[FA] Remove IGC_DisableLoopUnroll=1 (#2416)

No performance impact.

Signed-off-by: Whitney Tsang <[email protected]>

Author: whitneywhtsang

.github/workflows/triton-benchmarks.yml

@@ -222,7 +222,6 @@ jobs:
           TRITON_INTEL_ADVANCED_PATH=0 \
           TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
           IGC_VISAOptions=" -enableBCR -nolocalra -printregusage -DPASTokenReduction -enableHalfLSC -abiver 2" \
-          IGC_DisableLoopUnroll=1 \
           python flash_attention_fwd_benchmark.py --reports $REPORTS
 
           TAG=${{ inputs.tag || 'ci' }}-dflt
@@ -237,7 +236,6 @@ jobs:
           TRITON_INTEL_ENABLE_INSTR_SCHED=1 \
           TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
           IGC_VISAOptions=" -enableBCR -nolocalra -printregusage -DPASTokenReduction -enableHalfLSC -abiver 2" \
-          IGC_DisableLoopUnroll=1 \
           python flash_attention_fwd_benchmark.py --reports $REPORTS
 
           TAG=${{ inputs.tag || 'ci' }}-adv

scripts/test-triton.sh

@@ -280,15 +280,13 @@ run_benchmark_attention() {
   TRITON_INTEL_ADVANCED_PATH=0 \
     TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
     IGC_VISAOptions=" -enableBCR -nolocalra -printregusage -DPASTokenReduction -enableHalfLSC" \
-    IGC_DisableLoopUnroll=1 \
     python $TRITON_PROJ/benchmarks/triton_kernels_benchmark/flash_attention_fwd_benchmark.py
 
   echo "Advanced path:"
   TRITON_INTEL_ADVANCED_PATH=1 \
     TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
     TRITON_INTEL_ENABLE_INSTR_SCHED=1 \
     IGC_VISAOptions=" -enableBCR -nolocalra -printregusage -DPASTokenReduction -enableHalfLSC" \
-    IGC_DisableLoopUnroll=1 \
     python $TRITON_PROJ/benchmarks/triton_kernels_benchmark/flash_attention_fwd_benchmark.py
 }