binutils-gdb/gdb/amd-dbgapi-target.h

117 lines
3.5 KiB
C
Raw Normal View History

gdb: initial support for ROCm platform (AMDGPU) debugging This patch adds the foundation for GDB to be able to debug programs offloaded to AMD GPUs using the AMD ROCm platform [1]. The latest public release of the ROCm release at the time of writing is 5.4, so this is what this patch targets. The ROCm platform allows host programs to schedule bits of code for execution on GPUs or similar accelerators. The programs running on GPUs are typically referred to as `kernels` (not related to operating system kernels). Programs offloaded with the AMD ROCm platform can be written in the HIP language [2], OpenCL and OpenMP, but we're going to focus on HIP here. The HIP language consists of a C++ Runtime API and kernel language. Here's an example of a very simple HIP program: #include "hip/hip_runtime.h" #include <cassert> __global__ void do_an_addition (int a, int b, int *out) { *out = a + b; } int main () { int *result_ptr, result; /* Allocate memory for the device to write the result to. */ hipError_t error = hipMalloc (&result_ptr, sizeof (int)); assert (error == hipSuccess); /* Run `do_an_addition` on one workgroup containing one work item. */ do_an_addition<<<dim3(1), dim3(1), 0, 0>>> (1, 2, result_ptr); /* Copy result from device to host. Note that this acts as a synchronization point, waiting for the kernel dispatch to complete. */ error = hipMemcpyDtoH (&result, result_ptr, sizeof (int)); assert (error == hipSuccess); printf ("result is %d\n", result); assert (result == 3); return 0; } This program can be compiled with: $ hipcc simple.cpp -g -O0 -o simple ... where `hipcc` is the HIP compiler, shipped with ROCm releases. This generates an ELF binary for the host architecture, containing another ELF binary with the device code. The ELF for the device can be inspected with: $ roc-obj-ls simple 1 host-x86_64-unknown-linux file://simple#offset=8192&size=0 1 hipv4-amdgcn-amd-amdhsa--gfx906 file://simple#offset=8192&size=34216 $ roc-obj-extract 'file://simple#offset=8192&size=34216' $ file simple-offset8192-size34216.co simple-offset8192-size34216.co: ELF 64-bit LSB shared object, *unknown arch 0xe0* version 1, dynamically linked, with debug_info, not stripped ^ amcgcn architecture that my `file` doesn't know about ----´ Running the program gives the very unimpressive result: $ ./simple result is 3 While running, this host program has copied the device program into the GPU's memory and spawned an execution thread on it. The goal of this GDB port is to let the user debug host threads and these GPU threads simultaneously. Here's a sample session using a GDB with this patch applied: $ ./gdb -q -nx --data-directory=data-directory ./simple Reading symbols from ./simple... (gdb) break do_an_addition Function "do_an_addition" not defined. Make breakpoint pending on future shared library load? (y or [n]) y Breakpoint 1 (do_an_addition) pending. (gdb) r Starting program: /home/smarchi/build/binutils-gdb-amdgpu/gdb/simple [Thread debugging using libthread_db enabled] Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1". [New Thread 0x7ffff5db7640 (LWP 1082911)] [New Thread 0x7ffef53ff640 (LWP 1082913)] [Thread 0x7ffef53ff640 (LWP 1082913) exited] [New Thread 0x7ffdecb53640 (LWP 1083185)] [New Thread 0x7ffff54bf640 (LWP 1083186)] [Thread 0x7ffdecb53640 (LWP 1083185) exited] [Switching to AMDGPU Wave 2:2:1:1 (0,0,0)/0] Thread 6 hit Breakpoint 1, do_an_addition (a=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, b=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, out=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>) at simple.cpp:24 24 *out = a + b; (gdb) info inferiors Num Description Connection Executable * 1 process 1082907 1 (native) /home/smarchi/build/binutils-gdb-amdgpu/gdb/simple (gdb) info threads Id Target Id Frame 1 Thread 0x7ffff5dc9240 (LWP 1082907) "simple" 0x00007ffff5e9410b in ?? () from /opt/rocm-5.4.0/lib/libhsa-runtime64.so.1 2 Thread 0x7ffff5db7640 (LWP 1082911) "simple" __GI___ioctl (fd=3, request=3222817548) at ../sysdeps/unix/sysv/linux/ioctl.c:36 5 Thread 0x7ffff54bf640 (LWP 1083186) "simple" __GI___ioctl (fd=3, request=3222817548) at ../sysdeps/unix/sysv/linux/ioctl.c:36 * 6 AMDGPU Wave 2:2:1:1 (0,0,0)/0 do_an_addition ( a=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, b=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, out=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>) at simple.cpp:24 (gdb) bt Python Exception <class 'gdb.error'>: Unhandled dwarf expression opcode 0xe1 #0 do_an_addition (a=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, b=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, out=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>) at simple.cpp:24 (gdb) continue Continuing. result is 3 warning: Temporarily disabling breakpoints for unloaded shared library "file:///home/smarchi/build/binutils-gdb-amdgpu/gdb/simple#offset=8192&size=67208" [Thread 0x7ffff54bf640 (LWP 1083186) exited] [Thread 0x7ffff5db7640 (LWP 1082911) exited] [Inferior 1 (process 1082907) exited normally] One thing to notice is the host and GPU threads appearing under the same inferior. This is a design goal for us, as programmers tend to think of the threads running on the GPU as part of the same program as the host threads, so showing them in the same inferior in GDB seems natural. Also, the host and GPU threads share a global memory space, which fits the inferior model. Another thing to notice is the error messages when trying to read variables or printing a backtrace. This is expected for the moment, since the AMD GPU compiler produces some DWARF that uses some non-standard extensions: https://llvm.org/docs/AMDGPUDwarfExtensionsForHeterogeneousDebugging.html There were already some patches posted by Zoran Zaric earlier to make GDB support these extensions: https://inbox.sourceware.org/gdb-patches/20211105113849.118800-1-zoran.zaric@amd.com/ We think it's better to get the basic support for AMD GPU in first, which will then give a better justification for GDB to support these extensions. GPU threads are named `AMDGPU Wave`: a wave is essentially a hardware thread using the SIMT (single-instruction, multiple-threads) [3] execution model. GDB uses the amd-dbgapi library [4], included in the ROCm platform, for a few things related to AMD GPU threads debugging. Different components talk to the library, as show on the following diagram: +---------------------------+ +-------------+ +------------------+ | GDB | amd-dbgapi target | <-> | AMD | | Linux kernel | | +-------------------+ | Debugger | +--------+ | | | amdgcn gdbarch | <-> | API | <=> | AMDGPU | | | +-------------------+ | | | driver | | | | solib-rocm | <-> | (dbgapi.so) | +--------+---------+ +---------------------------+ +-------------+ - The amd-dbgapi target is a target_ops implementation used to control execution of GPU threads. While the debugging of host threads works by using the ptrace / wait Linux kernel interface (as usual), control of GPU threads is done through a special interface (dubbed `kfd`) exposed by the `amdgpu` Linux kernel module. GDB doesn't interact directly with `kfd`, but instead goes through the amd-dbgapi library (AMD Debugger API on the diagram). Since it provides execution control, the amd-dbgapi target should normally be a process_stratum_target, not just a target_ops. More on that later. - The amdgcn gdbarch (describing the hardware architecture of the GPU execution units) offloads some requests to the amd-dbgapi library, so that knowledge about the various architectures doesn't need to be duplicated and baked in GDB. This is for example for things like the list of registers. - The solib-rocm component is an solib provider that fetches the list of code objects loaded on the device from the amd-dbgapi library, and makes GDB read their symbols. This is very similar to other solib providers that handle shared libraries, except that here the shared libraries are the pieces of code loaded on the device. Given that Linux host threads are managed by the linux-nat target, and the GPU threads are managed by the amd-dbgapi target, having all threads appear in the same inferior requires the two targets to be in that inferior's target stack. However, there can only be one process_stratum_target in a given target