Breeze fixes (#11537)

Summary: Pull Request resolved: #11537 Reviewed By: xiaoxmeng Differential Revision: D66022781 Pulled By: Yuhta fbshipit-source-id: 2ac480e11fd1e75374f0252e0f3d1fe6acfaa296
facebookincubator · Nov 15, 2024 · 31ae379 · 31ae379
1 parent c21020c
commit 31ae379
Show file tree

Hide file tree

Showing 31 changed files with 143 additions and 86 deletions.
diff --git a/velox/experimental/breeze/CMakeLists.txt b/velox/experimental/breeze/CMakeLists.txt
@@ -120,7 +120,7 @@ endif()
 
 if(NOT DISABLE_GOOGLETEST)
   set(GTest_SOURCE AUTO)
-  resolve_dependency(GTest)
+  velox_resolve_dependency(GTest)
 endif()
 
 if(BUILD_TESTING)

diff --git a/velox/experimental/breeze/breeze/platforms/openmp.h b/velox/experimental/breeze/breeze/platforms/openmp.h
@@ -140,9 +140,13 @@ struct supported_type<float> {
   static float min() { return __FLT_MIN__; }
   static float max() { return __FLT_MAX__; }
   static float atomic_load(float *address) {
-    unsigned value = __atomic_load_n(reinterpret_cast<unsigned *>(address),
-                                     __ATOMIC_RELAXED);
-    return *reinterpret_cast<float *>(&value);
+    union {
+      float value;
+      unsigned raw;
+    } content;
+    content.raw = __atomic_load_n(reinterpret_cast<unsigned *>(address),
+                                  __ATOMIC_RELAXED);
+    return content.value;
   }
   static void atomic_store(float *address, float value) {
     __atomic_store_n(reinterpret_cast<unsigned *>(address),
@@ -166,9 +170,13 @@ struct supported_type<double> {
   static double min() { return __DBL_MIN__; }
   static double max() { return __DBL_MAX__; }
   static double atomic_load(double *address) {
-    unsigned long long value = __atomic_load_n(
+    union {
+      double value;
+      unsigned long long raw;
+    } content;
+    content.raw = __atomic_load_n(
         reinterpret_cast<unsigned long long *>(address), __ATOMIC_RELAXED);
-    return *reinterpret_cast<double *>(&value);
+    return content.value;
   }
   static void atomic_store(double *address, double value) {
     __atomic_store_n(reinterpret_cast<unsigned long long *>(address),

diff --git a/velox/experimental/breeze/breeze/utils/trace.h b/velox/experimental/breeze/breeze/utils/trace.h
@@ -29,6 +29,12 @@
 
 #if TRACING
 
+// Workaround for CUDA/HIP PERFETTO_NO_INLINE conflict
+#ifdef __noinline__
+#undef __noinline__
+#define __noinline__ noinline
+#endif
+
 #include <perfetto.h>
 
 #include <condition_variable>

diff --git a/velox/experimental/breeze/cmake/cuda.cmake b/velox/experimental/breeze/cmake/cuda.cmake
@@ -21,10 +21,6 @@ if(NOT DEFINED CMAKE_NVCC_FLAGS)
     set(CMAKE_NVCC_FLAGS
         "$ENV{NVCCFLAGS}"
         CACHE STRING "NVCC flags")
-  else()
-    set(CMAKE_NVCC_FLAGS
-        "-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_90,code=sm_90"
-        CACHE STRING "NVCC flags")
   endif()
 endif()
 if(BREEZE_BUILD_TYPE MATCHES Debug)
@@ -54,7 +50,7 @@ endif()
 # use PTX specialization by default for CUDA
 if(NOT DEFINED CUDA_PLATFORM_SPECIALIZATION_HEADER)
   set(CUDA_PLATFORM_SPECIALIZATION_HEADER
-      platforms/specialization/cuda-ptx.cuh
+      breeze/platforms/specialization/cuda-ptx.cuh
       CACHE STRING "CUDA platform specialization header")
 endif()
 
