gpu_command_processor.hh

Go to the documentation of this file.

/*
 * Copyright (c) 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.
 */
 
#ifndef __DEV_HSA_GPU_COMMAND_PROCESSOR_HH__
#define __DEV_HSA_GPU_COMMAND_PROCESSOR_HH__
 
#include <cstdint>
#include <functional>
 
#include "arch/amdgpu/vega/gpu_registers.hh"
#include "base/logging.hh"
#include "base/trace.hh"
#include "base/types.hh"
#include "debug/GPUCommandProc.hh"
#include "dev/dma_virt_device.hh"
#include "dev/hsa/hsa_packet_processor.hh"
#include "dev/hsa/hsa_signal.hh"
#include "gpu-compute/dispatcher.hh"
#include "gpu-compute/gpu_compute_driver.hh"
#include "gpu-compute/hsa_queue_entry.hh"
#include "params/GPUCommandProcessor.hh"
#include "sim/full_system.hh"
 
namespace gem5
{
 
struct GPUCommandProcessorParams;
class GPUComputeDriver;
class GPUDispatcher;
class Shader;
 
class GPUCommandProcessor : public DmaVirtDevice
{
  public:
    typedef GPUCommandProcessorParams Params;
    typedef std::function<void(const uint64_t &)> HsaSignalCallbackFunction;
 
    GPUCommandProcessor() = delete;
    GPUCommandProcessor(const Params &p);
 
    HSAPacketProcessor& hsaPacketProc();
    RequestorID vramRequestorId();
 
    void setGPUDevice(AMDGPUDevice *gpu_device);
    void setShader(Shader *shader);
    Shader* shader();
    GPUComputeDriver* driver();
 
    struct KernelDispatchData
    {
        AMDKernelCode *akc;
        void *raw_pkt;
        uint32_t queue_id;
        Addr host_pkt_addr;
        PacketPtr readPkt;
    };
    struct KernelDispatchData {…};
 
    std::list<struct KernelDispatchData> kernelDispatchList;
 
    enum AgentCmd
    {
      Nop = 0,
      Steal = 1
    };
    enum AgentCmd {…};
 
    void performTimingRead(PacketPtr pkt);
 
    void completeTimingRead();
 
    void submitAgentDispatchPkt(void *raw_pkt, uint32_t queue_id,
                           Addr host_pkt_addr);
    void submitDispatchPkt(void *raw_pkt, uint32_t queue_id,
                           Addr host_pkt_addr);
    void submitVendorPkt(void *raw_pkt, uint32_t queue_id,
                         Addr host_pkt_addr);
    void attachDriver(GPUComputeDriver *driver);
 
    void dispatchKernelObject(AMDKernelCode *akc, void *raw_pkt,
                              uint32_t queue_id, Addr host_pkt_addr);
    void dispatchPkt(HSAQueueEntry *task);
    void signalWakeupEvent(uint32_t event_id);
 
    Tick write(PacketPtr pkt) override { return 0; }
    Tick read(PacketPtr pkt) override { return 0; }
    AddrRangeList getAddrRanges() const override;
    System *system();
 
    void sendCompletionSignal(Addr signal_handle);
    void updateHsaSignal(Addr signal_handle, uint64_t signal_value,
                         HsaSignalCallbackFunction function =
                            [] (const uint64_t &) { });
    void updateHsaSignalAsync(Addr signal_handle, int64_t diff);
    void updateHsaSignalData(Addr value_addr, int64_t diff,
                             uint64_t *prev_value);
    void updateHsaSignalDone(uint64_t *signal_value);
    void updateHsaMailboxData(Addr signal_handle, uint64_t *mailbox_value);
    void updateHsaEventData(Addr signal_handle, uint64_t *event_value);
    void updateHsaEventTs(Addr signal_handle, amd_event_t *event_value);
 
    uint64_t functionalReadHsaSignal(Addr signal_handle);
 
    Addr getHsaSignalValueAddr(Addr signal_handle)
    {
        return signal_handle + offsetof(amd_signal_t, value);
    }
    Addr getHsaSignalValueAddr(Addr signal_handle) {…}
 
    Addr getHsaSignalMailboxAddr(Addr signal_handle)
    {
        return signal_handle + offsetof(amd_signal_t, event_mailbox_ptr);
    }
    Addr getHsaSignalMailboxAddr(Addr signal_handle) {…}
 
    Addr getHsaSignalEventAddr(Addr signal_handle)
    {
        return signal_handle + offsetof(amd_signal_t, event_id);
    }
    Addr getHsaSignalEventAddr(Addr signal_handle) {…}
 
  private:
    Shader *_shader;
    GPUDispatcher &dispatcher;
    GPUComputeDriver *_driver;
    AMDGPUDevice *gpuDevice;
    VegaISA::Walker *walker;
 
