Update

.github/workflows/triton-benchmarks.yml

@@ -340,9 +340,7 @@ jobs:
           rm -rf ./tt_cache
           TRITON_CACHE_DIR=./tt_cache \
           TRITON_INTEL_ENABLE_ADDRESS_PAYLOAD_OPT=1 \
-          TRITON_INTEL_ENABLE_BLOCK_PTR=1 \
-          TRITON_INTEL_PREFETCH_DISTANCE=2 \
-          TRITON_INTEL_SPLIT_BARRIER=1 \
+          TRITON_INTEL_ADVANCED_PATH=1 \
           IGC_VISAOptions=" -TotalGRFNum 256 -enableBCR -nolocalra -printregusage -DPASTokenReduction -enableHalfLSC -abiver 2" \
           IGC_ForcePrefetchToL1Cache=1 \
           IGC_VATemp=1 \

benchmarks/triton_kernels_benchmark/gemm_benchmark.py

@@ -328,14 +328,16 @@ def matmul(a, b):
         # argument names to use as an x-axis for the plot
         x_names=['B', 'M', 'K', 'N'],
         # different possible values for `x_name`
-        x_vals=[[1, 4096, 4096, 4096], [1, 8192, 8192, 8192], [1, 1, 5120, 13824], [1, 1024, 28672, 8192],
-                [1, 3072, 4096, 3072], [1, 4, 4096, 12288]],
+        x_vals=[[1, 256 * i, 256 * i, 256 * i] for i in range(1, 17)] +  #
+        [[4, 32768, 128, 4096],  #
+         [4, 32768, 4096, 128],  #
+         [32, 4096, 4096, 128]],
         line_arg='provider',
         # argument name whose value corresponds to a different line in the plot
         # possible values for `line_arg``
-        line_vals=['triton'],
+        line_vals=['onednn', 'triton', 'xetla'],
         # label name for the lines
-        line_names=["Triton"],
+        line_names=["oneDNN", "Triton", "XeTLA"],
         # line styles
         styles=[('green', '-'), ('green', '--'), ('blue', '-'), ('blue', '--')],
         ylabel=["GB/s", "TFlops"],  # label name for the y-axis

benchmarks/xetla_kernel/CMakeLists.txt

@@ -13,4 +13,4 @@ target_include_directories(xetla_kernel PUBLIC "${TORCH_IPEX_INCLUDE_DIRS}")
 target_include_directories(xetla_kernel PUBLIC "${XeTLALibrary_INCLUDE_DIR}")
 
 add_subdirectory(softmax)
-add_subdirectory(gemm)
+add_subdirectory(gemm)