@@ -87,8 +83,8 @@ function(breeze_add_cuda_test target source)
     "FLAGS;LIBS;DEPENDS"
     ${ARGN})
   list(APPEND arg_FLAGS -I${gtest_SOURCE_DIR}/googletest/include)
-  list(APPEND arg_FLAGS -I${CMAKE_SOURCE_DIR}/test)
-  list(APPEND arg_FLAGS -I${CMAKE_CURRENT_BINARY_DIR})
+  list(APPEND arg_FLAGS -I${CMAKE_SOURCE_DIR})
+  list(APPEND arg_FLAGS -I${CMAKE_BINARY_DIR})
   breeze_add_cuda_object(
     ${target}
     ${source}

diff --git a/velox/experimental/breeze/cmake/hip.cmake b/velox/experimental/breeze/cmake/hip.cmake
@@ -29,10 +29,9 @@ function(breeze_add_hip_test target source)
     OUTPUT ${target}.o
     COMMAND
       ${HIPCC_EXECUTABLE} ${HIP_HIPCC_FLAGS} ${NDEBUG_DEFINE} -DPLATFORM_HIP
-      -I${CMAKE_SOURCE_DIR} -I${CMAKE_SOURCE_DIR}/test
-      -I${gtest_SOURCE_DIR}/googletest/include -I${CMAKE_CURRENT_BINARY_DIR}
-      ${CMAKE_CXX_FLAGS} ${COMPILER_WARN_FLAGS} ${OPT_FLAGS} -std=c++17 -c
-      ${source} -MD -MF ${target}.o.d -o ${target}.o
+      -I${CMAKE_SOURCE_DIR} -I${gtest_SOURCE_DIR}/googletest/include
+      -I${CMAKE_BINARY_DIR} ${CMAKE_CXX_FLAGS} ${COMPILER_WARN_FLAGS}
+      ${OPT_FLAGS} -std=c++17 -c ${source} -MD -MF ${target}.o.d -o ${target}.o
     DEPFILE ${target}.o.d
     DEPENDS ${arg_DEPENDS}
     COMMENT "Building HIP object ${target}.o")

diff --git a/velox/experimental/breeze/cmake/metal.cmake b/velox/experimental/breeze/cmake/metal.cmake
@@ -27,9 +27,8 @@ function(breeze_add_metal_test_kernels target source)
     OUTPUT ${target}.air
     COMMAND
       xcrun -sdk macosx metal -DPLATFORM_METAL -I${CMAKE_SOURCE_DIR}
-      -I${CMAKE_CURRENT_BINARY_DIR} ${WARN_FLAGS} ${OPT_FLAGS}
-      -Wno-c++17-extensions -std=metal3.0 -c ${source} -MD -MF ${target}.air.d
-      -o ${target}.air
+      -I${CMAKE_BINARY_DIR} ${WARN_FLAGS} ${OPT_FLAGS} -Wno-c++17-extensions
+      -std=metal3.0 -c ${source} -MD -MF ${target}.air.d -o ${target}.air
     COMMENT "Compile ${source} --> ${target}.air"
     DEPENDS ${arg_DEPENDS}
     DEPFILE ${target}.air.d)
@@ -44,7 +43,7 @@ endfunction()
 function(breeze_add_metal_test target source shaderlib)
   add_executable(${target} ${source} platforms/metal_test.mm)
   target_compile_features(${target} PRIVATE cxx_std_17)
-  target_include_directories(${target} PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
+  target_include_directories(${target} PRIVATE ${CMAKE_BINARY_DIR})
   target_compile_definitions(${target} PUBLIC PLATFORM_METAL=1
                                               SHADER_LIB=\"${shaderlib}\")
   target_compile_options(

