1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_arm(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_4x2c4__wasmsimd_arm(
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)]) XNN_DISABLE_TSAN
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 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min);
60 const v128_t vmax = wasm_v32x4_load_splat(¶ms->scalar.max);
61 do {
62 v128_t vacc0x0c4 = wasm_f32x4_replace_lane(wasm_f32x4_splat(0.0f), 0, w[0]);
63 v128_t vacc0x1c4 = wasm_f32x4_replace_lane(vacc0x0c4, 0, w[1]);
64 v128_t vacc1x0c4 = vacc0x0c4;
65 v128_t vacc1x1c4 = vacc0x1c4;
66 v128_t vacc2x0c4 = vacc0x0c4;
67 v128_t vacc2x1c4 = vacc0x1c4;
68 v128_t vacc3x0c4 = vacc0x0c4;
69 v128_t vacc3x1c4 = vacc0x1c4;
70 w += 2;
71
72 size_t k = kc;
73 for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
74 const v128_t va0 = wasm_v128_load(a0);
75 a0 += 4;
76 const v128_t va1 = wasm_v128_load(a1);
77 a1 += 4;
78 const v128_t va2 = wasm_v128_load(a2);
79 a2 += 4;
80 const v128_t va3 = wasm_v128_load(a3);
81 a3 += 4;
82
83 const v128_t vb0 = wasm_v128_load(w);
84 const v128_t vb1 = wasm_v128_load(w + 4);
85 w += 8;
86
87 vacc0x0c4 = wasm_f32x4_add(vacc0x0c4, wasm_f32x4_mul(va0, vb0));
88 vacc0x1c4 = wasm_f32x4_add(vacc0x1c4, wasm_f32x4_mul(va0, vb1));
89 vacc1x0c4 = wasm_f32x4_add(vacc1x0c4, wasm_f32x4_mul(va1, vb0));
90 vacc1x1c4 = wasm_f32x4_add(vacc1x1c4, wasm_f32x4_mul(va1, vb1));
91 vacc2x0c4 = wasm_f32x4_add(vacc2x0c4, wasm_f32x4_mul(va2, vb0));
92 vacc2x1c4 = wasm_f32x4_add(vacc2x1c4, wasm_f32x4_mul(va2, vb1));
93 vacc3x0c4 = wasm_f32x4_add(vacc3x0c4, wasm_f32x4_mul(va3, vb0));
94 vacc3x1c4 = wasm_f32x4_add(vacc3x1c4, wasm_f32x4_mul(va3, vb1));
95 }
96 if XNN_UNLIKELY(k != 0) {
97 const v128_t va0 = wasm_v128_load(a0);
98 a0 = (const float*) ((uintptr_t) a0 + k);
99 const v128_t va1 = wasm_v128_load(a1);
100 a1 = (const float*) ((uintptr_t) a1 + k);
101 const v128_t va2 = wasm_v128_load(a2);
102 a2 = (const float*) ((uintptr_t) a2 + k);
103 const v128_t va3 = wasm_v128_load(a3);
104 a3 = (const float*) ((uintptr_t) a3 + k);
105
106 const v128_t vb0 = wasm_v128_load(w);
107 const v128_t vb1 = wasm_v128_load(w + 4);
108 w += 8;
109
110 const v128_t vzero = wasm_f32x4_splat(0.0f);
111 const v128_t vmask0 = wasm_f32x4_eq(vb0, vzero);
112 const v128_t vmask1 = wasm_f32x4_eq(vb1, vzero);
113
114 vacc0x0c4 = wasm_f32x4_add(vacc0x0c4, wasm_f32x4_mul(wasm_v128_andnot(va0, vmask0), vb0));
115 vacc0x1c4 = wasm_f32x4_add(vacc0x1c4, wasm_f32x4_mul(wasm_v128_andnot(va0, vmask1), vb1));
116 vacc1x0c4 = wasm_f32x4_add(vacc1x0c4, wasm_f32x4_mul(wasm_v128_andnot(va1, vmask0), vb0));
