Develop stream 2025 05 06

projects/rocprim/.gitlab-ci.yml

@@ -48,7 +48,7 @@ workflow:
       when: never
     - if: $CI_MERGE_REQUEST_TITLE !~ /Draft:/
       variables:
-        ROCPRIM_TEST_RUNS: 3
+        ROCPRIM_TEST_RUNS: 1
     - if: $CI_MERGE_REQUEST_TITLE =~ /Draft:/
       variables:
         ROCPRIM_TEST_RUNS: 1
@@ -124,7 +124,7 @@ copyright-date:
       -D AMDGPU_TARGETS=$GPU_TARGETS
       -D CMAKE_C_COMPILER_LAUNCHER=phc_sccache_c
       -D CMAKE_CXX_COMPILER_LAUNCHER=phc_sccache_cxx
-      -D CMAKE_CXX_STANDARD=14
+      -D CMAKE_CXX_STANDARD=17
       -S $CI_PROJECT_DIR
       -B $BUILD_DIR
     - cmake
@@ -174,6 +174,7 @@ build:cmake-minimum-apt:
     - .rules:build
   variables:
     EXTRA_CMAKE_CXX_FLAGS: ""
+    BUILD_TOOL_ARGS: ""
   script:
     - mkdir -p $BUILD_DIR
     - cd $BUILD_DIR
@@ -196,7 +197,7 @@ build:cmake-minimum-apt:
       -D CMAKE_CXX_STANDARD="$BUILD_VERSION"
       -S $CI_PROJECT_DIR
       -B $BUILD_DIR
-    - cmake --build $BUILD_DIR
+    - cmake --build $BUILD_DIR -- ${BUILD_TOOL_ARGS}
   artifacts:
     paths:
       - $BUILD_DIR/.ninja_log
@@ -213,6 +214,31 @@ build:cmake-minimum-apt:
       - $BUILD_DIR/test/test_*
     expire_in: 1 day
 
+build:spirv:
+  stage: build
+  needs: []
+  extends:
+    - .cmake-minimum
+    - .build:common
+  variables:
+    # For unknown reasons spir-v builds ignore 'clang diagnostic' pragmas that
+    # we use to ignore internal deprecations.
+    EXTRA_CMAKE_CXX_FLAGS: "-Wno-deprecated-declarations -mf16c -DROCPRIM_EXPERIMENTAL_SPIRV"
+    # Since not all targets are expected to build, do not stop building other
+    # targets when any target fails.
+    BUILD_TOOL_ARGS: "-k 0"
+    GPU_TARGETS: "amdgcnspirv"
+  image: "registry.streamhpc.internal/unstable-rocm:main"
+  allow_failure: true
+  parallel:
+    # Debug builds disabled due to excessive build times for debug test builds
+    matrix:
+      - BUILD_TYPE: Release
+        BUILD_TARGET: [BENCHMARK, TEST]
+        BUILD_VERSION: 17
+  artifacts:
+    when: always
+
 build:cmake-latest:
   stage: build
   needs: []
@@ -224,7 +250,7 @@ build:cmake-latest:
     matrix:
       - BUILD_TYPE: Release
         BUILD_TARGET: [BENCHMARK, TEST]
-        BUILD_VERSION: [14, 17]
+        BUILD_VERSION: 17
 
 build:cmake-minimum:
   needs: []
@@ -235,7 +261,7 @@ build:cmake-minimum:
     matrix:
       - BUILD_TYPE: [Debug, Release]
         BUILD_TARGET: [BENCHMARK, TEST]
-        BUILD_VERSION: 14
+        BUILD_VERSION: 17
 
