AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[SWDEV-482060] Set execute permission for HSA allocated memory (#7)

Browse Source 
Set execute permission for HSA allocated memory

We need execute permission for HSA memory (req for IB buffers).
Enforcement is upcoming which will break counter collection (see ticket).

Change-Id: Id939bbdd8686915407bc60d0dfcd5110b0e91e90

Co-authored-by: Benjamin Welton <bewelton@amd.com>
Co-authored-by: Bhardwaj, Gopesh <Gopesh.Bhardwaj@amd.com>
This commit is contained in:
Welton, Benjamin
2025-02-06 05:54:06 -08:00
committed by GitHub
parent 20415eff0a
commit 6cc5501260
6 changed files with 19 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/core/hsa/packets/packets_generator.cpp
+6 -6
View File
@@ -335,7 +335,7 @@ std::unique_ptr<AQLPacketProfile> InitializeAqlPackets(
status = HSA_STATUS_ERROR;
size_t size = (profile->command_buffer.size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
status = hsasupport_singleton.GetAmdExtTable().hsa_amd_memory_pool_allocate_fn(
agentInfo.cpu_pool_, size, 0, reinterpret_cast<void**>(&(profile->command_buffer.ptr)));
agentInfo.cpu_pool_, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, reinterpret_cast<void**>(&(profile->command_buffer.ptr)));
if (status != HSA_STATUS_SUCCESS) {
profile->command_buffer.ptr = malloc(size);
/*numa_alloc_onnode(
@@ -356,7 +356,7 @@ std::unique_ptr<AQLPacketProfile> InitializeAqlPackets(
status = HSA_STATUS_ERROR;
size_t size = (profile->output_buffer.size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
status = hsasupport_singleton.GetAmdExtTable().hsa_amd_memory_pool_allocate_fn(
agentInfo.kernarg_pool_, size, 0, &profile->output_buffer.ptr
agentInfo.kernarg_pool_, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, &profile->output_buffer.ptr
);
if (status != HSA_STATUS_SUCCESS) {
profile->output_buffer.ptr = malloc(size);
@@ -425,7 +425,7 @@ hsa_ven_amd_aqlprofile_profile_t* InitializeDeviceProfilingAqlPackets(
if (size <= 0) return nullptr;
size = (size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
status = hsasupport_singleton.GetAmdExtTable().hsa_amd_memory_pool_allocate_fn(
agentInfo.cpu_pool_, size, 0, reinterpret_cast<void**>(&(profile->command_buffer.ptr)));
agentInfo.cpu_pool_, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, reinterpret_cast<void**>(&(profile->command_buffer.ptr)));
// Both the CPU and GPU can access the memory
if (status == HSA_STATUS_SUCCESS) {
status = hsasupport_singleton.GetAmdExtTable().hsa_amd_agents_allow_access_fn(
@@ -450,7 +450,7 @@ hsa_ven_amd_aqlprofile_profile_t* InitializeDeviceProfilingAqlPackets(
profile->output_buffer.ptr = nullptr;
size = (size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
status = hsasupport_singleton.GetAmdExtTable().hsa_amd_memory_pool_allocate_fn(
agentInfo.gpu_pool_, size, 0, reinterpret_cast<void**>(&(profile->output_buffer.ptr)));
agentInfo.gpu_pool_, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, reinterpret_cast<void**>(&(profile->output_buffer.ptr)));
CHECK_HSA_STATUS("Error: Can't Allocate Output Buffer", status);
// Both the CPU and GPU can access the kernel arguments
if (status == HSA_STATUS_SUCCESS) {
@@ -484,7 +484,7 @@ uint8_t* AllocateSysMemory(hsa_agent_t gpu_agent, size_t size, hsa_amd_memory_po
rocprofiler::HSASupport_Singleton& hsasupport_singleton =
rocprofiler::HSASupport_Singleton::GetInstance();
status = hsasupport_singleton.GetAmdExtTable().hsa_amd_memory_pool_allocate_fn(
