coding style formatting

[ROCm/rocprofiler commit: cb3a591826]
This commit is contained in:
Evgeny
2017-11-29 13:53:12 -06:00
parent 15471db275
commit 8b5a6b7fed
31 changed files with 734 additions and 702 deletions
@@ -10,15 +10,15 @@
#include "util/test_assert.h"
int main(int argc, char** argv) {
bool ret_val = false;
// HSA status
hsa_status_t status = HSA_STATUS_ERROR;
// Profiling context
rocprofiler_t* context = NULL;
// Profiling properties
rocprofiler_properties_t properties;
// Number of context invocation
uint32_t invocation = 0;
bool ret_val = false;
// HSA status
hsa_status_t status = HSA_STATUS_ERROR;
// Profiling context
rocprofiler_t* context = NULL;
// Profiling properties
rocprofiler_properties_t properties;
// Number of context invocation
uint32_t invocation = 0;
#if 0
// Profiling info objects
@@ -29,48 +29,49 @@ int main(int argc, char** argv) {
info[0].type = ROCPROFILER_TYPE_METRIC;
info[0].name = "SQ_WAVES";
#else
// Profiling info objects
const unsigned info_count = 3;
rocprofiler_info_t info[info_count];
// PMC events
memset(info, 0, sizeof(info));
info[0].type = ROCPROFILER_TYPE_METRIC;
info[0].name = "SQ_WAVES";
info[1].type = ROCPROFILER_TYPE_METRIC;
info[1].name = "SQ_ITEMS";
// Tracing parameters
const unsigned parameter_count = 2;
rocprofiler_parameter_t parameters[parameter_count];
info[2].name = "THREAD_TRACE";
info[2].type = ROCPROFILER_TYPE_TRACE;
info[2].parameters = parameters;
info[2].parameter_count = parameter_count;
parameters[0].parameter_name = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_MASK;
parameters[0].value = 0;
parameters[1].parameter_name = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK;
parameters[1].value = 0;
// Profiling info objects
const unsigned info_count = 3;
rocprofiler_info_t info[info_count];
// PMC events
memset(info, 0, sizeof(info));
info[0].type = ROCPROFILER_TYPE_METRIC;
info[0].name = "SQ_WAVES";
info[1].type = ROCPROFILER_TYPE_METRIC;
info[1].name = "SQ_ITEMS";
// Tracing parameters
const unsigned parameter_count = 2;
rocprofiler_parameter_t parameters[parameter_count];
info[2].name = "THREAD_TRACE";
info[2].type = ROCPROFILER_TYPE_TRACE;
info[2].parameters = parameters;
info[2].parameter_count = parameter_count;
parameters[0].parameter_name = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_MASK;
parameters[0].value = 0;
parameters[1].parameter_name = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK;
parameters[1].value = 0;
#endif
// Creating profiling context
properties = {};
properties.queue_depth = 128;
status = rocprofiler_open(TestHsa::HsaAgentId(), info, info_count, &context, ROCPROFILER_MODE_STANDALONE|ROCPROFILER_MODE_OWNQUEUE, &properties);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
// Creating profiling context
properties = {};
properties.queue_depth = 128;
status = rocprofiler_open(TestHsa::HsaAgentId(), info, info_count, &context,
ROCPROFILER_MODE_STANDALONE | ROCPROFILER_MODE_OWNQUEUE, &properties);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
TestHsa::SetQueue(properties.queue);
TestHsa::SetQueue(properties.queue);
// Adding dispatch observer
status = rocprofiler_dispatch_observer(rocprofiler_dispatch_callback, context);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
// Adding dispatch observer
status = rocprofiler_dispatch_observer(rocprofiler_dispatch_callback, context);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
// Querying the number of context invocation
status = rocprofiler_invocation(context, &invocation);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
// Querying the number of context invocation
status = rocprofiler_invocation(context, &invocation);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
// Dispatching profiled kernel n-times to collect all counter groups data
unsigned n = 0;
while(1) {
std::cout << "> " << n << "/" << invocation << std::endl;
// Dispatching profiled kernel n-times to collect all counter groups data
unsigned n = 0;
while (1) {