 build:package:
   stage: build
@@ -252,7 +278,7 @@ build:package:
       -G Ninja
       -D CMAKE_CXX_COMPILER="$AMDCLANG"
       -D CMAKE_BUILD_TYPE=Release
-      -D CMAKE_CXX_STANDARD=14
+      -D CMAKE_CXX_STANDARD=17
       -B $PACKAGE_DIR
       -S $CI_PROJECT_DIR
     - cd $PACKAGE_DIR
@@ -285,7 +311,7 @@ build:windows:
       -D CMAKE_CXX_COMPILER:PATH="${env:HIP_PATH}\bin\clang++.exe"
       -D CMAKE_PREFIX_PATH:PATH="${env:HIP_PATH}"
       -D CMAKE_BUILD_TYPE="$BUILD_TYPE"
-      -D CMAKE_CXX_STANDARD=14
+      -D CMAKE_CXX_STANDARD=17
     - cmake --build "$CI_PROJECT_DIR/build"
   artifacts:
     paths:
@@ -332,7 +358,7 @@ autotune:build:
       -D AMDGPU_TARGETS=$GPU_TARGETS
       -D CMAKE_C_COMPILER_LAUNCHER=phc_sccache_c
       -D CMAKE_CXX_COMPILER_LAUNCHER=phc_sccache_cxx
-      -D CMAKE_CXX_STANDARD=14
+      -D CMAKE_CXX_STANDARD=17
     - cmake --build . --target $BENCHMARK_TARGETS
     - 'rm -rf $BUILD_DIR/benchmark/benchmark*.parallel'
     # The autotune benchmarks get very large, above GitLabs upload limit. Fortunately they compress well.
@@ -359,7 +385,7 @@ autotune:build:
         matrix:
           - BUILD_TYPE: Release
             BUILD_TARGET: TEST
-            BUILD_VERSION: 14
+            BUILD_VERSION: 17
   script:
     - cd $BUILD_DIR
     - cmake
@@ -398,6 +424,62 @@ test:all-gpus:
     - .test:common
     - .rules:test
 
+.test:common-spirv:
+  stage: test
+  tags:
+    - rocm
+    - $GPU
+  extends:
+    - .cmake-minimum
+  allow_failure: true
+  timeout: 3h
+  needs:
+    - job: build:spirv
+      parallel:
+        matrix:
+          - BUILD_TYPE: Release
+            BUILD_TARGET: TEST
+            BUILD_VERSION: 17
+  image: "registry.streamhpc.internal/unstable-rocm:main"
+  script:
+    - cd $BUILD_DIR
+    - cmake
+      -D CMAKE_PREFIX_PATH=/opt/rocm
+      -P $CI_PROJECT_DIR/cmake/GenerateResourceSpec.cmake
+    - cat ./resources.json
+    # Parallel execution (with other AMDGPU processes) can oversubscribe the SDMA queue.
+    # This causes the hipMemcpy to fail, which is not reported as an error by HIP.
+    # As a temporary workaround, disable the SDMA for test stability.
+    - HSA_ENABLE_SDMA=0 ctest
+      --output-on-failure
+      --repeat-until-fail 2
+      --resource-spec-file ./resources.json
+      --parallel $PARALLEL_JOBS
+      --exclude-regex rocprim.device_partition
+
+test:any-gpu-spirv:
+  variables:
+    GPU: ""
+    PARALLEL_JOBS: 1
+  extends:
+    - .test:common-spirv
+  rules:
+    - if: $CI_MERGE_REQUEST_TITLE =~ /Draft:/ && $CI_MERGE_REQUEST_LABELS !~ /Arch::/
+
+test:label-arch-spirv:
+  extends:
+    - .gpus:rocm
+    - .test:common-spirv
+    - .rules:arch-labels
+
+test:all-gpus-spirv:
+  variables:
+    SHOULD_BE_UNDRAFTED: "true"
+  extends:
+    - .gpus:rocm
+    - .test:common-spirv
+    - .rules:test
+
 .test-windows-base:
   stage: test
   extends:
@@ -437,7 +519,7 @@ test-windows-release:
       -D CMAKE_CXX_COMPILER="$AMDCLANG"
       -D CMAKE_BUILD_TYPE=Release
       -D AMDGPU_TARGETS=$GPU_TARGETS
-      -D CMAKE_CXX_STANDARD=14
+      -D CMAKE_CXX_STANDARD=17
       -S "$CI_PROJECT_DIR/test/extra"
       -B "$CI_PROJECT_DIR/package_test"
     - cmake --build "$CI_PROJECT_DIR/package_test"
@@ -459,7 +541,7 @@ test:install:
       -G Ninja
       -D CMAKE_CXX_COMPILER="$AMDCLANG"
       -D CMAKE_BUILD_TYPE=Release
-      -D CMAKE_CXX_STANDARD=14
+      -D CMAKE_CXX_STANDARD=17
       -B build
       -S $CI_PROJECT_DIR
     # Preserve $PATH when sudoing
@@ -507,7 +589,7 @@ benchmark:
         matrix:
           - BUILD_TYPE: Release
             BUILD_TARGET: BENCHMARK
-            BUILD_VERSION: 14
+            BUILD_VERSION: 17
   extends:
     - .cmake-minimum
     - .gpus:rocm

