Further testing windows

.github/workflows/build-windows.yml

@@ -119,6 +119,7 @@ jobs:
 
       - name: Build NVBench
         env:
+          NVBENCH_WINDOWS_CUDA: ${{ inputs.cuda }}
           NVBENCH_WINDOWS_STD: ${{ inputs.std }}
           NVBENCH_WINDOWS_ARCH: ${{ inputs.arch }}
         run: |
@@ -131,7 +132,8 @@ jobs:
           @"
           `$ErrorActionPreference = 'Stop'
           git config --global --add safe.directory '$containerRepo'
-          & '$containerRepo/ci/windows/build_nvbench.ps1' -std '$env:NVBENCH_WINDOWS_STD' -arch '$env:NVBENCH_WINDOWS_ARCH'
+          & '$containerRepo/ci/windows/install_cuda_profiler_api.ps1' -cudaVersion "`$env:NVBENCH_WINDOWS_CUDA"
+          & '$containerRepo/ci/windows/build_nvbench.ps1' -std "`$env:NVBENCH_WINDOWS_STD" -arch "`$env:NVBENCH_WINDOWS_ARCH"
           exit `$LASTEXITCODE
           "@ | Set-Content -Path $script -Encoding UTF8
 
@@ -152,6 +154,9 @@ jobs:
             "--env", "GITHUB_REPOSITORY=$env:GITHUB_REPOSITORY",
             "--env", "GITHUB_RUN_ID=$env:GITHUB_RUN_ID",
             "--env", "GITHUB_SHA=$env:GITHUB_SHA",
+            "--env", "NVBENCH_WINDOWS_ARCH=$env:NVBENCH_WINDOWS_ARCH",
+            "--env", "NVBENCH_WINDOWS_CUDA=$env:NVBENCH_WINDOWS_CUDA",
+            "--env", "NVBENCH_WINDOWS_STD=$env:NVBENCH_WINDOWS_STD",
             "--env", "SCCACHE_BUCKET=$env:SCCACHE_BUCKET",
             "--env", "SCCACHE_IDLE_TIMEOUT=$env:SCCACHE_IDLE_TIMEOUT",
             "--env", "SCCACHE_REGION=$env:SCCACHE_REGION",

.github/workflows/pr.yml

@@ -79,8 +79,6 @@ jobs:
 
   nvbench-windows:
     name: NVBench Windows CUDA${{ matrix.config.cuda }} ${{ matrix.config.host }} C++${{ matrix.config.std }}
-    # TODO: Re-enable after https://github.com/NVIDIA/nvbench/pull/354 fixes the Windows build.
-    if: false
     permissions:
       id-token: write
       contents: read

CMakeLists.txt

@@ -39,6 +39,9 @@ if (${CUDAToolkit_VERSION} VERSION_LESS 11.3)
 endif()
 
 option(BUILD_SHARED_LIBS "Build NVBench as a shared library" ON)
+if (WIN32 AND BUILD_SHARED_LIBS)
+  set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+endif()
 
 option(NVBench_ENABLE_NVML "Build with NVML support from the Cuda Toolkit." ON)
 option(NVBench_ENABLE_CUPTI "Build NVBench with CUPTI." ${cupti_default})

