Various chores (#166)

Suppress some spurious warnings using CUTE_INVALID_CONTROL_PATH Catch some potentially dangerous codepaths with same Improved ifdefs for CUDA arch specific stuff Remove an empty example file
codeplaysoftware · Jan 9, 2025 · 25b7875 · 25b7875
1 parent 01d4c32
commit 25b7875
Show file tree

Hide file tree

Showing 7 changed files with 153 additions and 36 deletions.
diff --git a/include/cute/arch/xe_copy_1B.hpp b/include/cute/arch/xe_copy_1B.hpp
@@ -37,13 +37,21 @@
 #ifdef __SYCL_DEVICE_ONLY__
 #define SYCL_DEVICE_BUILTIN(x) SYCL_EXTERNAL extern "C" x
 #else
-#define SYCL_DEVICE_BUILTIN(x) inline x { assert(false); }
+#define SYCL_DEVICE_BUILTIN(x)                                                 \
+  inline x {                                                                   \
+    CUTE_INVALID_CONTROL_PATH(                                                 \
+        "Attempting to use a device built-in in host code.");                  \
+  }
 #endif
 
 #ifdef __SYCL_DEVICE_ONLY__
 #define SYCL_DEVICE_OCL(x) SYCL_EXTERNAL x
 #else
-#define SYCL_DEVICE_OCL(x) inline x { assert(false); }
+#define SYCL_DEVICE_OCL(x)                                                     \
+  inline x {                                                                   \
+    CUTE_INVALID_CONTROL_PATH(                                                 \
+        "Attempting to use a device built-in in host code.");                  \
+  }
 #endif
 
 using namespace cute;

diff --git a/include/cute/arch/xe_copy_2B.hpp b/include/cute/arch/xe_copy_2B.hpp
@@ -38,13 +38,21 @@
 #ifdef __SYCL_DEVICE_ONLY__
 #define SYCL_DEVICE_BUILTIN(x) SYCL_EXTERNAL extern "C" x
 #else
-#define SYCL_DEVICE_BUILTIN(x) inline x { assert(false); }
+#define SYCL_DEVICE_BUILTIN(x)                                                 \
+  inline x {                                                                   \
+    CUTE_INVALID_CONTROL_PATH(                                                 \
+        "Attempting to use a device built-in in host code.");                  \
+  }
 #endif
 
 #ifdef __SYCL_DEVICE_ONLY__
 #define SYCL_DEVICE_OCL(x) SYCL_EXTERNAL x
 #else
-#define SYCL_DEVICE_OCL(x) inline x { assert(false); }
+#define SYCL_DEVICE_OCL(x)                                                     \
+  inline x {                                                                   \
+    CUTE_INVALID_CONTROL_PATH(                                                 \
+        "Attempting to use a device built-in in host code.");                  \
+  }
 #endif
 
 using namespace cute;

diff --git a/include/cute/arch/xe_copy_4B.hpp b/include/cute/arch/xe_copy_4B.hpp
@@ -37,13 +37,21 @@
 #ifdef __SYCL_DEVICE_ONLY__
 #define SYCL_DEVICE_BUILTIN(x) SYCL_EXTERNAL extern "C" x
 #else
-#define SYCL_DEVICE_BUILTIN(x) inline x { assert(false); }
+#define SYCL_DEVICE_BUILTIN(x)                                                 \
+  inline x {                                                                   \
+    CUTE_INVALID_CONTROL_PATH(                                                 \
+        "Attempting to use a device built-in in host code.");                  \
+  }
 #endif
 
 #ifdef __SYCL_DEVICE_ONLY__
 #define SYCL_DEVICE_OCL(x) SYCL_EXTERNAL x
 #else
-#define SYCL_DEVICE_OCL(x) inline x { assert(false); }
+#define SYCL_DEVICE_OCL(x)                                                     \
+  inline x {                                                                   \
+    CUTE_INVALID_CONTROL_PATH(                                                 \
+        "Attempting to use a device built-in in host code.");                  \
+  }
 #endif
 
 enum class CacheControl {

diff --git a/include/cute/arch/xe_copy_8B.hpp b/include/cute/arch/xe_copy_8B.hpp
@@ -37,13 +37,21 @@
 #ifdef __SYCL_DEVICE_ONLY__
 #define SYCL_DEVICE_BUILTIN(x) SYCL_EXTERNAL extern "C" x
 #else
-#define SYCL_DEVICE_BUILTIN(x) inline x { assert(false); }
+#define SYCL_DEVICE_BUILTIN(x)                                                 \
+  inline x {                                                                   \
+    CUTE_INVALID_CONTROL_PATH(                                                 \
+        "Attempting to use a device built-in in host code.");                  \
+  }
 #endif
 
 #ifdef __SYCL_DEVICE_ONLY__
 #define SYCL_DEVICE_OCL(x) SYCL_EXTERNAL x
 #else
-#define SYCL_DEVICE_OCL(x) inline x { assert(false); }
+#define SYCL_DEVICE_OCL(x)                                                     \
+  inline x {                                                                   \
+    CUTE_INVALID_CONTROL_PATH(                                                 \
+        "Attempting to use a device built-in in host code.");                  \
+  }
 #endif
 
 using namespace cute;

diff --git a/include/cutlass/arch/arch.h b/include/cutlass/arch/arch.h
@@ -35,6 +35,7 @@
 #pragma once
 
 #include "cutlass/cutlass.h"
+#include "cutlass/platform/platform.h"
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -47,20 +48,24 @@ namespace arch {
 CUTLASS_DEVICE
 int LaneId() {
   int ret;
-#if !defined(CUTLASS_ENABLE_SYCL) || defined(__SYCL_CUDA_ARCH__)
+#if defined(__CUDA_ARCH__) || defined(__SYCL_CUDA_ARCH__)
   asm ("mov.u32 %0, %%laneid;" : "=r"(ret) : );
-#endif
   return ret;
+#else
+  CUTLASS_INVALID_CONTROL_PATH("Attempting to use Nvidia-specific code path on non-Nvidia hardware.");
+#endif
 }
 
 /// Computes SM number the thread is running on
 CUTLASS_DEVICE
 int SmId() {
   int ret;
-#if !defined(CUTLASS_ENABLE_SYCL) || defined(__SYCL_CUDA_ARCH__)
+#if defined(__CUDA_ARCH__) || defined(__SYCL_CUDA_ARCH__)
   asm ("mov.u32 %0, %%smid;" : "=r"(ret) : );
-#endif
   return ret;
+#else
+  CUTLASS_INVALID_CONTROL_PATH("Attempting to use Nvidia-specific code path on non-Nvidia hardware.");
+#endif
 }
 
 #endif