/external/XNNPACK/src/qs8-igemm/gen/ |
D | 8x16c4-minmax-neondot.c | 412 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 413 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 414 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 415 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 416 int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 417 int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 418 int8x16_t vout6x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc6x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 419 int8x16_t vout7x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc7x01234567), vacc7x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 438 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 439 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() [all …]
|
D | 6x16c4-minmax-neondot.c | 332 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 333 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 334 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 335 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 336 int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 337 int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 352 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 353 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 354 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 355 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() [all …]
|
D | 4x16c4-minmax-neondot.c | 252 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 253 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 254 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 255 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 266 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 267 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 268 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 269 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
|
D | 8x8c4-minmax-neondot.c | 286 … int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() 287 … int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() 288 … int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() 289 … int8x16_t vout6x01234567_7x01234567 = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc7x01234567); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() 300 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x012… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() 301 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x012… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() 302 …int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x012… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() 303 …int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc7x012… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
|
D | 6x8c4-minmax-neondot.c | 236 … int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567); in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot() 237 … int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567); in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot() 238 … int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567); in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot() 247 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x012… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot() 248 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x012… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot() 249 …int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x012… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
|
D | 4x8c4-minmax-neondot.c | 186 … int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567); in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot() 187 … int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567); in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot() 194 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x012… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot() 195 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x012… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 8x16c4-minmax-neondot.c | 388 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 389 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 390 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 391 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 392 int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 393 int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 394 int8x16_t vout6x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc6x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 395 int8x16_t vout7x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc7x01234567), vacc7x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 414 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 415 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() [all …]
|
D | 6x16c4-minmax-neondot.c | 312 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 313 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 314 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 315 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 316 int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 317 int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 332 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 333 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 334 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 335 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() [all …]
|
D | 4x16c4-minmax-neondot.c | 236 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 237 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 238 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 239 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 250 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 251 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 252 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 253 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
|
D | 8x8c4-minmax-neondot.c | 262 … int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567); in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() 263 … int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567); in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() 264 … int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567); in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() 265 … int8x16_t vout6x01234567_7x01234567 = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc7x01234567); in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() 276 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x012… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() 277 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x012… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() 278 …int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x012… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() 279 …int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc7x012… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
|
D | 6x8c4-minmax-neondot.c | 216 … int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567); in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot() 217 … int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567); in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot() 218 … int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567); in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot() 227 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x012… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot() 228 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x012… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot() 229 …int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x012… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
|
D | 4x8c4-minmax-neondot.c | 170 … int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567); in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot() 171 … int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567); in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot() 178 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x012… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot() 179 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x012… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
|
/external/XNNPACK/src/qs8-vaddc/gen/ |
D | minmax-neon-ld64-x32.c | 82 … int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32() 83 … int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32() 111 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32()
|
D | minmax-neon-ld64-x24.c | 73 … int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x24() 74 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x24() 102 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x24()
|
D | minmax-neon-ld64-x16.c | 64 … int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16() 89 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
|
/external/XNNPACK/src/qs8-requantization/ |
D | q31-neon.c | 85 const int8x16_t xyzw_packed = vqmovn_high_s16(vqmovn_s16(xy_packed), zw_packed); in xnn_qs8_requantize_q31__neon() 89 const int8x16_t xyzw_packed = vcombine_s8(vqmovn_s16(xy_packed), vqmovn_s16(zw_packed)); in xnn_qs8_requantize_q31__neon()
|
/external/XNNPACK/src/qs8-vadd/gen/ |
D | minmax-neon-ld64-x32.c | 94 … int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32() 95 … int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32() 128 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32()
|
D | minmax-neon-ld64-x24.c | 81 … int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24() 82 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24() 115 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24()
|
D | minmax-neon-ld64-x16.c | 68 … int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16() 98 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16()
|
/external/XNNPACK/src/qs8-dwconv/ |
D | unipass-neon-mul16.c.in | 82 …int8x16_t vout${ABC[C:C+16]} = vqmovn_high_s16(vqmovn_s16(vacc${ABC[C:C+8]}), vacc${ABC[C+8:C+16]}… 84 int8x8_t vout${ABC[C:C+8]} = vqmovn_s16(vacc${ABC[C:C+8]}); 91 …int8x16_t vout${ABC[C:C+16]} = vcombine_s8(vqmovn_s16(vacc${ABC[C:C+8]}), vqmovn_s16(vacc${ABC[C+8… 93 int8x8_t vout${ABC[C:C+8]} = vqmovn_s16(vacc${ABC[C:C+8]}); 158 int8x8_t vout${ABC[0:8]} = vqmovn_s16(vacc${ABC[0:8]}); 162 int8x8_t vout${ABC[0:8]} = vqmovn_s16(vacc${ABC[0:8]});
|
/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7x-minmax-neon-c24-acc2.c | 204 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 205 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 218 … int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 219 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 292 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 299 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2()
|
D | 7x-minmax-neon-c32-acc2.c | 244 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 245 int8x16_t voutGHIJKLMNOPQRSTUV = vqmovn_high_s16(vqmovn_s16(vaccGHIJKLMN), vaccOPQRSTUV); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 261 … int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 262 … int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 335 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 342 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2()
|
/external/tensorflow/tensorflow/lite/kernels/internal/optimized/integer_ops/ |
D | mul.h | 117 const int8x16_t p = vcombine_s8(vqmovn_s16(p_part2), vqmovn_s16(p_part1)); in MulElementwise() 210 const int8x16_t p = vcombine_s8(vqmovn_s16(p_part2), vqmovn_s16(p_part1)); in MulSimpleBroadcast()
|
/external/XNNPACK/src/qs8-vaddc/ |
D | neon-ld64.c.in | 60 …int8x16_t vout${ABC[N:N+16]} = vcombine_s8(vqmovn_s16(vacc${ABC[N:N+8]}), vqmovn_s16(vacc${ABC[N+8… 62 int8x8_t vout${ABC[N:N+8]} = vqmovn_s16(vacc${ABC[N:N+8]}); 102 int8x8_t vout${ABC[0:8]} = vqmovn_s16(vacc${ABC[0:8]});
|
/external/XNNPACK/src/qs8-igemm/ |
D | MRxNRc4-neondot.c.in | 138 …int8x16_t vout${M}x${ABC[N:N+16]} = vqmovn_high_s16(vqmovn_s16(vacc${M}x${ABC[N:N+8]}), vacc${M}x$… 140 …int8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vqmovn_high_s16(vqmovn_s16(vacc${M-1}x${AB… 142 int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]}); 151 …int8x16_t vout${M}x${ABC[N:N+16]} = vcombine_s8(vqmovn_s16(vacc${M}x${ABC[N:N+8]}), vqmovn_s16(vac… 153 …M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vcombine_s8(vqmovn_s16(vacc${M-1}x${ABC[N:N+8]}), vqmovn_s… 155 int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
|