projects/rocprim/CHANGELOG.md

@@ -4,6 +4,11 @@ Full documentation for rocPRIM is available at [https://rocm.docs.amd.com/projec
 
 ## rocPRIM 3.6.0 for ROCm 7.0
 
+### Added
+Added `rocprim::accumulator_t` to ensure parity with CCCL.
+Added test for `rocprim::accumulator_t`
+Added `rocprim::invoke_result_r` to ensure parity with CCCL.
+
 ### Changed
 
 * The default scan accumulator types for device-level scan algorithms have changed. This is a breaking change.
@@ -24,6 +29,68 @@ This is a complete list of affected functions and how their default accumulator
     * past default: `class AccType = detail::input_type_t<InitValueType>>`
     * new default: `class AccType = rocprim::invoke_result_binary_op_t<rocprim::detail::input_type_t<InitValueType>, BinaryFunction>`
 
+* Added function `is_build_in` into `rocprim::traits::get`.
+* Changed the parameters `long_radix_bits` and `LongRadixBits` from `segmented_radix_sort` to `radix_bits` and `RadixBits` respectively.
+* Marked the initialisation constructor of `rocprim::reverse_iterator<Iter>` `explicit`, use `rocprim::make_reverse_iterator`.
+* Dropped `c++14` support for rocprim.
+* Merged `radix_key_codec` into type_traits system.
+* Renamed `type_traits_interface.hpp` to `type_traits.hpp`, rename the original `type_traits.hpp` to `type_traits_functions.hpp`.
+* Removed `radix_sort.hpp`
+* Added virtual shared memory as a fallback option in `rocprim::device_merge` when it exceeds shared memory capacity, similar to `rocprim::device_select`, `rocprim::device_partition`, and `rocprim::device_merge_sort`, which already include this feature.
+* Added initial value support to device level inclusive scans.
+* Made new optimization for `device_transform` when the input and output are pointers.
+* Added `LoadType` to `transform_config`, which is used for the `device_transform` when the input and output are pointers.
+* Changed the default accumulator type for various device-level scan algorithms:
+  * `rocprim::inclusive_scan`
+    * Previous default: `class AccType = typename std::iterator_traits<InputIterator>::value_type>`
+    * Current default: `class AccType = rocprim::accumulator_t<BinaryFunction, typename std::iterator_traits<InputIterator>::value_type>`
+  * `rocprim::deterministic_inclusive_scan`
+    * Previous default: `class AccType = typename std::iterator_traits<InputIterator>::value_type>`
+    * Current default: `class AccType = rocprim::accumulator_t<BinaryFunction, typename std::iterator_traits<InputIterator>::value_type>`
+  * `rocprim::exclusive_scan`
+    * Previous default: `class AccType = detail::input_type_t<InitValueType>>`
+    * Current default: `class AccType = rocprim::accumulator_t<BinaryFunction, rocprim::detail::input_type_t<InitValueType>>`
+  * `rocprim::deterministic_exclusive_scan`
+    * Previous default: `class AccType = detail::input_type_t<InitValueType>>`
+    * Current default: `class AccType = rocprim::accumulator_t<BinaryFunction, rocprim::detail::input_type_t<InitValueType>>`
+* Added `rocprim:device_transform` for n-ary transform operations API with as input `n` number of iterators inside a `rocprim::tuple`.
+
+### Deprecations
+
+* `rocprim::invoke_result_binary_op` and `rocprim::invoke_result_binary_op_t` are deprecated. Use `rocprim::accumulator_t` now.
