derek/gem5
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
501d72f121987900d7d96f0bc00ea0783cd6176f
gem5/src/gpu-compute/gpu_compute_driver.cc
Gabe Black f4209bbdee misc: Remove lingering uses of TheISA::. 
Change-Id: Ie55e0d79867fbc8f75a993fb456a58c84de5def4
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/62196
Reviewed-by: Giacomo Travaglini <giacomo.travaglini@arm.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Maintainer: Giacomo Travaglini <giacomo.travaglini@arm.com>
2022-08-20 07:30:16 +00:00

1040 lines
41 KiB
C++
Raw Blame History

/*
* Copyright (c) 2015-2018 Advanced Micro Devices, Inc.
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* 3. Neither the name of the copyright holder nor the names of its
* contributors may be used to endorse or promote products derived from this
* software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
#include "gpu-compute/gpu_compute_driver.hh"
#include <memory>
#include "arch/x86/page_size.hh"
#include "base/compiler.hh"
#include "base/logging.hh"
#include "base/trace.hh"
#include "cpu/thread_context.hh"
#include "debug/GPUDriver.hh"
#include "debug/GPUShader.hh"
#include "dev/hsa/hsa_packet_processor.hh"
#include "dev/hsa/kfd_event_defines.h"
#include "dev/hsa/kfd_ioctl.h"
#include "gpu-compute/gpu_command_processor.hh"
#include "gpu-compute/shader.hh"
#include "mem/port_proxy.hh"
#include "mem/se_translating_port_proxy.hh"
#include "mem/translating_port_proxy.hh"
#include "params/GPUComputeDriver.hh"
#include "sim/full_system.hh"
#include "sim/process.hh"
#include "sim/se_workload.hh"
#include "sim/syscall_emul_buf.hh"
namespace gem5
{
GPUComputeDriver::GPUComputeDriver(const Params &p)
: EmulatedDriver(p), device(p.device), queueId(0),
isdGPU(p.isdGPU), gfxVersion(p.gfxVersion), dGPUPoolID(p.dGPUPoolID),
eventPage(0), eventSlotIndex(0)
{
device->attachDriver(this);
DPRINTF(GPUDriver, "Constructing KFD: device\n");
// Convert the 3 bit mtype specified in Shader.py to the proper type
// used for requests.
std::bitset<MtypeFlags::NUM_MTYPE_BITS> mtype(p.m_type);
if (mtype.test(MtypeFlags::SHARED)) {
defaultMtype.set(Request::SHARED);
}
if (mtype.test(MtypeFlags::READ_WRITE)) {
defaultMtype.set(Request::READ_WRITE);
}
if (mtype.test(MtypeFlags::CACHED)) {
defaultMtype.set(Request::CACHED);
}
}
const char*
GPUComputeDriver::DriverWakeupEvent::description() const
{
return "DriverWakeupEvent";
}
/**
* Create an FD entry for the KFD inside of the owning process.
*/
int
GPUComputeDriver::open(ThreadContext *tc, int mode, int flags)
{
DPRINTF(GPUDriver, "Opened %s\n", filename);
auto process = tc->getProcessPtr();
auto device_fd_entry = std::make_shared<DeviceFDEntry>(this, filename);
int tgt_fd = process->fds->allocFD(device_fd_entry);
return tgt_fd;
}
/**
* Currently, mmap() will simply setup a mapping for the associated
* device's packet processor's doorbells and creates the event page.
*/
Addr
GPUComputeDriver::mmap(ThreadContext *tc, Addr start, uint64_t length,
int prot, int tgt_flags, int tgt_fd, off_t offset)
{
auto process = tc->getProcessPtr();
auto mem_state = process->memState;
Addr pg_off = offset >> PAGE_SHIFT;
Addr mmap_type = pg_off & KFD_MMAP_TYPE_MASK;
DPRINTF(GPUDriver, "amdkfd mmap (start: %p, length: 0x%x,"
"offset: 0x%x)\n", start, length, offset);
switch(mmap_type) {
case KFD_MMAP_TYPE_DOORBELL:
DPRINTF(GPUDriver, "amdkfd mmap type DOORBELL offset\n");
start = mem_state->extendMmap(length);
process->pTable->map(start, device->hsaPacketProc().pioAddr,
length, false);
break;
case KFD_MMAP_TYPE_EVENTS:
DPRINTF(GPUDriver, "amdkfd mmap type EVENTS offset\n");
panic_if(start != 0,
"Start address should be provided by KFD\n");
panic_if(length != 8 * KFD_SIGNAL_EVENT_LIMIT,
"Requested length %d, expected length %d; length "
"mismatch\n", length, 8* KFD_SIGNAL_EVENT_LIMIT);
/**
* We don't actually access these pages. We just need to reserve
* some VA space. See commit id 5ce8abce for details on how
* events are currently implemented.
