| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up2x25-scalar.c | 87 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x25__scalar() local
96 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
105 vacc0p0 += vi1x0 * vk1x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
114 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
123 vacc0p0 += vi3x0 * vk3x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
132 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
141 vacc0p0 += vi5x0 * vk5x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
150 vacc0p0 += vi6x0 * vk6x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
159 vacc0p0 += vi7x0 * vk7x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
168 vacc0p0 += vi8x0 * vk8x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
[all …]
|
| D | up2x25-wasm.c | 87 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x25__wasm() local
96 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
105 vacc0p0 += vi1x0 * vk1x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
114 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
123 vacc0p0 += vi3x0 * vk3x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
132 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
141 vacc0p0 += vi5x0 * vk5x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
150 vacc0p0 += vi6x0 * vk6x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
159 vacc0p0 += vi7x0 * vk7x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
168 vacc0p0 += vi8x0 * vk8x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
[all …]
|
| D | up1x25-wasm.c | 87 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x25__wasm() local
91 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x25__wasm()
95 vacc0p0 += vi1 * vk1; in xnn_f32_dwconv_ukernel_up1x25__wasm()
99 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x25__wasm()
103 vacc0p0 += vi3 * vk3; in xnn_f32_dwconv_ukernel_up1x25__wasm()
107 vacc0p0 += vi4 * vk4; in xnn_f32_dwconv_ukernel_up1x25__wasm()
111 vacc0p0 += vi5 * vk5; in xnn_f32_dwconv_ukernel_up1x25__wasm()
115 vacc0p0 += vi6 * vk6; in xnn_f32_dwconv_ukernel_up1x25__wasm()
119 vacc0p0 += vi7 * vk7; in xnn_f32_dwconv_ukernel_up1x25__wasm()
123 vacc0p0 += vi8 * vk8; in xnn_f32_dwconv_ukernel_up1x25__wasm()
[all …]
|
| D | up1x25-scalar.c | 87 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x25__scalar() local
91 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x25__scalar()
95 vacc0p0 += vi1 * vk1; in xnn_f32_dwconv_ukernel_up1x25__scalar()
99 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x25__scalar()
103 vacc0p0 += vi3 * vk3; in xnn_f32_dwconv_ukernel_up1x25__scalar()
107 vacc0p0 += vi4 * vk4; in xnn_f32_dwconv_ukernel_up1x25__scalar()
111 vacc0p0 += vi5 * vk5; in xnn_f32_dwconv_ukernel_up1x25__scalar()
115 vacc0p0 += vi6 * vk6; in xnn_f32_dwconv_ukernel_up1x25__scalar()
119 vacc0p0 += vi7 * vk7; in xnn_f32_dwconv_ukernel_up1x25__scalar()
123 vacc0p0 += vi8 * vk8; in xnn_f32_dwconv_ukernel_up1x25__scalar()
[all …]
|
| D | up2x9-wasm.c | 55 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x9__wasm() local
64 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
73 vacc0p0 += vi1x0 * vk1x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
82 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
91 vacc0p0 += vi3x0 * vk3x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
100 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
109 vacc0p0 += vi5x0 * vk5x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
118 vacc0p0 += vi6x0 * vk6x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
127 vacc0p0 += vi7x0 * vk7x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
136 vacc0p0 += vi8x0 * vk8x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
[all …]
|
| D | up2x9-scalar.c | 55 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x9__scalar() local
64 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
73 vacc0p0 += vi1x0 * vk1x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
82 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
91 vacc0p0 += vi3x0 * vk3x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
