/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x16c8-minmax-neon-mlal-padal.c | 245 int16x8_t vprod2x11 = vmull_s8(vb11x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 248 vprod2x11 = vmlal_s8(vprod2x11, vb11x1, va2x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() 251 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() 382 const int16x8_t vprod2x11 = vmull_s8(vb11, va2); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 385 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 230 int16x8_t vprod2x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 233 vprod2x11 = vmlal_s8(vprod2x11, vget_high_s8(vb11), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() 236 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 302 int16x8_t vprod2x11 = vmull_s8(vb11x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 306 vprod2x11 = vmlal_s8(vprod2x11, vb11x1, va2x1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() 310 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() 477 const int16x8_t vprod2x11 = vmull_s8(vb11, va2); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 481 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 286 int16x8_t vprod2x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 290 vprod2x11 = vmlal_s8(vprod2x11, vget_high_s8(vb11), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() 294 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 193 const int16x8_t vprod2x11 = vmull_s8(vb11, va2); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 196 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 238 const int16x8_t vprod2x11 = vmull_s8(vb11, va2); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 242 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x16c8-minmax-neon-mlal-padal.c | 260 int16x8_t vprod2x11 = vmull_s8(vb11x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 263 vprod2x11 = vmlal_s8(vprod2x11, vb11x1, va2x1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() 266 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() 397 const int16x8_t vprod2x11 = vmull_s8(vb11, va2); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 400 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 245 int16x8_t vprod2x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 248 vprod2x11 = vmlal_s8(vprod2x11, vget_high_s8(vb11), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() 251 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 319 int16x8_t vprod2x11 = vmull_s8(vb11x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 323 vprod2x11 = vmlal_s8(vprod2x11, vb11x1, va2x1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() 327 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() 494 const int16x8_t vprod2x11 = vmull_s8(vb11, va2); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 498 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 303 int16x8_t vprod2x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 307 vprod2x11 = vmlal_s8(vprod2x11, vget_high_s8(vb11), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() 311 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 208 const int16x8_t vprod2x11 = vmull_s8(vb11, va2); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 211 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 255 const int16x8_t vprod2x11 = vmull_s8(vb11, va2); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 259 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|