117 vacc1x1c4 = wasm_f32x4_add(vacc1x1c4, wasm_f32x4_mul(wasm_v128_andnot(va1, vmask1), vb1));
118 vacc2x0c4 = wasm_f32x4_add(vacc2x0c4, wasm_f32x4_mul(wasm_v128_andnot(va2, vmask0), vb0));
119 vacc2x1c4 = wasm_f32x4_add(vacc2x1c4, wasm_f32x4_mul(wasm_v128_andnot(va2, vmask1), vb1));
120 vacc3x0c4 = wasm_f32x4_add(vacc3x0c4, wasm_f32x4_mul(wasm_v128_andnot(va3, vmask0), vb0));
121 vacc3x1c4 = wasm_f32x4_add(vacc3x1c4, wasm_f32x4_mul(wasm_v128_andnot(va3, vmask1), vb1));
122 }
123
124 const v128_t vacc0x01c2 = wasm_f32x4_add(
125 wasm_v32x4_shuffle(vacc0x0c4, vacc0x1c4, 0, 4, 1, 5),
126 wasm_v32x4_shuffle(vacc0x0c4, vacc0x1c4, 2, 6, 3, 7));
127 const v128_t vacc1x01c2 = wasm_f32x4_add(
128 wasm_v32x4_shuffle(vacc1x0c4, vacc1x1c4, 0, 4, 1, 5),
129 wasm_v32x4_shuffle(vacc1x0c4, vacc1x1c4, 2, 6, 3, 7));
130 const v128_t vacc2x01c2 = wasm_f32x4_add(
131 wasm_v32x4_shuffle(vacc2x0c4, vacc2x1c4, 0, 4, 1, 5),
132 wasm_v32x4_shuffle(vacc2x0c4, vacc2x1c4, 2, 6, 3, 7));
133 const v128_t vacc3x01c2 = wasm_f32x4_add(
134 wasm_v32x4_shuffle(vacc3x0c4, vacc3x1c4, 0, 4, 1, 5),
135 wasm_v32x4_shuffle(vacc3x0c4, vacc3x1c4, 2, 6, 3, 7));
136
137 v128_t vacc01x01 = wasm_f32x4_add(
138 wasm_v32x4_shuffle(vacc0x01c2, vacc1x01c2, 0, 1, 4, 5),
139 wasm_v32x4_shuffle(vacc0x01c2, vacc1x01c2, 2, 3, 6, 7));
140 v128_t vacc23x01 = wasm_f32x4_add(
141 wasm_v32x4_shuffle(vacc2x01c2, vacc3x01c2, 0, 1, 4, 5),
142 wasm_v32x4_shuffle(vacc2x01c2, vacc3x01c2, 2, 3, 6, 7));
143
144 vacc01x01 = wasm_f32x4_max(vacc01x01, vmin);
145 vacc23x01 = wasm_f32x4_max(vacc23x01, vmin);
146
147 vacc01x01 = wasm_f32x4_min(vacc01x01, vmax);
148 vacc23x01 = wasm_f32x4_min(vacc23x01, vmax);
149
150 if XNN_LIKELY(nc >= 2) {
151 *((double*) c2) = wasm_f64x2_extract_lane(vacc23x01, 0);
152 c2 = (float*) ((uintptr_t) c2 + cn_stride);
153 a2 = (const float*) ((uintptr_t) a2 - kc);
154 *((double*) c3) = wasm_f64x2_extract_lane(vacc23x01, 1);
155 c3 = (float*) ((uintptr_t) c3 + cn_stride);
156 a3 = (const float*) ((uintptr_t) a3 - kc);
157 *((double*) c0) = wasm_f64x2_extract_lane(vacc01x01, 0);
158 c0 = (float*) ((uintptr_t) c0 + cn_stride);
159 a0 = (const float*) ((uintptr_t) a0 - kc);
160 *((double*) c1) = wasm_f64x2_extract_lane(vacc01x01, 1);
161 c1 = (float*) ((uintptr_t) c1 + cn_stride);
162 a1 = (const float*) ((uintptr_t) a1 - kc);
163
164 nc -= 2;
165 } else {
166 assert(nc == 1);
167 *c2 = wasm_f32x4_extract_lane(vacc23x01, 0);
168 *c3 = wasm_f32x4_extract_lane(vacc23x01, 2);
169 *c0 = wasm_f32x4_extract_lane(vacc01x01, 0);
170 *c1 = wasm_f32x4_extract_lane(vacc01x01, 2);
171
172 nc = 0;
173 }
174 } while (nc != 0);
175 }
176