Remove USE_DISPATCH_HSA_KERNEL=0 path.

Этот коммит содержится в:
Ben Sander
2016-12-17 07:20:30 -06:00
родитель 3f9404d0e1
Коммит 06d382bc6d
3 изменённых файлов: 0 добавлений и 113 удалений
-51
Просмотреть файл
@@ -387,57 +387,6 @@ void ihipStream_t::lockclose_postKernelCommand(const char * kernelName, hc::acce
};
#if USE_DISPATCH_HSA_KERNEL==0
// Precursor: the stream is already locked,specifically so this routine can enqueue work into the specified av.
void ihipStream_t::launchModuleKernel(
hc::accelerator_view av,
hsa_signal_t signal,
uint32_t blockDimX,
uint32_t blockDimY,
uint32_t blockDimZ,
uint32_t gridDimX,
uint32_t gridDimY,
uint32_t gridDimZ,
uint32_t groupSegmentSize,
uint32_t privateSegmentSize,
void *kernarg,
size_t kernSize,
uint64_t kernel){
hsa_status_t status;
hsa_queue_t *Queue = (hsa_queue_t*)av.get_hsa_queue();
const uint32_t queue_mask = Queue->size-1;
uint32_t packet_index = hsa_queue_load_write_index_relaxed(Queue);
hsa_kernel_dispatch_packet_t *dispatch_packet = &(((hsa_kernel_dispatch_packet_t*)(Queue->base_address))[packet_index & queue_mask]);
dispatch_packet->completion_signal = signal;
dispatch_packet->workgroup_size_x = blockDimX;
dispatch_packet->workgroup_size_y = blockDimY;
dispatch_packet->workgroup_size_z = blockDimZ;
dispatch_packet->grid_size_x = blockDimX * gridDimX;
dispatch_packet->grid_size_y = blockDimY * gridDimY;
dispatch_packet->grid_size_z = blockDimZ * gridDimZ;
dispatch_packet->group_segment_size = groupSegmentSize;
dispatch_packet->private_segment_size = privateSegmentSize;
dispatch_packet->kernarg_address = kernarg;
dispatch_packet->kernel_object = kernel;
uint16_t header = (HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE) |
(1 << HSA_PACKET_HEADER_BARRIER) |
(HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
(HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);
uint16_t setup = 3 << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS;
uint32_t header32 = header | (setup << 16);
__atomic_store_n((uint32_t*)(dispatch_packet), header32, __ATOMIC_RELEASE);
hsa_queue_store_write_index_relaxed(Queue, packet_index + 1);
hsa_signal_store_relaxed(Queue->doorbell_signal, packet_index);
}
#endif
//=============================================================================
// Recompute the peercnt and the packed _peerAgents whenever a peer is added or deleted.
// The packed _peerAgents can efficiently be used on each memory allocation.
-2
Просмотреть файл
@@ -33,9 +33,7 @@ THE SOFTWARE.
#error("This version of HIP requires a newer version of HCC.");
#endif
#define USE_DISPATCH_HSA_KERNEL 1
#define USE_IPC 0
//
//---
-60
Просмотреть файл
@@ -300,7 +300,6 @@ hipError_t hipModuleLaunchKernel(hipFunction_t f,
grid_launch_parm lp;
hStream = ihipPreLaunchKernel(hStream, 0, 0, &lp, "ModuleKernel");
#if USE_DISPATCH_HSA_KERNEL
hsa_kernel_dispatch_packet_t aql;
@@ -325,65 +324,6 @@ hipError_t hipModuleLaunchKernel(hipFunction_t f,
(HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);
lp.av->dispatch_hsa_kernel(&aql, config[1] /* kernarg*/, kernArgSize);
#else
/*
Create signal
*/
hsa_signal_t signal;
status = hsa_signal_create(1, 0, NULL, &signal);
if(status != HSA_STATUS_SUCCESS){
return ihipLogStatus(hipErrorNotFound);
}
/*
Allocate kernarg
*/
void *kern = nullptr;
hsa_amd_memory_pool_t *pool = reinterpret_cast<hsa_amd_memory_pool_t*>(lp.av->get_hsa_kernarg_region());
status = hsa_amd_memory_pool_allocate(*pool, kernArgSize, 0, &kern);
if(status != HSA_STATUS_SUCCESS){
return ihipLogStatus(hipErrorNotFound);
}
status = hsa_amd_agents_allow_access(1, (hsa_agent_t*)lp.av->get_hsa_agent(), 0, kern);
if(status != HSA_STATUS_SUCCESS){
return ihipLogStatus(hipErrorNotFound);
}
memcpy(kern, config[1], kernArgSize);
/*
Launch AQL packet
*/
hStream->launchModuleKernel(*lp.av, signal, blockDimX, blockDimY, blockDimZ,
gridDimX, gridDimY, gridDimZ, groupSegmentSize, privateSegmentSize, kern, kernArgSize, f->_object);
/*
Wait for signal
*/
hsa_signal_value_t value = hsa_signal_wait_acquire(signal, HSA_SIGNAL_CONDITION_LT, 1, UINT64_MAX, HSA_WAIT_STATE_BLOCKED);
/*
Destroy kernarg
*/
status = hsa_amd_memory_pool_free(kern);
if(status != HSA_STATUS_SUCCESS){
return ihipLogStatus(hipErrorNotFound);
}
/*
Destroy the signal
*/
status = hsa_signal_destroy(signal);
if(status != HSA_STATUS_SUCCESS){
return ihipLogStatus(hipErrorNotFound);
}
#endif // USE_DISPATCH_HSA_KERNEL
ihipPostLaunchKernel("ModuleKernel", hStream, lp);