*cpu_pool, size, 0, reinterpret_cast<void**>(&buffer));
*cpu_pool, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, reinterpret_cast<void**>(&buffer));
// Both the CPU and GPU can access the memory
if (status == HSA_STATUS_SUCCESS) {
status = hsasupport_singleton.GetAmdExtTable().hsa_amd_agents_allow_access_fn(
@@ -502,7 +502,7 @@ uint8_t* AllocateLocalMemory(size_t size, hsa_amd_memory_pool_t* gpu_pool) {
uint8_t* buffer = NULL;
size = (size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
status = hsasupport_singleton.GetAmdExtTable().hsa_amd_memory_pool_allocate_fn(
*gpu_pool, size, 0, reinterpret_cast<void**>(&buffer));
*gpu_pool, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, reinterpret_cast<void**>(&buffer));
uint8_t* ptr = (status == HSA_STATUS_SUCCESS) ? buffer : NULL;
return ptr;
}
src/util/hsa_rsrc_factory.cpp
+3 -3
View File
@@ -526,7 +526,7 @@ uint8_t* HsaRsrcFactory::AllocateLocalMemory(const AgentInfo* agent_info, size_t
hsa_status_t status = HSA_STATUS_ERROR;
uint8_t* buffer = nullptr;
size = (size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
status = hsa_api_.hsa_amd_memory_pool_allocate(agent_info->gpu_pool, size, 0,
status = hsa_api_.hsa_amd_memory_pool_allocate(agent_info->gpu_pool, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG,
reinterpret_cast<void**>(&buffer));
uint8_t* ptr = (status == HSA_STATUS_SUCCESS) ? buffer : nullptr;
return ptr;
@@ -542,7 +542,7 @@ uint8_t* HsaRsrcFactory::AllocateKernArgMemory(const AgentInfo* agent_info, size
uint8_t* buffer = nullptr;
if (!cpu_agents_.empty()) {
size = (size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
status = hsa_api_.hsa_amd_memory_pool_allocate(*kern_arg_pool_, size, 0,
status = hsa_api_.hsa_amd_memory_pool_allocate(*kern_arg_pool_, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG,
reinterpret_cast<void**>(&buffer));
// Both the CPU and GPU can access the kernel arguments
if (status == HSA_STATUS_SUCCESS) {
@@ -563,7 +563,7 @@ uint8_t* HsaRsrcFactory::AllocateSysMemory(const AgentInfo* agent_info, size_t s
uint8_t* buffer = nullptr;
size = (size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
if (!cpu_agents_.empty()) {
status = hsa_api_.hsa_amd_memory_pool_allocate(*cpu_pool_, size, 0,
status = hsa_api_.hsa_amd_memory_pool_allocate(*cpu_pool_, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG,
reinterpret_cast<void**>(&buffer));
// Both the CPU and GPU can access the memory
if (status == HSA_STATUS_SUCCESS) {
test/util/hsa_rsrc_factory.cpp
+3 -3
View File
@@ -493,7 +493,7 @@ uint8_t* HsaRsrcFactory::AllocateLocalMemory(const AgentInfo* agent_info, size_t
hsa_status_t status = HSA_STATUS_ERROR;
uint8_t* buffer = NULL;
size = (size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
status = hsa_api_.hsa_amd_memory_pool_allocate(agent_info->gpu_pool, size, 0,
status = hsa_api_.hsa_amd_memory_pool_allocate(agent_info->gpu_pool, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG,
reinterpret_cast<void**>(&buffer));
uint8_t* ptr = (status == HSA_STATUS_SUCCESS) ? buffer : NULL;
return ptr;
@@ -509,7 +509,7 @@ uint8_t* HsaRsrcFactory::AllocateKernArgMemory(const AgentInfo* agent_info, size
uint8_t* buffer = NULL;
if (!cpu_agents_.empty()) {
size = (size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
status = hsa_api_.hsa_amd_memory_pool_allocate(*kern_arg_pool_, size, 0,
status = hsa_api_.hsa_amd_memory_pool_allocate(*kern_arg_pool_, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG,
