Updated rocprofiler.h for v2 (#18)

* Update and rename rocprofiler.h to rocprofiler.h.in - Removing Service IDs - Fixing agent_id to be agent * [0/N] New rocprofiler headers - created rocprofiler/defines.h - ppdef macros - created rocprofiler/hip.h - HIP specific types - created rocprofiler/hsa.h - HSA specific types - created rocprofiler/marker.h - Marker (ROCTx) specific types - create version.h.in - file containing version info - updated source/lib/rocprofiler/CMakeLists.txt - set DEFINE_SYMBOL - compile defs provided by rocprofiler::rocprofiler-headers * [1/N] Update rocprofiler.h - pragma once - removed some ppdefs (in version.h.in and defines.h) - extern "C" after includes - added *_NONE and *_LAST enum values to all enums - provided some rocprofiler_status_t enums - tweaked rocprofiler_agent_type_t enum fields - tweaked rocprofiler_agent_info_t enum fields - provided rocprofiler_tracer_activity_domain_t - added missing rocprofiler_counter_instance_id_t typedef - may not be correct - provided rocprofiler_record_header_t struct - provided rocprofiler_record_tracer_t struct - add ROCPROFILER_NONNULL attribute where appropriate - CMakeLists.txt: add subdirectories for hsa, hip, and marker - defines.h: remove ROCPROFILER_CALL ppdef - rocprofiler.h - ROCPROFILER_STATUS_ERROR_NOT_IMPLEMENTED - extend rocprofiler_agent_t - modify rocprofiler_query_available_agents signature to callback - rocprofiler_pc_sampling_config_array_t - update rocprofiler_buffer_callback_t to include context id - update rocprofiler_create_buffer to accept context - rocprofiler_plugin.h - non-const rocprofiler_record_header_t** * [2/N] Update include/rocprofiler/rocprofiler_plugin.h - change prototype of rocprofiler_plugin_write_buffer_records to resemble rocprofiler_buffer_callback_t * [3/N] Update include/rocprofiler/hsa - Update hsa.h - Details in hsa subfolder * [4/N] Update include/rocprofiler/hip - Update hip.h - Details in hip subfolder * [5/N] Update include/rocprofiler/marker - Update marker.h - Details in marker subfolder * [6/N] Update samples/pc_sampling - fix issues with macros - fix API changes --------- Co-authored-by: Jonathan Madsen <jrmadsen@users.noreply.github.com> [ROCm/rocprofiler-sdk commit: 39b209c2a7]
2023-08-24 15:39:56 -07:00
commit cafeaadb8c
@@ -1 +1,7 @@
+#
+#
+#
+project(rocprofiler-samples LANGUAGES C CXX)

+# add_subdirectory(api_tracing)
+add_subdirectory(pc_sampling)
@@ -0,0 +1,16 @@
+#
+#
+#
+project(rocprofiler-samples-pc-sampling LANGUAGES C CXX)
+
+add_executable(pc_sampling_single-user-host-trap)
+target_sources(pc_sampling_single-user-host-trap PRIVATE common.h
+                                                         single-user-host-trap.cpp)
+target_link_libraries(pc_sampling_single-user-host-trap
+                      PRIVATE rocprofiler::rocprofiler-library)
+
+add_executable(pc_sampling_single-user-host-trap-retry)
+target_sources(pc_sampling_single-user-host-trap-retry
+               PRIVATE common.h single-user-host-trap-retries-service-instantiation.cpp)
+target_link_libraries(pc_sampling_single-user-host-trap-retry
+                      PRIVATE rocprofiler::rocprofiler-library)
@@ -0,0 +1,124 @@
+#ifndef PC_SAMPLING_COMMON_H
+#define PC_SAMPLING_COMMON_H
+
+#include <rocprofiler/rocprofiler.h>
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstddef>
+#include <cstdint>
+
+constexpr size_t BUFFER_SIZE_BYTES = 4096;
+constexpr size_t WATERMARK         = (BUFFER_SIZE_BYTES / 2);
+
+#define ROCPROFILER_CALL(result, msg)                                                              \
+    {                                                                                              \
+        rocprofiler_status_t status = result;                                                      \
+        if(status != ROCPROFILER_STATUS_SUCCESS)                                                   \
+        {                                                                                          \
+            puts(#result " failed");                                                               \
+        }                                                                                          \
+    }
+
+// We might want to test the calls that fails
+// e.g. calling `rocprofiler_configure_pc_sampling_service `
+// after previous initialization.
