tensorflow/tfjs tfjs-backend-webgpu-v0.0.1-alpha.8

on GitHub

tfjs-backend-webgpu (0.0.1-alpha.7 ==> 0.0.1-alpha.8)

Features

• [webgpu] Add support for scatterND (#5643). Thanks, @haoyunfeix.
• [webgpu] Add support for LogicalNot (#5644). Thanks, @haoyunfeix.
• [webgpu] Add support for split (#5645). Thanks, @haoyunfeix.
• [webgpu] Add support for sin, cos, sinh and cosh (#5642). Thanks, @haoyunfeix.
• [webgpu] Add depthToSpace op (#5585). Thanks, @haoyunfeix.
• [webgpu] Add halfPixelCenters support for resizeBilinear op (#5553). Thanks, @haoyunfeix.
• [webgpu] Enable bool and add op logicalAnd (#5190). Thanks, @axinging.

Bug fixes

• [webgpu] Fix wrongly used u32 (#5559). Thanks, @axinging.
• webgpu: Fix the memory leak for Cast (#5521). Thanks, @qjia7.
• Create a separate 'link-package-core' for packages that only need core (#5436).
• : [[webgpu] Add op tile] (#5186). Thanks, @axinging.

Performance

• webgpu: Optimize ScatterNd (#5667). Thanks, @qjia7.
• webgpu: Optimize depthwise conv2d (#5254). Thanks, @qjia7.
• webgpu: Optimize depthwise conv2d (#5209). Thanks, @qjia7.

Misc

• Update tfjs-backend-webgpu to 0.0.1-alpha.8. (#5681).
• [webgpu] Mark build-npm.sh as executable (#5676).
• [wasm] Fix kernels in wasm backend about how int32 inputs are handled (#5662).
• webgpu: Use a larger cpu forward threshold (#5665). Thanks, @qjia7.
• Use the release docker for all cloudbuild steps (#5669).
• [wgsl] Refine pow (#5622). Thanks, @axinging.
• webgpu: Batch several dispatching calls into one pass (#5598). Thanks, @qjia7.
• [webgpu] Fix matmul_small_output program assert error (#5646). Thanks, @haoyunfeix.
• webgpu: Remove glsl code (#5654). Thanks, @qjia7.
• [webgpu] Add support for SparseToDense (#5629). Thanks, @haoyunfeix.
• Add cases for Conv2DMMVec4Program and Conv2dNaiveProgram (#5637). Thanks, @axinging.
• webgpu: add flipLeftRight and rotate operators (#5649). Thanks, @xhcao.
• [webgpu] Add WGSL shader for Tanh (#5647). Thanks, @haoyunfeix.
• [webgpu] Use WGSL default (#5621). Thanks, @axinging.
• [webgpu] Use i32 as coords type in WGSL (#5628). Thanks, @axinging.
• [webgpu] Create FromPixles program as required (#5610). Thanks, @axinging.
• [webgpu] Only stride is used in PoolWithFilterSizeEqualsOneProgram (#5593). Thanks, @axinging.
• [webgpu] Fix WGSL pow case fail (#5589). Thanks, @axinging.
• [webgpu] Add MatMulSmallOutputSizeProgram WGSL support (#5567). Thanks, @axinging.
• [webgpu] Add conv2dBackprop WGSL support (#5565). Thanks, @axinging.
• [webgpu] Use module constant as workgroup size in WGSL (#5579). Thanks, @axinging.
• webgpu: Hide importExternalTexture behind flag (#5587). Thanks, @shaoboyan.
• Throw errors when any index is out of bound in GatherV2 (#5439).
• [webgpu] Add conv2d/conv2dNaive/matmul/im2col WGSL support (#5527). Thanks, @axinging.
• webgpu: Add wgsl support for FromPixels (#5552). Thanks, @shaoboyan.
• [webgpu] Add depthwise3x3/argminmax/reduce/resize_nearest_neighbor WGSL support (#5535). Thanks, @axinging.
• webgpu: optimize matmul with small output size (#5342). Thanks, @xhcao.
• [webgpu] Add time query profiling for fromPixels op (#5503). Thanks, @haoyunfeix.
• [webgpu] Fix exceed dispatch size on WGSL (#5519). Thanks, @axinging.
• [webgpu] update warning message to guide user enable timestamp-query extension (#5505). Thanks, @haoyunfeix.
• [webgpu] Fix warning that buffer binding size cannot be zero. (#5526). Thanks, @qjia7.
• [webgpu] Fix clipByValue vec4 error (#5511). Thanks, @axinging.
• webgpu: fix dispatch size exceeds the limits (#5432). Thanks, @xhcao.
• [webgpu] Add depthwise and other 21 programs WGSL support (#5403). Thanks, @axinging.
• [webgpu] Add conv2d vec4 and matmul vec4 WGSL support (#5369). Thanks, @axinging.
• Fix FromPixels type checker to avoid error in worker (#5472). Thanks, @shaoboyan.
• [webgpu] Fix wrongly modified NaN type (#5462). Thanks, @axinging.
• [webgpu] Use ImportToTexture to handle video in FromPixels (#5430). Thanks, @shaoboyan.
• [webgpu] Add binary WGSL support (#5322). Thanks, @axinging.
• [webgpu] Remove tfjs-backend-webgl and tfjs-converter dependency (#5428). Thanks, @haoyunfeix.
• [webgpu] Fix that exceeding limitation errors (#5385). Thanks, @qjia7.
• Build tfjs-backend-webgl with Bazel (#5339).
• [webgpu] Fix PadV2 with empty size (#5363). Thanks, @qjia7.
• Use Bazel and ts_library to build core and backend-cpu (#5133).
• Add 'WEBGPU_USE_PROFILE_TOOL' flag (#5300). Thanks, @shaoboyan.
• [webgpu] Fix nan fail dut to tint (#5321). Thanks, @axinging.
• [webgpu] Fix floorDiv vec4 compile fail (#5307). Thanks, @axinging.
• [webgpu] Update @webgpu/types to version 0.1.4 (#5290). Thanks, @haoyunfeix.
• [webgpu] Add WGSL support (#5252). Thanks, @axinging.
• [webgpu] Add case for matmul and conv2d to test special path (#5253). Thanks, @axinging.
• [webgpu] Check if runtime support WebGPU before initial a WebGPU backend (#5218). Thanks, @haoyunfeix.
• [webgpu] Upgrade clang-format version from 1.2.4 to 1.5.0 (#5225). Thanks, @haoyunfeix.
• [webgpu] Simplify shader key of binary op and unary op (#5222). Thanks, @axinging.
• [webgpu] Simplify activation shader key (#5208). Thanks, @axinging.
• Use CopyExternalImageToTexture() in FromPixels (#5202). Thanks, @shaoboyan.