std::cout << "> " << n << "/" << invocation << std::endl;
#if 0
status = rocprofiler_start(context);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
@@ -78,49 +79,49 @@ int main(int argc, char** argv) {
status = rocprofiler_stop(context);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
#else
ret_val = RunKernel<SimpleConvolution, TestAql>(argc, argv);
ret_val = RunKernel<SimpleConvolution, TestAql>(argc, argv);
#endif
status = rocprofiler_sample(context);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
for (rocprofiler_info_t* p = info; p < info + info_count; ++p) {
std::cout << (p - info) << ": " << p->name;
switch (p->data.kind) {
case ROCPROFILER_DATA_KIND_INT64:
std::cout << std::dec << " result64 (" << p->data.result64 << ")" << std::endl;
break;
case ROCPROFILER_BYTES: {
const char* ptr = reinterpret_cast<const char*>(p->data.result_bytes.ptr);
uint64_t size = 0;
for (unsigned i = 0; i < p->data.result_bytes.instance_count; ++i) {
size = *reinterpret_cast<const uint64_t*>(ptr);
const char* data = ptr + sizeof(size);
std::cout << std::endl;
std::cout << std::hex << " data (" << (void*)data << ")" << std::endl;
std::cout << std::dec << " size (" << size << ")" << std::endl;
ptr = data + size;
}
break;
}
default:
std::cout << "result kind (" << p->data.kind << ")" << std::endl;
TEST_ASSERT(false);
}
}
++n;
if (n < invocation) {
status = rocprofiler_next(context);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
continue;
}
break;
}
// Finishing cleanup
// Deleting profiling context will delete all allocated resources
status = rocprofiler_close(context);
status = rocprofiler_sample(context);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
return (ret_val) ? 0 : 1;
for (rocprofiler_info_t* p = info; p < info + info_count; ++p) {
std::cout << (p - info) << ": " << p->name;
switch (p->data.kind) {
case ROCPROFILER_DATA_KIND_INT64:
std::cout << std::dec << " result64 (" << p->data.result64 << ")" << std::endl;
break;
case ROCPROFILER_BYTES: {
const char* ptr = reinterpret_cast<const char*>(p->data.result_bytes.ptr);
uint64_t size = 0;
for (unsigned i = 0; i < p->data.result_bytes.instance_count; ++i) {
size = *reinterpret_cast<const uint64_t*>(ptr);
const char* data = ptr + sizeof(size);
std::cout << std::endl;
std::cout << std::hex << " data (" << (void*)data << ")" << std::endl;
std::cout << std::dec << " size (" << size << ")" << std::endl;
ptr = data + size;
}
break;
}
default:
std::cout << "result kind (" << p->data.kind << ")" << std::endl;
TEST_ASSERT(false);
}
}
++n;
if (n < invocation) {
status = rocprofiler_next(context);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
continue;
}
break;
}
// Finishing cleanup
// Deleting profiling context will delete all allocated resources
status = rocprofiler_close(context);
TEST_STATUS(status == HSA_STATUS_SUCCESS);
return (ret_val) ? 0 : 1;
}
+3 -1
View File
@@ -37,7 +37,9 @@ OF THE POSSIBILITY OF SUCH DAMAGE.
class TestAql {
public:
explicit TestAql(TestAql* t = 0) : test_(t) {}
virtual ~TestAql() { if (test_) delete test_; }
virtual ~TestAql() {
if (test_) delete test_;
}
TestAql* Test() { return test_; }
virtual AgentInfo* GetAgentInfo() { return (test_) ? test_->GetAgentInfo() : 0; }
+4 -2
View File
@@ -59,7 +59,7 @@ HsaRsrcFactory* TestHsa::HsaInstantiate(const uint32_t agent_ind) {
// Create an instance of Aql Queue
if (hsa_queue_ == NULL) {
uint32_t num_pkts = 128;
if(hsa_rsrc_->CreateQueue(agent_info_, num_pkts, &hsa_queue_) == false) {
if (hsa_rsrc_->CreateQueue(agent_info_, num_pkts, &hsa_queue_) == false) {
hsa_queue_ = NULL;
}
}
@@ -67,7 +67,9 @@ HsaRsrcFactory* TestHsa::HsaInstantiate(const uint32_t agent_ind) {
return hsa_rsrc_;
}
void TestHsa::HsaShutdown() { if (hsa_rsrc_) hsa_rsrc_->Destroy(); }
void TestHsa::HsaShutdown() {
if (hsa_rsrc_) hsa_rsrc_->Destroy();
}
bool TestHsa::Initialize(int arg_cnt, char** arg_list) {
std::clog << "TestHsa::Initialize :" << std::endl;
@@ -35,8 +35,8 @@ OF THE POSSIBILITY OF SUCH DAMAGE.
