Merge pull request #1661 from allisonvacanti/if_target_cdp

Update CDP support macros for if-target compatibility

Author: alliepiper

dependencies/cub

@@ -84,6 +84,24 @@ if (NOT valid_count EQUAL 5)
     "Matched ${valid_count} times, expected 5.")
 endif()
 
+################################################################################
+# Legacy macro checks.
+# Check all files in Thrust to make sure that they aren't using the legacy
+# CUB_RUNTIME_ENABLED and __THRUST_HAS_CUDART__ macros.
+#
+# These macros depend on __CUDA_ARCH__ and are not compatible with NV_IF_TARGET.
+# They are provided for legacy purposes and should be replaced with
+# [THRUST|CUB]_RDC_ENABLED and NV_IF_TARGET in Thrust/CUB code.
+#
+#
+set(legacy_macro_header_exclusions
+  # This header defines a legacy CUDART macro:
+  thrust/system/cuda/config.h
+)
+
+set(cub_legacy_macro_regex "CUB_RUNTIME_ENABLED")
+set(thrust_legacy_macro_regex "__THRUST_HAS_CUDART__")
+
 ################################################################################
 # Read source files:
 foreach(src ${thrust_srcs})
@@ -145,6 +163,21 @@ foreach(src ${thrust_srcs})
       set(found_errors 1)
     endif()
   endif()
+
+  if (NOT ${src} IN_LIST legacy_macro_header_exclusions)
+    count_substrings("${src_contents}" "${thrust_legacy_macro_regex}" thrust_count)
+    count_substrings("${src_contents}" "${cub_legacy_macro_regex}" cub_count)
+
+    if (NOT thrust_count EQUAL 0)
+      message("'${src}' uses __THRUST_HAS_CUDART__. Replace with THRUST_RDC_ENABLED and NV_IF_TARGET.")
+      set(found_errors 1)
+    endif()
+
+    if (NOT cub_count EQUAL 0)
+      message("'${src}' uses CUB_RUNTIME_ENABLED. Replace with CUB_RDC_ENABLED and NV_IF_TARGET.")
+      set(found_errors 1)
+    endif()
+  endif()
 endforeach()
 
 if (NOT found_errors EQUAL 0)

testing/cmake/check_source_files.cmake

@@ -6,6 +6,10 @@ file(GLOB test_srcs
 
 # These tests always build with RDC, so make sure that the sm_XX flags are
 # compatible. See note in ThrustCudaConfig.cmake.
+# TODO once we're using CUDA_ARCHITECTURES, we can setup non-rdc fallback
+# tests to build for non-rdc arches. But for now, all files in a given directory
+# must build with the same `CMAKE_CUDA_FLAGS` due to CMake constraints around
+# how CUDA_FLAGS works.
 set(CMAKE_CUDA_FLAGS "${THRUST_CUDA_FLAGS_BASE} ${THRUST_CUDA_FLAGS_RDC}")
 
 foreach(thrust_target IN LISTS THRUST_TARGETS)
@@ -18,11 +22,11 @@ foreach(thrust_target IN LISTS THRUST_TARGETS)
     get_filename_component(test_name "${test_src}" NAME_WLE)
     string(PREPEND test_name "cuda.")
 
-    thrust_add_test(test_target ${test_name} "${test_src}" ${thrust_target})
-
-    # All in testing/cuda will test device-side launch (aka calling parallel
-    # algorithms from device code), which requires the CUDA device-side runtime,
-    # which requires RDC, so these always need to be built with RDC.
-    thrust_enable_rdc_for_cuda_target(${test_target})
+    # Create two targets, one with RDC enabled, the other without. This tests
+    # both device-side behaviors -- the CDP kernel launch with RDC, and the
+    # serial fallback path without RDC.
+    thrust_add_test(seq_test_target ${test_name}.cdp_0 "${test_src}" ${thrust_target})
+    thrust_add_test(cdp_test_target ${test_name}.cdp_1 "${test_src}" ${thrust_target})
+    thrust_enable_rdc_for_cuda_target(${cdp_test_target})
   endforeach()
 endforeach()

testing/cuda/CMakeLists.txt

@@ -32,22 +32,47 @@
 // older releases. This header will always pull in version info:
 #include <cub/util_namespace.cuh>
 