100 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
109 vacc0p0 += vi5x0 * vk5x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
118 vacc0p0 += vi6x0 * vk6x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
127 vacc0p0 += vi7x0 * vk7x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
136 vacc0p0 += vi8x0 * vk8x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
[all …]
|
| D | up2x25-scalar-acc2.c | 87 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2() local
96 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
114 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
132 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
150 vacc0p0 += vi6x0 * vk6x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
168 vacc0p0 += vi8x0 * vk8x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
186 vacc0p0 += vi10x0 * vk10x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
204 vacc0p0 += vi12x0 * vk12x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
222 vacc0p0 += vi14x0 * vk14x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
240 vacc0p0 += vi16x0 * vk16x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
[all …]
|
| D | up2x25-wasm-acc2.c | 87 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2() local
96 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
114 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
132 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
150 vacc0p0 += vi6x0 * vk6x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
168 vacc0p0 += vi8x0 * vk8x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
186 vacc0p0 += vi10x0 * vk10x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
204 vacc0p0 += vi12x0 * vk12x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
222 vacc0p0 += vi14x0 * vk14x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
240 vacc0p0 += vi16x0 * vk16x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
[all …]
|
| D | up2x9-scalar-acc2.c | 55 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2() local
64 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
82 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
100 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
118 vacc0p0 += vi6x0 * vk6x0; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
136 vacc0p0 += vi8x0 * vk8x0; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
143 vacc0p0 = vacc0p0 + vacc0p1; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
146 float vacc0 = math_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
157 float vacc0p0 = *w++; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2() local
161 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
[all …]
|
| D | up2x9-wasm-acc2.c | 55 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2() local
64 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
82 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
100 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
118 vacc0p0 += vi6x0 * vk6x0; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
136 vacc0p0 += vi8x0 * vk8x0; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
143 vacc0p0 = vacc0p0 + vacc0p1; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
146 float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
157 float vacc0p0 = *w++; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2() local
161 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
[all …]
|
| D | up2x4-scalar.c | 45 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x4__scalar() local
54 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x4__scalar()
63 vacc0p0 += vi1x0 * vk1x0; in xnn_f32_dwconv_ukernel_up2x4__scalar()
72 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x4__scalar()
81 vacc0p0 += vi3x0 * vk3x0; in xnn_f32_dwconv_ukernel_up2x4__scalar()
88 float vacc0 = math_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up2x4__scalar()
99 float vacc0p0 = *w++; in xnn_f32_dwconv_ukernel_up2x4__scalar() local
103 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up2x4__scalar()
106 vacc0p0 += vi1 * vk1; in xnn_f32_dwconv_ukernel_up2x4__scalar()
109 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up2x4__scalar()
[all …]
|
