Fixes the ROCm (HIP) build of CUDACachingAllocator.cpp which was broken by two categories of errors:

1. Embedded preprocessor directives inside macro arguments — The #if ROCM_VERSION >= 70100 / #endif block was nested inside a C10_CUDA_CHECK(hipMemImportFromShareableHandle(...)) macro call. Clang rejects this with -Werror,-Wembedded-directive. Fixed by hoisting the conditional into a separate variable (myfd_ptr) before the macro invocation.
2. Void pointer arithmetic — The HIP API uses void* for ptr_, unlike the CUDA driver API which uses CUdeviceptr (an integer type). Arithmetic on void* is undefined behavior and rejected by Clang. Fixed by casting ptr_ to char* via reinterpret_cast in the three call sites (hipMemSetAccess, hipMemMap, hipMemUnmap).

Also fixes requestedHandleTypes → requestedHandleType for the HIP CUmemAllocationProp struct, which uses a different field name than the CUDA equivalent.

This is a FF for issues introduced in pytorch#173330

Pull Request resolved: pytorch#176136
Approved by: https://github.com/malfet

Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>

Use `ptr()` helper (returns `char*`) instead of raw `ptr_` for pointer
arithmetic in `cuMemSetAccess`, `cuMemMap`, and `cuMemUnmap` calls.

After hipify converts `CUdeviceptr` (unsigned long long) to
`hipDeviceptr_t` (void*), arithmetic on `ptr_` becomes void pointer
arithmetic which Clang rejects with `-Werror`. Using `ptr()` +
`reinterpret_cast` avoids this while keeping CUDA driver API
compatibility.

Also improves the USE_ROCM blocks to use `ptr()` instead of
`reinterpret_cast<char*>(ptr_)` for consistency.

This is a forward fix for void* arithmetic errors introduced by #173330
that #176136 attempted to fix via `#ifdef USE_ROCM` blocks.

Co-Authored-By: Claude <noreply@anthropic.com>

Use `ptr()` helper (returns `char*`) instead of raw `ptr_` for pointer
arithmetic in `cuMemSetAccess`, `cuMemMap`, and `cuMemUnmap` calls.

After hipify converts `CUdeviceptr` (unsigned long long) to
`hipDeviceptr_t` (void*), arithmetic on `ptr_` becomes void pointer
arithmetic which Clang rejects with `-Werror`. Using `ptr()` +
`reinterpret_cast` avoids this while keeping CUDA driver API
compatibility.

Also improves the USE_ROCM blocks to use `ptr()` instead of
`reinterpret_cast<char*>(ptr_)` for consistency.

This is a forward fix for void* arithmetic errors introduced by #173330
that #176136 attempted to fix via `#ifdef USE_ROCM` blocks.

Co-Authored-By: Claude <noreply@anthropic.com>

This reverts commit 4bdfaaa.

Reverted #176136 on behalf of https://github.com/wdvr due to reverting to put back on top of the trunk and develop a full forward fix ([comment](#176136 (comment)))

This reverts commit d49571b.

Reverted #173330 on behalf of https://github.com/wdvr due to sorry having issues with fixing build issues internally -- will revert and reland on top of master, and have a combined improved build fix instead of #176136 ([comment](#173330 (comment)))

Pull Request resolved: #173330
Approved by: https://github.com/jeffdaily

Co-authored-by: Jeff Daily <jeff.daily@amd.com>

Fixes the ROCm (HIP) build of CUDACachingAllocator.cpp which was broken by two categories of errors:

1. Embedded preprocessor directives inside macro arguments — The #if ROCM_VERSION >= 70100 / #endif block was nested inside a C10_CUDA_CHECK(hipMemImportFromShareableHandle(...)) macro call. Clang rejects this with -Werror,-Wembedded-directive. Fixed by hoisting the conditional into a separate variable (myfd_ptr) before the macro invocation.
2. Void pointer arithmetic — The HIP API uses void* for ptr_, unlike the CUDA driver API which uses CUdeviceptr (an integer type). Arithmetic on void* is undefined behavior and rejected by Clang. Fixed by casting ptr_ to char* via reinterpret_cast in the three call sites (hipMemSetAccess, hipMemMap, hipMemUnmap).

Also fixes requestedHandleTypes → requestedHandleType for the HIP CUmemAllocationProp struct, which uses a different field name than the CUDA equivalent.

This is a FF for issues introduced in pytorch#173330

Pull Request resolved: pytorch#176136
Approved by: https://github.com/malfet

Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>

Use `ptr()` helper (returns `char*`) instead of raw `ptr_` for pointer
arithmetic in `cuMemSetAccess`, `cuMemMap`, and `cuMemUnmap` calls.

After hipify converts `CUdeviceptr` (unsigned long long) to
`hipDeviceptr_t` (void*), arithmetic on `ptr_` becomes void pointer
arithmetic which Clang rejects with `-Werror`. Using `ptr()` +
`reinterpret_cast` avoids this while keeping CUDA driver API
compatibility.

Also improves the USE_ROCM blocks to use `ptr()` instead of
`reinterpret_cast<char*>(ptr_)` for consistency.

This is a forward fix for void* arithmetic errors introduced by pytorch#173330
that pytorch#176136 attempted to fix via `#ifdef USE_ROCM` blocks.

Co-Authored-By: Claude <noreply@anthropic.com>

This reverts commit 4bdfaaa.

Reverted pytorch#176136 on behalf of https://github.com/wdvr due to reverting to put back on top of the trunk and develop a full forward fix ([comment](pytorch#176136 (comment)))

This reverts commit d49571b.

Reverted pytorch#173330 on behalf of https://github.com/wdvr due to sorry having issues with fixing build issues internally -- will revert and reland on top of master, and have a combined improved build fix instead of pytorch#176136 ([comment](pytorch#173330 (comment)))