27 if (
const char *
str = getenv(name)) {
47 if (min_num_states < max_num_states) {
48 size_t total_memory = 0;
49 size_t free_memory = 0;
52 const size_t reserve_memory = total_memory * size_t(params.
reserve_percent) / 100;
53 const size_t budget_memory = free_memory - std::min(reserve_memory, free_memory);
56 size_t fit_num_states = budget_memory / state_size;
57 if (fit_num_states > baseline) {
68 LOG_INFO <<
"Adjusted concurrent states to available memory: " << baseline <<
" -> "
74 if (
const char *factor_str = getenv(
"CYCLES_CONCURRENT_STATES_FACTOR")) {
75 const float factor =
float(atof(factor_str));
80 LOG_TRACE <<
"CYCLES_CONCURRENT_STATES_FACTOR evaluated to 0";
98 stats_sorted.push_back(stat);
101 sort(stats_sorted.begin(),
103 [](
const pair<DeviceKernelMask, double> &a,
const pair<DeviceKernelMask, double> &
b) {
104 return a.second > b.second;
108 double total_time = 0.0;
109 for (
const auto &[
mask, time] : stats_sorted) {
111 LOG_TRACE <<
" " << std::setfill(
' ') << std::setw(10) << std::fixed << std::setprecision(5)
116 LOG_TRACE <<
"GPU queue total time: " << std::fixed << std::setprecision(5) << total_time;
150 const double new_time =
time_dt();
152 LOG_TRACE <<
"GPU queue synchronize, elapsed " << std::setw(10) << elapsed_time <<
"s";
void sort(btMatrix3x3 &U, btVector3 &sigma, btMatrix3x3 &V, int t)
Helper function of 3X3 SVD for sorting singular values.
DeviceKernelMask last_kernels_enqueued_
bool is_per_kernel_performance_
virtual void get_memory_info(size_t &total, size_t &free) const =0
virtual int num_concurrent_states(const size_t state_size) const
virtual ConcurrentStatesParams concurrent_states_params() const =0
map< DeviceKernelMask, double > stats_kernel_time_
string debug_active_kernels()
virtual bool synchronize()=0
void debug_init_execution()
DeviceQueue(Device *device)
void debug_enqueue_begin(DeviceKernel kernel, const int work_size)
#define CCL_NAMESPACE_END
string device_kernel_mask_as_string(DeviceKernelMask mask)
const char * device_kernel_as_string(DeviceKernel kernel)
ccl_gpu_kernel_postfix const ccl_global int ccl_global float const int work_size
ccl_device_inline float2 mask(const MaskType mask, const float2 a)
static void env_override(const char *name, int &value)
KERNEL_FEATURE_LIGHT_PASSES KERNEL_FEATURE_PATH_GUIDING float
string string_human_readable_size(size_t size)
CCL_NAMESPACE_BEGIN double time_dt()
ccl_device_inline size_t round_down(const size_t x, const size_t multiple)