+#define ROCPROFILER_CALL_FAILS(result, msg)                                                        \
+    {                                                                                              \
+        rocprofiler_status_t status = result;                                                      \
+        if(status == ROCPROFILER_STATUS_SUCCESS)                                                   \
+        {                                                                                          \
+            puts(#result " succeeded");                                                            \
+        }                                                                                          \
+    }
+
+static rocprofiler_status_t
+find_first_gpu_agent_impl(rocprofiler_agent_t** agents, size_t num_agents, void* data)
+{
+    // data is required
+    if(!data) return ROCPROFILER_STATUS_ERROR;
+
+    auto* _out_agent = static_cast<rocprofiler_agent_t*>(data);
+    // find the first GPU agent
+    for(size_t i = 0; i < num_agents; i++)
+    {
+        if(agents[i]->type == ROCPROFILER_AGENT_TYPE_GPU)
+        {
+            *_out_agent = *agents[i];
+            printf("[%s] %s :: id=%zu, type=%i, num pc sample configs=%zu\n",
+                   __FUNCTION__,
+                   _out_agent->name,
+                   _out_agent->id.handle,
+                   _out_agent->type,
+                   _out_agent->pc_sampling_configs.size);
+            return ROCPROFILER_STATUS_SUCCESS;
+        }
+        else
+        {
+            printf("[%s] %s :: id=%zu, type=%i, num pc sample configs=%zu\n",
+                   __FUNCTION__,
+                   agents[i]->name,
+                   agents[i]->id.handle,
+                   agents[i]->type,
+                   agents[i]->pc_sampling_configs.size);
+        }
+    }
+    return ROCPROFILER_STATUS_ERROR;
+}
+
+static rocprofiler_agent_t
+find_first_gpu_agent()
+{
+    // This function returns the first gpu agent it encounters.
+    // TODO: write the better function querying information about the agent,
+    //   and return if the agent is MI200.
+    rocprofiler_agent_t gpu_agent;
+
+    ROCPROFILER_CALL(rocprofiler_query_available_agents(&find_first_gpu_agent_impl,
+                                                        sizeof(rocprofiler_agent_t),
+                                                        static_cast<void*>(&gpu_agent)),
+                     "Failed to find GPU agents");
+
+    return gpu_agent;
+}
+
+static void
+rocprofiler_pc_sampling_callback(rocprofiler_context_id_t /*context_id*/,
+                                 rocprofiler_buffer_id_t /*buffer_id*/,
+                                 rocprofiler_record_header_t** headers,
+                                 size_t                        num_headers,
+                                 void* /*data*/,
+                                 uint64_t drop_count)
+{
+    // Vladimir: I am not sure if this is the right way of iterating over PC sampling records.
+    printf(
+        "The number of delivered samples is: %zu, while the number of dropped samples is: %lu.\n",
+        num_headers,
+        drop_count);
+
+    for(size_t i = 0; i < num_headers; i++)
+    {
+        auto* cur_header = headers[i];
+        if(cur_header->kind == 0)
+        {
+            auto* pc_sample = static_cast<rocprofiler_pc_sampling_record_t*>(cur_header->payload);
+            printf("--- pc: %lx, dispatch_id: %lx, timestamp: %lu, hardware_id: %lu\n",
+                   pc_sample->pc,
+                   pc_sample->dispatch_id,
+                   pc_sample->timestamp,
+                   pc_sample->hardware_id);
+            // Vladimir: How to parse the remaining part of the `rocprofiler_pc_sampling_record_t`
+            // struct?
+        }
+    }
+    // Vladimr: We might want to add somewhere in the documentation that headars actually contain PC
+    // samples.
+}
+
+static void
+run_HIP_app()
+{
+    // TODO: provide the simple HIP app
+}
+
+#endif
@@ -0,0 +1,165 @@
+// Vladimir: The example of using Host-trap PC sampling on a system with single MI200/300 by two
+// users. The first user initiates Host-Trap sampling with the configuration A. The second user
+// tries initiaiting stochastic sampling with configuration B and fails. Then it queries available
+// configurations and observes only the configuration A. It accepts it and starts PC sampling.
+// Vladimir: Currently, this example is written as a single-threaded program.
+// Decide whether to move the second user to a separate thread or process
+
+#include <rocprofiler/rocprofiler.h>
+
+#include "common.h"
+
+#include <cassert>
+
+#define HOST_TRAP_INTERVAL 1000
+
+rocprofiler_pc_sampling_method_t host_trap_sampling_method;
+rocprofiler_pc_sampling_unit_t   host_trap_sampling_unit_time;
+uint64_t                         host_trap_interval;
+
+void
+second_user()
+{
+    // creating a context
+    rocprofiler_context_id_t context_id2;
+    ROCPROFILER_CALL(rocprofiler_create_context(&context_id2),
+                     "Cannot create context for the second user\n");
+
+    rocprofiler_agent_t gpu_agent = find_first_gpu_agent();
+
+    // creating a buffer that will hold pc sampling information
+    rocprofiler_buffer_policy_t lossless_buffer_action = ROCPROFILER_BUFFER_POLICY_LOSSLESS;
+    rocprofiler_buffer_id_t     buffer_id2;
+    ROCPROFILER_CALL(rocprofiler_create_buffer(context_id2,
+                                               BUFFER_SIZE_BYTES,
+                                               WATERMARK,
+                                               lossless_buffer_action,
+                                               rocprofiler_pc_sampling_callback,
+                                               nullptr,
+                                               &buffer_id2),
+                     "Cannot create pc sampling buffer for the second user");
+
+    // The second user tries to create another pc sampling service with different configuration,
+    // but the rocprofiler rejects it.