*/
if (!eventPage) {
eventPage = mem_state->extendMmap(length);
start = eventPage;
}
break;
default:
warn_once("Unrecognized kfd mmap type %llx\n", mmap_type);
break;
}
return start;
}
/**
* Forward relevant parameters to packet processor; queueId
* is used to link doorbell. The queueIDs are not re-used
* in current implementation, and we allocate only one page
* (4096 bytes) for doorbells, so check if this queueID can
* be mapped into that page.
*/
void
GPUComputeDriver::allocateQueue(PortProxy &mem_proxy, Addr ioc_buf)
{
TypedBufferArg<kfd_ioctl_create_queue_args> args(ioc_buf);
args.copyIn(mem_proxy);
if ((doorbellSize() * queueId) > 4096) {
fatal("%s: Exceeded maximum number of HSA queues allowed\n", name());
}
args->doorbell_offset = (KFD_MMAP_TYPE_DOORBELL |
KFD_MMAP_GPU_ID(args->gpu_id)) << PAGE_SHIFT;
// for vega offset needs to include exact value of doorbell
if (doorbellSize())
args->doorbell_offset += queueId * doorbellSize();
args->queue_id = queueId++;
auto &hsa_pp = device->hsaPacketProc();
hsa_pp.setDeviceQueueDesc(args->read_pointer_address,
args->ring_base_address, args->queue_id,
args->ring_size, doorbellSize(), gfxVersion);
args.copyOut(mem_proxy);
}
void
GPUComputeDriver::DriverWakeupEvent::scheduleWakeup(Tick wakeup_delay)
{
assert(driver);
driver->schedule(this, curTick() + wakeup_delay);
}
void
GPUComputeDriver::signalWakeupEvent(uint32_t event_id)
{
panic_if(event_id >= eventSlotIndex,
"Trying wakeup on an event that is not yet created\n");
if (ETable[event_id].threadWaiting) {
panic_if(!ETable[event_id].tc,
"No thread context to wake up\n");
ThreadContext *tc = ETable[event_id].tc;
DPRINTF(GPUDriver,
"Signal event: Waking up CPU %d\n", tc->cpuId());
// Remove events that can wakeup this thread
TCEvents[tc].clearEvents();
// Now wakeup this thread
tc->activate();
} else {
// This may be a race condition between an ioctl call asking to wait on
// this event and this signalWakeupEvent. Taking care of this race
// condition here by setting the event here. The ioctl call should take
// the necessary action when waiting on an already set event. However,
// this may be a genuine instance in which the runtime has decided not
// to wait on this event. But since we cannot distinguish this case with
// the race condition, we are any way setting the event.
ETable[event_id].setEvent = true;
}
}
void
GPUComputeDriver::DriverWakeupEvent::process()
{
DPRINTF(GPUDriver,
"Timer event: Waking up CPU %d\n", tc->cpuId());
// Remove events that can wakeup this thread
driver->TCEvents[tc].clearEvents();
// Now wakeup this thread
tc->activate();
}
int
GPUComputeDriver::ioctl(ThreadContext *tc, unsigned req, Addr ioc_buf)
{
TranslatingPortProxy fs_proxy(tc);
SETranslatingPortProxy se_proxy(tc);
PortProxy &virt_proxy = FullSystem ? fs_proxy : se_proxy;
auto process = tc->getProcessPtr();
auto mem_state = process->memState;
switch (req) {
case AMDKFD_IOC_GET_VERSION:
{
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_GET_VERSION\n");
TypedBufferArg<kfd_ioctl_get_version_args> args(ioc_buf);
args->major_version = KFD_IOCTL_MAJOR_VERSION;
args->minor_version = KFD_IOCTL_MINOR_VERSION;
args.copyOut(virt_proxy);
}
break;
case AMDKFD_IOC_CREATE_QUEUE:
{
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_CREATE_QUEUE\n");
allocateQueue(virt_proxy, ioc_buf);
DPRINTF(GPUDriver, "Creating queue %d\n", queueId);
}
break;
case AMDKFD_IOC_DESTROY_QUEUE:
{
TypedBufferArg<kfd_ioctl_destroy_queue_args> args(ioc_buf);
args.copyIn(virt_proxy);
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_DESTROY_QUEUE;" \
"queue offset %d\n", args->queue_id);
device->hsaPacketProc().unsetDeviceQueueDesc(args->queue_id,
doorbellSize());
}
break;
case AMDKFD_IOC_SET_MEMORY_POLICY:
{
/**
* This is where the runtime requests MTYPE from an aperture.
* Basically, the globally memory aperture is divided up into
* a default aperture and an alternate aperture each of which have
* their own MTYPE policies. This is done to mark a small piece
* of the global memory as uncacheable. Host memory mappings will
* be carved out of this uncacheable aperture, which is how they
* implement 'coherent' host/device memory on dGPUs.
