Home
last modified time | relevance | path

Searched refs:vkernel_zero_point (Results 1 – 25 of 66) sorted by relevance

123

/external/XNNPACK/src/qu8-dwconv/gen/
Dup4x25-minmax-fp32-scalar-fmagic.c36 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic() local
180 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic()
181 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic()
182 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic()
183 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic()
196 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic()
197 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic()
198 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic()
199 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic()
212 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic()
[all …]
Dup4x25-minmax-fp32-scalar-lrintf.c36 const int32_t vkernel_zero_point = params->fp32_scalar_lrintf.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf() local
180 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf()
181 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf()
182 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf()
183 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf()
196 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf()
197 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf()
198 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf()
199 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf()
212 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf()
[all …]
Dup4x25-minmax-fp32-scalar-imagic.c36 const int32_t vkernel_zero_point = params->fp32_scalar_imagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic() local
180 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic()
181 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic()
182 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic()
183 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic()
196 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic()
197 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic()
198 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic()
199 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic()
212 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic()
[all …]
Dup4x25-minmax-fp32-wasm-fmagic.c36 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__wasm_fmagic() local
180 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__wasm_fmagic()
181 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__wasm_fmagic()
182 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__wasm_fmagic()
183 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__wasm_fmagic()
196 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__wasm_fmagic()
197 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__wasm_fmagic()
198 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__wasm_fmagic()
199 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__wasm_fmagic()
212 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__wasm_fmagic()
[all …]
Dup2x25-minmax-fp32-scalar-lrintf.c37 const int32_t vkernel_zero_point = params->fp32_scalar_lrintf.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf() local
177 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf()
178 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf()
187 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf()
188 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf()
197 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf()
198 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf()
207 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf()
208 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf()
217 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf()
[all …]
Dup2x25-minmax-fp32-scalar-fmagic.c37 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic() local
177 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic()
178 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic()
187 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic()
188 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic()
197 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic()
198 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic()
207 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic()
208 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic()
217 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic()
[all …]
Dup2x25-minmax-fp32-scalar-imagic.c37 const int32_t vkernel_zero_point = params->fp32_scalar_imagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic() local
177 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic()
178 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic()
187 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic()
188 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic()
197 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic()
198 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic()
207 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic()
208 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic()
217 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic()
[all …]
Dup2x25-minmax-fp32-wasm-fmagic.c37 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic() local
177 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic()
178 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic()
187 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic()
188 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic()
197 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic()
198 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic()
207 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic()
208 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic()
217 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic()
[all …]
Dup4x9-minmax-fp32-scalar-fmagic.c36 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic() local
100 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic()
101 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic()
102 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic()
103 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic()
116 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic()
117 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic()
118 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic()
119 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic()
132 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic()
[all …]
Dup4x9-minmax-fp32-scalar-lrintf.c36 const int32_t vkernel_zero_point = params->fp32_scalar_lrintf.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf() local
100 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf()
101 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf()
102 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf()
103 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf()
116 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf()
117 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf()
118 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf()
119 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf()
132 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf()
[all …]
Dup4x9-minmax-fp32-scalar-imagic.c36 const int32_t vkernel_zero_point = params->fp32_scalar_imagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic() local
100 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic()
101 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic()
102 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic()
103 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic()
116 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic()
117 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic()
118 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic()
119 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic()
132 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic()
[all …]
Dup4x9-minmax-fp32-wasm-fmagic.c36 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic() local
100 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic()
101 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic()
102 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic()
103 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic()
116 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic()
117 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic()
118 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic()
119 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic()
132 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic()
[all …]
Dup32x25-minmax-fp32-neonv8-mul16.c33 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->fp32_neonv8.kernel_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
180 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
182 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
184 …const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
186 …const int16x8_t vk0xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
198 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
200 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
202 …const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
204 …const int16x8_t vk1xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
216 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
[all …]
Dup2x9-minmax-fp32-scalar-lrintf.c37 const int32_t vkernel_zero_point = params->fp32_scalar_lrintf.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf() local
97 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf()
98 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf()
107 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf()
108 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf()
117 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf()
118 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf()
127 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf()
128 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf()
137 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf()
[all …]
Dup2x9-minmax-fp32-scalar-fmagic.c37 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic() local
97 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic()
98 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic()
107 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic()
108 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic()
117 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic()
118 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic()
127 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic()
128 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic()
137 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic()
[all …]
Dup2x9-minmax-fp32-wasm-fmagic.c37 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic() local
97 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic()
98 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic()
107 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic()
108 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic()
117 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic()
118 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic()
127 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic()
128 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic()
137 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic()
[all …]
Dup2x9-minmax-fp32-scalar-imagic.c37 const int32_t vkernel_zero_point = params->fp32_scalar_imagic.kernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic() local
97 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic()
98 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic()
107 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic()
108 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic()
117 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic()
118 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic()
127 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic()
128 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic()
137 …t32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8] - vkernel_zero_point; in xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic()
[all …]
Dup24x25-minmax-fp32-neonv8-mul16.c33 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->fp32_neonv8.kernel_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16() local
178 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16()
180 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16()
182 …const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16()
192 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16()
194 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16()
196 …const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16()
206 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16()
208 …const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16()
210 …const int16x8_t vk2xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16()
[all …]
Dup24x25-minmax-rndnu-neon-mul16.c32 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->rndnu_neon.kernel_zero_point); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x25__neon_mul16() local
179 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x25__neon_mul16()
181 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x25__neon_mul16()
183 …const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x25__neon_mul16()
193 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x25__neon_mul16()
195 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x25__neon_mul16()
197 …const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x25__neon_mul16()
207 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x25__neon_mul16()
209 …const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x25__neon_mul16()
211 …const int16x8_t vk2xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x25__neon_mul16()
[all …]
Dup24x25-minmax-fp32-neon-mul16.c32 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->fp32_neon.kernel_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neon_mul16() local
178 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neon_mul16()
180 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neon_mul16()
182 …const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neon_mul16()
192 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neon_mul16()
194 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neon_mul16()
196 …const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neon_mul16()
206 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neon_mul16()
208 …const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neon_mul16()
210 …const int16x8_t vk2xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__neon_mul16()
[all …]
Dup32x25-minmax-fp32-neon-mul16.c32 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->fp32_neon.kernel_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
180 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
182 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
184 …const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
186 …const int16x8_t vk0xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
198 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
200 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
202 …const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
204 …const int16x8_t vk1xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
216 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
[all …]
Dup32x25-minmax-rndnu-neon-mul16.c32 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->rndnu_neon.kernel_zero_point); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16() local
181 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
183 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
185 …const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
187 …const int16x8_t vk0xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
199 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
201 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
203 …const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
205 …const int16x8_t vk1xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
217 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
[all …]
Dup16x25-minmax-fp32-neon-mul16.c32 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->fp32_neon.kernel_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16() local
176 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
178 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
186 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
188 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
196 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
198 …const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
206 …const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
208 …const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
216 …const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
[all …]
Dup16x25-minmax-fp32-neonv8-mul16.c33 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->fp32_neonv8.kernel_zero_point); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16() local
176 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
178 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
186 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
188 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
196 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
198 …const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
206 …const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
208 …const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
216 …const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
[all …]
Dup16x25-minmax-rndnu-neon-mul16.c32 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->rndnu_neon.kernel_zero_point); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16() local
177 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
179 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
187 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
189 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
197 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
199 …const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
207 …const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
209 …const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
217 …const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
[all …]

123