derek/gem5
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

dev-hsa,gpu-compute: Agent Packet handler implemented.

Browse Source 
HSA packet processor will now accept and process agent packets.

Type field in packet is command type.
For now:
        AgentCmd::Nop = 0
        AgentCmd::Steal = 1

Steal command steals the completion signal for a running kernel.
This enables a benchmark to use hsa primitives to send an agent
packet to steal the signal, then wait on that signal.

Minimal working example to be added in gem5-resources.

Change-Id: I37f8a4b7ea1780b471559aecbf4af1050353b0b1
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/37015
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Reviewed-by: Matthew Poremba <matthew.poremba@amd.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
This commit is contained in:
Daniel Gerzhoy
2020-11-04 11:51:46 -05:00
parent a0cff8b945
commit 9a01d3e927
8 changed files with 153 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
src/gpu-compute/gpu_command_processor.cc
View File

@@ -93,6 +93,10 @@ GPUCommandProcessor::submitDispatchPkt(void *raw_pkt, uint32_t queue_id,
DPRINTF(GPUCommandProc, "GPU machine code is %lli bytes from start of the "
"kernel object\n", akc.kernel_code_entry_byte_offset);
DPRINTF(GPUCommandProc,"GPUCommandProc: Sending dispatch pkt to %lu\n",
(uint64_t)tc->cpuId());
Addr machine_code_addr = (Addr)disp_pkt->kernel_object
+ akc.kernel_code_entry_byte_offset;
@@ -166,6 +170,54 @@ GPUCommandProcessor::submitVendorPkt(void *raw_pkt, uint32_t queue_id,
hsaPP->finishPkt(raw_pkt, queue_id);
}
/**
* submitAgentDispatchPkt() is for accepting agent dispatch packets.
* These packets will control the dispatch of Wg on the device, and inform
* the host when a specified number of Wg have been executed on the device.
*
* For now it simply finishes the pkt.
*/
void
GPUCommandProcessor::submitAgentDispatchPkt(void *raw_pkt, uint32_t queue_id,
Addr host_pkt_addr)
{
//Parse the Packet, see what it wants us to do
_hsa_agent_dispatch_packet_t * agent_pkt =
(_hsa_agent_dispatch_packet_t *)raw_pkt;
if (agent_pkt->type == AgentCmd::Nop) {
DPRINTF(GPUCommandProc, "Agent Dispatch Packet NOP\n");
} else if (agent_pkt->type == AgentCmd::Steal) {
//This is where we steal the HSA Task's completion signal
int kid = agent_pkt->arg[0];
DPRINTF(GPUCommandProc,
"Agent Dispatch Packet Stealing signal handle for kernel %d\n",
kid);
HSAQueueEntry *task = dispatcher.hsaTask(kid);
uint64_t signal_addr = task->completionSignal();// + sizeof(uint64_t);
uint64_t return_address = agent_pkt->return_address;
DPRINTF(GPUCommandProc, "Return Addr: %p\n",return_address);
//*return_address = signal_addr;
Addr *new_signal_addr = new Addr;
*new_signal_addr = (Addr)signal_addr;
dmaWriteVirt(return_address, sizeof(Addr), nullptr, new_signal_addr, 0);
DPRINTF(GPUCommandProc,
"Agent Dispatch Packet Stealing signal handle from kid %d :" \
"(%x:%x) writing into %x\n",
kid,signal_addr,new_signal_addr,return_address);
} else
{
panic("The agent dispatch packet provided an unknown argument in" \
"arg[0],currently only 0(nop) or 1(return kernel signal) is accepted");
}
hsaPP->finishPkt(raw_pkt, queue_id);
}
/**
* Once the CP has finished extracting all relevant information about
* a task and has initialized the ABI state, we send a description of