| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 4x8-neon-dup-ld64.c | 97 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64() local
101 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64()
105 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64()
|
| D | 4x8-neonfma-dup-ld64.c | 97 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64() local
101 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64()
105 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64()
|
| D | 4x8-psimd-splat.c | 99 const psimd_f32 va2c1 = psimd_splat1_f32(va2); in xnn_f32_gemm_ukernel_4x8__psimd_splat() local
107 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_ukernel_4x8__psimd_splat()
111 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemm_ukernel_4x8__psimd_splat()
|
| D | 4x8-neonfma-dup-ld128.c | 99 const float32x4_t va2c1 = vdupq_lane_f32(vget_low_f32(va2), 1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld128() local
103 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld128()
107 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld128()
|
| D | 4x8-neon-dup-ld128.c | 99 const float32x4_t va2c1 = vdupq_lane_f32(vget_low_f32(va2), 1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld128() local
103 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld128()
107 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld128()
|
| D | 4x2-neonfma-lane-ld64.c | 100 const float32x2_t va2c1 = vdup_lane_f32(va2, 1); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64() local
104 vacc2x01 = vfma_f32(vacc2x01, va2c1, vb01c1); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
|
| D | 6x8-neonfma-dup-ld64.c | 121 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld64() local
127 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld64()
133 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld64()
|
| D | 6x8-neon-dup-ld64.c | 121 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld64() local
127 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld64()
133 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemm_ukernel_6x8__neon_dup_ld64()
|
| D | 6x8-psimd-splat.c | 125 const psimd_f32 va2c1 = psimd_splat1_f32(va2); in xnn_f32_gemm_ukernel_6x8__psimd_splat() local
135 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_ukernel_6x8__psimd_splat()
141 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemm_ukernel_6x8__psimd_splat()
|
| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 4x8-neon-dup-ld64.c | 119 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64() local
123 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64()
127 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64()
|
| D | 4x8-neonfma-dup-ld64.c | 119 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64() local
123 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64()
127 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64()
|
| D | 4x8-psimd-splat.c | 121 const psimd_f32 va2c1 = psimd_splat1_f32(va2); in xnn_f32_igemm_ukernel_4x8__psimd_splat() local
129 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_ukernel_4x8__psimd_splat()
133 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_igemm_ukernel_4x8__psimd_splat()
|
| D | 4x8-neonfma-dup-ld128.c | 121 const float32x4_t va2c1 = vdupq_lane_f32(vget_low_f32(va2), 1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld128() local
125 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld128()
129 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld128()
|
| D | 4x8-neon-dup-ld128.c | 121 const float32x4_t va2c1 = vdupq_lane_f32(vget_low_f32(va2), 1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld128() local
125 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld128()
129 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld128()
|
| D | 4x2-neonfma-lane-ld64.c | 122 const float32x2_t va2c1 = vdup_lane_f32(va2, 1); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64() local
126 vacc2x01 = vfma_f32(vacc2x01, va2c1, vb01c1); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
|
| D | 6x8-neonfma-dup-ld64.c | 149 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld64() local
155 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld64()
161 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld64()
|
| D | 6x8-neon-dup-ld64.c | 149 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld64() local
155 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld64()
161 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_igemm_ukernel_6x8__neon_dup_ld64()
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 4x8-neon-dup-ld64.c | 99 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64() local
103 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64()
107 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64()
|
| D | 4x8-neonfma-dup-ld64.c | 99 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64() local
103 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64()
107 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64()
|
| D | 4x8-neonfma-dup-ld128.c | 101 const float32x4_t va2c1 = vdupq_lane_f32(vget_low_f32(va2), 1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128() local
105 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128()
109 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128()
|
| D | 4x8-psimd-splat.c | 101 const psimd_f32 va2c1 = psimd_splat1_f32(va2); in xnn_f32_gemminc_ukernel_4x8__psimd_splat() local
109 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_ukernel_4x8__psimd_splat()
113 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemminc_ukernel_4x8__psimd_splat()
|
| D | 4x8-neon-dup-ld128.c | 101 const float32x4_t va2c1 = vdupq_lane_f32(vget_low_f32(va2), 1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128() local
105 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128()
109 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128()
|
| D | 6x8-neon-dup-ld64.c | 123 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64() local
129 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64()
135 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64()
|
| D | 6x8-neonfma-dup-ld64.c | 123 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64() local
129 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64()
135 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64()
|
| D | 6x8-psimd-splat.c | 127 const psimd_f32 va2c1 = psimd_splat1_f32(va2); in xnn_f32_gemminc_ukernel_6x8__psimd_splat() local
137 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_ukernel_6x8__psimd_splat()
143 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c1, vb4567c1); in xnn_f32_gemminc_ukernel_6x8__psimd_splat()
|