Revert "Test removal of TRITON_INTEL_ENABLE_INSTR_SCHED"

whitneywhtsang · whitneywhtsang · commit 55095c8cac51 · 2024-10-27T15:43:34.000Z
This reverts commit 35c3e09.
diff --git a/.github/workflows/triton-benchmarks.yml b/.github/workflows/triton-benchmarks.yml
@@ -111,27 +111,149 @@ jobs:
           cd benchmarks
           python setup.py install
 
-      - name: Run Triton FA kernel benchmark - advanced path (w/ TRITON_INTEL_ENABLE_INSTR_SCHED)
+      - name: Run Triton Softmax kernel benchmark
         if: ${{ steps.install.outcome == 'success' && !cancelled() }}
         run: |
           cd benchmarks/triton_kernels_benchmark
-          rm -rf ~/.triton/cache
+          python fused_softmax.py --reports $REPORTS
+          source ../../scripts/capture-hw-details.sh
+          python ../../scripts/build_report.py $REPORTS/softmax-performance.csv $REPORTS/softmax-triton-report.csv --benchmark softmax --compiler triton --param_cols "N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+          python ../../scripts/build_report.py $REPORTS/softmax-performance.csv $REPORTS/softmax-xetla-report.csv --benchmark softmax --compiler xetla --param_cols "N" --tflops_col XeTLA-TFlops --hbm_col "XeTLA-GB/s" --tag $TAG
+
+      - name: Run Triton GEMM kernel benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          python gemm_benchmark.py --reports $REPORTS
+          mv $REPORTS/matmul-performance.csv $REPORTS/matmul-performance-base.csv
+
+          source ../../scripts/capture-hw-details.sh
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-base.csv $REPORTS/gemm-triton-report.csv --benchmark gemm --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-base.csv $REPORTS/gemm-xetla-report.csv --benchmark gemm --compiler xetla --param_cols "B,M,K,N" --tflops_col XeTLA-TFlops --hbm_col "XeTLA-GB/s" --tag $TAG
+
+      - name: Run Triton GEMM kernel benchmark - default path
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          # Default path:
+          TRITON_INTEL_ADVANCED_PATH=0 \
+          TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
+          IGC_VISAOptions=" -enableBCR -nolocalra" \
+          IGC_DisableLoopUnroll=1 \
+          python gemm_benchmark.py --reports $REPORTS
+          mv $REPORTS/matmul-performance.csv $REPORTS/matmul-performance-default-path.csv
+
+          source ../../scripts/capture-hw-details.sh
+          TAG="${TAG}-dflt"
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-default-path.csv $REPORTS/gemm-triton-default-report.csv --benchmark gemm --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+
+      - name: Run Triton GEMM kernel benchmark - advanced path
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          # Advanced path:
           TRITON_INTEL_ADVANCED_PATH=1 \
-          TRITON_INTEL_ENABLE_INSTR_SCHED=1 \
+          TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
+          IGC_VISAOptions=" -enableBCR -nolocalra" \
+          IGC_DisableLoopUnroll=1 \
+          python gemm_benchmark.py --reports $REPORTS
+          mv $REPORTS/matmul-performance.csv $REPORTS/matmul-performance-adv-path.csv
+
+          source ../../scripts/capture-hw-details.sh
+          TAG="${TAG}-adv"
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-adv-path.csv $REPORTS/gemm-triton-advanced-report.csv --benchmark gemm --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+
+      - name: Run Triton GEMM (A@B^t) kernel benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          TRANSPOSE_B=1 python gemm_benchmark.py --reports $REPORTS
+          mv $REPORTS/matmul-performance.csv $REPORTS/matmul-performance-bt.csv
+          source ../../scripts/capture-hw-details.sh
+
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-bt.csv $REPORTS/gemm-bt-triton-report.csv --benchmark gemm-bt --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-bt.csv $REPORTS/gemm-bt-onednn-report.csv --benchmark gemm-bt --compiler onednn --param_cols "B,M,K,N" --tflops_col onednn-TFlops --hbm_col "onednn-GB/s" --tag $TAG
+
+      - name: Run Triton GEMM (A^t@B) kernel benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          TRANSPOSE_A=1 python gemm_benchmark.py --reports $REPORTS
+          mv $REPORTS/matmul-performance.csv $REPORTS/matmul-performance-at.csv
+          source ../../scripts/capture-hw-details.sh
+
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-at.csv $REPORTS/gemm-at-triton-report.csv --benchmark gemm-at --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-at.csv $REPORTS/gemm-at-onednn-report.csv --benchmark gemm-at --compiler onednn --param_cols "B,M,K,N" --tflops_col onednn-TFlops --hbm_col "onednn-GB/s" --tag $TAG
+
+      - name: Run Triton GEMM (stream-k) kernel benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          python gemm_streamk_benchmark.py --reports $REPORTS
+          source ../../scripts/capture-hw-details.sh
+          python ../../scripts/build_report.py $REPORTS/matmul-streamk-performance.csv $REPORTS/gemm-streamk-triton-report.csv --benchmark gemm-streamk --compiler triton --param_cols "M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+
+      - name: Run Triton GEMM (split-k) kernel benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          python gemm_splitk_benchmark.py --reports $REPORTS
+          source ../../scripts/capture-hw-details.sh
+          python ../../scripts/build_report.py $REPORTS/matmul-splitk-performance.csv $REPORTS/gemm-splitk-triton-report.csv --benchmark gemm-splitk --compiler triton --param_cols "M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+
+      - name: Run Triton GEMM + PreOp (exp) kernel benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          python gemm_preop_exp_benchmark.py --reports $REPORTS
+          source ../../scripts/capture-hw-details.sh
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-preop-exp.csv $REPORTS/gemm-preop-exp-triton-report.csv --benchmark gemm-preop-exp --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+
+      - name: Run Triton GEMM + PostOp (Gelu) kernel benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          python gemm_postop_gelu_benchmark.py --reports $REPORTS
+          source ../../scripts/capture-hw-details.sh
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-postop-gelu.csv $REPORTS/gemm-postop-gelu-triton-report.csv --benchmark gemm-postop-gelu --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+
+      - name: Run Triton GEMM + PostOp (add matrix) kernel benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          python gemm_postop_addmatrix_benchmark.py --reports $REPORTS
+          source ../../scripts/capture-hw-details.sh
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-postop-addmatrix.csv $REPORTS/gemm-postop-addmatrix-triton-report.csv --benchmark gemm-postop-addmatrix --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+
+      - name: Run Triton FA kernel benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          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
+          python ../../scripts/build_report.py $REPORTS/attn-performance.csv $REPORTS/attn-xetla-report.csv --benchmark attn --compiler xetla --param_cols "Z,H,N_CTX,D_HEAD,CAUSAL" --tflops_col XeTLA-TFlops --hbm_col "XeTLA-GB/s" --tag $TAG
+
+      - name: Run Triton FA kernel benchmark - default path
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          TRITON_INTEL_ADVANCED_PATH=0 \
           TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
           IGC_VISAOptions=" -enableBCR" \
           python flash_attention_fwd_benchmark.py --reports $REPORTS
 
