/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x8c8-minmax-neon-mlal-padal.c | 85 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 121 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
|
D | 2x8c8-minmax-neon-mlal-padal.c | 113 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 170 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mlal-padal.c | 101 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 169 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
|
D | 1x8c8-minmax-neon-mull-padal.c | 72 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local
|
D | 1x8c16-minmax-neon-mlal-padal.c | 79 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mlal-padal.c | 141 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 219 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 2x8c8-minmax-neon-mull-padal.c | 95 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
|
D | 2x8c16-minmax-neon-mlal-padal.c | 106 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 169 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 268 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 95 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mull-padal.c | 80 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
|
D | 2x16c8-minmax-neon-mlal-padal.c | 137 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 250 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 133 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x8c8-minmax-neon-mlal-padal.c | 96 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 132 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
|
D | 2x8c8-minmax-neon-mlal-padal.c | 126 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 183 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mlal-padal.c | 112 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 180 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
|
D | 1x8c8-minmax-neon-mull-padal.c | 83 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local
|
D | 1x8c16-minmax-neon-mlal-padal.c | 90 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mlal-padal.c | 156 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 234 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 186 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 285 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mull-padal.c | 91 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 106 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
|
D | 2x8c16-minmax-neon-mlal-padal.c | 119 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
|
D | 2x8c8-minmax-neon-mull-padal.c | 108 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
|
D | 2x16c8-minmax-neon-mlal-padal.c | 150 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 263 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
|