Files
rocm-systems/tests/bin/hsa-code-object/hsa_code_object_app.cpp
T
Nagaraj, Sriraksha d4a51e4102 Adding att v3 support (#84)
* Adding att v3 support

* misc fix

* bug fix

* Python linting workflow and rules

* fix regex

* Adding temporary args

* fix temporary args

* fix format

* remove att_perfcounters from test input

* Review comments (#163)

Co-authored-by: Giovanni Baraldi <gbaraldi@amd.com>

* Revert "Review comments (#163)"

This reverts commit 9ef0f8e5a4489d5581255e1b70ced2aef5c1c1d0.

* Address review comments 2

* review changes

* review comments

* review

* cmake alias

* review

* review

* review

* review

* Enabling percounter in v3 script

* review

* formatting

* formatting

---------

Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>
Co-authored-by: Baraldi, Giovanni <Giovanni.Baraldi@amd.com>
Co-authored-by: Giovanni Baraldi <gbaraldi@amd.com>
2025-02-04 04:05:38 -06:00

493 строки
18 KiB
C++

// MIT License
//
// Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
/** ROC Profiler Multi Queue Dependency Test
*
* The goal of this test is to ensure ROC profiler does not go to deadlock
* when multiple queue are created and they are dependent on each other
*
*/
#include "hsa_code_object_app.h"
enum class storage_type
{
CODE_OBJECT_STORAGE_FILE,
CODE_OBJECT_STORAGE_MEMORY
};
void
code_object_load(MQDependencyTest& obj,
storage_type type,
MQDependencyTest::CodeObject& code_object)
{
hsa_status_t status;
obj.device_discovery();
char agent_name[64];
status = hsa_agent_get_info(obj.gpu[0].agent, HSA_AGENT_INFO_NAME, agent_name);
RET_IF_HSA_ERR(status)
if(type == storage_type::CODE_OBJECT_STORAGE_FILE)
{
std::string hasco_file_path = std::string(agent_name) + std::string("_copy.hsaco");
obj.search_hasco(fs::current_path(), hasco_file_path);
if(!obj.load_code_object(hasco_file_path, obj.gpu[0].agent, code_object))
{
printf("Kernel file not found or not usable with given agent.\n");
abort();
}
}
else
{
std::string hasco_file_path = std::string(agent_name) + std::string("_copy_memory.hsaco");
obj.search_hasco(fs::current_path(), hasco_file_path);
if(!obj.load_code_object_memory(hasco_file_path, obj.gpu[0].agent, code_object))
{
abort();
}
}
}
MQDependencyTest::Kernel
get_kernel(MQDependencyTest::CodeObject& code_object,
std::string kernel_name,
MQDependencyTest& obj)
{
MQDependencyTest::Kernel copy;
if(!obj.get_kernel(code_object, kernel_name, obj.gpu[0].agent, copy))
{
printf("Test %s not found.\n", kernel_name.c_str());
abort();
}
return copy;
}
int
main()
{
hsa_status_t status;
MQDependencyTest obj;
MQDependencyTest obj_memory = {};
MQDependencyTest::CodeObject code_object = {}, code_object_memory = {};
code_object_load(obj, storage_type::CODE_OBJECT_STORAGE_FILE, code_object);
code_object_load(obj_memory, storage_type::CODE_OBJECT_STORAGE_MEMORY, code_object_memory);
MQDependencyTest::Kernel copyA = get_kernel(code_object, "copyA", obj);
MQDependencyTest::Kernel copyB = get_kernel(code_object, "copyB", obj);
