igc-1.0.3899

Fixed Issues / Improvements

• Fix for setting Src1.Length in IGA IR.
• Small refactor to render taget write code generation.
• Avoid splitting RetVal variable into two.
• Not limiting the loop unroll to small loops for high register pressure shaders if constant folding
• Handling the math intrinsics in FindInterestingConstants pass.
• Improve MergeURBWrites to also merge URBWrite instructions with dynamic URB offsets.
• Add support to dump GRF usage chart per instruction. Fix some bugs with variable splitting.
• Add GTPin flags to support L0 driver path
• relax ACC substitution constraints
• Improve ConstantBufferCount estimate
• Merge from IGA; includes new platforms and some IGA IR refactoring.
• Fix to handle a infinite loop stuck issue leading to compilation hang.
• Using attribute SpillMemOffset instead of option -spillMemOffset, so that it is encoded in isa binary.
• Refactor to avoid hard-coded simd size for vertex shader compilation.

Ubuntu 18.04 binary packages for LLVM10/Clang10 are included.

igc-1.0.3864

Fixed Issues / Improvements

• Fix get_sub_group_id builtin handling
• Remove unused GTPin vISA option
• Further refactoring of vISA attributes
• vISA will use dispatch SIMD size from an attribute instead of scanning all instructions
• Fix to handle a infinite loop stuck issue leading to compilation hang
• Fix the performance regression caused by scalar suppression
• Refactor pixel shader code gen to avoid hard-coding SIMD size and GRF size
• Fix to SIMD16 shuffle down
• Add possibility to control partial component packing via shader metadata
• Fix sub_group_clustered_reduce_* on 64-bit types
• Extend LVN to detect more instructions
• Use attributes to get attributes names instead of hard-coding them wherever possible
• Added SimdSize kernel attribute (not used yet)
• Added sin/cos to sinpi/cospi optimization to improve performance
• Fixed the BCR static check
• Added getPlatform() function to IR_Builder and G4_INST
• Fixes in L0 tests
• Added vISA documentation
• Minor cleanup for G4_Declare
• Fixes for getAttributeID (searching all attributes instead of only kernels')
• Fix for subgroup_scan_exclusive for char and short
• Updated fcl interface to latest
• Added include guards, needed overrides, and some miscellaneous path fixes
• Added a debug knob to change the max block push constants threshold.
• Other minor bugfixes and improvements

Ubuntu 18.04 binary packages for LLVM10/Clang10 are included.

igc-1.0.3826

Fixed Issues / Improvements

• Indirect call support in SWSB.
• Avoid hard-coded SIMD size checks for sampler payload coalescing.
• Enable EnableBCR and VISAOptions flags for linux release.
• Extend LVN to detect more instructions.
• Fix a break in subroutine returns.
• Avoid hard-coding execution size for vISA sampler messages.
• Emu64bit support for stack/indirect function calls.
• Minor fixes and improvements.

Ubuntu 18.04 binary packages for LLVM10/Clang10 are included.

igc-1.0.3771

Fixed issues/improvements:

• Fix send payload coalescing when a source is larger than two GRFs.
• Group instructions together for performance.
• Check dangling constant users for Function/Global symbols, before iterating through the user list.
• Add the debug flags to limit number of constants to fold.
• Fix 64-bit multiplication emulation where the result is one of the factors.
• Fix translation of operation parameter of OpGroupNonUniformBallotBitCount.
• Verification pass for explicit variable split phase. Pass can be enabled using -verifysplit switch.
• Enable integer mad by default for OCL.
• Support function symbols used by Constants. MemOpt may combine multiple function address symbols into a single constant for load/store. Updating codegen to support this.
• Optimize indirect calls with function groups using LLVM !callees metadata. Fixed a few existing bugs, and removed the pass from OptimizeIR.
• Checking in FindInterestingConstants changes to return constants in sorted order.
• Fix translation of BallotBitCOunt in SPIRV Reader.
• Skip bank conflict hint for split variables to increase coalescing chances.
• Fix ASR offset for signed extend 32-bit value to 64-bit.
• Fix incorrect specialization constants reuse by subsequent SPIR-V modules.
• Skip trivial RA when variable split is done to avoid generating extra movs.
• Fix specialization constants for linking path.
• Enable explicit variable splitting and disable existing local split.
• Promote to stateful if GEP is not existed and address type is DW-aligned.
• Bail out from mad optimization if one of the operand is INF.
• And other fixes and performance improvements.