*
* TODO: Need to reflect per-aperture MTYPE policies based on this
* call.
*
*/
warn("unimplemented ioctl: AMDKFD_IOC_SET_MEMORY_POLICY\n");
}
break;
case AMDKFD_IOC_GET_CLOCK_COUNTERS:
{
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_GET_CLOCK_COUNTERS\n");
TypedBufferArg<kfd_ioctl_get_clock_counters_args> args(ioc_buf);
args.copyIn(virt_proxy);
// Set nanosecond resolution
args->system_clock_freq = 1000000000;
/**
* Derive all clock counters based on the tick. All
* device clocks are identical and perfectly in sync.
*/
uint64_t elapsed_nsec = curTick() / sim_clock::as_int::ns;
args->gpu_clock_counter = elapsed_nsec;
args->cpu_clock_counter = elapsed_nsec;
args->system_clock_counter = elapsed_nsec;
args.copyOut(virt_proxy);
}
break;
case AMDKFD_IOC_GET_PROCESS_APERTURES:
{
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_GET_PROCESS_APERTURES\n");
TypedBufferArg<kfd_ioctl_get_process_apertures_args> args(ioc_buf);
args->num_of_nodes = 1;
/**
* Set the GPUVM/LDS/Scratch APEs exactly as they
* are in the real driver, see the KFD driver
* in the ROCm Linux kernel source:
* drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c
*/
for (int i = 0; i < args->num_of_nodes; ++i) {
/**
* While the GPU node numbers start at 0, we add 1
* to force the count to start at 1. This is to
* ensure that the base/limit addresses are
* calculated correctly.
*/
switch (gfxVersion) {
case GfxVersion::gfx801:
case GfxVersion::gfx803:
args->process_apertures[i].scratch_base =
scratchApeBase(i + 1);
args->process_apertures[i].lds_base =
ldsApeBase(i + 1);
break;
case GfxVersion::gfx900:
case GfxVersion::gfx902:
args->process_apertures[i].scratch_base =
scratchApeBaseV9();
args->process_apertures[i].lds_base =
ldsApeBaseV9();
break;
default:
fatal("Invalid gfx version\n");
}
// GFX8 and GFX9 set lds and scratch limits the same way
args->process_apertures[i].scratch_limit =
scratchApeLimit(args->process_apertures[i].scratch_base);
args->process_apertures[i].lds_limit =
ldsApeLimit(args->process_apertures[i].lds_base);
switch (gfxVersion) {
case GfxVersion::gfx801:
args->process_apertures[i].gpuvm_base =
gpuVmApeBase(i + 1);
args->process_apertures[i].gpuvm_limit =
gpuVmApeLimit(args->process_apertures[i].gpuvm_base);
break;
case GfxVersion::gfx803:
case GfxVersion::gfx900:
case GfxVersion::gfx902:
// Taken from SVM_USE_BASE in Linux kernel
args->process_apertures[i].gpuvm_base = 0x1000000ull;
// Taken from AMDGPU_GMC_HOLE_START in Linux kernel
args->process_apertures[i].gpuvm_limit =
0x0000800000000000ULL - 1;
break;
default:
fatal("Invalid gfx version");
}
// NOTE: Must match ID populated by hsaTopology.py
//
// https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver/
// blob/6a986c0943e9acd8c4c0cf2a9d510ff42167b43f/include/uapi/
// linux/kfd_ioctl.h#L564
//
// The gpu_id is a device identifier used by the driver for
// ioctls that allocate arguments. Each device has an unique
// id composed out of a non-zero base and an offset.
if (isdGPU) {
switch (gfxVersion) {
case GfxVersion::gfx803:
args->process_apertures[i].gpu_id = 50156;
break;
case GfxVersion::gfx900:
args->process_apertures[i].gpu_id = 22124;
break;
default:
fatal("Invalid gfx version for dGPU\n");
}
} else {
switch (gfxVersion) {
case GfxVersion::gfx801:
case GfxVersion::gfx902:
args->process_apertures[i].gpu_id = 2765;
break;
default:
fatal("Invalid gfx version for APU\n");
}
}
DPRINTF(GPUDriver, "GPUVM base for node[%i] = %#x\n", i,
args->process_apertures[i].gpuvm_base);
DPRINTF(GPUDriver, "GPUVM limit for node[%i] = %#x\n", i,
args->process_apertures[i].gpuvm_limit);
DPRINTF(GPUDriver, "LDS base for node[%i] = %#x\n", i,
args->process_apertures[i].lds_base);
DPRINTF(GPUDriver, "LDS limit for node[%i] = %#x\n", i,
args->process_apertures[i].lds_limit);
DPRINTF(GPUDriver, "Scratch base for node[%i] = %#x\n", i,
args->process_apertures[i].scratch_base);
DPRINTF(GPUDriver, "Scratch limit for node[%i] = %#x\n", i,
args->process_apertures[i].scratch_limit);
/**
* The CPU's 64b address space can only use the
* areas with VA[63:47] == 0x1ffff or VA[63:47] == 0,
* therefore we must ensure that the apertures do not
* fall in the CPU's address space.
