cuda: enable CUDA feature extraction on Windows (MSYS2/MinGW)

libvmaf/src/cuda/common.h

@@ -20,7 +20,6 @@
 #ifndef __VMAF_SRC_CUDA_COMMON_H__
 #define __VMAF_SRC_CUDA_COMMON_H__
 
-#include <pthread.h>
 #include <stdbool.h>
 
 #include "config.h"

libvmaf/src/cuda/cuda_helper.cuh

@@ -29,7 +29,11 @@
 
 #include "assert.h"
 #include "stdio.h"
+#ifdef DEVICE_CODE
+#include <cuda.h>
+#else
 #include <ffnvcodec/dynlink_loader.h>
+#endif
 
 #define DIV_ROUND_UP(x, y) (((x) + (y)-1) / (y))
 #define MAX(x, y) (((x) > (y)) ? (x) : (y))

libvmaf/src/feature/cuda/integer_adm/adm_cm.cu

@@ -16,7 +16,9 @@
  *     limitations under the License.
  *
  */
+#ifndef DEVICE_CODE
 #include "feature_collector.h"
+#endif
 #include "cuda/integer_adm_cuda.h"
 #include "common.h"
 #include "cuda_helper.cuh"

libvmaf/src/feature/cuda/integer_adm/adm_csf.cu

@@ -17,7 +17,9 @@
  *
  */
 
+#ifndef DEVICE_CODE
 #include "feature_collector.h"
+#endif
 #include "cuda/integer_adm_cuda.h"
 
 #include "common.h"

libvmaf/src/feature/cuda/integer_adm/adm_csf_den.cu

@@ -17,7 +17,9 @@
  *
  */
 
+#ifndef DEVICE_CODE
 #include "feature_collector.h"
+#endif
 #include "cuda/integer_adm_cuda.h"
 
 #include "common.h"

libvmaf/src/feature/cuda/integer_adm/adm_decouple.cu

@@ -17,7 +17,9 @@
  *
  */
 
+#ifndef DEVICE_CODE
 #include "feature_collector.h"
+#endif
 #include "cuda/integer_adm_cuda.h"
 
 #include "common.h"

libvmaf/src/feature/cuda/integer_adm/adm_dwt2.cu

@@ -17,7 +17,9 @@
  *
  */
 
+#ifndef DEVICE_CODE
 #include "feature_collector.h"
+#endif
 #include "cuda/integer_adm_cuda.h"
 
 #include "common.h"

libvmaf/src/feature/integer_adm.h

@@ -132,10 +132,12 @@ struct dwt_model_params {
 };
 
 // 0 -> Y, 1 -> Cb, 2 -> Cr
+#ifndef __CUDACC__
 static const struct dwt_model_params dwt_7_9_YCbCr_threshold[3] = {
     {.a = 0.495, .k = 0.466, .f0 = 0.401, .g = {1.501, 1.0, 0.534, 1.0}},
     {.a = 1.633, .k = 0.353, .f0 = 0.209, .g = {1.520, 1.0, 0.502, 1.0}},
     {.a = 0.944, .k = 0.521, .f0 = 0.404, .g = {1.868, 1.0, 0.516, 1.0}}};
+#endif
 
 /*
  * The following dwt basis function amplitudes, A(lambda,theta), are taken from

libvmaf/src/meson.build

@@ -34,8 +34,12 @@ is_asm_enabled = get_option('enable_asm') == true
 is_cuda_enabled = get_option('enable_cuda') == true
 is_avx512_enabled = get_option('enable_avx512') == true
 is_nvtx_enabled = get_option('enable_nvtx') == true
+if is_nvtx_enabled and host_machine.system() == 'windows'
+    warning('NVTX is not supported on Windows, disabling.')
+    is_nvtx_enabled = false
+endif
 
-if is_nvtx_enabled    
+if is_nvtx_enabled
     cdata.set10('HAVE_NVTX', is_nvtx_enabled)
 endif
 if is_cuda_enabled
@@ -310,23 +314,80 @@ if is_cuda_enabled
     ]
     gencode = []
     if get_option('enable_nvcc')
-        cuda_lang = add_languages('cuda', required : true)
-        cuda_compiler = meson.get_compiler('cuda')
         nvcc_exe = find_program('nvcc')
 
