Move compat to cute/util (#538)

rolandschulz · web-flow · commit b0cb10e655d8 · 2025-09-26T12:39:47.000-07:00
diff --git a/examples/cute/tutorial/bgemm_bmg_legacy.cpp b/examples/cute/tutorial/bgemm_bmg_legacy.cpp
@@ -31,7 +31,7 @@
  **************************************************************************************************/
 
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include <cute/tensor.hpp>
 
diff --git a/examples/cute/tutorial/sgemm_1_sycl.cpp b/examples/cute/tutorial/sgemm_1_sycl.cpp
@@ -32,7 +32,7 @@
  **************************************************************************************************/
 
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include <cute/tensor.hpp>
 
diff --git a/examples/cute/tutorial/sgemm_2_sycl.cpp b/examples/cute/tutorial/sgemm_2_sycl.cpp
@@ -32,7 +32,7 @@
  **************************************************************************************************/
 
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 
 #include <cute/tensor.hpp>
diff --git a/examples/cute/tutorial/sgemm_sm70_sycl.cpp b/examples/cute/tutorial/sgemm_sm70_sycl.cpp
@@ -32,7 +32,7 @@
  **************************************************************************************************/
 
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include <cute/tensor.hpp>
 
diff --git a/examples/cute/tutorial/sgemm_sm80_sycl.cpp b/examples/cute/tutorial/sgemm_sm80_sycl.cpp
@@ -32,7 +32,7 @@
  **************************************************************************************************/
 
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include <cute/tensor.hpp>
 
diff --git a/examples/cute/tutorial/tiled_copy_if_sycl.cpp b/examples/cute/tutorial/tiled_copy_if_sycl.cpp
@@ -30,7 +30,7 @@
  **************************************************************************************************/
 
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include <cute/tensor.hpp>
 
diff --git a/examples/cute/tutorial/tiled_copy_sycl.cpp b/examples/cute/tutorial/tiled_copy_sycl.cpp
@@ -32,7 +32,7 @@
  **************************************************************************************************/
 
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include <cute/tensor.hpp>
 
diff --git a/examples/cute/tutorial/xe_gemm.cpp b/examples/cute/tutorial/xe_gemm.cpp
@@ -30,7 +30,7 @@
 **************************************************************************************************/
 
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 #include <sycl/ext/intel/experimental/grf_size_properties.hpp>
 
 #include <cute/tensor.hpp>
diff --git a/include/cute/util/compat.hpp b/include/cute/util/compat.hpp
@@ -23,4 +23,4 @@
 
 #pragma once
 
-#include <compat/compat.hpp>
+#include <cute/util/compat/compat.hpp>
diff --git a/include/cute/util/compat/atomic.hpp b/include/cute/util/compat/atomic.hpp
@@ -31,6 +31,7 @@
 //===----------------------------------------------------------------------===//
 
 #pragma once
+#pragma GCC system_header
 
 #include <cassert>
 
@@ -39,7 +40,7 @@
 #include <sycl/memory_enums.hpp>
 #include <sycl/multi_ptr.hpp>
 
-#include <compat/traits.hpp>
+#include <cute/util/compat/traits.hpp>
 
 namespace compat {
 
diff --git a/include/cute/util/compat/compat.hpp b/include/cute/util/compat/compat.hpp
@@ -22,15 +22,16 @@
  **************************************************************************/
 
 #pragma once
+#pragma GCC system_header
 
-#include <compat/atomic.hpp>
-#include <compat/defs.hpp>
-#include <compat/device.hpp>
-#include <compat/dims.hpp>
-#include <compat/group_utils.hpp>
-#include <compat/id_query.hpp>
-#include <compat/kernel.hpp>
-#include <compat/launch.hpp>
-#include <compat/math.hpp>
-#include <compat/memory.hpp>
-#include <compat/util.hpp>
+#include <cute/util/compat/atomic.hpp>
+#include <cute/util/compat/defs.hpp>
+#include <cute/util/compat/device.hpp>
+#include <cute/util/compat/dims.hpp>
+#include <cute/util/compat/group_utils.hpp>
+#include <cute/util/compat/id_query.hpp>
+#include <cute/util/compat/kernel.hpp>
+#include <cute/util/compat/launch.hpp>
+#include <cute/util/compat/math.hpp>
+#include <cute/util/compat/memory.hpp>
+#include <cute/util/compat/util.hpp>
diff --git a/include/cute/util/compat/defs.hpp b/include/cute/util/compat/defs.hpp
@@ -32,6 +32,7 @@
 //===----------------------------------------------------------------------===//
 
 #pragma once
+#pragma GCC system_header
 
 #include <iostream>
 
diff --git a/include/cute/util/compat/device.hpp b/include/cute/util/compat/device.hpp
@@ -43,6 +43,7 @@
 //===----------------------------------------------------------------------===//
 
 #pragma once
+#pragma GCC system_header
 
 #include <algorithm>
 #include <cstring>
diff --git a/include/cute/util/compat/dims.hpp b/include/cute/util/compat/dims.hpp
@@ -22,6 +22,7 @@
  **************************************************************************/
 
