[FA] Remove TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT (#2919)

whitneywhtsang · web-flow · commit 24c95d9bc4af · 2024-12-04T09:16:16.000-05:00
Address payload optimization is already implemented in IGC. Removing `TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT` causes no performance impact. CI: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12152410219 Signed-off-by: Whitney Tsang <whitney.tsang@intel.com>
diff --git a/.github/workflows/triton-benchmarks.yml b/.github/workflows/triton-benchmarks.yml
@@ -250,7 +250,6 @@ jobs:
         run: |
           cd benchmarks/triton_kernels_benchmark
           TRITON_INTEL_ADVANCED_PATH=1 \
-          TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
           IGC_VISAOptions=" -enableBCR" \
           python flash_attention_fwd_benchmark.py --reports $REPORTS
 
diff --git a/scripts/test-triton.sh b/scripts/test-triton.sh
@@ -284,7 +284,6 @@ run_benchmark_attention() {
 
   echo "Advanced path:"
   TRITON_INTEL_ADVANCED_PATH=1 \
-    TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
     IGC_VISAOptions=" -enableBCR" \
     python $TRITON_PROJ/benchmarks/triton_kernels_benchmark/flash_attention_fwd_benchmark.py
 }

Original file line number	Diff line number	Diff line change
`@@ -284,7 +284,6 @@ run_benchmark_attention() {`
`284`	`284`
`285`	`285`	`echo "Advanced path:"`
`286`	`286`	`TRITON_INTEL_ADVANCED_PATH=1 \`
`287`		`- TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \`
`288`	`287`	`IGC_VISAOptions=" -enableBCR" \`
`289`	`288`	`python $TRITON_PROJ/benchmarks/triton_kernels_benchmark/flash_attention_fwd_benchmark.py`
`290`	`289`	`}`