-#if defined(__CUDACC__) || defined(_NVHPC_CUDA)
-#  if !defined(__CUDA_ARCH__) || defined(__CUDACC_RDC__)
-#    define __THRUST_HAS_CUDART__ 1
-#    define THRUST_RUNTIME_FUNCTION __host__ __device__ __forceinline__
-#  else
-#    define __THRUST_HAS_CUDART__ 0
-#    define THRUST_RUNTIME_FUNCTION __host__ __forceinline__
-#  endif
+#include <cub/detail/detect_cuda_runtime.cuh>
+
+/**
+ * \def THRUST_RUNTIME_FUNCTION
+ *
+ * Execution space for functions that can use the CUDA runtime API (`__host__`
+ * when RDC is off, `__host__ __device__` when RDC is on).
+ */
+#define THRUST_RUNTIME_FUNCTION CUB_RUNTIME_FUNCTION
+
+/**
+ * \def THRUST_RDC_ENABLED
+ *
+ * Defined if RDC is enabled.
+ */
+#ifdef CUB_RDC_ENABLED
+#define THRUST_RDC_ENABLED
+#endif
+
+/**
+ * \def __THRUST_HAS_CUDART__
+ *
+ * Whether or not the active compiler pass is allowed to invoke device kernels
+ * or methods from the CUDA runtime API.
+ *
+ * This macro should not be used in Thrust, as it depends on `__CUDA_ARCH__`
+ * and is not compatible with `NV_IF_TARGET`. It is provided for legacy
+ * purposes only.
+ *
+ * Replace any usages with `THRUST_RDC_ENABLED` and `NV_IF_TARGET`.
+ */
+#ifdef CUB_RUNTIME_ENABLED
+#define __THRUST_HAS_CUDART__ 1
 #else
-#  define __THRUST_HAS_CUDART__ 0
-#  define THRUST_RUNTIME_FUNCTION __host__ __forceinline__
+#define __THRUST_HAS_CUDART__ 0
 #endif
 
 // These definitions were intended for internal use only and are now obsolete.
 // If you relied on them, consider porting your code to use the functionality
 // in libcu++'s <nv/target> header.
+//
 // For a temporary workaround, define THRUST_PROVIDE_LEGACY_ARCH_MACROS to make
 // them available again. These should be considered deprecated and will be
 // fully removed in a future version.

thrust/system/cuda/config.h

@@ -29,12 +29,14 @@
 #include <thrust/detail/config.h>
 
 #if THRUST_DEVICE_COMPILER == THRUST_DEVICE_COMPILER_NVCC
+
 #include <thrust/detail/cstdint.h>
 #include <thrust/detail/minmax.h>
 #include <thrust/detail/temporary_array.h>
 #include <thrust/detail/type_traits.h>
 #include <thrust/functional.h>
 #include <thrust/system/cuda/config.h>
+#include <thrust/system/cuda/detail/cdp_dispatch.h>
 #include <thrust/system/cuda/detail/dispatch.h>
 #include <thrust/system/cuda/detail/par_to_seq.h>
 #include <thrust/system/cuda/detail/util.h>
@@ -260,27 +262,18 @@ adjacent_difference(execution_policy<Derived> &policy,
                     OutputIt                   result,
                     BinaryOp                   binary_op)
 {
-  OutputIt ret = result;
-  if (__THRUST_HAS_CUDART__)
-  {
-    ret = __adjacent_difference::adjacent_difference(policy,
-        first,
-        last,
-        result,
-        binary_op);
-  }
-  else
-  {
-#if !__THRUST_HAS_CUDART__
-    ret = thrust::adjacent_difference(cvt_to_seq(derived_cast(policy)),
-                                      first,
-                                      last,
-                                      result,
-                                      binary_op);
-#endif
-  }
-
-  return ret;
+  THRUST_CDP_DISPATCH(
+    (result = __adjacent_difference::adjacent_difference(policy,
+                                                         first,
+                                                         last,
+                                                         result,
+                                                         binary_op);),
+    (result = thrust::adjacent_difference(cvt_to_seq(derived_cast(policy)),
+                                          first,
+                                          last,
+                                          result,
+                                          binary_op);));
+  return result;
 }
 