    // Typedefing dmaRead and dmaWrite function pointer
    typedef void (DmaDevice::*DmaFnPtr)(Addr, int, Event*, uint8_t*, Tick);
    void initABI(HSAQueueEntry *task);
    void sanityCheckAKC(AMDKernelCode *akc);
    HSAPacketProcessor *hsaPP;
    TranslationGenPtr translate(Addr vaddr, Addr size) override;
 
    // Running counter of dispatched tasks
    int dynamic_task_id = 0;
 
    // Running counter of dispatched user (non-blit) kernels
    int non_blit_kernel_id = 0;
 
    // Skip all user (non-blit) kernels until reaching this kernel
    int target_non_blit_kernel_id = 0;
 
    // Keep track of start times for task dispatches.
    std::unordered_map<Addr, Tick> dispatchStartTime;
 
    void
    ReadDispIdOffsetDmaEvent(HSAQueueEntry *task,
                             const uint32_t &readDispIdOffset)
    {
        task->hostAMDQueueAddr = hsaPP->getQueueDesc(
            task->queueId())->hostReadIndexPtr - readDispIdOffset;
 
        auto *mqdDmaEvent = new DmaVirtCallback<int>(
            [ = ] (const int &) { MQDDmaEvent(task); });
 
        dmaReadVirt(task->hostAMDQueueAddr,
                    sizeof(_amd_queue_t), mqdDmaEvent, &task->amdQueue);
    }
    ReadDispIdOffsetDmaEvent(HSAQueueEntry *task, {…}
 
     void
     MQDDmaEvent(HSAQueueEntry *task)
     {
        if ((task->privMemPerItem() * VegaISA::NumVecElemPerVecReg) >
            task->amdQueue.compute_tmpring_size_wavesize * 1024) {
            // TODO: Raising this signal will potentially nuke scratch
            // space for in-flight kernels that were launched from this
            // queue.  We need to drain all kernels and deschedule the
            // queue before raising this signal. For now, just assert if
            // there are any in-flight kernels and tell the user that this
            // feature still needs to be implemented.
            fatal_if(hsaPP->inFlightPkts(task->queueId()) > 1,
                        "Needed more scratch, but kernels are in flight for "
                        "this queue and it is unsafe to reallocate scratch. "
                        "We need to implement additional intelligence in the "
                        "hardware scheduling logic to support CP-driven "
                        "queue draining and scheduling.");
            DPRINTF(GPUCommandProc, "Not enough scratch space to launch "
                    "kernel (%x available, %x requested bytes per "
                    "workitem). Asking host runtime to allocate more "
                    "space.\n",
                    task->amdQueue.compute_tmpring_size_wavesize * 1024,
                    task->privMemPerItem());
 
            updateHsaSignal(task->amdQueue.queue_inactive_signal.handle, 1,
                            [ = ] (const uint64_t &dma_buffer)
                                { WaitScratchDmaEvent(task, dma_buffer); });
 
        } else {
            DPRINTF(GPUCommandProc, "Sufficient scratch space, launching "
                    "kernel (%x available, %x requested bytes per "
                    "workitem).\n",
                    task->amdQueue.compute_tmpring_size_wavesize * 1024,
                    task->privMemPerItem());
            dispatchPkt(task);
        }
    }
     MQDDmaEvent(HSAQueueEntry *task) {…}
 
    void
    WaitScratchDmaEvent(HSAQueueEntry *task, const uint64_t &dmaBuffer)
    {
        if (dmaBuffer == 0) {
            DPRINTF(GPUCommandProc, "Host scratch allocation complete. "
                    "Attempting to re-read MQD\n");
            auto cb = new DmaVirtCallback<int>(
                [ = ] (const int &) { MQDDmaEvent(task); });
 
            dmaReadVirt(task->hostAMDQueueAddr, sizeof(_amd_queue_t), cb,
                        &task->amdQueue);
        } else {
            Addr value_addr = getHsaSignalValueAddr(
                task->amdQueue.queue_inactive_signal.handle);
            DPRINTF(GPUCommandProc, "Polling queue inactive signal at "
                    "%p.\n", value_addr);
            auto cb = new DmaVirtCallback<uint64_t>(
                [ = ] (const uint64_t &dma_buffer)
                { WaitScratchDmaEvent(task, dma_buffer); } );
 
            dmaReadVirt(value_addr, sizeof(Addr), cb, &cb->dmaBuffer, 1e9);
        }
    }
    WaitScratchDmaEvent(HSAQueueEntry *task, const uint64_t &dmaBuffer) {…}
};
class GPUCommandProcessor : public DmaVirtDevice {…};
 
} // namespace gem5
 
#endif // __DEV_HSA_GPU_COMMAND_PROCESSOR_HH__