/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x8c8-minmax-neon-mlal-padal.c | 133 int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 214 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 3x4c8-xw-minmax-wasmsimd.c | 121 const v128_t vprod2x3 = wasm_i16x8_mul(vxa2, vxb3); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__wasmsimd() local
|
D | 3x4c8-minmax-wasmsimd-ld128.c | 121 const v128_t vprod2x3 = wasm_i16x8_mul(vxb3, vxa2); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld128() local
|
D | 3x4c8-minmax-wasmsimd-ld64.c | 121 const v128_t vprod2x3 = wasm_i16x8_mul(vxa2, vxb3); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 158 int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 261 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 126 int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 113 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 165 int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 326 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 134 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|
D | 4x8c16-minmax-neon-mlal-padal.c | 150 int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 198 int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 405 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 137 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 158 int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x8c8-minmax-neon-mlal-padal.c | 148 int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 229 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 3x4c8-minmax-wasmsimd-ld64.c | 136 const v128_t vprod2x3 = wasm_i16x8_mul(vxa2, vxb3); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64() local
|
D | 3x4c8-minmax-wasmsimd-ld128.c | 136 const v128_t vprod2x3 = wasm_i16x8_mul(vxb3, vxa2); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld128() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 175 int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 278 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 128 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 141 int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 180 int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 341 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 151 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|
D | 4x8c16-minmax-neon-mlal-padal.c | 167 int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 215 int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 422 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 152 const int16x8_t vprod2x3 = vmull_s8(vb3, va2); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 173 int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|