ci/windows/install_cuda_profiler_api.ps1

@@ -0,0 +1,181 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+Param(
+    [Parameter(Mandatory = $false)]
+    [Alias("cudaVersion")]
+    [string]$CUDA_VERSION = ""
+)
+
+$ErrorActionPreference = "Stop"
+
+function Get-CudaVersionFromPath {
+    Param(
+        [Parameter(Mandatory = $false)]
+        [string]$Path = ""
+    )
+
+    if ($Path -and $Path -match "v(?<version>\d+\.\d+)[\\/]?$") {
+        return $Matches.version
+    }
+
+    return ""
+}
+
+function Assert-NvidiaAuthenticodeSignature {
+    Param(
+        [Parameter(Mandatory = $true)]
+        [ValidateNotNullOrEmpty()]
+        [string]$Path
+    )
+
+    $signature = Get-AuthenticodeSignature -FilePath $Path
+    if ($signature.Status -ne "Valid") {
+        throw "Invalid Authenticode signature for '$Path': $($signature.Status) $($signature.StatusMessage)"
+    }
+
+    $expectedPublisher = "NVIDIA Corporation"
+    $publisher = $signature.SignerCertificate.GetNameInfo(
+        [System.Security.Cryptography.X509Certificates.X509NameType]::SimpleName,
+        $false
+    )
+    if ($publisher -ne $expectedPublisher) {
+        throw "Unexpected signer for '$Path': $publisher"
+    }
+
+    Write-Host "Validated Authenticode signature for '$Path': $publisher"
+}
+
+function Get-HttpStatusCodeFromError {
+    Param(
+        [Parameter(Mandatory = $true)]
+        $ErrorRecord
+    )
+
+    $responseProperty = $ErrorRecord.Exception.PSObject.Properties["Response"]
+    if (-not $responseProperty) {
+        return $null
+    }
+
+    $response = $responseProperty.Value
+    if ($null -eq $response) {
+        return $null
+    }
+
+    $statusCodeProperty = $response.PSObject.Properties["StatusCode"]
+    if (-not $statusCodeProperty) {
+        return $null
+    }
+
+    return [int]$statusCodeProperty.Value
+}
+
+function Invoke-WebRequestWithRetry {
+    Param(
+        [Parameter(Mandatory = $true)]
+        [ValidateNotNullOrEmpty()]
+        [string]$Uri,
+
+        [Parameter(Mandatory = $true)]
+        [ValidateNotNullOrEmpty()]
+        [string]$OutFile,
+
+        [Parameter(Mandatory = $false)]
+        [ValidateRange(1, 10)]
+        [int]$MaxAttempts = 3
+    )
+
+    for ($attempt = 1; $attempt -le $MaxAttempts; $attempt++) {
+        try {
+            Remove-Item $OutFile -ErrorAction SilentlyContinue
+            Invoke-WebRequest -Uri $Uri -OutFile $OutFile -UseBasicParsing -TimeoutSec 300
+            return
+        } catch {
+            $statusCode = Get-HttpStatusCodeFromError -ErrorRecord $_
+            if ($statusCode -ge 400 -and $statusCode -lt 500) {
+                throw "Download failed with non-retryable HTTP status $statusCode from '$Uri'. $_"
+            }
+
+            if ($attempt -eq $MaxAttempts) {
+                throw
+            }
+
+            $delaySeconds = 5 * $attempt
+            Write-Warning "Download failed on attempt $attempt of $MaxAttempts. Retrying in $delaySeconds seconds. $_"
+            Start-Sleep -Seconds $delaySeconds
+        }
+    }
+}
+
+if (-not $CUDA_VERSION) {
+    $CUDA_VERSION = Get-CudaVersionFromPath -Path $env:CUDA_PATH
+    if (-not $CUDA_VERSION) {
+        throw "Could not determine CUDA version. Provide -cudaVersion or set CUDA_PATH to a path ending in v<major>.<minor>."
+    }
+}
+
+$version = [Version]$CUDA_VERSION
+$major = $version.Major
+$minor = $version.Minor
+$build = $version.Build
+
+if ($build -lt 0) {
+    $build = 0
+}
+
+$mmbVersionTag = "${major}.${minor}.${build}"
+$mmVersionTag = "${major}.${minor}"
+
+if ($env:CUDA_PATH) {
+    $cudaPathVersion = Get-CudaVersionFromPath -Path $env:CUDA_PATH
+    if (-not $cudaPathVersion) {
+        throw "CUDA_PATH is set but does not end in v<major>.<minor>: $env:CUDA_PATH"
+    }
+    if ($cudaPathVersion -ne $mmVersionTag) {
+        throw "CUDA_PATH points to CUDA $cudaPathVersion, but CUDA $mmVersionTag was requested."
+    }
+    $cudaRoot = $env:CUDA_PATH
+} else {
+    $cudaRoot = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v$mmVersionTag"
+}
+$profilerHeader = "$cudaRoot\include\cuda_profiler_api.h"
+
+if (Test-Path $profilerHeader) {
+    Write-Host "CUDA Profiler API is already installed: $profilerHeader"
+    return
+}
+
+$component = "cuda_profiler_api_$mmVersionTag"
+$cudaMajorUri = "${mmbVersionTag}/network_installers/cuda_${mmbVersionTag}_windows_network.exe"
+$cudaVersionUrl = "https://developer.download.nvidia.com/compute/cuda/$cudaMajorUri"
+$installer = Join-Path $env:TEMP "cuda_${mmbVersionTag}_windows_network.exe"
+
+Write-Host "Installing CUDA component: $component"
+Write-Host "Downloading CUDA network installer: $cudaVersionUrl"
+Invoke-WebRequestWithRetry -Uri $cudaVersionUrl -OutFile $installer
+Assert-NvidiaAuthenticodeSignature -Path $installer
+
+$installerTimeoutSeconds = 900
+$process = $null
+try {
+    $process = Start-Process -PassThru -FilePath $installer -ArgumentList @("-s", $component)
+    if (-not $process.WaitForExit($installerTimeoutSeconds * 1000)) {
+        Stop-Process -Id $process.Id -Force -ErrorAction SilentlyContinue
+        throw "CUDA network installer timed out after $installerTimeoutSeconds seconds."
+    }
+
+    if ($process.ExitCode -ne 0) {
+        throw "CUDA network installer failed with exit code $($process.ExitCode)."
+    }
+} finally {
+    if ($process) {
+        $process.Dispose()
+    }
+    Remove-Item $installer -ErrorAction SilentlyContinue
+}
+
+if (-not (Test-Path $profilerHeader)) {
+    throw "CUDA Profiler API installation completed, but header was not found: $profilerHeader"
+}
+
+Write-Host "CUDA Profiler API installed: $profilerHeader"