MQDependencyTest::Kernel copyC = get_kernel(code_object, "copyC", obj);
MQDependencyTest::Kernel copyD = get_kernel(code_object_memory, "copyD", obj_memory);
MQDependencyTest::Kernel copyE = get_kernel(code_object_memory, "copyE", obj_memory);
MQDependencyTest::Kernel copyF = get_kernel(code_object_memory, "copyF", obj_memory);
struct args_t
{
uint32_t* a = nullptr;
uint32_t* b = nullptr;
MQDependencyTest::OCLHiddenArgs hidden = {};
};
args_t* args = static_cast<args_t*>(obj.hsa_malloc(sizeof(args_t), obj.kernarg));
*args = {};
uint32_t* a = static_cast<uint32_t*>(obj.hsa_malloc(64 * sizeof(uint32_t), obj.kernarg));
uint32_t* b = static_cast<uint32_t*>(obj.hsa_malloc(64 * sizeof(uint32_t), obj.kernarg));
memset(a, 0, 64 * sizeof(uint32_t));
memset(b, 1, 64 * sizeof(uint32_t));
args_t* args_memory =
static_cast<args_t*>(obj_memory.hsa_malloc(sizeof(args_t), obj_memory.kernarg));
*args_memory = {};
uint32_t* c =
static_cast<uint32_t*>(obj_memory.hsa_malloc(64 * sizeof(uint32_t), obj_memory.kernarg));
uint32_t* d =
static_cast<uint32_t*>(obj_memory.hsa_malloc(64 * sizeof(uint32_t), obj_memory.kernarg));
memset(c, 0, 64 * sizeof(uint32_t));
memset(d, 1, 64 * sizeof(uint32_t));
// Create queue in gpu agent and prepare a kernel dispatch packet
hsa_queue_t* queue1 = nullptr;
status = hsa_queue_create(obj.gpu[0].agent,
1024,
HSA_QUEUE_TYPE_SINGLE,
nullptr,
nullptr,
UINT32_MAX,
UINT32_MAX,
&queue1);
RET_IF_HSA_ERR(status)
// Create a signal with a value of 1 and attach it to the first kernel
// dispatch packet
hsa_signal_t completion_signal_1 = {};
status = hsa_signal_create(1, 0, nullptr, &completion_signal_1);
RET_IF_HSA_ERR(status)
// First dispath packet on queue 1, Kernel A
{
MQDependencyTest::Aql packet{};
packet.header.type = HSA_PACKET_TYPE_KERNEL_DISPATCH;
packet.header.barrier = 1;
packet.header.acquire = HSA_FENCE_SCOPE_SYSTEM;
packet.header.release = HSA_FENCE_SCOPE_SYSTEM;
packet.dispatch.setup = 1;
packet.dispatch.workgroup_size_x = 64;
packet.dispatch.workgroup_size_y = 1;
packet.dispatch.workgroup_size_z = 1;
packet.dispatch.grid_size_x = 64;
packet.dispatch.grid_size_y = 1;
packet.dispatch.grid_size_z = 1;
packet.dispatch.group_segment_size = copyA.group;
packet.dispatch.private_segment_size = copyA.scratch;
packet.dispatch.kernel_object = copyA.handle;
packet.dispatch.kernarg_address = args;
packet.dispatch.completion_signal = completion_signal_1;
args->a = a;
args->b = b;
// Tell packet processor of A to launch the first kernel dispatch packet
obj.submit_packet(queue1, packet);
}
// Create a signal with a value of 1 and attach it to the second kernel
// dispatch packet
hsa_signal_t completion_signal_2 = {};
status = hsa_signal_create(1, 0, nullptr, &completion_signal_2);
RET_IF_HSA_ERR(status)
hsa_signal_t completion_signal_3 = {};
status = hsa_signal_create(1, 0, nullptr, &completion_signal_3);
RET_IF_HSA_ERR(status)
// Create barrier-AND packet that is enqueued in queue 1
{
