/external/XNNPACK/src/bf16-gemm/gen/ |
D | 1x4c8-minmax-neonfma-zip.c | 76 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3)); in xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_zip() local 119 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3)); in xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_zip() local
|
D | 1x4c8-minmax-neonfma-shland.c | 76 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask)); in xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_shland() local 119 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask)); in xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_shland() local
|
D | 2x4c8-minmax-neonfma-shland.c | 93 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask)); in xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_shland() local 153 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask)); in xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_shland() local
|
D | 2x4c8-minmax-neonfma-zip.c | 93 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3)); in xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_zip() local 153 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3)); in xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_zip() local
|
D | 3x4c8-minmax-neonfma-zip.c | 110 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3)); in xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_zip() local 187 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3)); in xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_zip() local
|
D | 3x4c8-minmax-neonfma-shland.c | 110 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask)); in xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_shland() local 187 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask)); in xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_shland() local
|
D | 4x4c8-minmax-neonfma-shland.c | 127 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask)); in xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_shland() local 221 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask)); in xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_shland() local
|
D | 4x4c8-minmax-neonfma-zip.c | 127 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3)); in xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_zip() local 221 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3)); in xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_zip() local
|
D | 5x4c8-minmax-neonfma-zip.c | 144 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_zip() local 255 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_zip() local
|
D | 5x4c8-minmax-neonfma-shland.c | 144 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland() local 255 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask)); in xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland() local
|