*/
assert(bits<Addr>(args->process_apertures[i].scratch_base, 63,
47) != 0x1ffff);
assert(bits<Addr>(args->process_apertures[i].scratch_base, 63,
47) != 0);
assert(bits<Addr>(args->process_apertures[i].scratch_limit, 63,
47) != 0x1ffff);
assert(bits<Addr>(args->process_apertures[i].scratch_limit, 63,
47) != 0);
assert(bits<Addr>(args->process_apertures[i].lds_base, 63,
47) != 0x1ffff);
assert(bits<Addr>(args->process_apertures[i].lds_base, 63,
47) != 0);
assert(bits<Addr>(args->process_apertures[i].lds_limit, 63,
47) != 0x1ffff);
assert(bits<Addr>(args->process_apertures[i].lds_limit, 63,
47) != 0);
}
args.copyOut(virt_proxy);
}
break;
case AMDKFD_IOC_UPDATE_QUEUE:
{
warn("unimplemented ioctl: AMDKFD_IOC_UPDATE_QUEUE\n");
}
break;
case AMDKFD_IOC_CREATE_EVENT:
{
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_CREATE_EVENT\n");
TypedBufferArg<kfd_ioctl_create_event_args> args(ioc_buf);
args.copyIn(virt_proxy);
if (args->event_type != KFD_IOC_EVENT_SIGNAL) {
warn("Signal events are only supported currently\n");
} else if (eventSlotIndex == SLOTS_PER_PAGE) {
fatal("Signal event wasn't created; signal limit reached\n");
}
// Currently, we allocate only one signal_page for events.
// Note that this signal page is of size 8 * KFD_SIGNAL_EVENT_LIMIT
uint64_t page_index = 0;
args->event_page_offset = (page_index | KFD_MMAP_TYPE_EVENTS);
args->event_page_offset <<= PAGE_SHIFT;
// TODO: Currently we support only signal events, hence using
// the same ID for both signal slot and event slot
args->event_slot_index = eventSlotIndex;
args->event_id = eventSlotIndex++;
args->event_trigger_data = args->event_id;
DPRINTF(GPUDriver, "amdkfd create events"
"(event_id: 0x%x, offset: 0x%x)\n",
args->event_id, args->event_page_offset);
// Since eventSlotIndex is increased everytime a new event is
// created ETable at eventSlotIndex(event_id) is guaranteed to be
// empty. In a future implementation that reuses deleted event_ids,
// we should check if event table at this
// eventSlotIndex(event_id) is empty before inserting a new event
// table entry
ETable.emplace(std::pair<uint32_t, ETEntry>(args->event_id, {}));
args.copyOut(virt_proxy);
}
break;
case AMDKFD_IOC_DESTROY_EVENT:
{
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_DESTROY_EVENT\n");
TypedBufferArg<kfd_ioctl_destroy_event_args> args(ioc_buf);
args.copyIn(virt_proxy);
DPRINTF(GPUDriver, "amdkfd destroying event %d\n", args->event_id);
fatal_if(ETable.count(args->event_id) == 0,
"Event ID invalid, cannot destroy this event\n");
ETable.erase(args->event_id);
}
break;
case AMDKFD_IOC_SET_EVENT:
{
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_SET_EVENTS\n");
TypedBufferArg<kfd_ioctl_set_event_args> args(ioc_buf);
args.copyIn(virt_proxy);
DPRINTF(GPUDriver, "amdkfd set event %d\n", args->event_id);
fatal_if(ETable.count(args->event_id) == 0,
"Event ID invlaid, cannot set this event\n");
ETable[args->event_id].setEvent = true;
signalWakeupEvent(args->event_id);
}
break;
case AMDKFD_IOC_RESET_EVENT:
{
warn("unimplemented ioctl: AMDKFD_IOC_RESET_EVENT\n");
}
break;
case AMDKFD_IOC_WAIT_EVENTS:
{
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_WAIT_EVENTS\n");
TypedBufferArg<kfd_ioctl_wait_events_args> args(ioc_buf);
args.copyIn(virt_proxy);
kfd_event_data *events =
(kfd_event_data *)args->events_ptr;
DPRINTF(GPUDriver, "amdkfd wait for events"
"(wait on all: %d, timeout : %d, num_events: %s)\n",
args->wait_for_all, args->timeout, args->num_events);
panic_if(args->wait_for_all != 0 && args->num_events > 1,
"Wait for all events not supported\n");
bool should_sleep = true;
if (TCEvents.count(tc) == 0) {
// This thread context trying to wait on an event for the first
// time, initialize it.
