rodinia/gaussian sycl and ndpx implementation

Author: roxx30198

.github/workflows/build_and_run.yml

@@ -168,5 +168,8 @@ jobs:
       - name: Run benchmarks
         run: dpbench -i ${{env.WORKLOADS}} run -r2 --no-print-results || exit 1
 
+      - name: Run rodinia benchmarks
+        run: dpbench -i ${{env.WORKLOADS}} run -r2 --no-print-results --rodinia --no-dpbench|| exit 1
+
       - name: Generate report
         run: dpbench -i ${{env.WORKLOADS}} report || exit 1

.github/workflows/conda-package.yml

@@ -195,6 +195,10 @@ jobs:
         run: |
           dpbench -i numpy -b azimint_hist run --npbench
 
+      - name: Run rodinia benchmark
+        run: |
+          dpbench -b gaussian run --rodinia --no-dpbench -r 1
+
   upload_anaconda:
     name: Upload dppy/label/dev ['${{ matrix.os }}', python='${{ matrix.python }}']
 

.pre-commit-config.yaml

@@ -50,7 +50,8 @@ repos:
     hooks:
     -   id: pydocstyle
         # TODO: add packages one by one to enforce pydocstyle eventually
-        files: (^dpbench/config/|^scripts/|^dpbench/console/|^dpbench/infrastructure/benchmark_runner.py|^dpbench/infrastructure/benchmark_validation.py)
+        files: (^dpbench/config/|^scripts/|^dpbench/console/|^dpbench/infrastructure/benchmark_runner.py|^dpbench/infrastructure/benchmark_validation.py|
+                ^dpbench/benchmarks/rodinia)
         args: ["--convention=google"]
         # D417 does not work properly:
         # https://github.com/PyCQA/pydocstyle/issues/459

dpbench/benchmarks/CMakeLists.txt

@@ -10,6 +10,7 @@ add_subdirectory(kmeans)
 add_subdirectory(knn)
 add_subdirectory(gpairs)
 add_subdirectory(dbscan)
+add_subdirectory(rodinia)
 
 # generate dpcpp version into config
 set(FILE ${CMAKE_SOURCE_DIR}/dpbench/configs/framework_info/dpcpp.toml)

dpbench/benchmarks/rodinia/CMakeLists.txt

@@ -0,0 +1,5 @@
+# SPDX-FileCopyrightText: 2022 - 2023 Intel Corporation
+#
+# SPDX-License-Identifier: Apache-2.0
+
+add_subdirectory(gaussian)

dpbench/benchmarks/rodinia/gaussian/CMakeLists.txt

@@ -0,0 +1,5 @@
+# SPDX-FileCopyrightText: 2022 - 2023 Intel Corporation
+#
+# SPDX-License-Identifier: Apache-2.0
+
+add_subdirectory(gaussian_sycl_native_ext)

dpbench/benchmarks/rodinia/gaussian/__init__.py

@@ -0,0 +1,28 @@
+# SPDX-FileCopyrightText: 2022 - 2023 Intel Corporation
+#
+# SPDX-License-Identifier: Apache-2.0
+
+"""
+
+Gaussian elimination implementation
+
+This is sycl and numba-dpex implementation for gaussian elimination
+
+Input
+---------
+size<int_64> : Forms an input matrix of dimensions (size x size)
+
+Output
+
+--------
+
+result<array<float>> : Result of the given set of linear equations using
+                        gaussian elimination.
+
+Method:
+
+The gaussian transformations are applied to the input matrix to form the
+diagonal matrix in forward elimination, and then the equations are solved
+to find the result in back substitution.
+
+"""

dpbench/benchmarks/rodinia/gaussian/gaussian_initialize.py

@@ -0,0 +1,34 @@
+# SPDX-FileCopyrightText: 2022 - 2023 Intel Corporation
+#
+# SPDX-License-Identifier: Apache-2.0
+
+LAMBDA = -0.01
+
+
+def initialize(size, types_dict):
+    import math
+
+    import numpy as np
+
+    dtype = types_dict["float"]
+
+    coe = np.empty((2 * size - 1), dtype=dtype)
+    a = np.empty((size * size), dtype=dtype)
+
+    for i in range(size):
+        coe_i = 10 * math.exp(LAMBDA * i)
+        j = size - 1 + i
+        coe[j] = coe_i
+        j = size - 1 - i
+        coe[j] = coe_i
+
+    for i in range(size):
+        for j in range(size):
+            a[i * size + j] = coe[size - 1 - i + j]
+
+    return (
+        a,
+        np.ones(size, dtype=dtype),
+        np.zeros((size * size), dtype=dtype),
+        np.zeros(size, dtype=dtype),
+    )