stack, since there can be only one target per slot. To achieve it, we therefore resort the hack^W solution of placing the amd-dbgapi target in the arch_stratum slot of the target stack, on top of the linux-nat target. Doing so allows the amd-dbgapi target to intercept target calls and handle them if they concern GPU threads, and offload to beneath otherwise. See amd_dbgapi_target::fetch_registers for a simple example: void amd_dbgapi_target::fetch_registers (struct regcache *regcache, int regno) { if (!ptid_is_gpu (regcache->ptid ())) { beneath ()->fetch_registers (regcache, regno); return; } // handle it } ptids of GPU threads are crafted with the following pattern: (pid, 1, wave id) Where pid is the inferior's pid and "wave id" is the wave handle handed to us by the amd-dbgapi library (in practice, a monotonically incrementing integer). The idea is that on Linux systems, the combination (pid != 1, lwp == 1) is not possible. lwp == 1 would always belong to the init process, which would also have pid == 1 (and it's improbable for the init process to offload work to the GPU and much less for the user to debug it). We can therefore differentiate GPU and non-GPU ptids this way. See ptid_is_gpu for more details. Note that we believe that this scheme could break down in the context of containers, where the initial process executed in a container has pid 1 (in its own pid namespace). For instance, if you were to execute a ROCm program in a container, then spawn a GDB in that container and attach to the process, it will likely not work. This is a known limitation. A workaround for this is to have a dummy process (like a shell) fork and execute the program of interest. The amd-dbgapi target watches native inferiors, and "attaches" to them using amd_dbgapi_process_attach, which gives it a notifier fd that is registered in the event loop (see enable_amd_dbgapi). Note that this isn't the same "attach" as in PTRACE_ATTACH, but being ptrace-attached is a precondition for amd_dbgapi_process_attach to work. When the debugged process enables the ROCm runtime, the amd-dbgapi target gets notified through that fd, and pushes itself on the target stack of the inferior. The amd-dbgapi target is then able to intercept target_ops calls. If the debugged process disables the ROCm runtime, the amd-dbgapi target unpushes itself from the target stack. This way, the amd-dbgapi target's footprint stays minimal when debugging a process that doesn't use the AMD ROCm platform, it does not intercept target calls. The amd-dbgapi library is found using pkg-config. Since enabling support for the amdgpu architecture (amdgpu-tdep.c) depends on the amd-dbgapi library being present, we have the following logic for the interaction with --target and --enable-targets: - if the user explicitly asks for amdgcn support with --target=amdgcn-*-* or --enable-targets=amdgcn-*-*, we probe for the amd-dbgapi and fail if not found - if the user uses --enable-targets=all, we probe for amd-dbgapi, enable amdgcn support if found, disable amdgcn support if not found - if the user uses --enable-targets=all and --with-amd-dbgapi=yes, we probe for amd-dbgapi, enable amdgcn if found and fail if not found - if the user uses --enable-targets=all and --with-amd-dbgapi=no, we do not probe for amd-dbgapi, disable amdgcn support - otherwise, amd-dbgapi is not probed for and support for amdgcn is not enabled Finally, a simple test is included. It only tests hitting a breakpoint in device code and resuming execution, pretty much like the example shown above. [1] https://docs.amd.com/category/ROCm_v5.4 [2] https://docs.amd.com/bundle/HIP-Programming-Guide-v5.4 [3] https://en.wikipedia.org/wiki/Single_instruction,_multiple_threads [4] https://docs.amd.com/bundle/ROCDebugger-API-Guide-v5.4 Change-Id: I591edca98b8927b1e49e4b0abe4e304765fed9ee Co-Authored-By: Zoran Zaric <zoran.zaric@amd.com> Co-Authored-By: Laurent Morichetti <laurent.morichetti@amd.com> Co-Authored-By: Tony Tye <Tony.Tye@amd.com> Co-Authored-By: Lancelot SIX <lancelot.six@amd.com> Co-Authored-By: Pedro Alves <pedro@palves.net>
2023-01-04 04:07:07 +08:00
/* Target used to communicate with the AMD Debugger API.