+    rocprofiler_pc_sampling_method_t sampling_method2 = ROCPROFILER_PC_SAMPLING_METHOD_STOCHASTIC;
+    rocprofiler_pc_sampling_unit_t   sampling_unit2   = ROCPROFILER_PC_SAMPLING_UNIT_CYCLES;
+    uint64_t interval2 = 2048;  // I assumed micro secs, so this should be 1ms
+    // The following function returns an error code indicating the PC sampling has already been
+    // configured.
+    ROCPROFILER_CALL_FAILS(
+        rocprofiler_configure_pc_sampling_service(
+            context_id2, gpu_agent, sampling_method2, sampling_unit2, interval2, buffer_id2),
+        "Instantiation of the PC sampling service should fail");
+
+    // After failure, the second user queries available configuration and observes the one chosen by
+    // the first user.
+    rocprofiler_pc_sampling_configuration_t* configs;
+    size_t                                   config_count;
+    ROCPROFILER_CALL(
+        rocprofiler_query_pc_sampling_agent_configurations(gpu_agent, configs, &config_count),
+        "The second user cannot query available configurations");
+
+    // Only one configuration should be listed, and its parameters should match the parameters set
+    // by the first user. Vladimir: Is it ok to use assertions? In the release mode, they might be
+    // ignored.
+    assert(config_count == 1);
+    rocprofiler_pc_sampling_configuration_t first_user_config = configs[0];
+    assert(first_user_config.method == host_trap_sampling_method);
+    assert(first_user_config.unit == host_trap_sampling_unit_time);
+    // Vladimir: Should the min_interval and max_interval have the same value at this point (the PC
+    // sampling is alredy configured)??
+    assert(first_user_config.min_interval == host_trap_interval &&
+           first_user_config.min_interval == first_user_config.max_interval);
+
+    // Vladimir: Do we need to explicitly free queried configurations?
+    free(configs);
+
+    // Reuse the same configuration set by the first user.
+    // The second user is satisfied with the configuration chosen by the first user, so it
+    // starts PC sampling.
+    ROCPROFILER_CALL(rocprofiler_configure_pc_sampling_service(context_id2,
+                                                               gpu_agent,
+                                                               first_user_config.method,
+                                                               first_user_config.unit,
+                                                               first_user_config.min_interval,
+                                                               buffer_id2),
+                     "The second user cannot share already created PC sampling configuration");
+
+    // Starting the context that should trigger PC sampling?
+    ROCPROFILER_CALL(rocprofiler_start_context(context_id2),
+                     "Cannot start PC sampling context for the second user");
+
+    // Running the applicaiton
+    run_HIP_app();
+
+    // Stop the context that should stop PC sampling?
+    ROCPROFILER_CALL(rocprofiler_stop_context(context_id2),
+                     "Cannot start PC sampling context for the second user");
+
+    // Explicit buffer flush, before destroying it
+    ROCPROFILER_CALL(rocprofiler_flush_buffer(buffer_id2),
+                     "Cannot destroy the second user's buffer");
+    // Destroying the buffer
+    ROCPROFILER_CALL(rocprofiler_destroy_buffer(buffer_id2), "Cannot destroy the second user's");
+}
+
+int
+main(int /*argc*/, char** /*argv*/)
+{
+    rocprofiler_status_t status;
+
+    // creating a context
+    rocprofiler_context_id_t context_id;
+    ROCPROFILER_CALL(rocprofiler_create_context(&context_id), "Cannot create context\n");
+
+    rocprofiler_agent_t gpu_agent = find_first_gpu_agent();
+
+    // creating a buffer that will hold pc sampling information
+    rocprofiler_buffer_policy_t drop_buffer_action = ROCPROFILER_BUFFER_POLICY_DISCARD;
+    rocprofiler_buffer_id_t     buffer_id;
+    ROCPROFILER_CALL(rocprofiler_create_buffer(context_id,
+                                               BUFFER_SIZE_BYTES,
+                                               WATERMARK,
+                                               drop_buffer_action,
+                                               rocprofiler_pc_sampling_callback,
+                                               nullptr,
+                                               &buffer_id),
+                     "Cannot create pc sampling buffer");
+
+    // PC sampling service configuration
+    rocprofiler_pc_sampling_method_t host_trap_sampling_method =
+        ROCPROFILER_PC_SAMPLING_METHOD_HOST_TRAP;
+    rocprofiler_pc_sampling_unit_t host_trap_sampling_unit_time = ROCPROFILER_PC_SAMPLING_UNIT_TIME;
+    // Vladimir: What units are we using for time? ms, micro secs, ns?