dpbench/benchmarks/rodinia/gaussian/gaussian_numba_dpex_k.py

@@ -0,0 +1,109 @@
+# SPDX-FileCopyrightText: 2022 - 2023 Intel Corporation
+#
+# SPDX-License-Identifier: Apache-2.0
+
+import dpctl
+import numba_dpex
+
+BLOCK_size_XY = 4
+
+
+@numba_dpex.kernel()
+def gaussian_kernel_1(m, a, size, t):
+    if (
+        numba_dpex.get_local_id(2)
+        + numba_dpex.get_group_id(2) * numba_dpex.get_local_size(2)
+        >= size - 1 - t
+    ):
+        return
+
+    m[
+        size
+        * (
+            numba_dpex.get_local_size(2) * numba_dpex.get_group_id(2)
+            + numba_dpex.get_local_id(2)
+            + t
+            + 1
+        )
+        + t
+    ] = (
+        a[
+            size
+            * (
+                numba_dpex.get_local_size(2) * numba_dpex.get_group_id(2)
+                + numba_dpex.get_local_id(2)
+                + t
+                + 1
+            )
+            + t
+        ]
+        / a[size * t + t]
+    )
+
+
+@numba_dpex.kernel()
+def gaussian_kernel_2(m, a, b, size, t):
+    if (
+        numba_dpex.get_local_id(2)
+        + numba_dpex.get_group_id(2) * numba_dpex.get_local_size(2)
+        >= size - 1 - t
+    ):
+        return
+
+    if (
+        numba_dpex.get_local_id(1)
+        + numba_dpex.get_group_id(1) * numba_dpex.get_local_size(1)
+        >= size - t
+    ):
+        return
+
+    xidx = numba_dpex.get_group_id(2) * numba_dpex.get_local_size(
+        2
+    ) + numba_dpex.get_local_id(2)
+    yidx = numba_dpex.get_group_id(1) * numba_dpex.get_local_size(
+        1
+    ) + numba_dpex.get_local_id(1)
+
+    a[size * (xidx + 1 + t) + (yidx + t)] -= (
+        m[size * (xidx + 1 + t) + t] * a[size * t + (yidx + t)]
+    )
+    if yidx == 0:
+        b[xidx + 1 + t] -= m[size * (xidx + 1 + t) + (yidx + t)] * b[t]
+
+
+def gaussian(a, b, m, size, result):
+    device = dpctl.SyclDevice()
+    block_size = device.max_work_group_size
+    grid_size = int((size / block_size) + 0 if not (size % block_size) else 1)
+
+    blocksize2d = BLOCK_size_XY
+    gridsize2d = int(
+        (size / blocksize2d) + (0 if not (size % blocksize2d) else 1)
+    )
+
+    global_range = numba_dpex.Range(1, 1, grid_size * block_size)
+    local_range = numba_dpex.Range(1, 1, block_size)
+
+    dim_blockXY = numba_dpex.Range(1, blocksize2d, blocksize2d)
+    dim_gridXY = numba_dpex.Range(
+        1, gridsize2d * blocksize2d, gridsize2d * blocksize2d
+    )
+
+    for t in range(size - 1):
+        gaussian_kernel_1[numba_dpex.NdRange(global_range, local_range)](
+            m, a, size, t
+        )
+
+        gaussian_kernel_2[numba_dpex.NdRange(dim_gridXY, dim_blockXY)](
+            m, a, b, size, t
+        )
+
+    for i in range(size):
+        result[size - i - 1] = b[size - i - 1]
+        for j in range(i):
+            result[size - i - 1] -= (
+                a[size * (size - i - 1) + (size - j - 1)] * result[size - j - 1]
+            )
+        result[size - i - 1] = (
+            result[size - i - 1] / a[size * (size - i - 1) + (size - i - 1)]
+        )

dpbench/benchmarks/rodinia/gaussian/gaussian_python.py

@@ -0,0 +1,24 @@
+# SPDX-FileCopyrightText: 2022 - 2023 Intel Corporation
+#
+# SPDX-License-Identifier: Apache-2.0
+
+
+def gaussian(a, b, m, size, result):
+    # Forward Elimination
+    for t in range(size - 1):
+        for i in range(t + 1, size):
+            m = a[i * size + t] / a[t * size + t]
+            for j in range(t, size):
+                a[i * size + j] = a[i * size + j] - m * a[t * size + j]
+            b[i] = b[i] - m * b[t]
+
+    # Back Substitution
+    for i in range(size):
+        result[size - i - 1] = b[size - i - 1]
+        for j in range(i):
+            result[size - i - 1] -= (
+                a[size * (size - i - 1) + (size - j - 1)] * result[size - j - 1]
+            )
+        result[size - i - 1] = (
+            result[size - i - 1] / a[size * (size - i - 1) + (size - i - 1)]
+        )