#include "util/test_assert.h"
hsa_status_t TestPGenRocpCallback(hsa_ven_amd_aqlprofile_info_type_t info_type,
hsa_ven_amd_aqlprofile_info_data_t* info_data,
void* callback_data) {
hsa_ven_amd_aqlprofile_info_data_t* info_data,
void* callback_data) {
hsa_status_t status = HSA_STATUS_SUCCESS;
typedef std::vector<hsa_ven_amd_aqlprofile_info_data_t> passed_data_t;
reinterpret_cast<passed_data_t*>(callback_data)->push_back(*info_data);
+1 -2
View File
@@ -48,8 +48,7 @@ bool TestPMgr::AddPacketGfx9(const packet_t* packet) {
*slot = aql_packet;
// After AQL packet is fully copied into queue buffer
// update packet header from invalid state to valid state
auto header_atomic_ptr =
reinterpret_cast<std::atomic<uint16_t>*>(&slot->header);
auto header_atomic_ptr = reinterpret_cast<std::atomic<uint16_t>*>(&slot->header);
header_atomic_ptr->store(header, std::memory_order_release);
// Increment the write index and ring the doorbell to dispatch the kernel.
+42 -26
View File
@@ -48,11 +48,14 @@ void check_status(hsa_status_t status) {
}
}
unsigned align_size(unsigned size, unsigned alignment) { return ((size + alignment - 1) & ~(alignment - 1)); }
unsigned align_size(unsigned size, unsigned alignment) {
return ((size + alignment - 1) & ~(alignment - 1));
}
void print_info(FILE* file, const rocprofiler_info_t* info, const unsigned info_count, const char* str) {
void print_info(FILE* file, const rocprofiler_info_t* info, const unsigned info_count,
const char* str) {
if (str) fprintf(file, "%s:\n", str);
for (unsigned i= 0; i < info_count; ++i) {
for (unsigned i = 0; i < info_count; ++i) {
const rocprofiler_info_t* p = &info[i];
fprintf(file, " %s ", p->name);
switch (p->data.kind) {
@@ -81,29 +84,30 @@ void print_info(FILE* file, const rocprofiler_info_t* info, const unsigned info_
void print_group(FILE* file, const rocprofiler_group_t* group, const char* str) {
if (str) fprintf(file, "%s:\n", str);
for (unsigned i= 0; i < group->info_count; ++i) {
for (unsigned i = 0; i < group->info_count; ++i) {
print_info(file, group->info[i], 1, NULL);
}
}
void store_context(context_entry_t context_entry) {
if(pthread_mutex_lock(&mutex) != 0) {
if (pthread_mutex_lock(&mutex) != 0) {
perror("pthread_mutex_lock");
exit(1);
}
if ((context_array == NULL) || (context_array_index >= context_array_size)) {
context_array_size *= 2;
context_array = reinterpret_cast<context_entry_t*>(realloc(context_array, context_array_size * sizeof(context_entry_t)));
context_array = reinterpret_cast<context_entry_t*>(
realloc(context_array, context_array_size * sizeof(context_entry_t)));
}
context_array_index += 1;
context_array[context_array_index - 1] = context_entry;
if(pthread_mutex_unlock(&mutex) != 0) {
if (pthread_mutex_unlock(&mutex) != 0) {
perror("pthread_mutex_unlock");
exit(1);
}
}
void dump_context(FILE *file, unsigned index) {
void dump_context(FILE* file, unsigned index) {
hsa_status_t status = HSA_STATUS_ERROR;
if (pthread_mutex_lock(&mutex) != 0) {
@@ -122,7 +126,7 @@ void dump_context(FILE *file, unsigned index) {
status = rocprofiler_get_group_data(group);
check_status(status);
//print_group(file, group, "Group[0] data");
// print_group(file, group, "Group[0] data");
status = rocprofiler_get_metrics(group->context);
check_status(status);
@@ -134,10 +138,8 @@ void dump_context(FILE *file, unsigned index) {
}
// Provided standard profiling callback
hsa_status_t dispatch_callback(
const rocprofiler_callback_data_t* callback_data,
void* user_data,
rocprofiler_group_t** group) {
hsa_status_t dispatch_callback(const rocprofiler_callback_data_t* callback_data, void* user_data,
rocprofiler_group_t** group) {
hsa_status_t status = HSA_STATUS_ERROR;
// Passed tool data
dispatch_data_t* tool_data = reinterpret_cast<dispatch_data_t*>(user_data);
@@ -173,17 +175,30 @@ void* dumping_data(void*) {
CONSTRUCTOR_API void constructor() {
