Skip to content

Commit

Permalink
Enhancing mem-bw testcase to print individual bank name and throughpu…
Browse files Browse the repository at this point in the history 
…t. (#8101)

Signed-off-by: ManojTakasi <[email protected]>
Co-authored-by: ManojTakasi <[email protected]>
  • Loading branch information
@ManojTakasi
ManojTakasi and ManojTakasi authored Apr 24, 2024
1 parent e9e2080 commit 371020b
Showing 1 changed file with 42 additions and 15 deletions.
57 changes: 42 additions & 15 deletions src/runtime_src/core/tools/common/tests/TestBandwidthKernel.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -42,7 +42,7 @@ TestBandwidthKernel::run(std::shared_ptr<xrt_core::device> dev)
}

static void
marshal_build_metadata(std::string test_path, int* num_kernel, int* num_kernel_ddr, bool* chk_hbm_mem)
marshal_build_metadata(std::string test_path, int* num_kernel, int* num_kernel_ddr, bool* chk_hbm_mem, std::vector<std::string>& bank_names)
{
static const std::string filename = "platform.json";
auto platform_json = std::filesystem::path(test_path) / filename;
Expand All @@ -59,7 +59,17 @@ marshal_build_metadata(std::string test_path, int* num_kernel, int* num_kernel_d
auto sValue = pt_mem_entry.get<std::string>("type");
if (sValue == "HBM")
*chk_hbm_mem = true;

else if (sValue == "DDR") {
auto banks = pt_mem_entry.get_child("banks");

for (const auto&bank : banks) {
auto bank_name = bank.second.get<std::string>("name");
bank_names.push_back(bank_name);
}
}
}

if (*chk_hbm_mem) {
// As HBM is part of platform, number of ddr kernels is total count reduced by 1(single HBM)
*num_kernel_ddr = *num_kernel - 1;
Expand Down Expand Up @@ -117,9 +127,9 @@ initialize_output_host_hbm(unsigned int data_size)
}

static double
calculate_max_throughput(std::chrono::time_point<std::chrono::high_resolution_clock> time_end,
calculate_throughput(std::chrono::time_point<std::chrono::high_resolution_clock> time_end,
std::chrono::time_point<std::chrono::high_resolution_clock> time_start,
unsigned int data_size, int num_bank, double max_throughput)
unsigned int data_size, int num_bank)
{
double usduration =
(double)(std::chrono::duration_cast<std::chrono::nanoseconds>(time_end - time_start).count() / reps);
Expand All @@ -129,16 +139,16 @@ calculate_max_throughput(std::chrono::time_point<std::chrono::high_resolution_cl
double bpersec = (data_size * num_bank) / dsduration;
double mbpersec = (2 * bpersec) / ((double)1024 * 1024); // Convert b/sec to mb/sec

if (mbpersec > max_throughput)
max_throughput = mbpersec;

return max_throughput;
return mbpersec;
}

static double
static std::pair<double, std::vector<double>>
test_bandwidth_ddr(xrt::device device, std::vector<xrt::kernel> krnls, int num_kernel_ddr)
{
double max_throughput = 0;
double mbpersec = 0;
std::vector<double> throughput_per_kernel(num_kernel_ddr, 0);
// Starting at 4K and going up to 16M with increments of power of 2
for (uint32_t a = 4 * 1024; a <= 16 * 1024 * 1024; a *= 2) {
unsigned int data_size = a;
Expand Down Expand Up @@ -166,12 +176,22 @@ test_bandwidth_ddr(xrt::device device, std::vector<xrt::kernel> krnls, int num_k

auto time_start = std::chrono::high_resolution_clock::now();
std::vector<xrt::run> runs(num_kernel_ddr);
for (int i = 0; i < num_kernel_ddr; i++)
std::vector<std::chrono::high_resolution_clock::time_point> start_time_per_kernel(num_kernel_ddr);
std::vector<std::chrono::high_resolution_clock::time_point> end_time_per_kernel(num_kernel_ddr);

