* [dGPU] Enable user scratchpad mode.
* Reuse intermediate buffer.
* Add own id to the memory dependencies at the c-tor of program_node
+ Allocate intermediate memory with memory_pool::get_memory() function.
+ Assign scratchpad memory desc in load() function for onednn primitive
serialization
* Allocate device mem for onednn scratchpad mem
* [dGPU] Enable stable diffusion
+ Prevent to fuse swish into oneDNN reorder.
+ Makes concat explicitly if batch size is greater than 1 and the siblings are oneDNN impl.
* [GPU] Fix a bug of permute optimization
For int8 models, if there is FakeQuantize between permute and convolution, an operation like data type casting could be fused to permute. In this case, do not optimize permute.
* [GPU] Optimize permute for acdb format
Target subgraphs to be optimized-out
- input(bfyx) - permute(byxf) - conv
- conv(byxf) - permute(bfyx) - output
+ Fix test_device_mem_usage_estimation unit test failed.
* [GPU] Support implicit crop in input transposition.
+ Make the crop in front of quantize implicit by changing output format to bfyx.
+ Use implicit concat after quantize nodes.
* Add unit test for implicit crop and concat.
+ remove unnecessary code.
+ Fix colorization-sig accuracy issue using oneDNN
Memory crash in case reuse_eltwise_sum_post in oneDNN and memory_pool
And print node in/out gpu_usm_mem addr at OV_GPU_Verbose >= 1
+ Check the size of z spatial axis for checking fulltensor.
+ Remove program_helpers's functions.
Co-authored-by: hyunback <hyunback.kim@intel.com>
+ cldnn supports hard sigmoid activation function but onednn doesn't.
+ split it into eltwise linear and eltwise clip in
add_onednn_optimization_attributes pass.
* Fix some bugs of cldnn eltwise kernel at bs_fs_yx_bsv32_fsv16 format.
+ Add a condition so that the eltwise_simple_vload8 kernel is not selected when the tensor is not aligned as bsv32_fsv16 or bsv32_fsv32.
+ Optimize gws/lws of eltwise_ref kernel for bsv32_fsv16 format.
* Check feature align of b_fs_yx_fsv32 for eltwise vload8 kernel
+ Minor fix for OV_GPU_Help option
* Add DumpLayersLimitBatch config to GPU debug utils.
+ Support OV_GPU_Help config
+ Only run first inference if OV_GPU_DumpLayersPath is set.
+ Fix dump graph bug.
* Apply some comments
* Remove unnecessary code.
+ Prevent to fuse PRelu if convolution is onednn impl.
+ Prevent to remove input reorder if onednn conv converts format from bfyx to b_fs_yx_fsv32.
+ Add a value to optimal lws values for eltwise kernel selector.
+ Fix a condition of remove redundant reorder. (try_fuse_reorder_bfyx_to_fsv32)
* [GPU] Merge mvn kernel updates
+ Support bs_fs_yx_bsv32_fsv32 format for i8 and u8 dtype.
+ Support bs_fs_yx_bfv32_fsv16 format for fp16 dtype.
* [GPU] Fix a bug of resample kernel selector.
* [GPU] fix a build error of mvn unit tests
Jade Cho