TCEvents.emplace(std::piecewise_construct, std::make_tuple(tc),
std::make_tuple(this, tc));
DPRINTF(GPUDriver, "\tamdkfd creating event list"
" for thread %d\n", tc->cpuId());
}
panic_if(TCEvents[tc].signalEvents.size() != 0,
"There are %d events that put this thread to sleep,"
" this thread should not be running\n",
TCEvents[tc].signalEvents.size());
for (int i = 0; i < args->num_events; i++) {
panic_if(!events,
"Event pointer invalid\n");
Addr eventDataAddr = (Addr)(events + i);
TypedBufferArg<kfd_event_data> EventData(
eventDataAddr, sizeof(kfd_event_data));
EventData.copyIn(virt_proxy);
DPRINTF(GPUDriver,
"\tamdkfd wait for event %d\n", EventData->event_id);
panic_if(ETable.count(EventData->event_id) == 0,
"Event ID invalid, cannot set this event\n");
if (ETable[EventData->event_id].threadWaiting)
warn("Multiple threads waiting on the same event\n");
if (ETable[EventData->event_id].setEvent) {
// If event is already set, the event has already happened.
// Just unset the event and dont put this thread to sleep.
ETable[EventData->event_id].setEvent = false;
should_sleep = false;
}
if (should_sleep) {
// Put this thread to sleep
ETable[EventData->event_id].threadWaiting = true;
ETable[EventData->event_id].tc = tc;
TCEvents[tc].signalEvents.insert(EventData->event_id);
}
}
// TODO: Return the correct wait_result back. Currently, returning
// success for both KFD_WAIT_TIMEOUT and KFD_WAIT_COMPLETE.
// Ideally, this needs to be done after the event is triggered and
// after the thread is woken up.
args->wait_result = 0;
args.copyOut(virt_proxy);
if (should_sleep) {
// Put this thread to sleep
sleepCPU(tc, args->timeout);
} else {
// Remove events that tried to put this thread to sleep
TCEvents[tc].clearEvents();
}
}
break;
case AMDKFD_IOC_DBG_REGISTER:
{
warn("unimplemented ioctl: AMDKFD_IOC_DBG_REGISTER\n");
}
break;
case AMDKFD_IOC_DBG_UNREGISTER:
{
warn("unimplemented ioctl: AMDKFD_IOC_DBG_UNREGISTER\n");
}
break;
case AMDKFD_IOC_DBG_ADDRESS_WATCH:
{
warn("unimplemented ioctl: AMDKFD_IOC_DBG_ADDRESS_WATCH\n");
}
break;
case AMDKFD_IOC_DBG_WAVE_CONTROL:
{
warn("unimplemented ioctl: AMDKFD_IOC_DBG_WAVE_CONTROL\n");
}
break;
case AMDKFD_IOC_SET_SCRATCH_BACKING_VA:
{
warn("unimplemented ioctl: AMDKFD_IOC_SET_SCRATCH_BACKING_VA\n");
}
break;
case AMDKFD_IOC_GET_TILE_CONFIG:
{
warn("unimplemented ioctl: AMDKFD_IOC_GET_TILE_CONFIG\n");
}
break;
case AMDKFD_IOC_SET_TRAP_HANDLER:
{
warn("unimplemented ioctl: AMDKFD_IOC_SET_TRAP_HANDLER\n");
}
break;
case AMDKFD_IOC_GET_PROCESS_APERTURES_NEW:
{
DPRINTF(GPUDriver,
"ioctl: AMDKFD_IOC_GET_PROCESS_APERTURES_NEW\n");
TypedBufferArg<kfd_ioctl_get_process_apertures_new_args>
ioc_args(ioc_buf);
ioc_args.copyIn(virt_proxy);
ioc_args->num_of_nodes = 1;
for (int i = 0; i < ioc_args->num_of_nodes; ++i) {
TypedBufferArg<kfd_process_device_apertures> ape_args
(ioc_args->kfd_process_device_apertures_ptr);
switch (gfxVersion) {
case GfxVersion::gfx801:
case GfxVersion::gfx803:
ape_args->scratch_base = scratchApeBase(i + 1);
ape_args->lds_base = ldsApeBase(i + 1);
break;
case GfxVersion::gfx900:
case GfxVersion::gfx902:
ape_args->scratch_base = scratchApeBaseV9();
ape_args->lds_base = ldsApeBaseV9();
break;
default:
fatal("Invalid gfx version\n");
}
// GFX8 and GFX9 set lds and scratch limits the same way
ape_args->scratch_limit =
scratchApeLimit(ape_args->scratch_base);
ape_args->lds_limit = ldsApeLimit(ape_args->lds_base);
switch (gfxVersion) {
