The current trap handler has 2 limitations: 1) If it receives a HOST_TRAP, it clears the corresponding bit and notifies the host, when it should not. 2) When it is entered because of a debug trap (s_trap 3) and the debugger is not attached, it returns unconditionally. However, if another exception is reported at the same time as the trap handler is entered for the debug trap (a memory violation for example), that other exception ends-up being ignored. This patch addresses both of those issues. It makes it so host traps and debug traps are ignored when necessary. If any other exception is reported to the wave, we halt the wave and notify the host, and if no other exception is reported (i.e. we entered the trap handler because of host trap or debug trap), we return to shader code. Other minor defects are also fixed during this refactor: - Fixed SQ_WAVE_EXCP_FLAG_PRIV_XNACK_ERROR_SHIFT which had an incorrect value - Host traps can be sent at any time, including after we have halted a wave. In such case, the old approach would have: 1) cleared the trap ID saved in ttmp6 2) clobbered ttmp10 where part of the actual wave's PC is saved. Change-Id: I9ecd341f4967e686233dec182b3e5b0388ef19bd
ROCR Runtime
This ROCm Runtime (ROCr) repo combines 2 previously separate repos into a single repo:
- The HSA Runtime (
hsa-runtime) for AMD GPU application development and - The ROCt Thunk Library (
libhsakmt), a "thunk" interface to the ROCm kernel driver (ROCk), used by the runtime.
Infrastructure
The HSA runtime is a thin, user-mode API that exposes the necessary interfaces to access and interact with graphics hardware driven by the AMDGPU driver set and the ROCK kernel driver. Together they enable programmers to directly harness the power of AMD discrete graphics devices by allowing host applications to launch compute kernels directly to the graphics hardware.
The capabilities expressed by the HSA Runtime API are:
- Error handling
- Runtime initialization and shutdown
- System and agent information
- Signals and synchronization
- Architected dispatch
- Memory management
- HSA runtime fits into a typical software architecture stack.
The HSA runtime provides direct access to the graphics hardware to give the programmer more control of the execution. An example of low level hardware access is the support of one or more user mode queues provides programmers with a low-latency kernel dispatch interface, allowing them to develop customized dispatch algorithms specific to their application.
The HSA Architected Queuing Language is an open standard, defined by the HSA Foundation, specifying the packet syntax used to control supported AMD/ATI Radeon (c) graphics devices. The AQL language supports several packet types, including packets that can command the hardware to automatically resolve inter-packet dependencies (barrier AND & barrier OR packet), kernel dispatch packets and agent dispatch packets.
In addition to user mode queues and AQL, the HSA runtime exposes various virtual address ranges that can be accessed by one or more of the system's graphics devices, and possibly the host. The exposed virtual address ranges either support a fine grained or a coarse grained access. Updates to memory in a fine grained region are immediately visible to all devices that can access it, but only one device can have access to a coarse grained allocation at a time. Ownership of a coarse grained region can be changed using the HSA runtime memory APIs, but this transfer of ownership must be explicitly done by the host application.
Programmers should consult the HSA Runtime Programmer's Reference Manual for a full description of the HSA Runtime APIs, AQL and the HSA memory policy.
Known issues
- Each HSA process creates an internal DMA queue, but there is a system-wide limit of four DMA queues. When the limit is reached HSA processes will use internal kernels for copies.
Artifacts produced by the build
- libhsakmt (ROCt) - User-mode API interfaces for interacting with the ROCk driver
- Runtime (ROCr) - Core runtime supporting HSA standards
- rocrtst - Runtime test suites for HSA implementation validation and performance testing
- kfdtest - Validation tests for ROCt
Building the ROCR Runtime
Target platform requirements
Please see the ROCm System requirements (Linux).
Ensure you have the following installed:
- CMake 3.7 or higher
libelf-devg++libdrm-amdgpu-devorlibdrm-devrocm-core
ROCr & ROCt Build Instructions
- Clone this repository and cd into its root
- Prepare the build directory
mkdir build && cd build - Configure the build (example)
cmake -DCMAKE_INSTALL_PREFIX=/opt/<rocm install dir>/llvm .. - Compile the project
make - Install the runtime
make install - (Optional) Build packages
make package
Non-default CMake Build Options
-
Produce a release build instead of debug
-DCMAKE_BUILD_TYPE=Release -
Control whether libhsakmt and libhsa-runtime are shared or static The default (when
BUILD_SHARED_LIBSis unset) is for libhsakmt to be static and libhsa-runtime to be shared. SettingBUILD_SHARED_LIBStoOFFwill make both static and setting it toONwill make them both shared.-DBUILD_SHARED_LIBS=OFF # or ON for both libs shared
Building the tests
rocrtst
- Go to rocrtst root
cd <rocr-runtime>/rocrtst/suites/test_common - Prepare the build directory
mkdir build && cd build - Configure the build
Example configuration:
cmake \ -DCMAKE_PREFIX_PATH="<rocm install root>;<llvm install root>" \ -DROCM_DIR="$ROCM_INSTALL_PATH" \ -DOPENCL_DIR="<rocm install root>" \ .. - Compile the project
make make rocrtst_kernels - ** Run the tests
Make sure libhsa-runtime.so is in the library path; e.g.,
$ LD_LIBRARY_PATH=<rocm install root> ./rocrtst -h # See help options
kfdtest
- Go to kfdtest root
cd <rocr-runtime>/libhsakmt/tests/kfdtest - Prepare the build directory
mkdir build && cd build - Configure the build
Example configuration:
cmake \ -DCMAKE_PREFIX_PATH="<rocm install root>" \ -DROCM_DIR="$ROCM_INSTALL_PATH" \ .. - Compile the project
make
Using the ROCR Runtime
After installation, you can link against the runtime by using the provided CMake package configurations. For example, to use the ROCR runtime in your project:
find_package(hsa-runtime64 1.0 REQUIRED)
add_executable(MyApp main.cpp)
target_link_libraries(MyApp PRIVATE hsa-runtime64::hsa-runtime64)
Disclaimer
The information contained herein is for informational purposes only, and is subject to change without notice. While every precaution has been taken in the preparation of this document, it may contain technical inaccuracies, omissions and typographical errors, and AMD is under no obligation to update or otherwise correct this information. Advanced Micro Devices, Inc. makes no representations or warranties with respect to the accuracy or completeness of the contents of this document, and assumes no liability of any kind, including the implied warranties of noninfringement, merchantability or fitness for particular purposes, with respect to the operation or use of AMD hardware, software or other products described herein. No license, including implied or arising by estoppel, to any intellectual property rights is granted by this document. Terms and limitations applicable to the purchase or use of AMD's products are as set forth in a signed agreement between the parties or in AMD's Standard Terms and Conditions of Sale.
AMD, the AMD Arrow logo, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies.
Copyright © 2014-2024 Advanced Micro Devices, Inc. All rights reserved.