std::map<std::string, hsa_ven_amd_aqlprofile_parameter_name_t> parameters_dict;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_COMPUTE_UNIT_TARGET"] = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_COMPUTE_UNIT_TARGET;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_VM_ID_MASK"] = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_VM_ID_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_MASK"] = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK"] = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK2"] = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK2;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_COMPUTE_UNIT_TARGET"] =
HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_COMPUTE_UNIT_TARGET;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_VM_ID_MASK"] =
HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_VM_ID_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_MASK"] =
HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK"] =
HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK2"] =
HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK2;
#ifdef TOOL_THREAD
int err = pthread_attr_init(&thr_attr);
if (err) { errno = err; perror("pthread_attr_init"); exit(1); }
if (err) {
errno = err;
perror("pthread_attr_init");
exit(1);
}
err = pthread_create(&thread, &thr_attr, dumping_data, NULL);
if (err) { errno = err; perror("pthread_create"); exit(1); }
if (err) {
errno = err;
perror("pthread_create");
exit(1);
}
#endif
// Set output file
@@ -194,7 +209,8 @@ CONSTRUCTOR_API void constructor() {
perror("fopen");
exit(1);
}
} else file_handle = stdout;
} else
file_handle = stdout;
// Getting input
const char* xml_name = getenv("ROCP_INPUT");
@@ -211,7 +227,7 @@ CONSTRUCTOR_API void constructor() {
for (auto* entry : metrics_list) {
const std::string entry_str = entry->opts["name"];
size_t pos1 = 0;
while(pos1 < entry_str.length()) {
while (pos1 < entry_str.length()) {
const size_t pos2 = entry_str.find(",", pos1);
const std::string metric_name = entry_str.substr(pos1, pos2 - pos1);
metrics_vec.push_back(metric_name);
@@ -224,10 +240,10 @@ CONSTRUCTOR_API void constructor() {
auto traces_list = xml->GetNodes("top.trace");
const unsigned info_count = metrics_vec.size() + traces_list.size();
rocprofiler_info_t* info= new rocprofiler_info_t[info_count];
rocprofiler_info_t* info = new rocprofiler_info_t[info_count];
memset(info, 0, info_count * sizeof(rocprofiler_info_t));
printf(" %d metrics\n", (int) metrics_vec.size());
printf(" %d metrics\n", (int)metrics_vec.size());
for (unsigned i = 0; i < metrics_vec.size(); ++i) {
const std::string& name = metrics_vec[i];
printf("%s%s", (i == 0) ? " " : ", ", name.c_str());
@@ -237,7 +253,7 @@ CONSTRUCTOR_API void constructor() {
}
if (metrics_vec.size()) printf("\n");
printf(" %d traces\n", (int) traces_list.size());
printf(" %d traces\n", (int)traces_list.size());
unsigned index = metrics_vec.size();
for (auto* entry : traces_list) {
auto params_list = xml->GetNodes("top.trace.parameters");
@@ -253,7 +269,7 @@ CONSTRUCTOR_API void constructor() {
for (auto* params : params_list) {
const unsigned parameter_count = params->opts.size();
rocprofiler_parameter_t *parameters = new rocprofiler_parameter_t[parameter_count];
rocprofiler_parameter_t* parameters = new rocprofiler_parameter_t[parameter_count];
unsigned p_index = 0;
for (auto& v : params->opts) {
const std::string parameter_name = v.first;
+42 -34
View File
@@ -65,14 +65,15 @@ void check_status(hsa_status_t status) {
context_entry_t* alloc_context_entry() {
context_entry_t* ptr = 0;
if(pthread_mutex_lock(&mutex) != 0) {
if (pthread_mutex_lock(&mutex) != 0) {
perror("pthread_mutex_lock");
exit(1);
}
if ((context_array == NULL) || (context_array_count >= context_array_size)) {
context_array_size *= 2;
context_array = reinterpret_cast<context_entry_t*>(realloc(context_array, context_array_size * sizeof(context_entry_t)));
