Separate HW commands from class declaration and interface

Change-Id: Ia49098171f4b7814c42a35686354713a322c9df7

Author: zzdanowicz

runtime/command_queue/gpgpu_walker.h

@@ -40,6 +40,8 @@
 
 namespace OCLRT {
 
+using WALKER_HANDLE = void *;
+
 constexpr int32_t NUM_ALU_INST_FOR_READ_MODIFY_WRITE = 4;
 
 constexpr int32_t L3SQC_BIT_LQSC_RO_PERF_DIS = 0x08000000;
@@ -135,7 +137,7 @@ class GpgpuWalkerHelper {
     static size_t getSizeForWADisableLSQCROPERFforOCL(const Kernel *pKernel);
 
     static size_t setGpgpuWalkerThreadData(
-        typename GfxFamily::GPGPU_WALKER *pCmd,
+        WALKER_HANDLE pCmdData,
         const size_t globalOffsets[3],
         const size_t startWorkGroups[3],
         const size_t numWorkGroups[3],

runtime/command_queue/gpgpu_walker.inl

@@ -116,13 +116,14 @@ void GpgpuWalkerHelper<GfxFamily>::addAluReadModifyWriteRegister(
 
 template <typename GfxFamily>
 inline size_t GpgpuWalkerHelper<GfxFamily>::setGpgpuWalkerThreadData(
-    typename GfxFamily::GPGPU_WALKER *pCmd,
+    WALKER_HANDLE pCmdData,
     const size_t globalOffsets[3],
     const size_t startWorkGroups[3],
     const size_t numWorkGroups[3],
     const size_t localWorkSizesIn[3],
     uint32_t simd) {
-    typedef typename GfxFamily::GPGPU_WALKER GPGPU_WALKER;
+    using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
+    GPGPU_WALKER *pCmd = static_cast<GPGPU_WALKER *>(pCmdData);
 
     auto localWorkSize = localWorkSizesIn[0] * localWorkSizesIn[1] * localWorkSizesIn[2];
 

runtime/device_queue/device_queue_hw.h

@@ -34,10 +34,7 @@ namespace OCLRT {
 template <typename GfxFamily>
 class DeviceQueueHw : public DeviceQueue {
     using BaseClass = DeviceQueue;
-    using MEDIA_STATE_FLUSH = typename GfxFamily::MEDIA_STATE_FLUSH;
-    using MEDIA_INTERFACE_DESCRIPTOR_LOAD = typename GfxFamily::MEDIA_INTERFACE_DESCRIPTOR_LOAD;
     using PIPE_CONTROL = typename GfxFamily::PIPE_CONTROL;
-    using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
     using MI_BATCH_BUFFER_START = typename GfxFamily::MI_BATCH_BUFFER_START;
     using MI_BATCH_BUFFER_END = typename GfxFamily::MI_BATCH_BUFFER_END;
     using INTERFACE_DESCRIPTOR_DATA = typename GfxFamily::INTERFACE_DESCRIPTOR_DATA;
@@ -48,7 +45,6 @@ class DeviceQueueHw : public DeviceQueue {
     using MI_LOAD_REGISTER_IMM = typename GfxFamily::MI_LOAD_REGISTER_IMM;
     using MI_MATH = typename GfxFamily::MI_MATH;
     using MI_MATH_ALU_INST_INLINE = typename GfxFamily::MI_MATH_ALU_INST_INLINE;
-    using MEDIA_VFE_STATE = typename GfxFamily::MEDIA_VFE_STATE;
 
   public:
     DeviceQueueHw(Context *context,

runtime/device_queue/device_queue_hw.inl

@@ -103,6 +103,10 @@ void DeviceQueueHw<GfxFamily>::resetDeviceQueue() {
 
 template <typename GfxFamily>
 size_t DeviceQueueHw<GfxFamily>::getMinimumSlbSize() {
+    using MEDIA_STATE_FLUSH = typename GfxFamily::MEDIA_STATE_FLUSH;
+    using MEDIA_INTERFACE_DESCRIPTOR_LOAD = typename GfxFamily::MEDIA_INTERFACE_DESCRIPTOR_LOAD;
+    using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
+
     return sizeof(MEDIA_STATE_FLUSH) +
            sizeof(MEDIA_INTERFACE_DESCRIPTOR_LOAD) +
            sizeof(PIPE_CONTROL) +
@@ -124,6 +128,10 @@ void DeviceQueueHw<GfxFamily>::initPipeControl(PIPE_CONTROL *pc) {
 
 template <typename GfxFamily>
 void DeviceQueueHw<GfxFamily>::buildSlbDummyCommands() {
+    using MEDIA_STATE_FLUSH = typename GfxFamily::MEDIA_STATE_FLUSH;
+    using MEDIA_INTERFACE_DESCRIPTOR_LOAD = typename GfxFamily::MEDIA_INTERFACE_DESCRIPTOR_LOAD;
+    using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
+
     auto igilCmdQueue = reinterpret_cast<IGIL_CommandQueue *>(queueBuffer->getUnderlyingBuffer());
     auto slbEndOffset = igilCmdQueue->m_controls.m_SLBENDoffsetInBytes;
     size_t commandsSize = getMinimumSlbSize() + getWaCommandsSize();
@@ -291,6 +299,7 @@ IndirectHeap *DeviceQueueHw<GfxFamily>::getIndirectHeap(IndirectHeap::Type type)
 
 template <typename GfxFamily>
 void DeviceQueueHw<GfxFamily>::setupIndirectState(IndirectHeap &surfaceStateHeap, IndirectHeap &dynamicStateHeap, Kernel *parentKernel, uint32_t parentIDCount) {
+    using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
     void *pDSH = dynamicStateHeap.getCpuBase();
 
     // Set scheduler ID to last entry in first table, it will have ID == 0, blocks will have following entries.
@@ -426,6 +435,7 @@ void DeviceQueueHw<GfxFamily>::addMediaStateClearCmds() {
 
 template <typename GfxFamily>
 size_t DeviceQueueHw<GfxFamily>::getMediaStateClearCmdsSize() {
+    using MEDIA_VFE_STATE = typename GfxFamily::MEDIA_VFE_STATE;
     // PC with GenreicMediaStateClear + WA PC
     size_t size = 2 * sizeof(PIPE_CONTROL);