diff --git a/velox/experimental/breeze/cmake/opencl.cmake b/velox/experimental/breeze/cmake/opencl.cmake
@@ -29,7 +29,7 @@ function(breeze_add_opencl_test_kernels target source)
     OUTPUT ${target}.so
     COMMAND
       ${CLCC_EXECUTABLE} -o ${target}.so ${source} -MD -MF ${target}.clcpp.d
-      -DPLATFORM_OPENCL -I${CMAKE_SOURCE_DIR} -I${CMAKE_CURRENT_BINARY_DIR}
+      -DPLATFORM_OPENCL -I${CMAKE_SOURCE_DIR} -I${CMAKE_BINARY_DIR}
       ${COMPILER_WARN_FLAGS} ${OPT_FLAGS}
     COMMENT "Building OpenCL kernels ${source} --> ${target}.so"
     DEPENDS ${arg_DEPENDS}
@@ -41,7 +41,7 @@ function(breeze_add_opencl_test target source shaderlib)
   add_executable(${target} ${source})
   target_compile_features(${target} PRIVATE cxx_std_17)
   target_include_directories(${target} PRIVATE ${CMAKE_OPENCL_INCLUDE}
-                                               ${CMAKE_CURRENT_BINARY_DIR})
+                                               ${CMAKE_BINARY_DIR})
   target_compile_definitions(${target} PUBLIC PLATFORM_OPENCL=1
                                               SHADER_LIB=\"${shaderlib}\")
   target_compile_options(${target} PRIVATE ${WARN_FLAGS} ${OPT_FLAGS}

diff --git a/velox/experimental/breeze/cmake/sycl.cmake b/velox/experimental/breeze/cmake/sycl.cmake
@@ -29,10 +29,10 @@ function(breeze_add_sycl_test target source)
     OUTPUT ${target}.o
     COMMAND
       ${SYCLCC_EXECUTABLE} ${NDEBUG_DEFINE} -DPLATFORM_SYCL
-      -I${CMAKE_SOURCE_DIR} -I${CMAKE_SOURCE_DIR}/test
-      -I${gtest_SOURCE_DIR}/googletest/include -I${CMAKE_CURRENT_BINARY_DIR}
-      ${CMAKE_CXX_FLAGS} ${OPT_FLAGS} ${SANITIZE_COMPILE_FLAGS} -std=c++17 -c
-      ${source} -MD -MF ${target}.o.d -o ${target}.o
+      -I${CMAKE_SOURCE_DIR} -I${gtest_SOURCE_DIR}/googletest/include
+      -I${CMAKE_BINARY_DIR} ${CMAKE_CXX_FLAGS} ${OPT_FLAGS}
+      ${SANITIZE_COMPILE_FLAGS} -std=c++17 -c ${source} -MD -MF ${target}.o.d -o
+      ${target}.o
     DEPFILE ${target}.o.d
     DEPENDS ${arg_DEPENDS}
     COMMENT "Building SYCL object ${target}.o")

diff --git a/velox/experimental/breeze/perftest/algorithms/reduce_perftest.cu b/velox/experimental/breeze/perftest/algorithms/reduce_perftest.cu
@@ -25,10 +25,10 @@
 #include <limits>
 #include <vector>
 
-#include "algorithms/reduce.h"
+#include "breeze/algorithms/reduce.h"
+#include "breeze/platforms/platform.h"
+#include "breeze/utils/device_vector.h"
 #include "perftest/perftest.h"
-#include "platforms/platform.h"
-#include "utils/device_vector.h"
 
 namespace breeze {
 

diff --git a/velox/experimental/breeze/perftest/algorithms/scan_perftest.cu b/velox/experimental/breeze/perftest/algorithms/scan_perftest.cu
@@ -24,14 +24,14 @@
 
 #include <vector>
 
-#include "algorithms/scan.h"
-#include "functions/load.h"
-#include "functions/scan.h"
-#include "functions/store.h"
+#include "breeze/algorithms/scan.h"
+#include "breeze/functions/load.h"
+#include "breeze/functions/scan.h"
+#include "breeze/functions/store.h"
+#include "breeze/platforms/platform.h"
+#include "breeze/utils/block_details.h"
+#include "breeze/utils/device_vector.h"
 #include "perftest/perftest.h"
-#include "platforms/platform.h"
-#include "utils/block_details.h"
-#include "utils/device_vector.h"
 
 namespace breeze {
 

diff --git a/velox/experimental/breeze/perftest/algorithms/sort_perftest.cu b/velox/experimental/breeze/perftest/algorithms/sort_perftest.cu
@@ -25,11 +25,11 @@
 #include <climits>
 #include <vector>
 