Copyright (C) 2019-2024 Free Software Foundation, Inc.
gdb: initial support for ROCm platform (AMDGPU) debugging This patch adds the foundation for GDB to be able to debug programs offloaded to AMD GPUs using the AMD ROCm platform [1]. The latest public release of the ROCm release at the time of writing is 5.4, so this is what this patch targets. The ROCm platform allows host programs to schedule bits of code for execution on GPUs or similar accelerators. The programs running on GPUs are typically referred to as `kernels` (not related to operating system kernels). Programs offloaded with the AMD ROCm platform can be written in the HIP language [2], OpenCL and OpenMP, but we're going to focus on HIP here. The HIP language consists of a C++ Runtime API and kernel language. Here's an example of a very simple HIP program: #include "hip/hip_runtime.h" #include <cassert> __global__ void do_an_addition (int a, int b, int *out) { *out = a + b; } int main () { int *result_ptr, result; /* Allocate memory for the device to write the result to. */ hipError_t error = hipMalloc (&result_ptr, sizeof (int)); assert (error == hipSuccess); /* Run `do_an_addition` on one workgroup containing one work item. */ do_an_addition<<<dim3(1), dim3(1), 0, 0>>> (1, 2, result_ptr); /* Copy result from device to host. Note that this acts as a synchronization point, waiting for the kernel dispatch to complete. */ error = hipMemcpyDtoH (&result, result_ptr, sizeof (int)); assert (error == hipSuccess); printf ("result is %d\n", result); assert (result == 3); return 0; } This program can be compiled with: $ hipcc simple.cpp -g -O0 -o simple ... where `hipcc` is the HIP compiler, shipped with ROCm releases. This generates an ELF binary for the host architecture, containing another ELF binary with the device code. The ELF for the device can be inspected with: $ roc-obj-ls simple 1 host-x86_64-unknown-linux file://simple#offset=8192&size=0 1 hipv4-amdgcn-amd-amdhsa--gfx906 file://simple#offset=8192&size=34216 $ roc-obj-extract 'file://simple#offset=8192&size=34216' $ file simple-offset8192-size34216.co simple-offset8192-size34216.co: ELF 64-bit LSB shared object, *unknown arch 0xe0* version 1, dynamically linked, with debug_info, not stripped ^ amcgcn architecture that my `file` doesn't know about ----´ Running the program gives the very unimpressive result: $ ./simple result is 3 While running, this host program has copied the device program into the GPU's memory and spawned an execution thread on it. The goal of this GDB port is to let the user debug host threads and these GPU threads simultaneously. Here's a sample session using a GDB with this patch applied: $ ./gdb -q -nx --data-directory=data-directory ./simple Reading symbols from ./simple... (gdb) break do_an_addition Function "do_an_addition" not defined. Make breakpoint pending on future shared library load? (y or [n]) y Breakpoint 1 (do_an_addition) pending. (gdb) r Starting program: /home/smarchi/build/binutils-gdb-amdgpu/gdb/simple [Thread debugging using libthread_db enabled] Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1". [New Thread 0x7ffff5db7640 (LWP 1082911)] [New Thread 0x7ffef53ff640 (LWP 1082913)] [Thread 0x7ffef53ff640 (LWP 1082913) exited] [New Thread 0x7ffdecb53640 (LWP 1083185)] [New Thread 0x7ffff54bf640 (LWP 1083186)] [Thread 0x7ffdecb53640 (LWP 1083185) exited] [Switching to AMDGPU Wave 2:2:1:1 (0,0,0)/0] Thread 6 hit Breakpoint 1, do_an_addition (a=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, b=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, out=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>) at simple.cpp:24 24 *out = a + b; (gdb) info inferiors Num Description Connection Executable * 1 process 1082907 1 (native) /home/smarchi/build/binutils-gdb-amdgpu/gdb/simple (gdb) info threads Id Target Id Frame 1 Thread 0x7ffff5dc9240 (LWP 1082907) "simple" 0x00007ffff5e9410b in ?? () from /opt/rocm-5.4.0/lib/libhsa-runtime64.so.1 2 Thread 0x7ffff5db7640 (LWP 1082911) "simple" __GI___ioctl (fd=3, request=3222817548) at ../sysdeps/unix/sysv/linux/ioctl.c:36 5 Thread 0x7ffff54bf640 (LWP 1083186) "simple" __GI___ioctl (fd=3, request=3222817548) at ../sysdeps/unix/sysv/linux/ioctl.c:36 * 6 AMDGPU Wave 2:2:1:1 (0,0,0)/0 do_an_addition ( a=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, b=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, out=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>) at simple.cpp:24 (gdb) bt Python Exception <class 'gdb.error'>: Unhandled dwarf expression opcode 0xe1 #0 do_an_addition (a=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, b=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>, out=<error reading variable: DWARF-2 expression error: `DW_OP_regx' operations must be used either alone or in conjunction with DW_OP_piece or DW_OP_bit_piece.>) at simple.cpp:24 (gdb) continue Continuing. result is 3 warning: Temporarily disabling breakpoints for unloaded shared library "file:///home/smarchi/build/binutils-gdb-amdgpu/gdb/simple#offset=8192&size=67208" [Thread 0x7ffff54bf640 (LWP 1083186) exited] [Thread 0x7ffff5db7640 (LWP 1082911) exited] [Inferior 1 (process 1082907) exited normally] One thing to notice is the host and GPU threads appearing under the same inferior. This is a design goal for us, as programmers tend to think of the threads running on the GPU as part of the same program as the host threads, so showing them in the same inferior in GDB seems natural. Also, the host and GPU threads share a global memory space, which fits the inferior model. Another thing to notice is the error messages when trying to read variables or printing a backtrace. This is expected for the moment, since the AMD GPU compiler produces some DWARF that uses some non-standard extensions: https://llvm.org/docs/AMDGPUDwarfExtensionsForHeterogeneousDebugging.html There were already some patches posted by Zoran Zaric earlier to make GDB support these extensions: https://inbox.sourceware.org/gdb-patches/20211105113849.118800-1-zoran.zaric@amd.com/ We think it's better to get the basic support for AMD GPU in first, which will then give a better justification for GDB to support these extensions. GPU threads are named `AMDGPU Wave`: a wave is essentially a hardware thread using the SIMT (single-instruction, multiple-threads) [3] execution model. GDB uses the amd-dbgapi library [4], included in the ROCm platform, for a few things related to AMD GPU threads debugging. Different components talk to the library, as show on the following diagram: +---------------------------+ +-------------+ +------------------+ | GDB | amd-dbgapi target | <-> | AMD | | Linux kernel | | +-------------------+ | Debugger | +--------+ | | | amdgcn gdbarch | <-> | API | <=> | AMDGPU | | | +-------------------+ | | | driver | | | | solib-rocm | <-> | (dbgapi.so) | +--------+---------+ +---------------------------+ +-------------+ - The amd-dbgapi target is a target_ops implementation used to control execution of GPU threads. While the debugging of host threads works by using the ptrace / wait Linux kernel interface (as usual), control of GPU threads is done through a special interface (dubbed `kfd`) exposed by the `amdgpu` Linux kernel module. GDB doesn't interact directly with `kfd`, but instead goes through the amd-dbgapi library (AMD Debugger API on the diagram). Since it provides execution control, the amd-dbgapi target should normally be a process_stratum_target, not just a target_ops. More on that later. - The amdgcn gdbarch (describing the hardware architecture of the GPU execution units) offloads some requests to the amd-dbgapi library, so that knowledge about the various architectures doesn't need to be duplicated and baked in GDB. This is for example for things like the list of registers. - The solib-rocm component is an solib provider that fetches the list of code objects loaded on the device from the amd-dbgapi library, and makes GDB read their symbols. This is very similar to other solib providers that handle shared libraries, except that here the shared libraries are the pieces of code loaded on the device. Given that Linux host threads are managed by the linux-nat target, and the GPU threads are managed by the amd-dbgapi target, having all threads appear in the same inferior requires the two targets to be in that inferior's target stack. However, there can only