Ubuntu 18.04 binary packages for LLVM10/Clang10 are included.

igc-1.0.3698

Fixed issues/improvements:

• IGC now links with LLVM10,
• Emulation of 64-bit subgroup arithmetic and clustered non-uniform functions for platforms not supporting 64-it type natively,
• Fixes to cl_khr_subgroup_shuffle_relative and cl_khr_subgroup_clustered_reduce,
• Skip prefixes when parsing build options,
• Indirect address access is not allowed when dst is accumulator,
• Fix for SubgroupLtMask,
• Using local fences for local atomics,
• Fix sub_group_shuffle_down for long and double,
• Fix local block read builtins,
• Other fixes and improvements.

Ubuntu 18.04 binary packages for LLVM10/Clang10 are included.

igc-1.0.3627

Fixed Issues / Improvements

• Fixed kernel args iterator that causes compile error.
• Following llvm style for print and dump function.
• Fix for derived constants optimization.
• Fix FNeg intrinsic-function for the vector type.
• Emulation of 64-bit subgroup arithmetic and clustered non-uniform functions for platforms not supporting 64-bit type natively.
• SWSB: Ignore sync's dependency.
• Update GED for compact encoding.
• Minor fixes and improvements.

Ubuntu 18.04 binary packages for LLVM9/Clang9 are included.

igc-1.0.3586

• Support for extended subgroup functions.
• Improve IGCConstProp pass to support constant folding in rounding-mode-specific GenISA intrinsics.
• Add sat and overflow operation support for add and sub.
• Fixes realted to cl_khr_fp16 and cl_khr_fp64 types.
• Introduce inline asm support for SPIRV reader.
• Handling of splitting of rare regions like r[a0.0]<2,0> correctly.
• Small CMake path change
• Refactor IEEE FP32 divide/sqrt macro to avoid hard-coded execution size.
• Fixing the implementations of FENCE_PRE_OP and FENCE_POST_OP. The OCL’s GPU memory model problem was originally identified and reported by Tyler Sorensen and Jake Kirkham from Princeton University.
• Add address register initialization on ICL platform.
• Implement scratch space allocation for CMC thread-private memory.
• Increase OCL inlining threshold to 4500 to resolve the performance regression reported.
• And other bug fixes.

Ubuntu 18.04 binary packages for LLVM9/Clang9 are included.

igc-1.0.3529

Changelog

• Add VF immediate support for divm/sqrtm/pow,
• Emit scalar atomic load,
• Fix vISA options printing in the .asm file,
• Set loop unroll threshold for high register pressure shaders via IGC flag,
• Control the register pressure threshold used to limit the loop unroll to smaller loops via IGC flag,
• Support struct type return values for caller/callee,
• Dumping basic-blocks type by name,
• Support for cl_khr_subgroup_extended_types,
• Fixed referencing null iterator when KernelArgs size is 0,
• Added an ocl internal option -cl-intel-vector-coalesing=<0-5> to control vector coalescing,
• Fix alignment for MemCpy instructions,
• Added SPIRV-Tools configuration.

Ubuntu 18.04 binary packages for LLVM9/Clang9 are included.

igc-1.0.3471

Fixed Issues / Improvements

• Fix for gtpin when stackcall is enabled. rera pass is off, free GRFs are not computed.
• Add gen11 and gen 12 debug SIPs.
• Adding new compile time macros.
• Add a legalization check on scalar byte to half conversion.
• Minor fixes and improvements.

Ubuntu 18.04 binary packages for LLVM9/Clang9 are included.

igc-1.0.3445

• Handle the splitting of Vx1
• Modify spill/fill intrinsics' offset units.
• Support for hardware native atomic subtract.
• Add interface with dualsubslice information.
• Adding a configuration flag for enforcing double precision floating point emulation on unsupported platforms as an experimental feature.
• Fix live-interval in debug info when trivial RA is run.
• Fixing regression caused by wrong rounding.
• Don't do integer mad if thre are more than 1 immediate sources.
• Avoid splitting an instruction with byte dst into even/odd when one of its sources has vertical stride 32.
• Cache debug source line information during assembly dump.
• Fix for subgroup function resolution.
• Add Induction Variable Elimination pass.
• Remove redundant goto instructions.

Ubuntu 18.04 binary packages for LLVM9/Clang9 are included.