/external/XNNPACK/src/f32-gemm/gen/ |
D | 4x8-neon-dup-ld64.c | 96 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64() local 100 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64() 104 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64()
|
D | 4x8-neonfma-dup-ld64.c | 96 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64() local 100 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64() 104 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64()
|
D | 4x8-psimd-splat.c | 98 const psimd_f32 va1c1 = psimd_splat1_f32(va1); in xnn_f32_gemm_ukernel_4x8__psimd_splat() local 106 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_ukernel_4x8__psimd_splat() 110 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemm_ukernel_4x8__psimd_splat()
|
D | 4x8-neonfma-dup-ld128.c | 98 const float32x4_t va1c1 = vdupq_lane_f32(vget_low_f32(va1), 1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld128() local 102 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld128() 106 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld128()
|
D | 4x8-neon-dup-ld128.c | 98 const float32x4_t va1c1 = vdupq_lane_f32(vget_low_f32(va1), 1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld128() local 102 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld128() 106 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld128()
|
D | 4x2-neonfma-lane-ld64.c | 99 const float32x2_t va1c1 = vdup_lane_f32(va1, 1); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64() local 103 vacc1x01 = vfma_f32(vacc1x01, va1c1, vb01c1); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
|
D | 6x8-neonfma-dup-ld64.c | 120 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld64() local 126 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld64() 132 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld64()
|
D | 6x8-neon-dup-ld64.c | 120 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld64() local 126 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld64() 132 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld64()
|
D | 6x8-psimd-splat.c | 124 const psimd_f32 va1c1 = psimd_splat1_f32(va1); in xnn_f32_gemm_ukernel_6x8__psimd_splat() local 134 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_ukernel_6x8__psimd_splat() 140 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemm_ukernel_6x8__psimd_splat()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 4x8-neon-dup-ld64.c | 118 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64() local 122 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64() 126 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64()
|
D | 4x8-neonfma-dup-ld64.c | 118 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64() local 122 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64() 126 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64()
|
D | 4x8-psimd-splat.c | 120 const psimd_f32 va1c1 = psimd_splat1_f32(va1); in xnn_f32_igemm_ukernel_4x8__psimd_splat() local 128 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_ukernel_4x8__psimd_splat() 132 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_igemm_ukernel_4x8__psimd_splat()
|
D | 4x8-neonfma-dup-ld128.c | 120 const float32x4_t va1c1 = vdupq_lane_f32(vget_low_f32(va1), 1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld128() local 124 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld128() 128 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld128()
|
D | 4x8-neon-dup-ld128.c | 120 const float32x4_t va1c1 = vdupq_lane_f32(vget_low_f32(va1), 1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld128() local 124 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld128() 128 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld128()
|
D | 4x2-neonfma-lane-ld64.c | 121 const float32x2_t va1c1 = vdup_lane_f32(va1, 1); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64() local 125 vacc1x01 = vfma_f32(vacc1x01, va1c1, vb01c1); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
|
D | 6x8-neonfma-dup-ld64.c | 148 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld64() local 154 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld64() 160 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld64()
|
D | 6x8-neon-dup-ld64.c | 148 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld64() local 154 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld64() 160 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld64()
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 4x8-neon-dup-ld64.c | 98 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64() local 102 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64() 106 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64()
|
D | 4x8-neonfma-dup-ld64.c | 98 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64() local 102 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64() 106 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64()
|
D | 4x8-neonfma-dup-ld128.c | 100 const float32x4_t va1c1 = vdupq_lane_f32(vget_low_f32(va1), 1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128() local 104 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128() 108 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128()
|
D | 4x8-psimd-splat.c | 100 const psimd_f32 va1c1 = psimd_splat1_f32(va1); in xnn_f32_gemminc_ukernel_4x8__psimd_splat() local 108 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_ukernel_4x8__psimd_splat() 112 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemminc_ukernel_4x8__psimd_splat()
|
D | 4x8-neon-dup-ld128.c | 100 const float32x4_t va1c1 = vdupq_lane_f32(vget_low_f32(va1), 1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128() local 104 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128() 108 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128()
|
D | 6x8-neon-dup-ld64.c | 122 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64() local 128 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64() 134 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64()
|
D | 6x8-neonfma-dup-ld64.c | 122 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64() local 128 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64() 134 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64()
|
D | 6x8-psimd-splat.c | 126 const psimd_f32 va1c1 = psimd_splat1_f32(va1); in xnn_f32_gemminc_ukernel_6x8__psimd_splat() local 136 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_ukernel_6x8__psimd_splat() 142 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c1, vb4567c1); in xnn_f32_gemminc_ukernel_6x8__psimd_splat()
|