/external/XNNPACK/src/f32-gemm/gen/ |
D | 4x8-neon-dup-ld64.c | 82 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64() local 86 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64() 90 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64()
|
D | 4x8-neonfma-dup-ld64.c | 82 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64() local 86 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64() 90 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64()
|
D | 4x8-psimd-splat.c | 83 const psimd_f32 va2c0 = psimd_splat0_f32(va2); in xnn_f32_gemm_ukernel_4x8__psimd_splat() local 91 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_ukernel_4x8__psimd_splat() 95 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_ukernel_4x8__psimd_splat()
|
D | 4x8-neonfma-dup-ld128.c | 83 const float32x4_t va2c0 = vdupq_lane_f32(vget_low_f32(va2), 0); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld128() local 87 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld128() 91 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld128()
|
D | 4x8-neon-dup-ld128.c | 83 const float32x4_t va2c0 = vdupq_lane_f32(vget_low_f32(va2), 0); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld128() local 87 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld128() 91 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld128()
|
D | 4x2-neonfma-lane-ld64.c | 83 const float32x2_t va2c0 = vdup_lane_f32(va2, 0); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64() local 87 vacc2x01 = vfma_f32(vacc2x01, va2c0, vb01c0); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
|
D | 6x8-neonfma-dup-ld64.c | 100 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld64() local 106 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld64() 112 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld64()
|
D | 6x8-neon-dup-ld64.c | 100 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld64() local 106 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld64() 112 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld64()
|
D | 6x8-psimd-splat.c | 103 const psimd_f32 va2c0 = psimd_splat0_f32(va2); in xnn_f32_gemm_ukernel_6x8__psimd_splat() local 113 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_ukernel_6x8__psimd_splat() 119 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_ukernel_6x8__psimd_splat()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 4x8-neon-dup-ld64.c | 104 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64() local 108 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64() 112 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64()
|
D | 4x8-neonfma-dup-ld64.c | 104 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64() local 108 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64() 112 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64()
|
D | 4x8-psimd-splat.c | 105 const psimd_f32 va2c0 = psimd_splat0_f32(va2); in xnn_f32_igemm_ukernel_4x8__psimd_splat() local 113 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_ukernel_4x8__psimd_splat() 117 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_ukernel_4x8__psimd_splat()
|
D | 4x8-neonfma-dup-ld128.c | 105 const float32x4_t va2c0 = vdupq_lane_f32(vget_low_f32(va2), 0); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld128() local 109 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld128() 113 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld128()
|
D | 4x8-neon-dup-ld128.c | 105 const float32x4_t va2c0 = vdupq_lane_f32(vget_low_f32(va2), 0); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld128() local 109 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld128() 113 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld128()
|
D | 4x2-neonfma-lane-ld64.c | 105 const float32x2_t va2c0 = vdup_lane_f32(va2, 0); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64() local 109 vacc2x01 = vfma_f32(vacc2x01, va2c0, vb01c0); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
|
D | 6x8-neonfma-dup-ld64.c | 128 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld64() local 134 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld64() 140 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld64()
|
D | 6x8-neon-dup-ld64.c | 128 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld64() local 134 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld64() 140 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld64()
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 4x8-neon-dup-ld64.c | 84 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64() local 88 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64() 92 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64()
|
D | 4x8-neonfma-dup-ld64.c | 84 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64() local 88 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64() 92 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64()
|
D | 4x8-neonfma-dup-ld128.c | 85 const float32x4_t va2c0 = vdupq_lane_f32(vget_low_f32(va2), 0); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128() local 89 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128() 93 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128()
|
D | 4x8-psimd-splat.c | 85 const psimd_f32 va2c0 = psimd_splat0_f32(va2); in xnn_f32_gemminc_ukernel_4x8__psimd_splat() local 93 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_ukernel_4x8__psimd_splat() 97 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_ukernel_4x8__psimd_splat()
|
D | 4x8-neon-dup-ld128.c | 85 const float32x4_t va2c0 = vdupq_lane_f32(vget_low_f32(va2), 0); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128() local 89 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128() 93 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128()
|
D | 6x8-neon-dup-ld64.c | 102 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64() local 108 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64() 114 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64()
|
D | 6x8-neonfma-dup-ld64.c | 102 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64() local 108 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64() 114 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64()
|
D | 6x8-psimd-splat.c | 105 const psimd_f32 va2c0 = psimd_splat0_f32(va2); in xnn_f32_gemminc_ukernel_6x8__psimd_splat() local 115 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_ukernel_6x8__psimd_splat() 121 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_ukernel_6x8__psimd_splat()
|