Home
last modified time | relevance | path

Searched refs:vacc89ABCDEFp0 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup16x25-minmax-fma3.c168 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
178 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
187 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi1x89ABCDEF, vk1x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
196 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
205 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi3x89ABCDEF, vk3x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
214 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi4x89ABCDEF, vk4x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
223 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi5x89ABCDEF, vk5x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
232 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
241 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi7x89ABCDEF, vk7x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
250 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi8x89ABCDEF, vk8x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
[all …]
Dup16x25-minmax-avx.c168 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
178 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi0x89ABCDEF, vk0x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
187 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
196 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
205 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi3x89ABCDEF, vk3x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
214 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi4x89ABCDEF, vk4x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
223 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi5x89ABCDEF, vk5x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
232 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi6x89ABCDEF, vk6x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
241 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi7x89ABCDEF, vk7x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
250 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi8x89ABCDEF, vk8x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
[all …]
Dup16x9-minmax-fma3.c88 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
98 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
107 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi1x89ABCDEF, vk1x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
116 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
125 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi3x89ABCDEF, vk3x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
134 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi4x89ABCDEF, vk4x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
143 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi5x89ABCDEF, vk5x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
152 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
161 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi7x89ABCDEF, vk7x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
170 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi8x89ABCDEF, vk8x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
[all …]
Dup16x9-minmax-avx.c88 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local
98 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi0x89ABCDEF, vk0x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
107 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
116 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
125 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi3x89ABCDEF, vk3x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
134 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi4x89ABCDEF, vk4x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
143 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi5x89ABCDEF, vk5x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
152 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi6x89ABCDEF, vk6x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
161 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi7x89ABCDEF, vk7x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
170 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi8x89ABCDEF, vk8x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
[all …]
Dup16x4-minmax-fma3.c63 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local
73 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
82 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi1x89ABCDEF, vk1x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
91 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
100 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi3x89ABCDEF, vk3x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
106 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
Dup16x4-minmax-avx.c63 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx() local
73 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi0x89ABCDEF, vk0x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx()
82 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx()
91 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx()
100 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi3x89ABCDEF, vk3x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx()
106 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx()
Dup16x9-minmax-fma3-acc2.c88 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
98 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
116 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
134 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi4x89ABCDEF, vk4x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
152 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
170 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi8x89ABCDEF, vk8x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
176 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, vacc89ABCDEFp1); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
179 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
Dup16x25-minmax-fma3-acc2.c168 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
178 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
196 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
214 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi4x89ABCDEF, vk4x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
232 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
250 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi8x89ABCDEF, vk8x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
268 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi10x89ABCDEF, vk10x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
286 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi12x89ABCDEF, vk12x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
304 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi14x89ABCDEF, vk14x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
322 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi16x89ABCDEF, vk16x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
[all …]
Dup16x25-minmax-avx-acc2.c168 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
178 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi0x89ABCDEF, vk0x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
196 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
214 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi4x89ABCDEF, vk4x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
232 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi6x89ABCDEF, vk6x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
250 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi8x89ABCDEF, vk8x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
268 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi10x89ABCDEF, vk10x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
286 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi12x89ABCDEF, vk12x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
304 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi14x89ABCDEF, vk14x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
322 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi16x89ABCDEF, vk16x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
[all …]
Dup16x4-minmax-fma3-acc2.c63 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2() local
73 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2()
91 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2()
106 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, vacc89ABCDEFp1); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2()
109 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2()
Dup16x4-minmax-avx-acc2.c63 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2() local
73 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi0x89ABCDEF, vk0x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2()
91 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2()
106 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, vacc89ABCDEFp1); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2()
109 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup16x25-minmax-neonfp16arith.c168 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
176 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
183 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
190 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
197 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi3x89ABCDEF, vk3x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
204 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi4x89ABCDEF, vk4x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
211 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi5x89ABCDEF, vk5x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
218 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
225 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi7x89ABCDEF, vk7x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
232 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi8x89ABCDEF, vk8x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
[all …]
Dup16x9-minmax-neonfp16arith.c88 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local
96 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
103 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
110 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
117 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi3x89ABCDEF, vk3x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
124 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi4x89ABCDEF, vk4x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
131 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi5x89ABCDEF, vk5x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
138 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
145 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi7x89ABCDEF, vk7x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
152 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi8x89ABCDEF, vk8x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
[all …]
Dup16x4-minmax-neonfp16arith.c63 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local
71 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
78 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
85 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
92 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi3x89ABCDEF, vk3x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
96 float16x8_t vacc89ABCDEF = vmaxq_f16(vacc89ABCDEFp0, vmin); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
Dup16x25-minmax-neonfp16arith-acc2.c168 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
176 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
190 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
204 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi4x89ABCDEF, vk4x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
218 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
232 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi8x89ABCDEF, vk8x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
246 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi10x89ABCDEF, vk10x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
260 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi12x89ABCDEF, vk12x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
274 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi14x89ABCDEF, vk14x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
288 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi16x89ABCDEF, vk16x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
[all …]
Dup16x9-minmax-neonfp16arith-acc2.c88 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local
96 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
110 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
124 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi4x89ABCDEF, vk4x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
138 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
152 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi8x89ABCDEF, vk8x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
156 vacc89ABCDEFp0 = vaddq_f16(vacc89ABCDEFp0, vacc89ABCDEFp1); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
159 float16x8_t vacc89ABCDEF = vmaxq_f16(vacc89ABCDEFp0, vmin); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
Dup16x4-minmax-neonfp16arith-acc2.c63 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2() local
71 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
85 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
96 vacc89ABCDEFp0 = vaddq_f16(vacc89ABCDEFp0, vacc89ABCDEFp1); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
99 float16x8_t vacc89ABCDEF = vmaxq_f16(vacc89ABCDEFp0, vmin); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()