| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 6x8-psimd-splat.c | 171 const psimd_f32 va4c3 = psimd_splat3_f32(va4); in xnn_f32_gemm_ukernel_6x8__psimd_splat() local
181 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_gemm_ukernel_6x8__psimd_splat()
187 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_gemm_ukernel_6x8__psimd_splat()
|
| D | 6x8-neonfma-dup-ld128.c | 169 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128() local
175 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
181 vacc4x4567 = vfmaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
|
| D | 6x8-neon-dup-ld128.c | 169 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld128() local
175 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld128()
181 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld128()
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 6x8-psimd-splat.c | 173 const psimd_f32 va4c3 = psimd_splat3_f32(va4); in xnn_f32_gemminc_ukernel_6x8__psimd_splat() local
183 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_gemminc_ukernel_6x8__psimd_splat()
189 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_gemminc_ukernel_6x8__psimd_splat()
|
| D | 6x8-neon-dup-ld128.c | 171 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld128() local
177 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld128()
183 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld128()
|
| D | 6x8-neonfma-dup-ld128.c | 171 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld128() local
177 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld128()
183 vacc4x4567 = vfmaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld128()
|
| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 6x8-neonfma-dup-ld128.c | 197 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128() local
203 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
209 vacc4x4567 = vfmaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
|
| D | 6x8-psimd-splat.c | 199 const psimd_f32 va4c3 = psimd_splat3_f32(va4); in xnn_f32_igemm_ukernel_6x8__psimd_splat() local
209 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_igemm_ukernel_6x8__psimd_splat()
215 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_igemm_ukernel_6x8__psimd_splat()
|
| D | 6x8-neon-dup-ld128.c | 197 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld128() local
203 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld128()
209 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld128()
|
| /external/XNNPACK/src/f16-gemm/gen/ |
| D | 6x8-neonfp16arith-ld64.c | 175 const float16x8_t va4c3 = vdupq_lane_f16(va4, 3); in xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64() local
182 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3); in xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64()
|
| D | 8x8-neonfp16arith-ld64.c | 211 const float16x8_t va4c3 = vdupq_lane_f16(va4, 3); in xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64() local
220 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3); in xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64()
|