be one process_stratum_target in a given target stack, since there can be only one target per slot. To achieve it, we therefore resort the hack^W solution of placing the amd-dbgapi target in the arch_stratum slot of the target stack, on top of the linux-nat target. Doing so allows the amd-dbgapi target to intercept target calls and handle them if they concern GPU threads, and offload to beneath otherwise. See amd_dbgapi_target::fetch_registers for a simple example: void amd_dbgapi_target::fetch_registers (struct regcache *regcache, int regno) { if (!ptid_is_gpu (regcache->ptid ())) { beneath ()->fetch_registers (regcache, regno); return; } // handle it } ptids of GPU threads are crafted with the following pattern: (pid, 1, wave id) Where pid is the inferior's pid and "wave id" is the wave handle handed to us by the amd-dbgapi library (in practice, a monotonically incrementing integer). The idea is that on Linux systems, the combination (pid != 1, lwp == 1) is not possible. lwp == 1 would always belong to the init process, which would also have pid == 1 (and it's improbable for the init process to offload work to the GPU and much less for the user to debug it). We can therefore differentiate GPU and non-GPU ptids this way. See ptid_is_gpu for more details. Note that we believe that this scheme could break down in the context of containers, where the initial process executed in a container has pid 1 (in its own pid namespace). For instance, if you were to execute a ROCm program in a container, then spawn a GDB in that container and attach to the process, it will likely not work. This is a known limitation. A workaround for this is to have a dummy process (like a shell) fork and execute the program of interest. The amd-dbgapi target watches native inferiors, and "attaches" to them using amd_dbgapi_process_attach, which gives it a notifier fd that is registered in the event loop (see enable_amd_dbgapi). Note that this isn't the same "attach" as in PTRACE_ATTACH, but being ptrace-attached is a precondition for amd_dbgapi_process_attach to work. When the debugged process enables the ROCm runtime, the amd-dbgapi target gets notified through that fd, and pushes itself on the target stack of the inferior. The amd-dbgapi target is then able to intercept target_ops calls. If the debugged process disables the ROCm runtime, the amd-dbgapi target unpushes itself from the target stack. This way, the amd-dbgapi target's footprint stays minimal when debugging a process that doesn't use the AMD ROCm platform, it does not intercept target calls. The amd-dbgapi library is found using pkg-config. Since enabling support for the amdgpu architecture (amdgpu-tdep.c) depends on the amd-dbgapi library being present, we have the following logic for the interaction with --target and --enable-targets: - if the user explicitly asks for amdgcn support with --target=amdgcn-*-* or --enable-targets=amdgcn-*-*, we probe for the amd-dbgapi and fail if not found - if the user uses --enable-targets=all, we probe for amd-dbgapi, enable amdgcn support if found, disable amdgcn support if not found - if the user uses --enable-targets=all and --with-amd-dbgapi=yes, we probe for amd-dbgapi, enable amdgcn if found and fail if not found - if the user uses --enable-targets=all and --with-amd-dbgapi=no, we do not probe for amd-dbgapi, disable amdgcn support - otherwise, amd-dbgapi is not probed for and support for amdgcn is not enabled Finally, a simple test is included. It only tests hitting a breakpoint in device code and resuming execution, pretty much like the example shown above. [1] https://docs.amd.com/category/ROCm_v5.4 [2] https://docs.amd.com/bundle/HIP-Programming-Guide-v5.4 [3] https://en.wikipedia.org/wiki/Single_instruction,_multiple_threads [4] https://docs.amd.com/bundle/ROCDebugger-API-Guide-v5.4 Change-Id: I591edca98b8927b1e49e4b0abe4e304765fed9ee Co-Authored-By: Zoran Zaric <zoran.zaric@amd.com> Co-Authored-By: Laurent Morichetti <laurent.morichetti@amd.com> Co-Authored-By: Tony Tye <Tony.Tye@amd.com> Co-Authored-By: Lancelot SIX <lancelot.six@amd.com> Co-Authored-By: Pedro Alves <pedro@palves.net>
2023-01-04 04:07:07 +08:00
This file is part of GDB.