+
+### Removed
+
+* Removed `rocprim::detail::float_bit_mask` and relative tests, use `rocprim::traits::float_bit_mask` instead.
+* Removed `rocprim::traits::is_fundamental`, please use `rocprim::traits::get<T>::is_fundamental()` directly.
+* Removed the deprecated parameters `short_radix_bits` and `ShortRadixBits` from the `segmented_radix_sort` config. They were unused, it is only an API change.
+* Removed the deprecated `operator<<` from the iterators.
+* Removed the deprecated `TwiddleIn` and `TwiddleOut`. Use `radix_key_codec` instead.
+* Removed the deprecated flags API of `block_adjacent_difference`. Use `subtract_left()` or `block_discontinuity::flag_heads()` instead.
+* Removed the deprecated `to_exclusive` functions in the warp scans.
+* Removed the `rocprim::load_cs` from the `cache_load_modifier` enum. Use `rocprim::load_nontemporal` instead.
+* Removed the `rocprim::store_cs` from the `cache_store_modifier` enum. Use `rocprim::store_nontemporal` instead.
+* Removed the deprecated header file `rocprim/detail/match_result_type.hpp`. Include `rocprim/type_traits.hpp` instead.
+  * This header included `rocprim::detail::invoke_result`. Use `rocprim::invoke_result` instead.
+  * This header included `rocprim::detail::invoke_result_binary_op`. Use `rocprim::invoke_result_binary_op` instead.
+  * This header included `rocprim::detail::match_result_type`. Use `rocprim::invoke_result_binary_op_t` instead.
+* Removed the deprecated `rocprim::detail::radix_key_codec` function. Use `rocprim::radix_key_codec` instead.
+* Due to the removal of `__AMDGCN_WAVEFRONT_SIZE` in the compiler, the following deprecated warp size-related symbols have been removed:
+  * `rocprim::device_warp_size()`
+    * For compile-time constants, this is replaced with `rocprim::arch::wavefront::min_size()` and `rocprim::arch::wavefront::max_size()`. Use this when allocating global or shared memory.
+    * For run-time constants, this is replaced with `rocprim::arch::wavefront::size().`
+  * `rocprim::warp_size()`
+    * Use `rocprim::host_warp_size()`, `rocprim::arch::wavefront::min_size()` or `rocprim::arch::wavefront::max_size()` instead.
+  * `ROCPRIM_WAVEFRONT_SIZE`
+    * Use `rocprim::arch::wavefront::min_size()` or `rocprim::arch::wavefront::max_size()` instead.
+  * `__AMDGCN_WAVEFRONT_SIZE`
+    * This was a fallback define for the compiler's removed symbol, having the same name. 
+
+### Resolved issues
+
+* Fixed `device_batch_memcpy` its reported benchmarking throughput being 2x lower than it was in reality.
+* Fixed `device_segmented_reduce` its reported autotuning throughput being 5x lower than it was in reality.
+
 ## rocPRIM 3.5.0 for ROCm 6.5.0
 
 ### Removed
@@ -40,6 +107,7 @@ This is a complete list of affected functions and how their default accumulator
 * Added the `rocprim::merge_inplace` function for merging in-place.
 * Added initial value support for warp- and block-level inclusive scan.
 * Added support for building tests with device-side random data generation, making them finish faster. This requires rocRAND, and is enabled with the `WITH_ROCRAND=ON` build flag.
+* Added tests and documentation to `lookback_scan_state`. It is still in the `detail` namespace.
 