MQDependencyTest::Aql packet{};
packet.header.type = HSA_PACKET_TYPE_BARRIER_AND;
packet.header.barrier = 1;
packet.header.acquire = HSA_FENCE_SCOPE_SYSTEM;
packet.header.release = HSA_FENCE_SCOPE_SYSTEM;
packet.barrier_and.dep_signal[0] = completion_signal_2;
obj.submit_packet(queue1, packet);
}
// Second dispath packet on queue 1, Kernel C
{
MQDependencyTest::Aql packet{};
packet.header.type = HSA_PACKET_TYPE_KERNEL_DISPATCH;
packet.header.barrier = 1;
packet.header.acquire = HSA_FENCE_SCOPE_SYSTEM;
packet.header.release = HSA_FENCE_SCOPE_SYSTEM;
packet.dispatch.setup = 1;
packet.dispatch.workgroup_size_x = 64;
packet.dispatch.workgroup_size_y = 1;
packet.dispatch.workgroup_size_z = 1;
packet.dispatch.grid_size_x = 64;
packet.dispatch.grid_size_y = 1;
packet.dispatch.grid_size_z = 1;
packet.dispatch.group_segment_size = copyC.group;
packet.dispatch.private_segment_size = copyC.scratch;
packet.dispatch.kernel_object = copyC.handle;
packet.dispatch.completion_signal = completion_signal_3;
packet.dispatch.kernarg_address = args;
args->a = a;
args->b = b;
// Tell packet processor to launch the second kernel dispatch packet
obj.submit_packet(queue1, packet);
}
// Create queue 2
hsa_queue_t* queue2 = nullptr;
status = hsa_queue_create(obj.gpu[0].agent,
1024,
HSA_QUEUE_TYPE_SINGLE,
nullptr,
nullptr,
UINT32_MAX,
UINT32_MAX,
&queue2);
RET_IF_HSA_ERR(status)
// Create barrier-AND packet that is enqueued in queue 2
{
MQDependencyTest::Aql packet{};
packet.header.type = HSA_PACKET_TYPE_BARRIER_AND;
packet.header.barrier = 1;
packet.header.acquire = HSA_FENCE_SCOPE_SYSTEM;
packet.header.release = HSA_FENCE_SCOPE_SYSTEM;
packet.barrier_and.dep_signal[0] = completion_signal_1;
obj.submit_packet(queue2, packet);
}
// Third dispath packet on queue 2, Kernel B
{
MQDependencyTest::Aql packet{};
packet.header.type = HSA_PACKET_TYPE_KERNEL_DISPATCH;
packet.header.barrier = 1;
packet.header.acquire = HSA_FENCE_SCOPE_SYSTEM;
packet.header.release = HSA_FENCE_SCOPE_SYSTEM;
packet.dispatch.setup = 1;
packet.dispatch.workgroup_size_x = 64;
packet.dispatch.workgroup_size_y = 1;
packet.dispatch.workgroup_size_z = 1;
packet.dispatch.grid_size_x = 64;
packet.dispatch.grid_size_y = 1;
packet.dispatch.grid_size_z = 1;
packet.dispatch.group_segment_size = copyB.group;
packet.dispatch.private_segment_size = copyB.scratch;
packet.dispatch.kernel_object = copyB.handle;
packet.dispatch.kernarg_address = args;
packet.dispatch.completion_signal = completion_signal_2;
args->a = a;
args->b = b;
// Tell packet processor to launch the third kernel dispatch packet
obj.submit_packet(queue2, packet);
}
// Create a signal with a value of 1 and attach it to the first kernel
// dispatch packet
hsa_signal_t completion_signal_4 = {};
status = hsa_signal_create(1, 0, nullptr, &completion_signal_4);
RET_IF_HSA_ERR(status)
// First dispath packet on queue 1, Kernel D
{
[[maybe_unused]] MQDependencyTest::Aql packet{};
packet.header.type = HSA_PACKET_TYPE_KERNEL_DISPATCH;
