AI/rocm-systems
2
0 Rozštěpit 0
Zdrojový kód Úkoly Pull requesty Akce Balíčky Projekty Vydání Wiki Aktivita

Fix rocprofiler_iterate_callback_tracing_kind_operation_args for HIP compiler callbacks (#532)

Procházet zdrojové kódy 
* Fix HIP compiler iterate args

- `include/rocprofiler-sdk/hip/api_args.h`
  - replace struct fields named "f" with "func"
  - replace hip stream fields named "hStream" with "stream"
- `lib/rocprofiler-sdk/callback_tracing.cpp`
  - iterate_args for HIP compiler table
- `lib/rocprofiler-sdk/registration.cpp`
  - fix warning about roctx num_tables
- `lib/rocprofiler-sdk/hip/hip.def.cpp`
  - replace struct fields named "f" with "func"
  - replace hip stream fields named "hStream" with "stream"
- `lib/rocprofiler-sdk/{hip,hsa,marker}/utils.hpp`
  - improve `stringize_impl`
- `lib/rocprofiler-sdk/hsa/code_object.cpp`
  - remove stale commented out code
- `lib/rocprofiler-sdk/hsa/queue_controller.*`
  - destory_queue -> destroy_queue
- `tests/tools/json-tool.cpp`
  - improve parallelism in tool_tracing_callback
  - serialize the marker api args
  - only invoke rocprofiler_iterate_callback_tracing_kind_operation_args in exit phase
- `samples/counter_collection/CMakeLists.txt`
  - reduce timeout on tests to 120 seconds

* Update lib/rocprofiler-sdk/hsa/utils.hpp

- disable dereference of double pointer in stringize_impl

* Update lib/common

- indirection_level in mpl.hpp
- stringize_arg.hpp

* Rework rocprofiler_iterate_callback_tracing_kind_operation_args

- provide more information in rocprofiler_callback_tracing_operation_args_cb_t
- support specifying the dereference level to account for output paramters

[ROCm/rocprofiler-sdk commit: 1bb94add11]
Tento commit je obsažen v:
Jonathan R. Madsen authored and GitHub committed 2024-03-01 01:46:07 -06:00
1 rodič 15302ff11d
revize cb6b79c323
27 files changed
+419 -200
Stáhněte soubor záplaty Stáhněte rozdílový soubor Rozbalit všechny soubory Sbalit všechny soubory

