9# include <hiprt/hiprt.h>
19HIPRTDeviceQueue::HIPRTDeviceQueue(HIPRTDevice *device)
20 : HIPDeviceQueue((HIPDevice *)device), hiprt_device_(device)
24int HIPRTDeviceQueue::num_concurrent_states(
const size_t state_size)
const
34 if (hiprt_device_->have_error()) {
39 return HIPDeviceQueue::enqueue(kernel,
work_size, args);
42 const HIPContextScope scope(hiprt_device_);
43 const HIPDeviceKernel &hip_kernel = hiprt_device_->kernels.get(kernel);
50 const int num_stacks = num_blocks * num_threads_per_block;
51 hiprtGlobalStackBuffer &stack_buffer = hiprt_device_->global_stack_buffer;
52 if (!stack_buffer.stackData || stack_buffer.stackCount < num_stacks) {
53 const hiprtContext hiprt_context = hiprt_device_->get_hiprt_context();
55 if (stack_buffer.stackData) {
60 hiprtDestroyGlobalStackBuffer(hiprt_context, stack_buffer);
64 hiprtGlobalStackBufferInput stack_buffer_input{hiprtStackTypeGlobal,
65 hiprtStackEntryTypeInteger,
69 hiprtError rt_result = hiprtCreateGlobalStackBuffer(
70 hiprt_context, stack_buffer_input, stack_buffer);
72 if (rt_result != hiprtSuccess) {
73 LOG_ERROR <<
"Failed to create hiprt Global Stack Buffer";
82 (
void *)(&stack_buffer),
83 sizeof(hiprtGlobalStackBuffer));
85 int shared_mem_bytes = 0;
87 assert_success(hipModuleLaunchKernel(hip_kernel.function,
91 num_threads_per_block,
96 const_cast<void **
>(args_copy.
values),
102 return !(hiprt_device_->have_error());
#define CCL_NAMESPACE_END
#define HIPRT_THREAD_STACK_SIZE
#define HIPRT_THREAD_GROUP_SIZE
bool device_kernel_has_intersection(DeviceKernel kernel)
ccl_gpu_kernel_postfix const ccl_global int ccl_global float const int work_size
void add(const KernelFilmConvert *value)
ccl_device_inline size_t divide_up(const size_t x, const size_t y)