| D | up2x4-wasm.c | 45 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x4__wasm() local
54 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x4__wasm()
63 vacc0p0 += vi1x0 * vk1x0; in xnn_f32_dwconv_ukernel_up2x4__wasm()
72 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x4__wasm()
81 vacc0p0 += vi3x0 * vk3x0; in xnn_f32_dwconv_ukernel_up2x4__wasm()
88 float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up2x4__wasm()
99 float vacc0p0 = *w++; in xnn_f32_dwconv_ukernel_up2x4__wasm() local
103 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up2x4__wasm()
106 vacc0p0 += vi1 * vk1; in xnn_f32_dwconv_ukernel_up2x4__wasm()
109 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up2x4__wasm()
[all …]
|
| D | up2x4-wasm-acc2.c | 45 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2() local
54 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2()
72 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2()
88 vacc0p0 = vacc0p0 + vacc0p1; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2()
91 float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2()
102 float vacc0p0 = *w++; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2() local
106 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2()
112 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2()
118 vacc0p0 = vacc0p0 + vacc0p1; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2()
120 float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2()
|
| D | up2x4-scalar-acc2.c | 45 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2() local
54 vacc0p0 += vi0x0 * vk0x0; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2()
72 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2()
88 vacc0p0 = vacc0p0 + vacc0p1; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2()
91 float vacc0 = math_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2()
102 float vacc0p0 = *w++; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2() local
106 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2()
112 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2()
118 vacc0p0 = vacc0p0 + vacc0p1; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2()
120 float vacc0 = math_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2()
|
| D | up1x9-scalar.c | 55 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x9__scalar() local
59 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x9__scalar()
63 vacc0p0 += vi1 * vk1; in xnn_f32_dwconv_ukernel_up1x9__scalar()
67 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x9__scalar()
71 vacc0p0 += vi3 * vk3; in xnn_f32_dwconv_ukernel_up1x9__scalar()
75 vacc0p0 += vi4 * vk4; in xnn_f32_dwconv_ukernel_up1x9__scalar()
79 vacc0p0 += vi5 * vk5; in xnn_f32_dwconv_ukernel_up1x9__scalar()
83 vacc0p0 += vi6 * vk6; in xnn_f32_dwconv_ukernel_up1x9__scalar()
87 vacc0p0 += vi7 * vk7; in xnn_f32_dwconv_ukernel_up1x9__scalar()
91 vacc0p0 += vi8 * vk8; in xnn_f32_dwconv_ukernel_up1x9__scalar()
[all …]
|
| D | up1x9-wasm.c | 55 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x9__wasm() local
59 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x9__wasm()
63 vacc0p0 += vi1 * vk1; in xnn_f32_dwconv_ukernel_up1x9__wasm()
67 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x9__wasm()
71 vacc0p0 += vi3 * vk3; in xnn_f32_dwconv_ukernel_up1x9__wasm()
75 vacc0p0 += vi4 * vk4; in xnn_f32_dwconv_ukernel_up1x9__wasm()
79 vacc0p0 += vi5 * vk5; in xnn_f32_dwconv_ukernel_up1x9__wasm()
83 vacc0p0 += vi6 * vk6; in xnn_f32_dwconv_ukernel_up1x9__wasm()
87 vacc0p0 += vi7 * vk7; in xnn_f32_dwconv_ukernel_up1x9__wasm()
91 vacc0p0 += vi8 * vk8; in xnn_f32_dwconv_ukernel_up1x9__wasm()
[all …]
|
| D | up1x25-wasm-acc2.c | 87 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x25__wasm_acc2() local
91 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x25__wasm_acc2()
99 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x25__wasm_acc2()
107 vacc0p0 += vi4 * vk4; in xnn_f32_dwconv_ukernel_up1x25__wasm_acc2()
115 vacc0p0 += vi6 * vk6; in xnn_f32_dwconv_ukernel_up1x25__wasm_acc2()
123 vacc0p0 += vi8 * vk8; in xnn_f32_dwconv_ukernel_up1x25__wasm_acc2()
131 vacc0p0 += vi10 * vk10; in xnn_f32_dwconv_ukernel_up1x25__wasm_acc2()
