/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x8c8-minmax-neon-mlal-padal.c | 143 int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 221 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 171 int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 270 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 135 int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 120 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 175 int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 333 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 143 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|
D | 4x8c16-minmax-neon-mlal-padal.c | 162 int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 211 int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 414 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 144 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 167 int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 202 int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 175 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x8c8-minmax-neon-mlal-padal.c | 158 int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 236 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 188 int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 287 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 135 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 150 int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 190 int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 348 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 160 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|
D | 4x8c16-minmax-neon-mlal-padal.c | 179 int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 228 int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 431 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 159 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 182 int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 192 const int16x8_t vprod2x4 = vmull_s8(vb4, va2); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 219 int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|