for (int i = 0; i < num_kernel_ddr; i++) {
start_time_per_kernel[i] = std::chrono::high_resolution_clock::now();
runs[i] = krnls[i](input_buffer[i], output_buffer[i], data_size, reps);
for (int i = 0; i < num_kernel_ddr; i++)
}
for (int i = 0; i < num_kernel_ddr; i++) {
runs[i].wait();
end_time_per_kernel[i] = std::chrono::high_resolution_clock::now();
}
auto time_end = std::chrono::high_resolution_clock::now();

for (int i = 0; i < num_kernel_ddr; i++) {
throughput_per_kernel[i] = std::max(throughput_per_kernel[i], calculate_throughput(end_time_per_kernel[i], start_time_per_kernel[i], data_size, 1));
}
for (int i = 0; i < num_kernel_ddr; i++) {
output_buffer[i].sync(XCL_BO_SYNC_BO_FROM_DEVICE);
output_buffer[i].read(output_host[i].data());
Expand All @@ -186,15 +206,17 @@ test_bandwidth_ddr(xrt::device device, std::vector<xrt::kernel> krnls, int num_k
}
}

max_throughput = calculate_max_throughput(time_end, time_start, data_size, num_kernel_ddr, max_throughput);
mbpersec = calculate_throughput(time_end, time_start, data_size, num_kernel_ddr);
max_throughput = std::max(mbpersec, max_throughput);
}
return max_throughput;
return std::make_pair(max_throughput, throughput_per_kernel);
}

static double
test_bandwidth_hbm(xrt::device device, std::vector<xrt::kernel> krnls, int num_kernel)
{
double max_throughput = 0;
double mbpersec = 0;
// Starting at 4K and going up to 16M with increments of power of 2
for (uint32_t i = 4 * 1024; i <= 16 * 1024 * 1024; i *= 2) {
unsigned int data_size = i;
Expand Down Expand Up @@ -230,7 +252,8 @@ test_bandwidth_hbm(xrt::device device, std::vector<xrt::kernel> krnls, int num_k
}
}

max_throughput = calculate_max_throughput(time_end, time_start, data_size, 1, max_throughput);
mbpersec = calculate_throughput(time_end, time_start, data_size, 1);
max_throughput = std::max(mbpersec, max_throughput);
}
return max_throughput;
}
Expand Down Expand Up @@ -260,9 +283,9 @@ TestBandwidthKernel::runTest(std::shared_ptr<xrt_core::device> dev, boost::prope
int num_kernel = 0;
int num_kernel_ddr = 0;
bool chk_hbm_mem = false;

std::vector<std::string> bank_names;
try {
marshal_build_metadata(test_path, &num_kernel, &num_kernel_ddr, &chk_hbm_mem);
marshal_build_metadata(test_path, &num_kernel, &num_kernel_ddr, &chk_hbm_mem, bank_names);
} catch (const std::exception&) {
logger(ptree, "Error", "Bad JSON format detected while marshaling build metadata");
ptree.put("status", test_token_skipped);
Expand All @@ -281,8 +304,12 @@ TestBandwidthKernel::runTest(std::shared_ptr<xrt_core::device> dev, boost::prope

try {
if (num_kernel_ddr) {
double max_throughput = test_bandwidth_ddr(device, krnls, num_kernel_ddr);
auto throughputs = test_bandwidth_ddr(device, krnls, num_kernel_ddr);
double max_throughput = throughputs.first;
std::vector <double> throughput_per_kernel = throughputs.second;
logger(ptree, "Details", boost::str(boost::format("Throughput (Type: DDR) (Bank count: %d) : %.1f MB/s") % num_kernel_ddr % max_throughput));
for (int i = 0; i < num_kernel_ddr; i++)
logger(ptree, "Details", boost::str(boost::format("Throughput of Memory Tag: %s : %.1f MB/s") % bank_names[i] % throughput_per_kernel[i]));
}
if (chk_hbm_mem) {
double max_throughput = test_bandwidth_hbm(device, krnls, num_kernel);
Expand Down

0 comments on commit 371020b

Please sign in to comment.