+        # On Windows, nvcc requires MSVC's cl.exe as host compiler for preprocessing.
+        # We avoid adding cl.exe to PATH as that causes meson to pick MSVC as the
+        # default C compiler. Use vswhere + powershell to find cl.exe automatically.
+        if host_machine.system() == 'windows'
+            cl_find_result = run_command('powershell', '-NoProfile', '-Command',
+                '(Get-ChildItem -Path (& "${env:ProgramFiles(x86)}/Microsoft Visual Studio/Installer/vswhere.exe" -latest -products * -property installationPath) -Recurse -Filter cl.exe -ErrorAction SilentlyContinue | Where-Object { $_.DirectoryName -like "*HostX64*x64*" } | Select-Object -First 1).FullName',
+                check: false)
+            if cl_find_result.returncode() == 0 and cl_find_result.stdout().strip() != ''
+                cl_path = cl_find_result.stdout().strip()
+                message('Found MSVC cl.exe at: ' + cl_path)
+                nvcc_ccbin_flags = ['--allow-unsupported-compiler', '-ccbin', cl_path]
+            else
+                cl_exe = find_program('cl', required: false)
+                if cl_exe.found()
+                    nvcc_ccbin_flags = ['--allow-unsupported-compiler', '-ccbin', cl_exe.full_path()]
+                else
+                    error('MSVC cl.exe not found. nvcc on Windows requires Visual Studio Build Tools.')
+                endif
+            endif
+
+            # Find MSVC and Windows SDK include dirs so cl.exe can find system
+            # headers when invoked by nvcc outside a vcvars environment.
+            # Pass them as -I flags directly to nvcc.
+            # Derive MSVC include path from cl.exe's location to ensure version match.
+            msvc_include_result = run_command('powershell', '-NoProfile', '-Command',
+                '$clPath = "' + cl_path + '"; ' +
+                '$msvcRoot = ($clPath -replace "\\\\bin\\\\.*$", ""); ' +
+                'Write-Output $msvcRoot',
+                check: false)
+            winsdk_include_result = run_command('powershell', '-NoProfile', '-Command',
+                '$sdkInc = Get-ChildItem "${env:ProgramFiles(x86)}/Windows Kits/10/Include" -Directory | Sort-Object Name | Select-Object -Last 1; ' +
+                'Write-Output $sdkInc.FullName',
+                check: false)
+            nvcc_host_includes = []
+            if msvc_include_result.returncode() == 0 and winsdk_include_result.returncode() == 0
+                msvc_root_dir = msvc_include_result.stdout().strip()
+                winsdk_inc_dir = winsdk_include_result.stdout().strip()
+                nvcc_host_includes = [
+                    '-I', msvc_root_dir / 'include',
+                    '-I', winsdk_inc_dir / 'ucrt',
+                    '-I', winsdk_inc_dir / 'shared',
+                    '-I', winsdk_inc_dir / 'um',
+                ]
+                message('MSVC include: ' + msvc_root_dir / 'include')
+                message('UCRT include: ' + winsdk_inc_dir / 'ucrt')
+            else
+                warning('Could not find MSVC/Windows SDK include paths. nvcc compilation may fail.')
+            endif
+        else
+            nvcc_ccbin_flags = []
+            nvcc_host_includes = []
+        endif
+
+        # Detect CUDA version from nvcc directly instead of add_languages('cuda'),
+        # which requires MSVC on Windows and fails with MinGW.
+        # nvcc --version outputs: "...Cuda compilation tools, release 12.6, V12.6.85..."
+        nvcc_ver_out = run_command(nvcc_exe, '--version', check: true)
+        cuda_version = nvcc_ver_out.stdout().strip().split('release ')[1].split(',')[0]
+        message('Found CUDA version = @0@'.format(cuda_version))
+
         gencode = [
             '--fatbin',
             '-gencode=arch=compute_75,code=sm_75',
             '-gencode=arch=compute_80,code=sm_80',
         ]
-        message('Found CUDA version = @0@'.format(cuda_compiler.version()))
-        if cuda_compiler.version().version_compare('<13')
+        if cuda_version.version_compare('<13')
             gencode += '-gencode=arch=compute_50,code=compute_50'
         endif
         # always compile device code to enable quick startup on newer GPUs, for the last supported GPU also generate PTX for future compatibility
-        if cuda_compiler.version().version_compare('>11.8')
+        if cuda_version.version_compare('>11.8')
             gencode += '-gencode=arch=compute_90,code=sm_90'
-            if cuda_compiler.version().version_compare('>12.8')
+            if cuda_version.version_compare('>12.8')
                 gencode += [
                     '-gencode=arch=compute_100,code=sm_100',
                     '-gencode=arch=compute_120,code=sm_120',
@@ -353,7 +414,7 @@ if is_cuda_enabled
     message('CUDA gencode = @0@'.format(gencode))
 
 
-    cuda_flags = []
+    cuda_flags = ['-D_USE_MATH_DEFINES']
     if get_option('buildtype') == 'debug'
         cuda_flags += ['-DCUDA_DEBUG', '-G']
     else
@@ -378,7 +439,7 @@ if is_cuda_enabled
                 '-I', '../src/feature',
                 '-I', '../src/' + cuda_dir,
                 '-DDEVICE_CODE',
-            ] + cuda_flags
+            ] + cuda_flags + nvcc_ccbin_flags + nvcc_host_includes
         )
         ptx_files += {name : [t]}
     endforeach
@@ -414,7 +475,6 @@ if is_cuda_enabled
             cuda_inc,
         ],
         c_args : vmaf_cflags_common,
-        cuda_args: cuda_flags # + ['-gencode', 'arch=compute_86,code=sm_86' ] #, '--use_fast_math']
     )
 
     common_cuda_objects += cuda_static_lib.extract_all_objects()

libvmaf/src/picture.h

@@ -20,9 +20,14 @@
 #define __VMAF_SRC_PICTURE_H__
 
 #ifdef HAVE_CUDA
+#ifdef DEVICE_CODE
+#include <cuda.h>
+typedef struct VmafCudaState VmafCudaState;
+#else
 #include <ffnvcodec/dynlink_cuda.h>
 #include "libvmaf/libvmaf_cuda.h"
 #endif
+#endif
 #include "libvmaf/picture.h"
 
 enum VmafPictureBufferType {