case GfxVersion::gfx801:
ape_args->gpuvm_base = gpuVmApeBase(i + 1);
ape_args->gpuvm_limit =
gpuVmApeLimit(ape_args->gpuvm_base);
break;
case GfxVersion::gfx803:
case GfxVersion::gfx900:
case GfxVersion::gfx902:
// Taken from SVM_USE_BASE in Linux kernel
ape_args->gpuvm_base = 0x1000000ull;
// Taken from AMDGPU_GMC_HOLE_START in Linux kernel
ape_args->gpuvm_limit = 0x0000800000000000ULL - 1;
break;
default:
fatal("Invalid gfx version\n");
}
// NOTE: Must match ID populated by hsaTopology.py
if (isdGPU) {
switch (gfxVersion) {
case GfxVersion::gfx803:
ape_args->gpu_id = 50156;
break;
case GfxVersion::gfx900:
ape_args->gpu_id = 22124;
break;
default:
fatal("Invalid gfx version for dGPU\n");
}
} else {
switch (gfxVersion) {
case GfxVersion::gfx801:
case GfxVersion::gfx902:
ape_args->gpu_id = 2765;
break;
default:
fatal("Invalid gfx version for APU\n");
}
}
assert(bits<Addr>(ape_args->scratch_base, 63, 47) != 0x1ffff);
assert(bits<Addr>(ape_args->scratch_base, 63, 47) != 0);
assert(bits<Addr>(ape_args->scratch_limit, 63, 47) != 0x1ffff);
assert(bits<Addr>(ape_args->scratch_limit, 63, 47) != 0);
assert(bits<Addr>(ape_args->lds_base, 63, 47) != 0x1ffff);
assert(bits<Addr>(ape_args->lds_base, 63, 47) != 0);
assert(bits<Addr>(ape_args->lds_limit, 63, 47) != 0x1ffff);
assert(bits<Addr>(ape_args->lds_limit, 63, 47) != 0);
ape_args.copyOut(virt_proxy);
}
ioc_args.copyOut(virt_proxy);
}
break;
case AMDKFD_IOC_ACQUIRE_VM:
{
warn("unimplemented ioctl: AMDKFD_IOC_ACQUIRE_VM\n");
}
break;
/**
* In real hardware, this IOCTL maps host memory, dGPU memory, or dGPU
* doorbells into GPUVM space. Essentially, ROCm implements SVM by
* carving out a region of free VA space that both the host and GPUVM
* can agree upon. The entire GPU VA space is reserved on the host
* using a fixed mmap at a low VA range that is also directly
* accessable by the GPU's limited number of VA bits. When we actually
* call memory allocation later in the program, this IOCTL is invoked
* to create BOs/VMAs in the driver and bind them to physical
* memory/doorbells.
*
* For gem5, we don't need to carve out any GPUVM space here (we don't
* support GPUVM and use host page tables on the GPU directly). We can
* can just use the existing host SVM region. We comment on each memory
* type seperately.
*/
case AMDKFD_IOC_ALLOC_MEMORY_OF_GPU:
{
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_ALLOC_MEMORY_OF_GPU\n");
TypedBufferArg<kfd_ioctl_alloc_memory_of_gpu_args> args(ioc_buf);
args.copyIn(virt_proxy);
assert(isdGPU || gfxVersion == GfxVersion::gfx902);
assert((args->va_addr % X86ISA::PageBytes) == 0);
[[maybe_unused]] Addr mmap_offset = 0;
Request::CacheCoherenceFlags mtype = defaultMtype;
Addr pa_addr = 0;
int npages = divCeil(args->size, (int64_t)X86ISA::PageBytes);
bool cacheable = true;
if (KFD_IOC_ALLOC_MEM_FLAGS_VRAM & args->flags) {
DPRINTF(GPUDriver, "amdkfd allocation type: VRAM\n");
args->mmap_offset = args->va_addr;
// VRAM allocations are device memory mapped into GPUVM
// space.
//
// We can't rely on the lazy host allocator (fixupFault) to
// handle this mapping since it needs to be placed in dGPU
// framebuffer memory. The lazy allocator will try to place
// this in host memory.
//
// TODO: We don't have the appropriate bifurcation of the
// physical address space with different memory controllers
// yet. This is where we will explicitly add the PT maps to
// dGPU memory in the future.