 ### Changed
 
@@ -599,3 +667,5 @@ The following is the complete list of affected functions and how their default a
 
 * Switched to HIP-Clang as the default compiler
 * CMake searches for rocPRIM locally first; if t's not found, CMake downloads it from GitHub
+
+

projects/rocprim/CMakeLists.txt

@@ -41,11 +41,8 @@ set(CMAKE_HIP_STANDARD 14)
 set(CMAKE_HIP_STANDARD_REQUIRED ON)
 set(CMAKE_HIP_EXTENSIONS OFF)
 
-# Set CXX standard
-if (CMAKE_CXX_STANDARD EQUAL 14)
-  message(WARNING "C++14 will be deprecated in the next major release")
-elseif(NOT CMAKE_CXX_STANDARD EQUAL 17)
-  message(FATAL_ERROR "Only C++14 and C++17 are supported")
+if(NOT CMAKE_CXX_STANDARD EQUAL 17)
+  message(FATAL_ERROR "Only C++17 is supported")
 endif()
 
 if (CMAKE_CURRENT_SOURCE_DIR STREQUAL CMAKE_SOURCE_DIR)

projects/rocprim/benchmark/CMakeLists.txt

@@ -1,6 +1,6 @@
 # MIT License
 #
-# Copyright (c) 2017-2024 Advanced Micro Devices, Inc. All rights reserved.
+# Copyright (c) 2017-2025 Advanced Micro Devices, Inc. All rights reserved.
 #
 # Permission is hereby granted, free of charge, to any person obtaining a copy
 # of this software and associated documentation files (the "Software"), to deal
@@ -174,6 +174,7 @@ add_rocprim_benchmark(benchmark_device_segmented_radix_sort_keys.cpp)
 add_rocprim_benchmark(benchmark_device_segmented_radix_sort_pairs.cpp)
 add_rocprim_benchmark(benchmark_device_segmented_reduce.cpp)
 add_rocprim_benchmark(benchmark_device_transform.cpp)
+add_rocprim_benchmark(benchmark_device_transform_pointer.cpp)
 add_rocprim_benchmark(benchmark_predicate_iterator.cpp)
 add_rocprim_benchmark(benchmark_warp_exchange.cpp)
 add_rocprim_benchmark(benchmark_warp_reduce.cpp)

