/external/XNNPACK/src/f32-prelu/gen/ |
D | wasmsimd-minmax-4x16.c | 109 v128_t vacc2xCDEF = wasm_i32x4_max(vi2xCDEF, vzero); in xnn_f32_prelu_ukernel__wasmsimd_minmax_4x16() local
|
D | neon-4x16.c | 103 float32x4_t vacc2xCDEF = vmulq_f32(vi2xCDEF, vwCDEF); in xnn_f32_prelu_ukernel__neon_4x16() local
|
D | wasmsimd-bitselect-4x16.c | 109 v128_t vacc2xCDEF = wasm_f32x4_mul(vi2xCDEF, vwCDEF); in xnn_f32_prelu_ukernel__wasmsimd_bitselect_4x16() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x16c8-minmax-avx512skx.c | 78 __m512i vacc2xCDEF = vacc0xCDEF; in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx() local
|
D | 4x16c8-minmax-avx512skx.c | 84 __m512i vacc2xCDEF = vacc0xCDEF; in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx() local
|
D | 4x16c4-minmax-neondot.c | 75 int32x4_t vacc2xCDEF = vacc0xCDEF; in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 265 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 350 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF ); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 313 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 398 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF ); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 3x16-minmax-neon-mlal-lane.c | 66 int32x4_t vacc2xCDEF = vacc0xCDEF; in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane() local
|
D | 6x16c4-minmax-neondot.c | 87 int32x4_t vacc2xCDEF = vacc0xCDEF; in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() local
|
D | 4x16-minmax-neon-mlal-lane.c | 72 int32x4_t vacc2xCDEF = vacc0xCDEF; in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 392 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 481 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF ); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 328 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 417 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF ); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x16c8-minmax-avx512skx.c | 75 __m512i vacc2xCDEF = vacc0xCDEF; in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx() local
|
D | 4x16c4-minmax-neondot.c | 70 int32x4_t vacc2xCDEF = vacc0xCDEF; in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() local
|
D | 4x16c8-minmax-avx512skx.c | 79 __m512i vacc2xCDEF = vacc0xCDEF; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx() local
|
D | 6x16c4-minmax-neondot.c | 78 int32x4_t vacc2xCDEF = vacc0xCDEF; in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 283 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 368 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF ); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 331 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 416 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF ); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 3x16-minmax-neon-mlal-lane.c | 65 int32x4_t vacc2xCDEF = vacc0xCDEF; in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane() local
|
D | 8x16c4-minmax-neondot.c | 86 int32x4_t vacc2xCDEF = vacc0xCDEF; in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local
|
D | 4x16-minmax-neon-mlal-lane.c | 69 int32x4_t vacc2xCDEF = vacc0xCDEF; in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 348 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 437 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF ); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 412 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 501 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF ); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 472 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 557 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF ); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|