context_array = reinterpret_cast<context_entry_t*>(
realloc(context_array, context_array_size * sizeof(context_entry_t)));
}
ptr = &context_array[context_array_count];
*ptr = {};
@@ -98,7 +99,7 @@ void dump_sqtt_trace(const uint32_t chunk, const void* data, const uint32_t& siz
perror("result file fopen");
exit(1);
}
// Write the buffer in terms of shorts (16 bits)
const unsigned short* ptr = reinterpret_cast<const unsigned short*>(data);
for (uint32_t i = 0; i < (size / sizeof(short)); ++i) {
@@ -108,28 +109,30 @@ void dump_sqtt_trace(const uint32_t chunk, const void* data, const uint32_t& siz
}
// Trace data callback for getting trace data from GPU local mamory
hsa_status_t trace_data_cb(
hsa_ven_amd_aqlprofile_info_type_t info_type,
hsa_ven_amd_aqlprofile_info_data_t* info_data,
void* data)
{
hsa_status_t trace_data_cb(hsa_ven_amd_aqlprofile_info_type_t info_type,
hsa_ven_amd_aqlprofile_info_data_t* info_data, void* data) {
FILE* file = reinterpret_cast<FILE*>(data);
hsa_status_t status = HSA_STATUS_SUCCESS;
if (info_type == HSA_VEN_AMD_AQLPROFILE_INFO_SQTT_DATA) {
fprintf(file, " data ptr (%p), size(%u)\n", info_data->sqtt_data.ptr, info_data->sqtt_data.size);
fprintf(file, " data ptr (%p), size(%u)\n", info_data->sqtt_data.ptr,
info_data->sqtt_data.size);
dump_sqtt_trace(info_data->sample_id, info_data->sqtt_data.ptr, info_data->sqtt_data.size);
} else status = HSA_STATUS_ERROR;
} else
status = HSA_STATUS_ERROR;
return status;
}
// Align to specified alignment
unsigned align_size(unsigned size, unsigned alignment) { return ((size + alignment - 1) & ~(alignment - 1)); }
unsigned align_size(unsigned size, unsigned alignment) {
return ((size + alignment - 1) & ~(alignment - 1));
}
// Output profiling results for input features
void output_results(FILE* file, const rocprofiler_feature_t* features, const unsigned feature_count, rocprofiler_t* context, const char* str) {
void output_results(FILE* file, const rocprofiler_feature_t* features, const unsigned feature_count,
rocprofiler_t* context, const char* str) {
if (str) fprintf(file, "%s:\n", str);
for (unsigned i= 0; i < feature_count; ++i) {
for (unsigned i = 0; i < feature_count; ++i) {
const rocprofiler_feature_t* p = &features[i];
fprintf(file, " %s ", p->name);
switch (p->data.kind) {
@@ -173,7 +176,7 @@ void output_results(FILE* file, const rocprofiler_feature_t* features, const uns
// Output group intermeadate profiling results, created internally for complex metrics
void output_group(FILE* file, const rocprofiler_group_t* group, const char* str) {
if (str) fprintf(file, "%s:\n", str);
for (unsigned i= 0; i < group->feature_count; ++i) {
for (unsigned i = 0; i < group->feature_count; ++i) {
output_results(file, group->features[i], 1, group->context, NULL);
}
}
@@ -190,15 +193,15 @@ void dump_context(context_entry_t* entry) {
FILE* file_handle = entry->file_handle;
fprintf(file_handle, "Dispatch[%u], kernel_object(0x%lx):\n", index, entry->data.kernel_object);
status = rocprofiler_group_get_data(&group);
check_status(status);
//output_group(file, group, "Group[0] data");
// output_group(file, group, "Group[0] data");
status = rocprofiler_get_metrics(group.context);
check_status(status);
output_results(file_handle, features, feature_count, group.context, NULL);
// Finishing cleanup
// Deleting profiling context will delete all allocated resources
rocprofiler_close(group.context);
@@ -240,10 +243,8 @@ void handler(rocprofiler_group_t group, void* arg) {
}
// Kernel disoatch callback
hsa_status_t dispatch_callback(
const rocprofiler_callback_data_t* callback_data,
void* user_data,
rocprofiler_group_t* group) {
hsa_status_t dispatch_callback(const rocprofiler_callback_data_t* callback_data, void* user_data,
rocprofiler_group_t* group) {
// HSA status
hsa_status_t status = HSA_STATUS_ERROR;