cmake/NVBenchCUPTI.cmake

@@ -16,21 +16,84 @@ endif()
 # - The dll locations are not specified
 # - Dependent libraries nvperf_* are not linked.
 # So we create our own targets:
+function(nvbench_find_windows_cupti_runtime_library out_var dep_name library_path)
+  cmake_path(GET library_path PARENT_PATH library_dir)
+  set(runtime_search_dirs "${library_dir}")
+
+  if ("${library_dir}" MATCHES "/Library/lib/x64$")
+    cmake_path(GET library_dir PARENT_PATH conda_lib_dir)
+    cmake_path(GET conda_lib_dir PARENT_PATH conda_library_dir)
+    list(APPEND runtime_search_dirs "${conda_library_dir}/bin")
+  elseif ("${library_dir}" MATCHES "/Library/lib$")
+    cmake_path(GET library_dir PARENT_PATH conda_library_dir)
+    list(APPEND runtime_search_dirs "${conda_library_dir}/bin")
+  endif()
+
+  list(REMOVE_DUPLICATES runtime_search_dirs)
+
+  foreach(runtime_search_dir IN LISTS runtime_search_dirs)
+    if ("${dep_name}" STREQUAL "cupti")
+      file(GLOB runtime_libraries LIST_DIRECTORIES false
+        "${runtime_search_dir}/cupti64_*.dll"
+      )
+      if (NOT runtime_libraries)
+        file(GLOB runtime_libraries LIST_DIRECTORIES false
+          "${runtime_search_dir}/cupti.dll"
+        )
+      endif()
+    else()
+      file(GLOB runtime_libraries LIST_DIRECTORIES false
+        "${runtime_search_dir}/${dep_name}.dll"
+      )
+    endif()
+
+    if (runtime_libraries)
+      list(LENGTH runtime_libraries num_runtime_libraries)
+      if (num_runtime_libraries GREATER 1)
+        message(FATAL_ERROR
+          "Found multiple runtime DLLs for ${dep_name}: ${runtime_libraries}"
+        )
+      endif()
+
+      list(GET runtime_libraries 0 runtime_library)
+      set(${out_var} "${runtime_library}" PARENT_SCOPE)
+      return()
+    endif()
+  endforeach()
+
+  message(FATAL_ERROR
+    "Could not find the runtime DLL for ${dep_name}. "
+    "Searched these directories: ${runtime_search_dirs}"
+  )
+endfunction()
+
 function(nvbench_add_cupti_dep dep_name)
   string(TOLOWER ${dep_name} dep_name_lower)
   string(TOUPPER ${dep_name} dep_name_upper)
 
   add_library(nvbench::${dep_name_lower} SHARED IMPORTED)
 
   find_library(NVBench_${dep_name_upper}_LIBRARY ${dep_name_lower} REQUIRED
-    DOC "The full path to lib${dep_name_lower}.so from the CUDA Toolkit."
+    DOC "The import library for ${dep_name_lower} from the CUDA Toolkit."
     HINTS "${nvbench_cupti_root}/lib64"
   )
   mark_as_advanced(NVBench_${dep_name_upper}_LIBRARY)
 
-  set_target_properties(nvbench::${dep_name_lower} PROPERTIES
-    IMPORTED_LOCATION "${NVBench_${dep_name_upper}_LIBRARY}"
-  )
+  if (WIN32)
+    nvbench_find_windows_cupti_runtime_library(
+      NVBench_${dep_name_upper}_DLL
+      ${dep_name_lower}
+      "${NVBench_${dep_name_upper}_LIBRARY}"
+    )
+    set_target_properties(nvbench::${dep_name_lower} PROPERTIES
+      IMPORTED_IMPLIB "${NVBench_${dep_name_upper}_LIBRARY}"
+      IMPORTED_LOCATION "${NVBench_${dep_name_upper}_DLL}"
+    )
+  else()
+    set_target_properties(nvbench::${dep_name_lower} PROPERTIES
+      IMPORTED_LOCATION "${NVBench_${dep_name_upper}_LIBRARY}"
+    )
+  endif()
 endfunction()
 