//
// Bind the VA space to the dGPU physical memory pool. Mark
// this region as Uncacheable. The Uncacheable flag is only
// really used by the CPU and is ignored by the GPU. We mark
// this as uncacheable from the CPU so that we can implement
// direct CPU framebuffer access similar to what we currently
// offer in real HW through the so-called Large BAR feature.
pa_addr = process->seWorkload->allocPhysPages(
npages, dGPUPoolID);
//
// TODO: Uncacheable accesses need to be supported by the
// CPU-side protocol for this to work correctly. I believe
// it only works right now if the physical memory is MMIO
cacheable = false;
DPRINTF(GPUDriver, "Mapping VA %p to framebuffer PA %p size "
"%d\n", args->va_addr, pa_addr, args->size);
} else if (KFD_IOC_ALLOC_MEM_FLAGS_USERPTR & args->flags) {
DPRINTF(GPUDriver, "amdkfd allocation type: USERPTR\n");
mmap_offset = args->mmap_offset;
// USERPTR allocations are system memory mapped into GPUVM
// space. The user provides the driver with the pointer.
pa_addr = process->seWorkload->allocPhysPages(npages);
DPRINTF(GPUDriver, "Mapping VA %p to framebuffer PA %p size "
"%d\n", args->va_addr, pa_addr, args->size);
// If the HSA runtime requests system coherent memory, than we
// need to explicity mark this region as uncacheable from the
// perspective of the GPU.
if (args->flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
mtype.clear();
} else if (KFD_IOC_ALLOC_MEM_FLAGS_GTT & args->flags) {
DPRINTF(GPUDriver, "amdkfd allocation type: GTT\n");
args->mmap_offset = args->va_addr;
// GTT allocations are system memory mapped into GPUVM space.
// It's different than a USERPTR allocation since the driver
// itself allocates the physical memory on the host.
//
// We will lazily map it into host memory on first touch. The
// fixupFault will find the original SVM aperture mapped to the
// host.
pa_addr = process->seWorkload->allocPhysPages(npages);
DPRINTF(GPUDriver, "Mapping VA %p to framebuffer PA %p size "
"%d\n", args->va_addr, pa_addr, args->size);
// If the HSA runtime requests system coherent memory, than we
// need to explicity mark this region as uncacheable from the
// perspective of the GPU.
if (args->flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
mtype.clear();
// Note that for GTT the thunk layer needs to call mmap on the
// driver FD later if it wants the host to have access to this
// memory (which it probably does). This will be ignored.
} else if (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL & args->flags) {
DPRINTF(GPUDriver, "amdkfd allocation type: DOORBELL\n");
// DOORBELL allocations are the queue doorbells that are
// memory mapped into GPUVM space.
//
// Explicitly map this virtual address to our PIO doorbell
// interface in the page tables (non-cacheable)
pa_addr = device->hsaPacketProc().pioAddr;
cacheable = false;
}
DPRINTF(GPUDriver, "amdkfd allocation arguments: va_addr %p "
"size %lu, mmap_offset %p, gpu_id %d\n",
args->va_addr, args->size, mmap_offset, args->gpu_id);
// Bind selected physical memory to provided virtual address range
// in X86 page tables.
process->pTable->map(args->va_addr, pa_addr, args->size,
cacheable);
// We keep track of allocated regions of GPU mapped memory,
// just like the driver would. This allows us to provide the
// user with a unique handle for a given allocation. The user
// will only provide us with a handle after allocation and expect
// us to be able to use said handle to extract all the properties
// of the region.
//
// This is a simplified version of regular system VMAs, but for
// GPUVM space (none of the clobber/remap nonsense we find in real
// OS managed memory).
allocateGpuVma(mtype, args->va_addr, args->size);
// Used by the runtime to uniquely identify this allocation.
// We can just use the starting address of the VMA region.
args->handle= args->va_addr;
args.copyOut(virt_proxy);
}
break;
case AMDKFD_IOC_FREE_MEMORY_OF_GPU:
{
DPRINTF(GPUDriver, "ioctl: AMDKFD_IOC_FREE_MEMORY_OF_GPU\n");
TypedBufferArg<kfd_ioctl_free_memory_of_gpu_args> args(ioc_buf);
args.copyIn(virt_proxy);
assert(isdGPU);
DPRINTF(GPUDriver, "amdkfd free arguments: handle %p ",
args->handle);
// We don't recycle physical pages in SE mode
Addr size = deallocateGpuVma(args->handle);
process->pTable->unmap(args->handle, size);
// TODO: IOMMU and GPUTLBs do not seem to correctly support
// shootdown. This is also a potential issue for APU systems
// that perform unmap or remap with system memory.
tc->getMMUPtr()->flushAll();
args.copyOut(virt_proxy);
}
break;
/**
* Called to map an already allocated region of memory to this GPU's
* GPUVM VA space. We don't need to implement this in the simulator
* since we only have a single VM system. If the region has already
* been allocated somewhere like the CPU, then it's already visible
* to the device.