packet.header.barrier = 1;
packet.header.acquire = HSA_FENCE_SCOPE_SYSTEM;
packet.header.release = HSA_FENCE_SCOPE_SYSTEM;
packet.dispatch.setup = 1;
packet.dispatch.workgroup_size_x = 64;
packet.dispatch.workgroup_size_y = 1;
packet.dispatch.workgroup_size_z = 1;
packet.dispatch.grid_size_x = 64;
packet.dispatch.grid_size_y = 1;
packet.dispatch.grid_size_z = 1;
packet.dispatch.group_segment_size = copyD.group;
packet.dispatch.private_segment_size = copyD.scratch;
packet.dispatch.kernel_object = copyD.handle;
packet.dispatch.kernarg_address = args_memory;
packet.dispatch.completion_signal = completion_signal_4;
args_memory->a = c;
args_memory->b = d;
// Tell packet processor of A to launch the first kernel dispatch packet
obj_memory.submit_packet(queue1, packet);
}
// Create a signal with a value of 1 and attach it to the second kernel
// dispatch packet
hsa_signal_t completion_signal_5 = {};
status = hsa_signal_create(1, 0, nullptr, &completion_signal_5);
RET_IF_HSA_ERR(status)
hsa_signal_t completion_signal_6 = {};
status = hsa_signal_create(1, 0, nullptr, &completion_signal_6);
RET_IF_HSA_ERR(status)
// Create barrier-AND packet that is enqueued in queue 1
{
MQDependencyTest::Aql packet{};
packet.header.type = HSA_PACKET_TYPE_BARRIER_AND;
packet.header.barrier = 1;
packet.header.acquire = HSA_FENCE_SCOPE_SYSTEM;
packet.header.release = HSA_FENCE_SCOPE_SYSTEM;
packet.barrier_and.dep_signal[0] = completion_signal_5;
obj_memory.submit_packet(queue1, packet);
}
// Second dispath packet on queue 1, Kernel F
{
MQDependencyTest::Aql packet{};
packet.header.type = HSA_PACKET_TYPE_KERNEL_DISPATCH;
packet.header.barrier = 1;
packet.header.acquire = HSA_FENCE_SCOPE_SYSTEM;
packet.header.release = HSA_FENCE_SCOPE_SYSTEM;
packet.dispatch.setup = 1;
packet.dispatch.workgroup_size_x = 64;
packet.dispatch.workgroup_size_y = 1;
packet.dispatch.workgroup_size_z = 1;
packet.dispatch.grid_size_x = 64;
packet.dispatch.grid_size_y = 1;
packet.dispatch.grid_size_z = 1;
packet.dispatch.group_segment_size = copyF.group;
packet.dispatch.private_segment_size = copyF.scratch;
packet.dispatch.kernel_object = copyF.handle;
packet.dispatch.completion_signal = completion_signal_6;
packet.dispatch.kernarg_address = args_memory;
args_memory->a = c;
args_memory->b = d;
// Tell packet processor to launch the second kernel dispatch packet
obj_memory.submit_packet(queue1, packet);
}
// Create barrier-AND packet that is enqueued in queue 2
{
MQDependencyTest::Aql packet{};
packet.header.type = HSA_PACKET_TYPE_BARRIER_AND;
packet.header.barrier = 1;
packet.header.acquire = HSA_FENCE_SCOPE_SYSTEM;
packet.header.release = HSA_FENCE_SCOPE_SYSTEM;
packet.barrier_and.dep_signal[0] = completion_signal_4;
obj_memory.submit_packet(queue2, packet);
}
// Third dispath packet on queue 2, Kernel
{
MQDependencyTest::Aql packet{};
packet.header.type = HSA_PACKET_TYPE_KERNEL_DISPATCH;
packet.header.barrier = 1;
packet.header.acquire = HSA_FENCE_SCOPE_SYSTEM;