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>. */
#ifndef AMD_DBGAPI_TARGET_H
#define AMD_DBGAPI_TARGET_H 1
#include "gdbsupport/observable.h"
#include <amd-dbgapi/amd-dbgapi.h>
struct inferior;
namespace detail
{
template <typename T>
using is_amd_dbgapi_handle
= gdb::Or<std::is_same<T, amd_dbgapi_address_class_id_t>,
std::is_same<T, amd_dbgapi_address_space_id_t>,
std::is_same<T, amd_dbgapi_architecture_id_t>,
std::is_same<T, amd_dbgapi_agent_id_t>,
std::is_same<T, amd_dbgapi_breakpoint_id_t>,
std::is_same<T, amd_dbgapi_code_object_id_t>,
std::is_same<T, amd_dbgapi_dispatch_id_t>,
std::is_same<T, amd_dbgapi_displaced_stepping_id_t>,
std::is_same<T, amd_dbgapi_event_id_t>,
std::is_same<T, amd_dbgapi_process_id_t>,
std::is_same<T, amd_dbgapi_queue_id_t>,
std::is_same<T, amd_dbgapi_register_class_id_t>,
std::is_same<T, amd_dbgapi_register_id_t>,
std::is_same<T, amd_dbgapi_watchpoint_id_t>,
std::is_same<T, amd_dbgapi_wave_id_t>>;
} /* namespace detail */
/* Get the token of amd-dbgapi's inferior_created observer. */
const gdb::observers::token &
get_amd_dbgapi_target_inferior_created_observer_token ();
/* Comparison operators for amd-dbgapi handle types. */
template <typename T,
typename = gdb::Requires<detail::is_amd_dbgapi_handle<T>>>
bool
operator== (const T &lhs, const T &rhs)
{
return lhs.handle == rhs.handle;
}
template <typename T,
typename = gdb::Requires<detail::is_amd_dbgapi_handle<T>>>
bool
operator!= (const T &lhs, const T &rhs)
{
return !(lhs == rhs);
}
/* Return true if the given ptid is a GPU thread (wave) ptid. */
static inline bool
ptid_is_gpu (ptid_t ptid)
{
/* FIXME: Currently using values that are known not to conflict with other
processes to indicate if it is a GPU thread. ptid.pid 1 is the init
process and is the only process that could have a ptid.lwp of 1. The init
process cannot have a GPU. No other process can have a ptid.lwp of 1.
The GPU wave ID is stored in the ptid.tid. */
return ptid.pid () != 1 && ptid.lwp () == 1;
}
/* Return INF's amd_dbgapi process id. */
amd_dbgapi_process_id_t get_amd_dbgapi_process_id (inferior *inf);
/* Get the amd-dbgapi wave id for PTID. */
static inline amd_dbgapi_wave_id_t
get_amd_dbgapi_wave_id (ptid_t ptid)
{
gdb_assert (ptid_is_gpu (ptid));
return amd_dbgapi_wave_id_t {
static_cast<decltype (amd_dbgapi_wave_id_t::handle)> (ptid.tid ())
};
}
/* Get the textual version of STATUS.
Always returns non-nullptr, and asserts that STATUS has a valid value. */
static inline const char *
get_status_string (amd_dbgapi_status_t status)
{
const char *ret;
status = amd_dbgapi_get_status_string (status, &ret);
gdb_assert (status == AMD_DBGAPI_STATUS_SUCCESS);
return ret;
}
#endif /* AMD_DBGAPI_TARGET_H */