projects/rocprim/benchmark/ConfigAutotuneSettings.cmake

@@ -85,18 +85,23 @@ ${TUNING_TYPES};${LIMITED_TUNING_TYPES};using_warp_scan reduce_then_scan" PARENT
     set(list_across "\
 binary_search upper_bound lower_bound;${TUNING_TYPES};${LIMITED_TUNING_TYPES};64 128 256;1 2 4 8 16" PARENT_SCOPE)
     set(output_pattern_suffix "@SubAlgorithm@_@ValueType@_@OutputType@_@BlockSize@_@ItemsPerThread@" PARENT_SCOPE)
+  elseif(file STREQUAL "benchmark_device_search_n")
+    set(list_across_names "InputType;BlockSize;ItemsPerThread;Threshold" PARENT_SCOPE)
+    set(list_across "\
+${TUNING_TYPES};64 128 256 512 1024;1 2 4 8 16;4 8 12 16" PARENT_SCOPE)
+    set(output_pattern_suffix "@InputType@_@BlockSize@_@ItemsPerThread@_@Threshold@" PARENT_SCOPE)
   elseif(file STREQUAL "benchmark_device_segmented_radix_sort_keys")
     set(list_across_names "\
-KeyType;LongBits;BlockSize;ItemsPerThread;WarpSmallLWS;WarpSmallIPT;WarpSmallBS;WarpPartition;WarpMediumLWS;WarpMediumIPT;WarpMediumBS" PARENT_SCOPE)
+KeyType;RadixBits;BlockSize;ItemsPerThread;WarpSmallLWS;WarpSmallIPT;WarpSmallBS;WarpPartition;WarpMediumLWS;WarpMediumIPT;WarpMediumBS" PARENT_SCOPE)
     set(list_across "${TUNING_TYPES};8;256;4 8 16;8;4;256;64;16;8;256" PARENT_SCOPE)
     set(output_pattern_suffix "\
-@KeyType@_@LongBits@_@BlockSize@_@ItemsPerThread@_@WarpSmallLWS@_@WarpSmallIPT@_@WarpSmallBS@_@WarpPartition@_@WarpMediumLWS@_@WarpMediumIPT@_@WarpMediumBS@" PARENT_SCOPE)
+@KeyType@_@RadixBits@_@BlockSize@_@ItemsPerThread@_@WarpSmallLWS@_@WarpSmallIPT@_@WarpSmallBS@_@WarpPartition@_@WarpMediumLWS@_@WarpMediumIPT@_@WarpMediumBS@" PARENT_SCOPE)
   elseif(file STREQUAL "benchmark_device_segmented_radix_sort_pairs")
     set(list_across_names "\
-KeyType;ValueType;LongBits;BlockSize;ItemsPerThread;WarpSmallLWS;WarpSmallIPT;WarpSmallBS;WarpPartition;WarpMediumLWS;WarpMediumIPT;WarpMediumBS" PARENT_SCOPE)
+KeyType;ValueType;RadixBits;BlockSize;ItemsPerThread;WarpSmallLWS;WarpSmallIPT;WarpSmallBS;WarpPartition;WarpMediumLWS;WarpMediumIPT;WarpMediumBS" PARENT_SCOPE)
     set(list_across "${TUNING_TYPES};${LIMITED_TUNING_TYPES};8;256;4 8 16;8;4;256;64;16;8;256" PARENT_SCOPE)
     set(output_pattern_suffix "\
-@KeyType@_@ValueType@_@LongBits@_@BlockSize@_@ItemsPerThread@_@WarpSmallLWS@_@WarpSmallIPT@_@WarpSmallBS@_@WarpPartition@_@WarpMediumLWS@_@WarpMediumIPT@_@WarpMediumBS@" PARENT_SCOPE)
+@KeyType@_@ValueType@_@RadixBits@_@BlockSize@_@ItemsPerThread@_@WarpSmallLWS@_@WarpSmallIPT@_@WarpSmallBS@_@WarpPartition@_@WarpMediumLWS@_@WarpMediumIPT@_@WarpMediumBS@" PARENT_SCOPE)
 elseif(file STREQUAL "benchmark_device_segmented_reduce")
     set(list_across_names "DataType;BlockSize;ItemsPerThread" PARENT_SCOPE)
     set(list_across "\
@@ -108,6 +113,12 @@ DataType;BlockSize;" PARENT_SCOPE)
     set(list_across "${TUNING_TYPES};64 128 256 512 1024" PARENT_SCOPE)
     set(output_pattern_suffix "\
 @DataType@_@BlockSize@" PARENT_SCOPE)
+elseif(file STREQUAL "benchmark_device_transform_pointer")
+    set(list_across_names "\
+DataType;BlockSize;LoadType" PARENT_SCOPE)
+    set(list_across "${TUNING_TYPES};64 128 256 512 1024;rocprim::load_default rocprim::load_nontemporal" PARENT_SCOPE)
+    set(output_pattern_suffix "\
+@DataType@_@BlockSize@_@LoadType@" PARENT_SCOPE)
   elseif(file STREQUAL "benchmark_device_partition")
     set(list_across_names "DataType;BlockSize" PARENT_SCOPE)
     set(list_across "${TUNING_TYPES};128 192 256 384 512" PARENT_SCOPE)