-#include "algorithms/sort.h"
+#include "breeze/algorithms/sort.h"
+#include "breeze/platforms/platform.h"
+#include "breeze/utils/block_details.h"
+#include "breeze/utils/device_vector.h"
 #include "perftest/perftest.h"
-#include "platforms/platform.h"
-#include "utils/block_details.h"
-#include "utils/device_vector.h"
 
 namespace breeze {
 

diff --git a/velox/experimental/breeze/perftest/functions/reduce_perftest.cu b/velox/experimental/breeze/perftest/functions/reduce_perftest.cu
@@ -24,10 +24,10 @@
 
 #include <vector>
 
-#include "functions/reduce.h"
+#include "breeze/functions/reduce.h"
+#include "breeze/platforms/platform.h"
+#include "breeze/utils/device_vector.h"
 #include "perftest/perftest.h"
-#include "platforms/platform.h"
-#include "utils/device_vector.h"
 
 namespace breeze {
 

diff --git a/velox/experimental/breeze/perftest/functions/scan_perftest.cu b/velox/experimental/breeze/perftest/functions/scan_perftest.cu
@@ -24,10 +24,10 @@
 
 #include <vector>
 
-#include "functions/scan.h"
+#include "breeze/functions/scan.h"
+#include "breeze/platforms/platform.h"
+#include "breeze/utils/device_vector.h"
 #include "perftest/perftest.h"
-#include "platforms/platform.h"
-#include "utils/device_vector.h"
 
 namespace breeze {
 

diff --git a/velox/experimental/breeze/perftest/functions/sort_perftest.cu b/velox/experimental/breeze/perftest/functions/sort_perftest.cu
@@ -25,11 +25,11 @@
 #include <limits>
 #include <vector>
 
-#include "functions/sort.h"
+#include "breeze/functions/sort.h"
+#include "breeze/platforms/platform.h"
+#include "breeze/utils/device_vector.h"
+#include "breeze/utils/types.h"
 #include "perftest/perftest.h"
-#include "platforms/platform.h"
-#include "utils/device_vector.h"
-#include "utils/types.h"
 
 namespace breeze {
 

diff --git a/velox/experimental/breeze/perftest/perftest-cuda.cuh b/velox/experimental/breeze/perftest/perftest-cuda.cuh
@@ -30,8 +30,8 @@
 #include <regex>
 #include <string>
 
-#include "platforms/cuda.cuh"
-#include "utils/trace.h"
+#include "breeze/platforms/cuda.cuh"
+#include "breeze/utils/trace.h"
 
 namespace breeze {
 

diff --git a/velox/experimental/breeze/perftest/queries/caching_device_allocator.cuh b/velox/experimental/breeze/perftest/queries/caching_device_allocator.cuh
@@ -27,12 +27,12 @@
 #include <memory>
 
 #ifdef __EXCEPTIONS
-#include "utils/types.h"
+#include "breeze/utils/types.h"
 #else
 #include <iostream>
 #endif
 
-#include "platforms/cuda.cuh"
+#include "breeze/platforms/cuda.cuh"
 
 // caching device allocator that can be used to eliminate
 // allocation cost from benchmarks by handling all neccessary

diff --git a/velox/experimental/breeze/perftest/queries/device_column.h b/velox/experimental/breeze/perftest/queries/device_column.h
@@ -22,8 +22,8 @@
 
 #pragma once
 
-#include "utils/device_allocator.h"
-#include "utils/device_vector.h"
+#include "breeze/utils/device_allocator.h"
+#include "breeze/utils/device_vector.h"
 
 namespace breeze {
 

diff --git a/velox/experimental/breeze/perftest/queries/order_by_perftest.cu b/velox/experimental/breeze/perftest/queries/order_by_perftest.cu
@@ -28,17 +28,17 @@
 #include <string>
 #include <vector>
 