 nvbench_add_cupti_dep(cupti)

cmake/NVBenchConfigTarget.cmake

@@ -91,11 +91,25 @@ endif()
 if (CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
   # fmtlib uses llvm's _BitInt internally, which is not available when compiling through nvcc:
   target_compile_definitions(nvbench.build_interface INTERFACE "FMT_USE_BITINT=0")
+  if (MSVC)
+    # cudafe cannot evaluate fmtlib's UTF-8 literal check even when /utf-8 is passed to the host compiler:
+    target_compile_definitions(nvbench.build_interface INTERFACE
+      $<$<COMPILE_LANGUAGE:CUDA>:FMT_UNICODE=0>
+    )
+  endif()
 endif()
 
 target_compile_options(nvbench.build_interface INTERFACE
   $<$<COMPILE_LANG_AND_ID:CUDA,NVIDIA>:-Xcudafe=--display_error_number>
   $<$<COMPILE_LANG_AND_ID:CUDA,NVIDIA>:-Wno-deprecated-gpu-targets>
+  $<$<AND:$<COMPILE_LANG_AND_ID:CUDA,NVIDIA>,$<CXX_COMPILER_ID:MSVC>>:-Xcompiler=/utf-8>
+  # Suppress cudafe diagnostics triggered by fmtlib headers when compiled through MSVC+nvcc:
+  #   27: character value is out of range (char32_t sentinel values in lookup tables)
+  #  128: loop is not reachable (dead code in constexpr string comparison)
+  # 2417: constexpr constructor calls non-constexpr function (bigint default ctor)
+  $<$<AND:$<COMPILE_LANG_AND_ID:CUDA,NVIDIA>,$<CXX_COMPILER_ID:MSVC>>:-Xcudafe=--diag_suppress=27>
+  $<$<AND:$<COMPILE_LANG_AND_ID:CUDA,NVIDIA>,$<CXX_COMPILER_ID:MSVC>>:-Xcudafe=--diag_suppress=128>
+  $<$<AND:$<COMPILE_LANG_AND_ID:CUDA,NVIDIA>,$<CXX_COMPILER_ID:MSVC>>:-Xcudafe=--diag_suppress=2417>
 )
 if (NVBench_ENABLE_WERROR)
   target_compile_options(nvbench.build_interface INTERFACE
@@ -115,8 +129,8 @@ function(nvbench_config_target target_name)
   # the library path, other times they're in a subdirectory that isn't added to
   # the library path...
   # To simplify installed nvbench usage, add the CUPTI libraries path to the
-  # installed nvbench rpath:
-  if (NVBench_ENABLE_CUPTI AND nvbench_cupti_root)
+  # installed nvbench rpath (Unix only; Windows uses PATH for DLL lookup):
+  if (NVBench_ENABLE_CUPTI AND nvbench_cupti_root AND NOT WIN32)
     set_target_properties(${target_name} PROPERTIES
       INSTALL_RPATH "${nvbench_cupti_root}/lib64"
     )

nvbench/CMakeLists.txt

@@ -134,6 +134,10 @@ nvbench_config_target(nvbench.main)
 target_compile_definitions(nvbench.main PRIVATE NVBENCH_NO_IMPLICIT_SYSTEM_HEADER)
 # Propagate `nvbench` to consumers but keep NVBench's own build warning-visible.
 target_link_libraries(nvbench.main PUBLIC nvbench)
+if (MSVC)
+  # inform MSVC that library provides main
+  target_link_options(nvbench.main INTERFACE "/INCLUDE:main")
+endif()
 # Ensure CUDA/CUPTI/NVML include dirs are visible for nvbench.main's build.
 target_link_libraries(nvbench.main PRIVATE ${ctk_libraries})
 # Add NVBench's headers privately so the main library itself sees warnings.

nvbench/config.cuh.in

@@ -24,7 +24,11 @@
 // Defined if NVBench has been built with CUPTI support.
 #cmakedefine NVBENCH_HAS_CUPTI
 
+#if defined(_MSVC_LANG)
+#define NVBENCH_CPLUSPLUS _MSVC_LANG
+#else
 #define NVBENCH_CPLUSPLUS __cplusplus
+#endif
 
 // Detect current dialect:
 #if NVBENCH_CPLUSPLUS < 201703L

testing/axes_metadata.cu

@@ -24,6 +24,7 @@
 #include <fmt/format.h>
 
 #include <algorithm>
+#include <iterator>
 #include <string_view>
 
 #include "test_asserts.cuh"