No files matched your search

projects/rocprofiler-sdk/source/include/rocprofiler-sdk/callback_tracing.h
+59 -2
Zobrazit soubor
@@ -193,18 +193,26 @@ typedef int (*rocprofiler_callback_tracing_kind_operation_cb_t)(
* @param [in] kind domain
* @param [in] operation associated domain operation
* @param [in] arg_number the argument number, starting at zero
* @param [in] arg_value_addr the address of the argument stored by rocprofiler.
* @param [in] arg_indirection_count the total number of indirection levels for the argument, e.g.
* int == 0, int* == 1, int** == 2
* @param [in] arg_type the typeid name of the argument
* @param [in] arg_name the name of the argument in the prototype (or rocprofiler union)
* @param [in] arg_value_str conversion of the argument to a string, e.g. operator<< overload
* @param [in] arg_value_addr the address of the argument stored by rocprofiler.
* @param [in] arg_dereference_count the number of times the argument was dereferenced when it was
* converted to a string
* @param [in] data user data
*/
typedef int (*rocprofiler_callback_tracing_operation_args_cb_t)(
rocprofiler_callback_tracing_kind_t kind,
uint32_t operation,
uint32_t arg_number,
const void* const arg_value_addr,
int32_t arg_indirection_count,
const char* arg_type,
const char* arg_name,
const char* arg_value_str,
const void* const arg_value_addr,
int32_t arg_dereference_count,
void* data);
/**
@@ -322,14 +330,63 @@ rocprofiler_iterate_callback_tracing_kind_operations(
* particularly useful when tools want to annotate traces with the function arguments. See
* @example samples/api_callback_tracing/client.cpp for a usage example.
*
* It is recommended to use this function when the record phase is ::ROCPROFILER_CALLBACK_PHASE_EXIT
* or ::ROCPROFILER_CALLBACK_PHASE_NONE. When the phase is ::ROCPROFILER_CALLBACK_PHASE_ENTER, the
* function may have output parameters which have not set. In the case of an output parameter with
* one level of indirection, e.g. `int* output_len`, this is considered safe since the output
* parameter is either null or, in the worst case scenario, pointing to an uninitialized value which
* will result in garbage values to be stringified. However, if the output parameter has more than
* one level of indirection, e.g. `const char** output_name`, this can result in a segmentation
* fault because the dereferenced output parameter may be uninitialized and point to an invalid
address. E.g.:
*
* @code{.cpp}
* struct dim3
* {
* int x;
* int y;
* int z;
* };
*
* static dim3 default_dims = {.x = 1, .y = 1, .z = 1};
*
* void set_dim_x(int val, dim3* output_dims) { output_dims->x = val; }
*
* void get_default_dims(dim3** output_dims) { *output_dims = default_dims; }
*
* int main()
* {
* dim3 my_dims; // uninitialized value. x, y, and z may be set to random values
* dim3* current_dims; // uninitialized pointer. May be set to invalid address
*
* set_dim_x(3, &my_dims); // if rocprofiler-sdk wrapped this function and tried to stringify
* // in the enter phase, dereferencing my_dims is not problematic
* // since there is an actual dim3 allocation
*
* get_default_dims(&current_dims); // if rocprofiler-sdk wrapped this function,
* // and tried to stringify in the enter phase,
* // current_dims may point to an address outside
* // of the address space of this process and
* // cause a segfault
* }
* @endcode
*
*
* @param[in] record Record provided by service callback
* @param[in] callback The callback function which will be invoked for each argument
* @param[in] max_dereference_count In the callback enter phase, certain arguments may be output
* parameters which have not been set. When the output parameter has multiple levels of indirection,
* it may be invalid to dereference the output parameter more than once and doing so may result in a
* segmentation fault. Thus, it is recommended to set this parameter to a maximum value of 1 when
* the phase is ::ROCPROFILER_CALLBACK_PHASE_ENTER to ensure that output parameters which point to
* uninitialized pointers do not cause segmentation faults.
* @param[in] user_data Data to be passed to each invocation of the callback
*/
rocprofiler_status_t ROCPROFILER_API
rocprofiler_iterate_callback_tracing_kind_operation_args(
rocprofiler_callback_tracing_record_t record,
rocprofiler_callback_tracing_operation_args_cb_t callback,
int32_t max_dereference_count,
void* user_data) ROCPROFILER_NONNULL(2);
/** @} */
projects/rocprofiler-sdk/source/include/rocprofiler-sdk/hip/api_args.h
+17 -17
Zobrazit soubor
@@ -1298,7 +1298,7 @@ typedef union rocprofiler_hip_api_args_u
} hipIpcOpenMemHandle;
struct
{
hipFunction_t f;
hipFunction_t func;
} hipKernelNameRef;
struct
{
@@ -1311,7 +1311,7 @@ typedef union rocprofiler_hip_api_args_u
} hipLaunchByPtr;
struct
{
const void* f;
const void* func;
dim3 gridDim;
dim3 blockDimX;
void** kernelParams;
@@ -1960,7 +1960,7 @@ typedef union rocprofiler_hip_api_args_u
} hipModuleGetTexRef;
struct
{
hipFunction_t f;
hipFunction_t func;
unsigned int gridDimX;
unsigned int gridDimY;
unsigned int gridDimZ;
@@ -1979,7 +1979,7 @@ typedef union rocprofiler_hip_api_args_u
} hipModuleLaunchCooperativeKernelMultiDevice;
struct
{
hipFunction_t f;
hipFunction_t func;
unsigned int gridDimX;
unsigned int gridDimY;
unsigned int gridDimZ;
@@ -2012,14 +2012,14 @@ typedef union rocprofiler_hip_api_args_u
struct
{
int* numBlocks;
hipFunction_t f;
hipFunction_t func;
int blockSize;
size_t dynSharedMemPerBlk;
} hipModuleOccupancyMaxActiveBlocksPerMultiprocessor;
struct
{
int* numBlocks;
hipFunction_t f;
hipFunction_t func;
int blockSize;
size_t dynSharedMemPerBlk;
unsigned int flags;
@@ -2028,7 +2028,7 @@ typedef union rocprofiler_hip_api_args_u
{
int* gridSize;
int* blockSize;
hipFunction_t f;
hipFunction_t func;
size_t dynSharedMemPerBlk;
int blockSizeLimit;
} hipModuleOccupancyMaxPotentialBlockSize;
@@ -2036,7 +2036,7 @@ typedef union rocprofiler_hip_api_args_u
{
int* gridSize;
int* blockSize;
hipFunction_t f;
hipFunction_t func;
size_t dynSharedMemPerBlk;
int blockSizeLimit;
unsigned int flags;
@@ -2048,14 +2048,14 @@ typedef union rocprofiler_hip_api_args_u
struct
{
int* numBlocks;
const void* f;
const void* func;
int blockSize;
size_t dynSharedMemPerBlk;
} hipOccupancyMaxActiveBlocksPerMultiprocessor;
struct
{
int* numBlocks;
const void* f;
const void* func;
int blockSize;
size_t dynSharedMemPerBlk;
unsigned int flags;
@@ -2064,7 +2064,7 @@ typedef union rocprofiler_hip_api_args_u
{
int* gridSize;
int* blockSize;
const void* f;
const void* func;
size_t dynSharedMemPerBlk;
int blockSizeLimit;
} hipOccupancyMaxPotentialBlockSize;
@@ -2453,7 +2453,7 @@ typedef union rocprofiler_hip_api_args_u
} hipCreateChannelDesc;
struct
{
hipFunction_t f;
hipFunction_t func;
uint32_t globalWorkSizeX;
uint32_t globalWorkSizeY;
uint32_t globalWorkSizeZ;
@@ -2461,7 +2461,7 @@ typedef union rocprofiler_hip_api_args_u
uint32_t localWorkSizeY;
uint32_t localWorkSizeZ;
size_t sharedMemBytes;
hipStream_t hStream;
hipStream_t stream;
void** kernelParams;
void** extra;
hipEvent_t startEvent;
@@ -2470,7 +2470,7 @@ typedef union rocprofiler_hip_api_args_u
} hipExtModuleLaunchKernel;
struct
{
hipFunction_t f;
hipFunction_t func;
uint32_t globalWorkSizeX;
uint32_t globalWorkSizeY;
uint32_t globalWorkSizeZ;
@@ -2478,7 +2478,7 @@ typedef union rocprofiler_hip_api_args_u
uint32_t localWorkSizeY;
uint32_t localWorkSizeZ;
size_t sharedMemBytes;
hipStream_t hStream;
hipStream_t stream;
void** kernelParams;
void** extra;
hipEvent_t startEvent;
@@ -2699,12 +2699,12 @@ typedef union rocprofiler_hip_api_args_u
} hipEventRecord_spt;
struct
{
const void* f;
const void* func;
dim3 gridDim;
dim3 blockDim;
void** kernelParams;
uint32_t sharedMemBytes;
hipStream_t hStream;
hipStream_t stream;
} hipLaunchCooperativeKernel_spt;
struct
{