-#include "algorithms/sort.h"
-#include "functions/load.h"
-#include "functions/reduce.h"
-#include "functions/scan.h"
-#include "functions/store.h"
+#include "breeze/algorithms/sort.h"
+#include "breeze/functions/load.h"
+#include "breeze/functions/reduce.h"
+#include "breeze/functions/scan.h"
+#include "breeze/functions/store.h"
+#include "breeze/platforms/cuda.cuh"
+#include "breeze/utils/device_vector.h"
+#include "breeze/utils/types.h"
 #include "perftest/perftest.h"
 #include "perftest/queries/caching_device_allocator.cuh"
 #include "perftest/queries/device_column.h"
-#include "platforms/cuda.cuh"
-#include "utils/device_vector.h"
-#include "utils/types.h"
 
 namespace breeze {
 namespace test {

diff --git a/velox/experimental/breeze/test/CMakeLists.txt b/velox/experimental/breeze/test/CMakeLists.txt
@@ -16,6 +16,8 @@
 # Licensed under the Apache License, Version 2.0, see LICENSE for details.
 # SPDX-License-Identifier: Apache-2.0
 
+include_directories(${CMAKE_SOURCE_DIR})
+
 cxx_library(test_main "${cxx_strict}" test_main.cpp)
 target_link_libraries(
   test_main

diff --git a/velox/experimental/breeze/test/algorithms/kernels.clcpp b/velox/experimental/breeze/test/algorithms/kernels.clcpp
@@ -22,7 +22,7 @@
 
 enum { OPENCL_WARP_THREADS = 32 };
 
-#include "generated/algorithms/kernels-opencl.h"
+#include "test/generated/algorithms/kernels-opencl.h"
 
 // kernel specializations
 

diff --git a/velox/experimental/breeze/test/algorithms/kernels.metal b/velox/experimental/breeze/test/algorithms/kernels.metal
@@ -20,7 +20,7 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-#include "generated/algorithms/kernels-metal.h"
+#include "test/generated/algorithms/kernels-metal.h"
 
 // kernel specializations
 

diff --git a/velox/experimental/breeze/test/functions/kernels.clcpp b/velox/experimental/breeze/test/functions/kernels.clcpp
@@ -20,7 +20,7 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-#include "generated/functions/kernels-opencl.h"
+#include "test/generated/functions/kernels-opencl.h"
 
 // kernel specializations
 

diff --git a/velox/experimental/breeze/test/functions/kernels.metal b/velox/experimental/breeze/test/functions/kernels.metal
@@ -20,7 +20,7 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-#include "generated/functions/kernels-metal.h"
+#include "test/generated/functions/kernels-metal.h"
 
 // kernel specializations
 

diff --git a/velox/experimental/breeze/test/functions/reduce_unittest.cpp b/velox/experimental/breeze/test/functions/reduce_unittest.cpp
@@ -22,6 +22,7 @@
 
 #include <gtest/gtest.h>
 
+#include <algorithm>
 #include <numeric>
 
 #include "function_test.h"

diff --git a/velox/experimental/breeze/test/generated/algorithms/algorithm_test-openmp.h b/velox/experimental/breeze/test/generated/algorithms/algorithm_test-openmp.h
@@ -14,9 +14,16 @@
  * limitations under the License.
  */
 
-// Copyright (c) 2024 by Rivos Inc.
-// Licensed under the Apache License, Version 2.0, see LICENSE for details.
-// SPDX-License-Identifier: Apache-2.0
+/*
+ * Copyright (c) 2024 by Rivos Inc.
+ * Licensed under the Apache License, Version 2.0, see LICENSE for details.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/*
+ * This file is auto-generated from test_fixture_generator.py
+ * DO NOT EDIT!
+ */
 
 #include <gtest/gtest.h>
 #include <omp.h>