Enable long loads

Author: victor-eds

.github/workflows/triton-benchmarks.yml

@@ -233,7 +233,7 @@ jobs:
         if: ${{ steps.install.outcome == 'success' && !cancelled() && !contains(fromJson(inputs.skip_benchmarks || '[]'), 'flash_attention_fwd_benchmark.py') }}
         run: |
           cd benchmarks/triton_kernels_benchmark
-          python flash_attention_fwd_benchmark.py --reports $REPORTS
+          IGC_EnableVector8LoadStore=1 python flash_attention_fwd_benchmark.py --reports $REPORTS
 
           source ../../scripts/capture-hw-details.sh
           python ../../scripts/build_report.py $REPORTS/attn-performance.csv $REPORTS/attn-triton-report.csv --benchmark attn --compiler triton --param_cols "Z,H,N_CTX,D_HEAD,CAUSAL" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
@@ -245,7 +245,7 @@ jobs:
           cd benchmarks/triton_kernels_benchmark
           TRITON_INTEL_ADVANCED_PATH=0 \
           TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
-          IGC_VISAOptions=" -enableBCR" \
+          IGC_EnableVector8LoadStore=1 IGC_VISAOptions=" -enableBCR" \
           python flash_attention_fwd_benchmark.py --reports $REPORTS
 
           TAG="${TAG}-dflt"