igc-1.0.4521

Fixed Issues / Improvements

• Fix bugs in image/sampler tracking to properly track argument when compiling with -cl-opt-disable
• Move pass to erase redundant movs after other optimizations are done.
• Skip simd32 compilation for per-pixel dispatch with x16 samples.
• Revert new-reg-per-function behavior in VC RA
• Add new Relocation Type R_PER_THREAD_PAYLOAD_OFFSET_32 Also refactor vISA::RelocationEntry create API
• Moving IntrinsicGenISA.gen to build Config folder and adding proper dependency requirement
• Switch to using LLVMTargetMachine in VC. Initialized GenX pass in BackendPlugin.
• Add check for fp64 and i64 copy move if platform does not support 64b types.
• Program the correct response length for spill of a scalar variable used as send dst.
• Enable FP64 accumulator as mul instruction source.
• Add TGL emulation functions for DP and SP
• DWARF debugger location expressions
• Emit variable location off privateBase
• additional include guards to avoid re-defintion conflict of LARGE_INTEGER type
• Uniform analysis tuning for performance.
• For stack calls do not adjust the spill size by global scratch offset
• Relocations and symbols support in L0 binary in VC
• Update register numbering for debug info.
• Try to avoid bank conflict for Gen12 when scheduling
• Fix some excessive mov instructions emitted by VectorCompiler.
• Avoid unncessary llvm metadata regenerations to optimize compilation time

Ubuntu 18.04 binary packages for LLVM10/Clang10 are included.