 template <class Derived,

thrust/system/cuda/detail/adjacent_difference.h

@@ -0,0 +1,72 @@
+/*
+*  Copyright 2021-2022 NVIDIA Corporation
+*
+*  Licensed under the Apache License, Version 2.0 (the "License");
+*  you may not use this file except in compliance with the License.
+*  You may obtain a copy of the License at
+*
+*      http://www.apache.org/licenses/LICENSE-2.0
+*
+*  Unless required by applicable law or agreed to in writing, software
+*  distributed under the License is distributed on an "AS IS" BASIS,
+*  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+*  See the License for the specific language governing permissions and
+*  limitations under the License.
+*/
+
+/**
+ * \file
+ * Utilities for CUDA dynamic parallelism.
+ */
+
+#pragma once
+
+#include <cub/config.cuh>
+#include <cub/detail/detect_cuda_runtime.cuh>
+
+#include <nv/target>
+
+/**
+ * \def THRUST_CDP_DISPATCH
+ *
+ * If CUDA Dynamic Parallelism / CUDA Nested Parallelism is available, always
+ * run the parallel implementation. Otherwise, run the parallel implementation
+ * when called from the host, and fallback to the sequential implementation on
+ * the device.
+ *
+ * `par_impl` and `seq_impl` are blocks of C++ statements enclosed in
+ * parentheses, similar to NV_IF_TARGET blocks:
+ *
+ * \code
+ * THRUST_CDP_DISPATCH((launch_parallel_kernel();), (run_serial_impl();));
+ * \endcode
+ */
+
+#ifdef THRUST_RDC_ENABLED
+
+// seq_impl unused.
+#define THRUST_CDP_DISPATCH(par_impl, seq_impl)                                \
+  NV_IF_TARGET(NV_ANY_TARGET, par_impl)
+
+#else // THRUST_RDC_ENABLED
+
+// Special case for NVCC -- need to inform the device path about the kernels
+// that are launched from the host path.
+#if defined(__CUDACC__) && defined(__CUDA_ARCH__)
+
+// Device-side launch not supported, fallback to sequential in device code.
+#define THRUST_CDP_DISPATCH(par_impl, seq_impl)                                \
+  if (false)                                                                   \
+  { /* Without this, the device pass won't compile any kernels. */             \
+    NV_IF_TARGET(NV_ANY_TARGET, par_impl);                                     \
+  }                                                                            \
+  NV_IF_TARGET(NV_IS_HOST, par_impl, seq_impl)
+
+#else // NVCC device pass
+
+#define THRUST_CDP_DISPATCH(par_impl, seq_impl)                                \
+  NV_IF_TARGET(NV_IS_HOST, par_impl, seq_impl)
+
+#endif // NVCC device pass
+
+#endif // THRUST_RDC_ENABLED

thrust/system/cuda/detail/cdp_dispatch.h

@@ -28,7 +28,10 @@
 
 #include <thrust/detail/config.h>
 
+#include <thrust/advance.h>
+
 #include <thrust/system/cuda/config.h>
+#include <thrust/system/cuda/detail/cdp_dispatch.h>
 #include <thrust/system/cuda/detail/execution_policy.h>
 #include <thrust/system/cuda/detail/cross_system.h>
 
@@ -117,22 +120,11 @@ copy(execution_policy<System> &system,
      InputIterator             last,
      OutputIterator            result)
 {
-  OutputIterator ret = result;
-  if (__THRUST_HAS_CUDART__)
-  {
-    ret = __copy::device_to_device(system, first, last, result);
-  }
-  else
-  {
-#if !__THRUST_HAS_CUDART__
-    ret = thrust::copy(cvt_to_seq(derived_cast(system)),
-                       first,
-                       last,
-                       result);
-#endif
-  }
-
-  return ret;
+  THRUST_CDP_DISPATCH(
+    (result = __copy::device_to_device(system, first, last, result);),
+    (result =
+       thrust::copy(cvt_to_seq(derived_cast(system)), first, last, result);));
+  return result;
 }    // end copy()
 
 __thrust_exec_check_disable__
@@ -146,19 +138,14 @@ copy_n(execution_policy<System> &system,
        Size                      n,
        OutputIterator            result)
 {
-  OutputIterator ret = result;
-  if (__THRUST_HAS_CUDART__)
-  {
-    ret = __copy::device_to_device(system, first, first + n, result);
-  }
-  else
-  {
-#if !__THRUST_HAS_CUDART__
-    ret = thrust::copy_n(cvt_to_seq(derived_cast(system)), first, n, result);
-#endif
-  }
-
-  return ret;
+  THRUST_CDP_DISPATCH(
+    (result = __copy::device_to_device(system,
+                                       first,
+                                       thrust::next(first, n),
+                                       result);),
+    (result =
+       thrust::copy_n(cvt_to_seq(derived_cast(system)), first, n, result);));
+  return result;
 } // end copy_n()
 #endif