 #pragma once
+#pragma GCC system_header
 
 #include <tuple>
 
diff --git a/include/cute/util/compat/group_utils.hpp b/include/cute/util/compat/group_utils.hpp
@@ -31,13 +31,14 @@
 //===------------------------------------------------------------------===//
 
 #pragma once
+#pragma GCC system_header
 
 #include <iterator>
 #include <stdexcept>
 #include <sycl/sycl.hpp>
 
-#include <compat/defs.hpp>
-#include <compat/math.hpp>
+#include <cute/util/compat/defs.hpp>
+#include <cute/util/compat/math.hpp>
 
 namespace compat {
 namespace group {
diff --git a/include/cute/util/compat/id_query.hpp b/include/cute/util/compat/id_query.hpp
@@ -22,6 +22,7 @@
  **************************************************************************/
 
 #pragma once
+#pragma GCC system_header
 
 #include <sycl/ext/oneapi/free_function_queries.hpp>
 #include <sycl/nd_item.hpp>
diff --git a/include/cute/util/compat/kernel.hpp b/include/cute/util/compat/kernel.hpp
@@ -31,6 +31,7 @@
 //===----------------------------------------------------------------------===//
 
 #pragma once
+#pragma GCC system_header
 
 #ifdef _WIN32
 #include <unordered_set>
diff --git a/include/cute/util/compat/launch.hpp b/include/cute/util/compat/launch.hpp
@@ -22,6 +22,7 @@
  **************************************************************************/
 
 #pragma once
+#pragma GCC system_header
 
 #include <sycl/accessor.hpp>
 #include <sycl/event.hpp>
@@ -30,9 +31,9 @@
 #include <sycl/range.hpp>
 #include <sycl/reduction.hpp>
 
-#include <compat/device.hpp>
-#include <compat/dims.hpp>
-#include <compat/launch_policy.hpp>
+#include <cute/util/compat/device.hpp>
+#include <cute/util/compat/dims.hpp>
+#include <cute/util/compat/launch_policy.hpp>
 
 namespace compat {
 
diff --git a/include/cute/util/compat/launch_policy.hpp b/include/cute/util/compat/launch_policy.hpp
@@ -22,6 +22,7 @@
  **************************************************************************/
 
 #pragma once
+#pragma GCC system_header
 
 #include "sycl/ext/oneapi/experimental/enqueue_functions.hpp"
 #include "sycl/ext/oneapi/properties/properties.hpp"
@@ -30,10 +31,10 @@
 #include <sycl/queue.hpp>
 #include <sycl/range.hpp>
 
-#include <compat/defs.hpp>
-#include <compat/device.hpp>
-#include <compat/dims.hpp>
-#include <compat/traits.hpp>
+#include <cute/util/compat/defs.hpp>
+#include <cute/util/compat/device.hpp>
+#include <cute/util/compat/dims.hpp>
+#include <cute/util/compat/traits.hpp>
 
 namespace compat {
 namespace experimental {
diff --git a/include/cute/util/compat/math.hpp b/include/cute/util/compat/math.hpp
@@ -31,6 +31,7 @@
 //===----------------------------------------------------------------------===//
 
 #pragma once
+#pragma GCC system_header
 
 #include <limits>
 #include <sycl/feature_test.hpp>
@@ -45,7 +46,7 @@
 #include <sycl/ext/oneapi/experimental/bfloat16_math.hpp>
 #endif
 #include <sycl/ext/oneapi/experimental/complex/complex.hpp>
-#include <compat/traits.hpp>
+#include <cute/util/compat/traits.hpp>
 
 namespace compat {
 namespace detail {
diff --git a/include/cute/util/compat/memory.hpp b/include/cute/util/compat/memory.hpp
@@ -31,6 +31,7 @@
 //===----------------------------------------------------------------------===//
 
 #pragma once
+#pragma GCC system_header
 
 #include <cassert>
 #include <cstdint>
@@ -52,9 +53,9 @@
 #include <sycl/ext/intel/experimental/usm_properties.hpp>
 #endif
 
-#include <compat/device.hpp>
-#include <compat/traits.hpp>
-#include <compat/defs.hpp>
+#include <cute/util/compat/device.hpp>
+#include <cute/util/compat/traits.hpp>
+#include <cute/util/compat/defs.hpp>
 
 #if defined(__linux__)
 #include <sys/mman.h>
diff --git a/include/cute/util/compat/traits.hpp b/include/cute/util/compat/traits.hpp
@@ -22,6 +22,7 @@
  **************************************************************************/
 
 #pragma once
+#pragma GCC system_header
 
 #include <sycl/feature_test.hpp>
 #ifdef SYCL_EXT_ONEAPI_BFLOAT16_MATH_FUNCTIONS
diff --git a/include/cute/util/compat/util.hpp b/include/cute/util/compat/util.hpp
@@ -31,6 +31,7 @@
 //===----------------------------------------------------------------------===//
 