// Passed tool data
@@ -258,7 +259,8 @@ hsa_status_t dispatch_callback(
properties.handler_arg = (void*)entry;
// Open profiling context
status = rocprofiler_open(callback_data->agent, tool_data->features, tool_data->feature_count, &context, 0/*ROCPROFILER_MODE_SINGLEGROUP*/, &properties);
status = rocprofiler_open(callback_data->agent, tool_data->features, tool_data->feature_count,
&context, 0 /*ROCPROFILER_MODE_SINGLEGROUP*/, &properties);
check_status(status);
// Check that we have only one profiling group
@@ -284,11 +286,16 @@ hsa_status_t dispatch_callback(
// Tool constructor
CONSTRUCTOR_API void constructor() {
std::map<std::string, hsa_ven_amd_aqlprofile_parameter_name_t> parameters_dict;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_COMPUTE_UNIT_TARGET"] = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_COMPUTE_UNIT_TARGET;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_VM_ID_MASK"] = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_VM_ID_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_MASK"] = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK"] = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK2"] = HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK2;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_COMPUTE_UNIT_TARGET"] =
HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_COMPUTE_UNIT_TARGET;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_VM_ID_MASK"] =
HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_VM_ID_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_MASK"] =
HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK"] =
HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK;
parameters_dict["HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK2"] =
HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_TOKEN_MASK2;
// Set output file
result_prefix = getenv("ROCP_OUTPUT_DIR");
@@ -301,7 +308,8 @@ CONSTRUCTOR_API void constructor() {
perror("result file fopen");
exit(1);
}
} else file_handle = stdout;
} else
file_handle = stdout;
// Getting input
const char* xml_name = getenv("ROCP_INPUT");
@@ -318,7 +326,7 @@ CONSTRUCTOR_API void constructor() {
for (auto* entry : metrics_list) {
const std::string entry_str = entry->opts["name"];
size_t pos1 = 0;
while(pos1 < entry_str.length()) {
while (pos1 < entry_str.length()) {
const size_t pos2 = entry_str.find(",", pos1);
const std::string metric_name = entry_str.substr(pos1, pos2 - pos1);
metrics_vec.push_back(metric_name);
@@ -331,10 +339,10 @@ CONSTRUCTOR_API void constructor() {
auto traces_list = xml->GetNodes("top.trace");
const unsigned feature_count = metrics_vec.size() + traces_list.size();
rocprofiler_feature_t* features= new rocprofiler_feature_t[feature_count];
rocprofiler_feature_t* features = new rocprofiler_feature_t[feature_count];
memset(features, 0, feature_count * sizeof(rocprofiler_feature_t));
printf(" %d metrics\n", (int) metrics_vec.size());
printf(" %d metrics\n", (int)metrics_vec.size());
for (unsigned i = 0; i < metrics_vec.size(); ++i) {
const std::string& name = metrics_vec[i];
printf("%s%s", (i == 0) ? " " : ", ", name.c_str());
@@ -344,7 +352,7 @@ CONSTRUCTOR_API void constructor() {
}
if (metrics_vec.size()) printf("\n");
printf(" %d traces\n", (int) traces_list.size());
printf(" %d traces\n", (int)traces_list.size());
unsigned index = metrics_vec.size();
for (auto* entry : traces_list) {
auto params_list = xml->GetNodes("top.trace.parameters");
@@ -362,7 +370,7 @@ CONSTRUCTOR_API void constructor() {
for (auto* params : params_list) {
const unsigned parameter_count = params->opts.size();
rocprofiler_parameter_t *parameters = new rocprofiler_parameter_t[parameter_count];
rocprofiler_parameter_t* parameters = new rocprofiler_parameter_t[parameter_count];
unsigned p_index = 0;
for (auto& v : params->opts) {
const std::string parameter_name = v.first;