packet.header.release = HSA_FENCE_SCOPE_SYSTEM;
packet.dispatch.setup = 1;
packet.dispatch.workgroup_size_x = 64;
packet.dispatch.workgroup_size_y = 1;
packet.dispatch.workgroup_size_z = 1;
packet.dispatch.grid_size_x = 64;
packet.dispatch.grid_size_y = 1;
packet.dispatch.grid_size_z = 1;
packet.dispatch.group_segment_size = copyE.group;
packet.dispatch.private_segment_size = copyE.scratch;
packet.dispatch.kernel_object = copyE.handle;
packet.dispatch.kernarg_address = args_memory;
packet.dispatch.completion_signal = completion_signal_5;
args_memory->a = c;
args_memory->b = d;
// Tell packet processor to launch the third kernel dispatch packet
obj_memory.submit_packet(queue2, packet);
}
// Wait on the completion signal
hsa_signal_wait_relaxed(
completion_signal_1, HSA_SIGNAL_CONDITION_EQ, 0, UINT64_MAX, HSA_WAIT_STATE_BLOCKED);
// Wait on the completion signal
hsa_signal_wait_relaxed(
completion_signal_2, HSA_SIGNAL_CONDITION_EQ, 0, UINT64_MAX, HSA_WAIT_STATE_BLOCKED);
// Wait on the completion signal
hsa_signal_wait_relaxed(
completion_signal_3, HSA_SIGNAL_CONDITION_EQ, 0, UINT64_MAX, HSA_WAIT_STATE_BLOCKED);
// Wait on the completion signal
hsa_signal_wait_relaxed(
completion_signal_4, HSA_SIGNAL_CONDITION_EQ, 0, UINT64_MAX, HSA_WAIT_STATE_BLOCKED);
// Wait on the completion signal
hsa_signal_wait_relaxed(
completion_signal_5, HSA_SIGNAL_CONDITION_EQ, 0, UINT64_MAX, HSA_WAIT_STATE_BLOCKED);
// Wait on the completion signal
hsa_signal_wait_relaxed(
completion_signal_6, HSA_SIGNAL_CONDITION_EQ, 0, UINT64_MAX, HSA_WAIT_STATE_BLOCKED);
for(int i = 0; i < 64; i++)
{
if(a[i] != b[i])
{
printf("error at %d: expected %d, got %d\n", i, b[i], a[i]);
abort();
}
}
// Clearing data structures and memory
status = hsa_signal_destroy(completion_signal_1);
RET_IF_HSA_ERR(status)
status = hsa_signal_destroy(completion_signal_2);
RET_IF_HSA_ERR(status)
status = hsa_signal_destroy(completion_signal_3);
RET_IF_HSA_ERR(status)
// Clearing data structures and memory
status = hsa_signal_destroy(completion_signal_4);
RET_IF_HSA_ERR(status)
status = hsa_signal_destroy(completion_signal_5);
RET_IF_HSA_ERR(status)
status = hsa_signal_destroy(completion_signal_6);
RET_IF_HSA_ERR(status)
if(queue1 != nullptr)
{
status = hsa_queue_destroy(queue1);
RET_IF_HSA_ERR(status)
}
if(queue2 != nullptr)
{
status = hsa_queue_destroy(queue2);
RET_IF_HSA_ERR(status)
}
status = hsa_memory_free(a);
RET_IF_HSA_ERR(status)
status = hsa_memory_free(b);
RET_IF_HSA_ERR(status)
status = hsa_memory_free(c);
RET_IF_HSA_ERR(status)
status = hsa_memory_free(d);
RET_IF_HSA_ERR(status)
status = hsa_executable_destroy(code_object.executable);
RET_IF_HSA_ERR(status)
status = hsa_code_object_reader_destroy(code_object.code_obj_rdr);
RET_IF_HSA_ERR(status)
status = hsa_executable_destroy(code_object_memory.executable);
RET_IF_HSA_ERR(status)
status = hsa_code_object_reader_destroy(code_object_memory.code_obj_rdr);
RET_IF_HSA_ERR(status)
close(code_object.file);
close(code_object_memory.file);
}