Skip to content

Conversation

@doru1004
Copy link
Contributor

This patch updates the construction of packet headers to replace the usage of ACQUIRE/RELEASE with SCACQUIRE/SCRELEASE which is now recommended.
The patch also ensures consistency between kernel dispatches.

@llvmbot
Copy link
Member

llvmbot commented Mar 20, 2024

@llvm/pr-subscribers-libc

@llvm/pr-subscribers-backend-amdgpu

Author: Gheorghe-Teodor Bercea (doru1004)

Changes

This patch updates the construction of packet headers to replace the usage of ACQUIRE/RELEASE with SCACQUIRE/SCRELEASE which is now recommended.
The patch also ensures consistency between kernel dispatches.


Full diff: https://github.com/llvm/llvm-project/pull/85952.diff

3 Files Affected:

  • (modified) libc/utils/gpu/loader/amdgpu/Loader.cpp (+2-1)
  • (modified) openmp/libomptarget/plugins-nextgen/amdgpu/dynamic_hsa/hsa.h (+2-2)
  • (modified) openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp (+12-11)
diff --git a/libc/utils/gpu/loader/amdgpu/Loader.cpp b/libc/utils/gpu/loader/amdgpu/Loader.cpp index e3911eda2bd82a..7fd45acfd47e97 100644 --- a/libc/utils/gpu/loader/amdgpu/Loader.cpp +++ b/libc/utils/gpu/loader/amdgpu/Loader.cpp @@ -276,7 +276,8 @@ hsa_status_t launch_kernel(hsa_agent_t dev_agent, hsa_executable_t executable, (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE) | (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE); uint32_t header_word = header | (setup << 16u); - __atomic_store_n((uint32_t *)&packet->header, header_word, __ATOMIC_RELEASE); + __atomic_store_n(reinterpret_cast<uint32_t *>(packet), header_word, + __ATOMIC_RELEASE); hsa_signal_store_relaxed(queue->doorbell_signal, packet_id); // Wait until the kernel has completed execution on the device. Periodically diff --git a/openmp/libomptarget/plugins-nextgen/amdgpu/dynamic_hsa/hsa.h b/openmp/libomptarget/plugins-nextgen/amdgpu/dynamic_hsa/hsa.h index 64a1d3308aed0b..0ac7661d7e0076 100644 --- a/openmp/libomptarget/plugins-nextgen/amdgpu/dynamic_hsa/hsa.h +++ b/openmp/libomptarget/plugins-nextgen/amdgpu/dynamic_hsa/hsa.h @@ -216,8 +216,8 @@ typedef enum { HSA_FENCE_SCOPE_SYSTEM = 2 } hsa_fence_scope_t; typedef enum { HSA_PACKET_HEADER_TYPE = 0, - HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE = 9, - HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE = 11 + HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE = 9, + HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE = 11 } hsa_packet_header_t; typedef enum { diff --git a/openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp b/openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp index fce7454bf2800d..c147cefe58e9b4 100644 --- a/openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp +++ b/openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp @@ -826,15 +826,15 @@ struct AMDGPUQueueTy { /// Assumes the queue lock is acquired. void publishKernelPacket(uint64_t PacketId, uint16_t Setup, hsa_kernel_dispatch_packet_t *Packet) { - uint32_t *PacketPtr = reinterpret_cast<uint32_t *>(Packet); - - uint16_t Header = HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE; - Header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE; - Header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE; + uint16_t Header = + (HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE) | + (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE) | + (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE); // Publish the packet. Do not modify the package after this point. uint32_t HeaderWord = Header | (Setup << 16u); - __atomic_store_n(PacketPtr, HeaderWord, __ATOMIC_RELEASE); + __atomic_store_n(reinterpret_cast<uint32_t *>(Packet), HeaderWord, + __ATOMIC_RELEASE); // Signal the doorbell about the published packet. hsa_signal_store_relaxed(Queue->doorbell_signal, PacketId); @@ -845,15 +845,16 @@ struct AMDGPUQueueTy { /// barrier dependencies (signals) are satisfied. Assumes the queue is locked void publishBarrierPacket(uint64_t PacketId, hsa_barrier_and_packet_t *Packet) { - uint32_t *PacketPtr = reinterpret_cast<uint32_t *>(Packet); uint16_t Setup = 0; - uint16_t Header = HSA_PACKET_TYPE_BARRIER_AND << HSA_PACKET_HEADER_TYPE; - Header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE; - Header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE; + uint16_t Header = + (HSA_PACKET_TYPE_BARRIER_AND << HSA_PACKET_HEADER_TYPE) | + (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE) | + (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE); // Publish the packet. Do not modify the package after this point. uint32_t HeaderWord = Header | (Setup << 16u); - __atomic_store_n(PacketPtr, HeaderWord, __ATOMIC_RELEASE); + __atomic_store_n(reinterpret_cast<uint32_t *>(Packet), HeaderWord, + __ATOMIC_RELEASE); // Signal the doorbell about the published packet. hsa_signal_store_relaxed(Queue->doorbell_signal, PacketId); 
Copy link
Contributor

@jhuber6 jhuber6 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah, I see, dynamic HSA. I always forget about that.

@jhuber6
Copy link
Contributor

jhuber6 commented Mar 20, 2024

Personally I would've just fixed this with direct push that just adds these two enum members.

@doru1004
Copy link
Contributor Author

Ah, I see, dynamic HSA. I always forget about that.

Yep, had to revert the previous commit and had to open a new PR too.

@jhuber6
Copy link
Contributor

jhuber6 commented Mar 20, 2024

Ah, I see, dynamic HSA. I always forget about that.

Yep, had to revert the previous commit and had to open a new PR too.

Self-reviewed PRs muddy the waters. If you're going to self review just do a git push and be done with it, I think that's the LLVM policy.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

3 participants