Merge pull request #6706 from xq114/dev

fix find_cudadevices for CUDA 13

2 files changed, 39 insertions, 10 deletions

diff --git a/xmake/modules/lib/detect/find_cudadevices.lua b/xmake/modules/lib/detect/find_cudadevices.lua
index 2b75d725a..c780bf311 100644
--- a/xmake/modules/lib/detect/find_cudadevices.lua
+++ b/xmake/modules/lib/detect/find_cudadevices.lua
@@ -130,7 +130,7 @@ function _find_devices(verbose, opt)
     local results, errors = try
     {
         function ()
-            local args = { sourcefile, "-run", "-o", outfile , '-DPRINT_SUFFIX="' .. _PRINT_SUFFIX .. '"' }
+            local args = { sourcefile, "-run", "-lcuda", "-o", outfile , '-DPRINT_SUFFIX="' .. _PRINT_SUFFIX .. '"' }
             if opt.arch == "x86" then
                 table.insert(args, "-m32")
             elseif opt.arch == "x64" or opt.arch == "x86_64" then
@@ -247,6 +247,8 @@ function _order_by_flops(devices)
     ,   [80] =     64
     ,   [86] =    128
     ,   [87] =    128
+    ,   [89] =    128
+    ,   [90] =    128
     }
 
     for _, dev in ipairs(devices) do
@@ -254,7 +256,7 @@ function _order_by_flops(devices)
         if dev.major == 9999 and dev.minor == 9999 then
             sm_per_multiproc = 1
         else
-            sm_per_multiproc = ngpu_arch_cores_per_sm[dev.major * 10 + dev.minor] or 64;
+            sm_per_multiproc = ngpu_arch_cores_per_sm[dev.major * 10 + dev.minor] or 128;
         end
         dev["$flops"] = dev.multiProcessorCount * sm_per_multiproc * dev.clockRate
     end
diff --git a/xmake/scripts/find_cudadevices.cpp b/xmake/scripts/find_cudadevices.cpp
index fd002cc84..c8f082555 100644
--- a/xmake/scripts/find_cudadevices.cpp
+++ b/xmake/scripts/find_cudadevices.cpp
@@ -1,4 +1,5 @@
 #include <cuda_runtime.h>
+#include <cuda.h>
 #include <stdio.h>
 #include <stdlib.h>
 
@@ -19,6 +20,19 @@ inline void check(cudaError_t result)
     }
 }
 
+inline void check_driver(CUresult result)
+{
+    static char unknown[] = "unknown";
+    if (result != CUDA_SUCCESS)
+    {
+        const char *name = NULL, *msg = NULL;
+        cuGetErrorName(result, &name);
+        cuGetErrorString(result, &msg);
+        fprintf(stderr, PRINT_SUFFIX "%s (%s)", name ? name : unknown, msg ? msg : unknown);
+        exit(0);
+    }
+}
+
 inline void print_value(unsigned long long value)
 {
     printf("%llu", value);
@@ -115,21 +129,16 @@ inline void print_device(int id)
     PRINT_PROPERTY(maxThreadsPerBlock);
     PRINT_PROPERTY(maxThreadsDim);
     PRINT_PROPERTY(maxGridSize);
-    PRINT_PROPERTY(clockRate);
     PRINT_PROPERTY(totalConstMem);
     PRINT_PROPERTY(major);
     PRINT_PROPERTY(minor);
     PRINT_PROPERTY(textureAlignment);
     PRINT_PROPERTY(texturePitchAlignment);
-    PRINT_BOOL_PROPERTY(deviceOverlap);
     PRINT_PROPERTY(multiProcessorCount);
-    PRINT_BOOL_PROPERTY(kernelExecTimeoutEnabled);
     PRINT_BOOL_PROPERTY(integrated);
     PRINT_BOOL_PROPERTY(canMapHostMemory);
-    PRINT_PROPERTY(computeMode);
     PRINT_PROPERTY(maxTexture1D);
     PRINT_PROPERTY(maxTexture1DMipmap);
-    PRINT_PROPERTY(maxTexture1DLinear);
     PRINT_PROPERTY(maxTexture2D);
     PRINT_PROPERTY(maxTexture2DMipmap);
     PRINT_PROPERTY(maxTexture2DLinear);
@@ -156,7 +165,6 @@ inline void print_device(int id)
     PRINT_BOOL_PROPERTY(tccDriver);
     PRINT_PROPERTY(asyncEngineCount);
     PRINT_BOOL_PROPERTY(unifiedAddressing);
-    PRINT_PROPERTY(memoryClockRate);
     PRINT_PROPERTY(memoryBusWidth);
     PRINT_PROPERTY(l2CacheSize);
     PRINT_PROPERTY(maxThreadsPerMultiProcessor);
@@ -167,18 +175,37 @@ inline void print_device(int id)
     PRINT_PROPERTY(regsPerMultiprocessor);
     PRINT_BOOL_PROPERTY(isMultiGpuBoard);
     PRINT_PROPERTY(multiGpuBoardGroupID);
-    PRINT_PROPERTY(singleToDoublePrecisionPerfRatio);
     PRINT_BOOL_PROPERTY(pageableMemoryAccess);
     PRINT_BOOL_PROPERTY(concurrentManagedAccess);
     PRINT_BOOL_PROPERTY(managedMemory);
 
+#if MY_CUDA_VER < 1300
+    PRINT_PROPERTY(clockRate);
+    PRINT_PROPERTY(memoryClockRate);
+    PRINT_BOOL_PROPERTY(deviceOverlap);
+    PRINT_BOOL_PROPERTY(kernelExecTimeoutEnabled);
+    PRINT_PROPERTY(computeMode);
+    PRINT_PROPERTY(maxTexture1DLinear);
+    PRINT_PROPERTY(singleToDoublePrecisionPerfRatio);
+#else
+    int clockRate, memoryClockRate, computeMode;
+    check_driver(cuDeviceGetAttribute(&clockRate, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, id));
+    check_driver(cuDeviceGetAttribute(&memoryClockRate, CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE, id));
+    check_driver(cuDeviceGetAttribute(&computeMode, CU_DEVICE_ATTRIBUTE_COMPUTE_MODE, id));
+    print_property("clockRate", clockRate);
+    print_property("memoryClockRate", memoryClockRate);
+    print_property("computeMode", computeMode);
+#endif
+
 #if MY_CUDA_VER >= 900
     // Added in cuda 9.0
     PRINT_BOOL_PROPERTY(computePreemptionSupported);
     PRINT_BOOL_PROPERTY(canUseHostPointerForRegisteredMem);
     PRINT_BOOL_PROPERTY(cooperativeLaunch);
-    PRINT_BOOL_PROPERTY(cooperativeMultiDeviceLaunch);
     PRINT_PROPERTY(sharedMemPerBlockOptin);
+#if MY_CUDA_VER < 1300
+    PRINT_BOOL_PROPERTY(cooperativeMultiDeviceLaunch);
+#endif
 #endif
 
 #if MY_CUDA_VER >= 902