-          TAG="${TAG}-adv"
+          TAG="${TAG}-dflt"
           source ../../scripts/capture-hw-details.sh
-          python ../../scripts/build_report.py $REPORTS/attn-performance.csv $REPORTS/attn-triton-advanced-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
+          python ../../scripts/build_report.py $REPORTS/attn-performance.csv $REPORTS/attn-triton-default-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
 
-      - name: Run Triton FA kernel benchmark - advanced path (w/o TRITON_INTEL_ENABLE_INSTR_SCHED)
+      - name: Run Triton FA kernel benchmark - advanced path
         if: ${{ steps.install.outcome == 'success' && !cancelled() }}
         run: |
           cd benchmarks/triton_kernels_benchmark
-          rm -rf ~/.triton/cache
           TRITON_INTEL_ADVANCED_PATH=1 \
+          TRITON_INTEL_ENABLE_INSTR_SCHED=1 \
           TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
           IGC_VISAOptions=" -enableBCR" \
           python flash_attention_fwd_benchmark.py --reports $REPORTS
@@ -140,6 +262,20 @@ jobs:
           source ../../scripts/capture-hw-details.sh
           python ../../scripts/build_report.py $REPORTS/attn-performance.csv $REPORTS/attn-triton-advanced-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
 
+      - name: Run Prefix Sums kernel benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/triton_kernels_benchmark
+          python prefix_sums.py --reports $REPORTS
+          source ../../scripts/capture-hw-details.sh
+          python ../../scripts/build_report.py $REPORTS/prefix-sums.csv $REPORTS/prefix_sums-triton-report.csv --benchmark prefix_sums --compiler triton --param_cols "N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
+
+      - name: Run micro benchmark
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          cd benchmarks/micro_benchmarks
+          python run_benchmarks.py --reports $REPORTS
+
       - name: Save pip cache
         if: ${{ steps.pip-cache.outputs.status == 'miss' }}
         uses: ./.github/actions/save
