Wino:
- Tile-size: 6

Blocking:
- tail oc/ic handling

Build:
- arch-specific compiling option

Fusion:
- relu
- sum

OMP:
- kmp_malloc(?) allocate from thread-local heap
- omp_in_final()
- omp_get_num_procs

Modularity (2018-07-05): Done
- trans_input/trans_output: common function for offset computing
- code reorg, break transform kernels into small kernel files:
  elk_conv_wino_4x4_3x3_input.hxx
  ...
- Memory size calculation for different xopt


Plain format (nchw/oihw) support (2018-7-11): Done
- fused reorder
- scatter/gather
- double buffer (drop as no effect)
- enable plain fmt for 0xa0e0/0xa0e1
- enable plain fmt for 0xa073
- fix performance regression in block16: format as template parameter
- format-as-blocked (esp. to avoid false sharing in fused output reorder)


IC/OC != 16x (2018-7-16): Done
- Support DNN first layer (drop as it can not achieve good performance with Winograd)
  IC < 16, OC = 16x, nchw + Oihw16o => nChw16o
- Support blocked format with padded tensor
  IC|OC != 16x, nChw16c + OIhw16i16o => nChw16c
- Support plain format in format-as path
  IC|OC != 16x, nchw + oihw => nchw
- Support plain format in format-is path
  IC|OC != 16x, nchw + oihw => nchw

MD-Array (2018-7-22): Done
- Cross platform/compiler MD-array
- Improve MD-Array performance for ICC

Conv_1x1 (2018-8-1): Done
- Uni-stride: kernel=1x1, stride=1, padding=0
- Stride=2, padding=0
- Blocked format
- plain format, IC/OC != 16x
- TODO: code clean and Perf tuning
- TODO: padding support

GEMM kernel (2018-8-23): Done
- Rewrite gemm kernel with better readability and modularity
- Apply new gemm kernel to conv1x1
- Apply new gemm kernel to Winograd
- Perf tuning for Winograd: xopt-A072

- Refactoring elx_conv_wino_t