2, 6, 3, 7));
154
155 v128_t vacc01x01 = wasm_f32x4_add(
156 wasm_v32x4_shuffle(vacc0x01c2, vacc1x01c2, 0, 1, 4, 5),
157 wasm_v32x4_shuffle(vacc0x01c2, vacc1x01c2, 2, 3, 6, 7));
158 v128_t vacc23x01 = wasm_f32x4_add(
159 wasm_v32x4_shuffle(vacc2x01c2, vacc3x01c2, 0, 1, 4, 5),
160 wasm_v32x4_shuffle(vacc2x01c2, vacc3x01c2, 2, 3, 6, 7));
161
162 const v128_t vzero = wasm_f32x4_splat(0.0f);
163 vacc01x01 = wasm_i32x4_max(vacc01x01, vzero);
164 vacc23x01 = wasm_i32x4_max(vacc23x01, vzero);
165
166 if XNN_LIKELY(nc >= 2) {
167 *((double*) c3) = wasm_f64x2_extract_lane(vacc23x01, 1);
168 c3 = (float*) ((uintptr_t) c3 + cn_stride);
169 *((double*) c2) = wasm_f64x2_extract_lane(vacc23x01, 0);
170 c2 = (float*) ((uintptr_t) c2 + cn_stride);
171 *((double*) c1) = wasm_f64x2_extract_lane(vacc01x01, 1);
172 c1 = (float*) ((uintptr_t) c1 + cn_stride);
173 *((double*) c0) = wasm_f64x2_extract_lane(vacc01x01, 0);
174 c0 = (float*) ((uintptr_t) c0 + cn_stride);
175
176 a = (const float**restrict) ((uintptr_t) a - ks);
177 nc -= 2;
178 } else {
179 assert(nc == 1);
180 *c3 = wasm_f32x4_extract_lane(vacc23x01, 2);
181 *c2 = wasm_f32x4_extract_lane(vacc23x01, 0);
182 *c1 = wasm_f32x4_extract_lane(vacc01x01, 2);
183 *c0 = wasm_f32x4_extract_lane(vacc01x01, 0);
184
185 nc = 0;
186 }
187 } while (nc != 0);
188 }
189