139 vacc0p0 += vi12 * vk12; in xnn_f32_dwconv_ukernel_up1x25__wasm_acc2()
147 vacc0p0 += vi14 * vk14; in xnn_f32_dwconv_ukernel_up1x25__wasm_acc2()
155 vacc0p0 += vi16 * vk16; in xnn_f32_dwconv_ukernel_up1x25__wasm_acc2()
[all …]
|
| D | up1x25-scalar-acc2.c | 87 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x25__scalar_acc2() local
91 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x25__scalar_acc2()
99 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x25__scalar_acc2()
107 vacc0p0 += vi4 * vk4; in xnn_f32_dwconv_ukernel_up1x25__scalar_acc2()
115 vacc0p0 += vi6 * vk6; in xnn_f32_dwconv_ukernel_up1x25__scalar_acc2()
123 vacc0p0 += vi8 * vk8; in xnn_f32_dwconv_ukernel_up1x25__scalar_acc2()
131 vacc0p0 += vi10 * vk10; in xnn_f32_dwconv_ukernel_up1x25__scalar_acc2()
139 vacc0p0 += vi12 * vk12; in xnn_f32_dwconv_ukernel_up1x25__scalar_acc2()
147 vacc0p0 += vi14 * vk14; in xnn_f32_dwconv_ukernel_up1x25__scalar_acc2()
155 vacc0p0 += vi16 * vk16; in xnn_f32_dwconv_ukernel_up1x25__scalar_acc2()
[all …]
|
| D | up1x9-wasm-acc2.c | 55 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x9__wasm_acc2() local
59 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x9__wasm_acc2()
67 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x9__wasm_acc2()
75 vacc0p0 += vi4 * vk4; in xnn_f32_dwconv_ukernel_up1x9__wasm_acc2()
83 vacc0p0 += vi6 * vk6; in xnn_f32_dwconv_ukernel_up1x9__wasm_acc2()
91 vacc0p0 += vi8 * vk8; in xnn_f32_dwconv_ukernel_up1x9__wasm_acc2()
95 vacc0p0 += vacc0p1; in xnn_f32_dwconv_ukernel_up1x9__wasm_acc2()
97 float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up1x9__wasm_acc2()
|
| D | up1x9-scalar-acc2.c | 55 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x9__scalar_acc2() local
59 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x9__scalar_acc2()
67 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x9__scalar_acc2()
75 vacc0p0 += vi4 * vk4; in xnn_f32_dwconv_ukernel_up1x9__scalar_acc2()
83 vacc0p0 += vi6 * vk6; in xnn_f32_dwconv_ukernel_up1x9__scalar_acc2()
91 vacc0p0 += vi8 * vk8; in xnn_f32_dwconv_ukernel_up1x9__scalar_acc2()
95 vacc0p0 += vacc0p1; in xnn_f32_dwconv_ukernel_up1x9__scalar_acc2()
97 float vacc0 = math_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up1x9__scalar_acc2()
|
| D | up1x4-wasm.c | 45 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x4__wasm() local
49 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x4__wasm()
53 vacc0p0 += vi1 * vk1; in xnn_f32_dwconv_ukernel_up1x4__wasm()
57 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x4__wasm()
61 vacc0p0 += vi3 * vk3; in xnn_f32_dwconv_ukernel_up1x4__wasm()
66 float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up1x4__wasm()
|
| D | up1x4-scalar.c | 45 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x4__scalar() local
49 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x4__scalar()
53 vacc0p0 += vi1 * vk1; in xnn_f32_dwconv_ukernel_up1x4__scalar()
57 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x4__scalar()
61 vacc0p0 += vi3 * vk3; in xnn_f32_dwconv_ukernel_up1x4__scalar()
66 float vacc0 = math_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up1x4__scalar()
|
| D | up1x4-wasm-acc2.c | 45 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x4__wasm_acc2() local
49 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x4__wasm_acc2()
57 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x4__wasm_acc2()
65 vacc0p0 += vacc0p1; in xnn_f32_dwconv_ukernel_up1x4__wasm_acc2()
67 float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up1x4__wasm_acc2()
|
| D | up1x4-scalar-acc2.c | 45 float vacc0p0 = w[0]; in xnn_f32_dwconv_ukernel_up1x4__scalar_acc2() local
49 vacc0p0 += vi0 * vk0; in xnn_f32_dwconv_ukernel_up1x4__scalar_acc2()
57 vacc0p0 += vi2 * vk2; in xnn_f32_dwconv_ukernel_up1x4__scalar_acc2()
65 vacc0p0 += vacc0p1; in xnn_f32_dwconv_ukernel_up1x4__scalar_acc2()
67 float vacc0 = math_max_f32(vacc0p0, vmin); in xnn_f32_dwconv_ukernel_up1x4__scalar_acc2()
|
| /external/XNNPACK/src/f32-dwconv/ |
| D | up-scalar.c.in | 82 float vacc0p0 = *w++;
101 float vacc0 = ${MAX_F32}(vacc0p0, vmin);
107 float vacc0p0 = w[0];
126 float vacc0 = ${MAX_F32}(vacc0p0, vmin);
|