*/
case AMDKFD_IOC_MAP_MEMORY_TO_GPU:
{
warn("unimplemented ioctl: AMDKFD_IOC_MAP_MEMORY_TO_GPU\n");
}
break;
case AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU:
{
warn("unimplemented ioctl: AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU\n");
}
break;
case AMDKFD_IOC_SET_CU_MASK:
{
warn("unimplemented ioctl: AMDKFD_IOC_SET_CU_MASK\n");
}
break;
case AMDKFD_IOC_GET_QUEUE_WAVE_STATE:
{
warn("unimplemented ioctl: AMDKFD_IOC_GET_QUEUE_WAVE_STATE\n");
}
break;
case AMDKFD_IOC_GET_DMABUF_INFO:
{
warn("unimplemented ioctl: AMDKFD_IOC_GET_DMABUF_INFO\n");
}
break;
case AMDKFD_IOC_IMPORT_DMABUF:
{
warn("unimplemented ioctl: AMDKFD_IOC_IMPORT_DMABUF\n");
}
break;
case AMDKFD_IOC_ALLOC_QUEUE_GWS:
{
warn("unimplemented ioctl: AMDKFD_IOC_ALLOC_QUEUE_GWS\n");
}
break;
case AMDKFD_IOC_SMI_EVENTS:
{
warn("unimplemented ioctl: AMDKFD_IOC_SMI_EVENTS\n");
}
break;
default:
fatal("%s: bad ioctl %d\n", req);
break;
}
return 0;
}
void
GPUComputeDriver::sleepCPU(ThreadContext *tc, uint32_t milliSecTimeout)
{
// Convert millisecs to ticks
Tick wakeup_delay((uint64_t)milliSecTimeout * 1000000000);
assert(TCEvents.count(tc) == 1);
TCEvents[tc].timerEvent.scheduleWakeup(wakeup_delay);
tc->suspend();
DPRINTF(GPUDriver,
"CPU %d is put to sleep\n", tc->cpuId());
}
Addr
GPUComputeDriver::gpuVmApeBase(int gpuNum) const
{
return ((Addr)gpuNum << 61) + 0x1000000000000L;
}
Addr
GPUComputeDriver::gpuVmApeLimit(Addr apeBase) const
{
return (apeBase & 0xFFFFFF0000000000UL) | 0xFFFFFFFFFFL;
}
Addr
GPUComputeDriver::scratchApeBase(int gpuNum) const
{
return ((Addr)gpuNum << 61) + 0x100000000L;
}
// Used for GFX9 devices
// From drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c in the Linux kernel
Addr
GPUComputeDriver::scratchApeBaseV9() const
{
return ((Addr)0x1 << 48);
}
Addr
GPUComputeDriver::scratchApeLimit(Addr apeBase) const
{
return (apeBase & 0xFFFFFFFF00000000UL) | 0xFFFFFFFF;
}
Addr
GPUComputeDriver::ldsApeBase(int gpuNum) const
{
return ((Addr)gpuNum << 61) + 0x0;
}
//Used for GFX9 devices
// From drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c in the Linux kernel
Addr
GPUComputeDriver::ldsApeBaseV9() const
{
return ((Addr)0x2 << 48);
}
Addr
GPUComputeDriver::ldsApeLimit(Addr apeBase) const
{
return (apeBase & 0xFFFFFFFF00000000UL) | 0xFFFFFFFF;
}
void
GPUComputeDriver::allocateGpuVma(Request::CacheCoherenceFlags mtype,
Addr start, Addr length)
{
AddrRange range = AddrRange(start, start + length);
DPRINTF(GPUDriver, "Registering [%p - %p] with MTYPE %d\n",
range.start(), range.end(), mtype);
fatal_if(gpuVmas.insert(range, mtype) == gpuVmas.end(),
"Attempted to double register Mtypes for [%p - %p]\n",
range.start(), range.end());
}
Addr
GPUComputeDriver::deallocateGpuVma(Addr start)
{
auto vma = gpuVmas.contains(start);
assert(vma != gpuVmas.end());
assert((vma->first.start() == start));
Addr size = vma->first.size();
DPRINTF(GPUDriver, "Unregistering [%p - %p]\n", vma->first.start(),
vma->first.end());
gpuVmas.erase(vma);
return size;
}
void
GPUComputeDriver::setMtype(RequestPtr req)
{
// If we are a dGPU then set the MTYPE from our VMAs.
if (isdGPU) {
assert(!FullSystem);
AddrRange range = RangeSize(req->getVaddr(), req->getSize());
auto vma = gpuVmas.contains(range);
assert(vma != gpuVmas.end());
DPRINTF(GPUShader, "Setting req from [%p - %p] MTYPE %d\n"
"%d\n", range.start(), range.end(), vma->second);
req->setCacheCoherenceFlags(vma->second);
// APUs always get the default MTYPE
} else {
req->setCacheCoherenceFlags(defaultMtype);
}
}
} // namespace gem5
Reference in New Issue View Git Blame Copy Permalink