Searched refs:vsum4x2 (Results 1 – 4 of 4) sorted by relevance
/aosp_15_r20/external/XNNPACK/src/bf16-gemm/gen/ |
H A D | 5x4c8-minmax-neonbf16-bfdot.c | 214 const float32x2_t vsum4x2 = vadd_f32(vget_low_f32(vacc4x2), vget_high_f32(vacc4x2)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonbf16_bfdot() local 225 float32x4_t vacc4x0123 = vcombine_f32(vpadd_f32(vsum4x0, vsum4x1), vpadd_f32(vsum4x2, vsum4x3)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonbf16_bfdot()
|
H A D | 5x4c8-minmax-neonbf16-bfmlal.c | 255 const float32x2_t vsum4x2 = vadd_f32(vget_low_f32(vacc4x2), vget_high_f32(vacc4x2)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonbf16_bfmlal() local 266 float32x4_t vacc4x0123 = vcombine_f32(vpadd_f32(vsum4x0, vsum4x1), vpadd_f32(vsum4x2, vsum4x3)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonbf16_bfmlal()
|
H A D | 5x4c8-minmax-neonfma-zip.c | 332 const float32x2_t vsum4x2 = vadd_f32(vget_low_f32(vacc4x2), vget_high_f32(vacc4x2)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_zip() local 343 float32x4_t vacc4x0123 = vcombine_f32(vpadd_f32(vsum4x0, vsum4x1), vpadd_f32(vsum4x2, vsum4x3)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_zip()
|
H A D | 5x4c8-minmax-neonfma-shland.c | 332 const float32x2_t vsum4x2 = vadd_f32(vget_low_f32(vacc4x2), vget_high_f32(vacc4x2)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland() local 343 float32x4_t vacc4x0123 = vcombine_f32(vpadd_f32(vsum4x0, vsum4x1), vpadd_f32(vsum4x2, vsum4x3)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland()
|