/external/XNNPACK/src/bf16-gemm/gen/ |
D | 1x4c8-minmax-neonfma-zip.c | 75 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2)); in xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_zip() local 118 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2)); in xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_zip() local
|
D | 1x4c8-minmax-neonfma-shland.c | 75 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask)); in xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_shland() local 118 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask)); in xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_shland() local
|
D | 2x4c8-minmax-neonfma-shland.c | 92 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask)); in xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_shland() local 152 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask)); in xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_shland() local
|
D | 2x4c8-minmax-neonfma-zip.c | 92 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2)); in xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_zip() local 152 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2)); in xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_zip() local
|
D | 3x4c8-minmax-neonfma-zip.c | 109 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2)); in xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_zip() local 186 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2)); in xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_zip() local
|
D | 3x4c8-minmax-neonfma-shland.c | 109 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask)); in xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_shland() local 186 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask)); in xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_shland() local
|
D | 4x4c8-minmax-neonfma-shland.c | 126 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask)); in xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_shland() local 220 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask)); in xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_shland() local
|
D | 4x4c8-minmax-neonfma-zip.c | 126 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2)); in xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_zip() local 220 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2)); in xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_zip() local
|
D | 5x4c8-minmax-neonfma-zip.c | 143 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_zip() local 254 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_zip() local
|
D | 5x4c8-minmax-neonfma-shland.c | 143 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland() local 254 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland() local
|