 #pragma once
+#pragma GCC system_header
 
 #include <cassert>
 #include <type_traits>
@@ -39,9 +40,9 @@
 #include <sycl/group_barrier.hpp>
 #include <sycl/kernel_bundle.hpp>
 
-#include <compat/math.hpp>
-#include <compat/memory.hpp>
-#include <compat/dims.hpp>
+#include <cute/util/compat/math.hpp>
+#include <cute/util/compat/memory.hpp>
+#include <cute/util/compat/dims.hpp>
 
 #if defined(__NVPTX__)
 #include <sycl/ext/oneapi/experimental/cuda/masked_shuffles.hpp>
diff --git a/include/cute/util/debug.hpp b/include/cute/util/debug.hpp
@@ -38,7 +38,7 @@
 
 #if defined(CUTLASS_ENABLE_SYCL)
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 #else
 #include <cuda_runtime_api.h>
 #endif
diff --git a/include/cutlass/gpu_generics.h b/include/cutlass/gpu_generics.h
@@ -39,7 +39,7 @@
 
 #if defined(CUTLASS_ENABLE_SYCL)
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 #endif
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
diff --git a/test/unit/common/util.hpp b/test/unit/common/util.hpp
@@ -32,7 +32,7 @@
  **************************************************************************************************/
 
 #if defined(CUTLASS_ENABLE_SYCL)
-#include <compat/compat.hpp>
+#include <cute/util/compat/compat.hpp>
 
 #include <vector>
 #else
diff --git a/test/unit/cute/cooperative_gemm_common.hpp b/test/unit/cute/cooperative_gemm_common.hpp
@@ -167,7 +167,7 @@ void verify_gemm_correctness(cute::Tensor<EngineC, CLayout> const& h_c_out_tenso
 
 #if defined(CUTLASS_ENABLE_SYCL)
 #include <sycl/sycl.hpp>
-#include <compat/compat.hpp>
+#include <cute/util/compat/compat.hpp>
 #include <cutlass/sycl_vector_types.h>
 
 namespace sc = compat;
diff --git a/test/unit/cute/intel_xe/copy_1d.cpp b/test/unit/cute/intel_xe/copy_1d.cpp
@@ -34,7 +34,7 @@
 
 #include <cute/tensor.hpp>
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include "cutlass_unit_test.h"
 
diff --git a/test/unit/cute/intel_xe/copy_block.cpp b/test/unit/cute/intel_xe/copy_block.cpp
@@ -34,7 +34,7 @@
 
 #include <cute/tensor.hpp>
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include "cutlass_unit_test.h"
 
diff --git a/test/unit/cute/intel_xe/copy_scatter.cpp b/test/unit/cute/intel_xe/copy_scatter.cpp
@@ -34,7 +34,7 @@
 
 #include <cute/tensor.hpp>
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include "cutlass_unit_test.h"
 
diff --git a/test/unit/cute/intel_xe/copy_subgroup_block.cpp b/test/unit/cute/intel_xe/copy_subgroup_block.cpp
@@ -34,7 +34,7 @@
 
 #include <cute/tensor.hpp>
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include "cutlass_unit_test.h"
 
diff --git a/test/unit/cute/intel_xe/mma.cpp b/test/unit/cute/intel_xe/mma.cpp
@@ -34,7 +34,7 @@
 
 #include <cute/tensor.hpp>
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include "cutlass_unit_test.h"
 #include "utils.hpp"
diff --git a/test/unit/cute/intel_xe/utils.hpp b/test/unit/cute/intel_xe/utils.hpp
@@ -36,7 +36,7 @@
 
 #include <cute/tensor.hpp>
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include "cutlass_unit_test.h"
 
diff --git a/tools/copy_debug/copy_debug.cpp b/tools/copy_debug/copy_debug.cpp
@@ -32,10 +32,10 @@
 
 #include <cute/tensor.hpp>
 #include <sycl/sycl.hpp>
-#include <compat.hpp>
+#include <cute/util/compat.hpp>
 
 #include <cutlass/util/device_memory.h>
-#include <compat/compat.hpp>
+#include <cute/util/compat/compat.hpp>
 #include <cutlass/cutlass.h>
 
 using namespace compat::experimental;
diff --git a/tools/util/include/cutlass/util/initialize_block.hpp b/tools/util/include/cutlass/util/initialize_block.hpp
diff --git a/tools/util/include/cutlass/util/mixed_dtype_utils.hpp b/tools/util/include/cutlass/util/mixed_dtype_utils.hpp
diff --git a/tools/util/include/cutlass/util/sycl_timer.hpp b/tools/util/include/cutlass/util/sycl_timer.hpp

Original file line number	Diff line number	Diff line change
`@@ -23,4 +23,4 @@`
`23`	`23`
`24`	`24`	`#pragma once`
`25`	`25`
`26`		`-#include <compat/compat.hpp>`
	`26`	`+#include <cute/util/compat/compat.hpp>`