[fix] mlir: specify the mtriple in llvm target

[liamsemeria]

Motivation

#137 and #138 Revealed a regression in the mlir performance. The regression was due to the llvm update from 21 to 22 now requiring us to specify --mtriple in llc, in order for the memset() for the linalg.fill to lower optimally. With the old llvm, llc correctly lowered a large llvm.memset to a call to memset(). With the new llvm, the same input lowers to an inlined scalar byte-at-a-time store loop.

Description

The fix is explicitly passing -mtriple to llc for the native case too, queried from the LLVM install itself via llvm-config --host-target.

Using the Example from the Issue #137:

tvm performance: (same for both)

loop-explore  --threads 1 --peak-flops 65.6e9 --operator conv2d --backends tvm --use-tensor --strategy tile_prp --test 1 1 1 64 --save-temps
Schedule: tvm: [1; 1; 1; 64]: time: 1.63 msecs, peak perf: 110.28%

mlir peformance before:

loop-explore  --threads 1 --peak-flops 65.6e9 --operator conv2d --backends mlir --use-tensor --strategy tile_prp --test 1 1 1 64 --save-temps
Schedule: mlir: [1; 1; 1; 64]: time: 2.49 msecs, peak perf: 72.38%

mlir performance after:

loop-explore  --threads 1 --peak-flops 65.6e9 --operator conv2d --backends mlir --use-tensor --strategy tile_prp --test 1 1 1 64 --save-temps
Schedule: mlir: [1; 1; 1; 64]: time: 1.79 msecs, peak perf: 100.72%

For the compile time example in #138
tvm:

loop-explore  --threads 1 --peak-flops 65.6e9 --operator conv2d --backends tvm  --strategy tile7d --test 1 1 1 2 2 1 7 4 4 2 1 16 1 7 3 --save-temps --use-tensor
Schedule: tvm: [1; 1; 1; 2; 2; 1; 7; 4; 4; 2; 1; 16; 1; 7; 3]: time: 1.71 msecs, peak perf: 105.46%

mlir before:

loop-explore  --threads 1 --peak-flops 65.6e9 --operator conv2d --backends mlir  --strategy tile7d --test 1 1 1 2 2 1 7 4 4 2 1 16 1 7 3 --save-temps --use-tensor
Schedule: mlir: [1; 1; 1; 2; 2; 1; 7; 4; 4; 2; 1; 16; 1; 7; 3]: time: 2.46 msecs, peak perf: 73.07%

mlir after:

loop-explore  --threads 1 --peak-flops 65.6e9 --operator conv2d --backends mlir  --strategy tile7d --test 1 1 1 2 2 1 7 4 4 2 1 16 1 7 3 --save-temps --use-tensor
Schedule: mlir: [1; 1; 1; 2; 2; 1; 7; 4; 4; 2; 1; 16; 1; 7; 3]: time: 1.79 msecs, peak perf: 100.33%

[codecov]

Codecov Report

❌ Patch coverage is 70.58824% with 5 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/xtc/backends/mlir/MlirTarget/MlirLLVMTarget.py	70.58%	3 Missing and 2 partials ⚠️

📢 Thoughts on this report? Let us know!

[guillon]

Fixes issue #137

[guillon]

Just to clarify, this does not fix the compile time issue #138 mentionned above.