+    uint64_t host_trap_interval = HOST_TRAP_INTERVAL;
+    // Instantiating the first PC sampling service succeeds.
+    ROCPROFILER_CALL(rocprofiler_configure_pc_sampling_service(context_id,
+                                                               gpu_agent,
+                                                               host_trap_sampling_method,
+                                                               host_trap_sampling_unit_time,
+                                                               host_trap_interval,
+                                                               buffer_id),
+                     "Cannot create PC sampling service");
+
+    // Trigger the second user code.
+    // Vladimir: Discuss whether this should be put in a separate thread/process.
+    second_user();
+
+    // Starting the context that should trigger PC sampling?
+    ROCPROFILER_CALL(rocprofiler_start_context(context_id), "Cannot start PC sampling context");
+
+    // Running the applicaiton
+    run_HIP_app();
+
+    // Stop the context that should stop PC sampling?
+    ROCPROFILER_CALL(rocprofiler_stop_context(context_id), "Cannot start PC sampling context");
+
+    // Explicit buffer flush, before destroying it
+    ROCPROFILER_CALL(rocprofiler_flush_buffer(buffer_id), "Cannot destroy buffer");
+    // Destroying the buffer
+    ROCPROFILER_CALL(rocprofiler_destroy_buffer(buffer_id), "Cannot destroy buffer");
+
+    // Vladimir: Do we need to destroy context or a service?
+
+    return 0;
+}
@@ -0,0 +1,61 @@
+// Vladimir: The example of using Host-trap PC sampling exclusively on the system with single MI200.
+// If any of the rocprofiler calls returns status fail, we simply stop the application.
+
+#include <rocprofiler/rocprofiler.h>
+#include "common.h"
+
+int
+main(int /*argc*/, char** /*argv*/)
+{
+    rocprofiler_status_t status;
+
+    // creating a context
+    rocprofiler_context_id_t context_id;
+    ROCPROFILER_CALL(rocprofiler_create_context(&context_id), "Cannot create context\n");
+
+    rocprofiler_agent_t gpu_agent = find_first_gpu_agent();
+
+    // creating a buffer that will hold pc sampling information
+    rocprofiler_buffer_policy_t drop_buffer_action = ROCPROFILER_BUFFER_POLICY_DISCARD;
+    rocprofiler_buffer_id_t     buffer_id;
+    ROCPROFILER_CALL(rocprofiler_create_buffer(context_id,
+                                               BUFFER_SIZE_BYTES,
+                                               WATERMARK,
+                                               drop_buffer_action,
+                                               rocprofiler_pc_sampling_callback,
+                                               nullptr,
+                                               &buffer_id),
+                     "Cannot create pc sampling buffer");
+
+    // PC sampling service configuration
+    rocprofiler_pc_sampling_method_t sampling_method = ROCPROFILER_PC_SAMPLING_METHOD_HOST_TRAP;
+    rocprofiler_pc_sampling_unit_t   sampling_unit   = ROCPROFILER_PC_SAMPLING_UNIT_TIME;
+    // What units are we using for time? ms, micro secs, ns?
+    uint64_t interval = 1000;  // I assumed micro secs, so this should be 1ms
+    // Instantiating the PC sampling service
+    ROCPROFILER_CALL(
+        rocprofiler_configure_pc_sampling_service(
+            context_id, gpu_agent, sampling_method, sampling_unit, interval, buffer_id),
+        "Cannot create PC sampling service");
+
+    // Vladimir: Is this the place of retrying if someone already created the
+    // configuration and the previous call fails?
+
+    // Starting the context that should trigger PC sampling?
+    ROCPROFILER_CALL(rocprofiler_start_context(context_id), "Cannot start PC sampling context");
+
+    // Running the applicaiton
+    run_HIP_app();
+
+    // Stop the context that should stop PC sampling?
+    ROCPROFILER_CALL(rocprofiler_stop_context(context_id), "Cannot start PC sampling context");
+
+    // Explicit buffer flush, before destroying it
+    ROCPROFILER_CALL(rocprofiler_flush_buffer(buffer_id), "Cannot destroy buffer");
+    // Destroying the buffer
+    ROCPROFILER_CALL(rocprofiler_destroy_buffer(buffer_id), "Cannot destroy buffer");
+
+    // Vladimir: Do we need to destroy context or a service?
+
+    return 0;
+}