gpu-compute: Add methods to read GPU memory requestor ID

These methods are called from various places to override the requestor
ID of a request in order to determine which Ruby network a request
should be routed on.

Change-Id: Ic0270ddd7123f0457a13144e69ef9132204d4334
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/57651
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>

src/gpu-compute/compute_unit.cc

@@ -2009,6 +2009,15 @@ ComputeUnit::sendToLds(GPUDynInstPtr gpuDynInst)
     return ldsPort.sendTimingReq(newPacket);
 }
 
+/**
+ * Forward the VRAM requestor ID needed for device memory from shader.
+ */
+RequestorID
+ComputeUnit::vramRequestorId()
+{
+    return FullSystem ? shader->vramRequestorId() : requestorId();
+}
+
 /**
  * get the result of packets sent to the LDS when they return
  */

src/gpu-compute/compute_unit.hh

@@ -458,6 +458,7 @@ class ComputeUnit : public ClockedObject
     void updatePageDivergenceDist(Addr addr);
 
     RequestorID requestorId() { return _requestorId; }
+    RequestorID vramRequestorId();
 
     bool isDone() const;
     bool isVectorAluIdle(uint32_t simdId) const;

src/gpu-compute/gpu_command_processor.cc

@@ -65,6 +65,15 @@ GPUCommandProcessor::hsaPacketProc()
     return *hsaPP;
 }
 
+/**
+ * Forward the VRAM requestor ID needed for device memory from GPU device.
+ */
+RequestorID
+GPUCommandProcessor::vramRequestorId()
+{
+    return gpuDevice->vramRequestorId();
+}
+
 TranslationGenPtr
 GPUCommandProcessor::translate(Addr vaddr, Addr size)
 {

src/gpu-compute/gpu_command_processor.hh

@@ -76,6 +76,7 @@ class GPUCommandProcessor : public DmaVirtDevice
     GPUCommandProcessor(const Params &p);
 
     HSAPacketProcessor& hsaPacketProc();
+    RequestorID vramRequestorId();
 
     void setGPUDevice(AMDGPUDevice *gpu_device);
     void setShader(Shader *shader);

src/gpu-compute/shader.cc

@@ -523,6 +523,15 @@ Shader::notifyCuSleep() {
         stats.shaderActiveTicks += curTick() - _lastInactiveTick;
 }
 
+/**
+ * Forward the VRAM requestor ID needed for device memory from CP.
+ */
+RequestorID
+Shader::vramRequestorId()
+{
+    return gpuCmdProc.vramRequestorId();
+}
+
 Shader::ShaderStats::ShaderStats(statistics::Group *parent, int wf_size)
     : statistics::Group(parent),
       ADD_STAT(allLatencyDist, "delay distribution for all"),

src/gpu-compute/shader.hh

@@ -183,6 +183,8 @@ class Shader : public ClockedObject
         shHiddenPrivateBaseVmid = sh_hidden_base_new;
     }
 
+    RequestorID vramRequestorId();
+
     EventFunctionWrapper tickEvent;
 
     // is this simulation going to be timing mode in the memory?