reinterpret_cast<void**>(&buffer));
// Both the CPU and GPU can access the kernel arguments
if (status == HSA_STATUS_SUCCESS) {
@@ -530,7 +530,7 @@ uint8_t* HsaRsrcFactory::AllocateSysMemory(const AgentInfo* agent_info, size_t s
uint8_t* buffer = NULL;
size = (size + MEM_PAGE_MASK) & ~MEM_PAGE_MASK;
if (!cpu_agents_.empty()) {
status = hsa_api_.hsa_amd_memory_pool_allocate(*cpu_pool_, size, 0,
status = hsa_api_.hsa_amd_memory_pool_allocate(*cpu_pool_, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG,
reinterpret_cast<void**>(&buffer));
// Both the CPU and GPU can access the memory
if (status == HSA_STATUS_SUCCESS) {
tests-v2/featuretests/profiler/apps/async_mem_copy.cpp
+3 -3
View File
@@ -328,14 +328,14 @@ int main() {
// Allocate memory on each source/destination
if (twoGPUs) {
sz = lcm(sz, args.gpu2.granule);
err = hsa_amd_memory_pool_allocate(args.gpu2.pool, sz, 0,
err = hsa_amd_memory_pool_allocate(args.gpu2.pool, sz, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG,
reinterpret_cast<void**>(&args.gpu2.ptr));
RET_IF_HSA_ERR(err);
}
err = hsa_amd_memory_pool_allocate(args.cpu.pool, sz, 0, reinterpret_cast<void**>(&args.cpu.ptr));
err = hsa_amd_memory_pool_allocate(args.cpu.pool, sz, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, reinterpret_cast<void**>(&args.cpu.ptr));
RET_IF_HSA_ERR(err);
err =
hsa_amd_memory_pool_allocate(args.gpu1.pool, sz, 0, reinterpret_cast<void**>(&args.gpu1.ptr));
hsa_amd_memory_pool_allocate(args.gpu1.pool, sz, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, reinterpret_cast<void**>(&args.gpu1.ptr));
RET_IF_HSA_ERR(err);
char name[64];
err = hsa_agent_get_info(args.cpu.dev, HSA_AGENT_INFO_NAME, &name);
tests-v2/featuretests/profiler/apps/multiqueue_testapp.h
+1 -1
View File
@@ -221,7 +221,7 @@ class MQDependencyTest {
void* hsaMalloc(size_t size, const Device::Memory& mem) {
void* ret;
hsa_status_t err = hsa_amd_memory_pool_allocate(mem.pool, size, 0, &ret);
hsa_status_t err = hsa_amd_memory_pool_allocate(mem.pool, size, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, &ret);
ASSERT_EQ(err, HSA_STATUS_SUCCESS);
err = hsa_amd_agents_allow_access(Device::all_devices.size(), &Device::all_devices[0], nullptr,
tests-v2/featuretests/tracer/apps/copy_on_engine.cpp
+3 -3
View File
@@ -275,10 +275,10 @@ int main() {
// potential sources and destinations of the copy
size_t sz = sizeof(uint32_t);
// Allocate memory on each source/destination
err = hsa_amd_memory_pool_allocate(args.cpu.pool, sz, 0, reinterpret_cast<void**>(&args.cpu.ptr));
err = hsa_amd_memory_pool_allocate(args.cpu.pool, sz, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, reinterpret_cast<void**>(&args.cpu.ptr));
RET_IF_HSA_ERR(err);
err =
hsa_amd_memory_pool_allocate(args.gpu1.pool, sz, 0, reinterpret_cast<void**>(&args.gpu1.ptr));
hsa_amd_memory_pool_allocate(args.gpu1.pool, sz, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG, reinterpret_cast<void**>(&args.gpu1.ptr));
RET_IF_HSA_ERR(err);
char name[64];
err = hsa_agent_get_info(args.cpu.dev, HSA_AGENT_INFO_NAME, &name);
@@ -299,7 +299,7 @@ int main() {
fprintf(stdout, "Success!\n");
if (twoGPUs) {
err = hsa_amd_memory_pool_allocate(args.gpu2.pool, sz, 0,
err = hsa_amd_memory_pool_allocate(args.gpu2.pool, sz, HSA_AMD_MEMORY_POOL_EXECUTABLE_FLAG,
reinterpret_cast<void**>(&args.gpu2.ptr));
RET_IF_HSA_ERR(err);
fprintf(stdout, "Copying %lu bytes from gpu1 memory to gpu2 memory...\n", sz);