Program hardware context in L0 command queue

Related-To: NEO-4577

Change-Id: I204a5e86ad3b23b71071bbbfd58c23a408f6865f
Signed-off-by: Mateusz Hoppe <[email protected]>

Author: HoppeMateusz

level_zero/core/source/cmdqueue/cmdqueue_hw.h

@@ -22,6 +22,7 @@ namespace L0 {
 template <GFXCORE_FAMILY gfxCoreFamily>
 struct CommandQueueHw : public CommandQueueImp {
     using CommandQueueImp::CommandQueueImp;
+    using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
 
     ze_result_t createFence(const ze_fence_desc_t *desc, ze_fence_handle_t *phFence) override;
     ze_result_t destroy() override;

level_zero/core/source/cmdqueue/cmdqueue_hw.inl

@@ -10,6 +10,7 @@
 #include "shared/source/built_ins/built_ins.h"
 #include "shared/source/built_ins/sip.h"
 #include "shared/source/command_container/command_encoder.h"
+#include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/linear_stream.h"
 #include "shared/source/command_stream/preemption.h"
 #include "shared/source/command_stream/thread_arbitration_policy.h"
@@ -126,6 +127,7 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(
     }
 
     size_t linearStreamSizeEstimate = totalCmdBuffers * sizeof(MI_BATCH_BUFFER_START);
+    linearStreamSizeEstimate += csr->getCmdsSizeForHardwareContext();
 
     if (directSubmissionEnabled) {
         linearStreamSizeEstimate += sizeof(MI_BATCH_BUFFER_START);
@@ -175,6 +177,13 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(
     reserveLinearStreamSize(alignedSize);
     NEO::LinearStream child(commandStream->getSpace(alignedSize), alignedSize);
 
+    const auto globalFenceAllocation = csr->getGlobalFenceAllocation();
+    if (globalFenceAllocation) {
+        residencyContainer.push_back(globalFenceAllocation);
+    }
+
+    csr->programHardwareContext(child);
+
     if (!isCopyOnlyCommandQueue) {
         if (!gpgpuEnabled) {
             programPipelineSelect(child);

level_zero/core/source/gen11/icllp/cmdqueue_icllp.cpp

@@ -15,7 +15,7 @@
 #include "igfxfmid.h"
 
 namespace L0 {
-
+template struct CommandQueueHw<IGFX_GEN11_CORE>;
 static CommandQueuePopulateFactory<IGFX_ICELAKE_LP, CommandQueueHw<IGFX_GEN11_CORE>> populateICLLP;
 
 } // namespace L0

level_zero/core/source/gen12lp/tgllp/cmdqueue_tgllp.cpp

@@ -15,7 +15,7 @@
 #include "igfxfmid.h"
 
 namespace L0 {
-
+template struct CommandQueueHw<IGFX_GEN12LP_CORE>;
 static CommandQueuePopulateFactory<IGFX_TIGERLAKE_LP, CommandQueueHw<IGFX_GEN12LP_CORE>>
     populateTGLLP;
 

level_zero/core/source/gen9/cfl/cmdqueue_cfl.cpp

@@ -15,7 +15,7 @@
 #include "igfxfmid.h"
 
 namespace L0 {
-
+template struct CommandQueueHw<IGFX_GEN9_CORE>;
 static CommandQueuePopulateFactory<IGFX_COFFEELAKE, CommandQueueHw<IGFX_GEN9_CORE>> populateCFL;
 
 } // namespace L0

level_zero/core/source/gen9/glk/cmdqueue_glk.cpp

@@ -15,7 +15,7 @@
 #include "igfxfmid.h"
 
 namespace L0 {
-
+template struct CommandQueueHw<IGFX_GEN9_CORE>;
 static CommandQueuePopulateFactory<IGFX_GEMINILAKE, CommandQueueHw<IGFX_GEN9_CORE>> populateGLK;
 
 } // namespace L0

level_zero/core/source/gen9/kbl/cmdqueue_kbl.cpp

@@ -15,7 +15,7 @@
 #include "igfxfmid.h"
 
 namespace L0 {
-
+template struct CommandQueueHw<IGFX_GEN9_CORE>;
 static CommandQueuePopulateFactory<IGFX_KABYLAKE, CommandQueueHw<IGFX_GEN9_CORE>> populateKBL;
 
 } // namespace L0

level_zero/core/source/gen9/skl/cmdqueue_skl.cpp

@@ -15,7 +15,7 @@
 #include "igfxfmid.h"
 
 namespace L0 {
-
+template struct CommandQueueHw<IGFX_GEN9_CORE>;
 static CommandQueuePopulateFactory<IGFX_SKYLAKE, CommandQueueHw<IGFX_GEN9_CORE>> populateSKL;
 
 } // namespace L0

level_zero/core/test/unit_tests/sources/cmdqueue/test_cmdqueue.cpp

@@ -8,8 +8,8 @@
 #include "shared/source/helpers/state_base_address.h"
 #include "shared/test/unit_test/helpers/debug_manager_state_restore.h"
 #include "shared/test/unit_test/helpers/default_hw_info.h"
+#include "shared/test/unit_test/mocks/mock_command_stream_receiver.h"
 
-#include "opencl/test/unit_test/mocks/mock_graphics_allocation.h"
 #include "test.h"
 
 #include "level_zero/core/source/driver/driver_handle_imp.h"
@@ -217,5 +217,45 @@ HWTEST2_F(CommandQueueDestroy, whenCommandQueueDestroyIsCalledPrintPrintfOutputI
     EXPECT_EQ(1u, kernel.printPrintfOutputCalledTimes);
 }
 
+using CommandQueueCommands = Test<DeviceFixture>;
+HWTEST_F(CommandQueueCommands, givenCommandQueueWhenExecutingCommandListsThenHardwareContextIsProgrammedAndGlobalAllocationResident) {
+    const ze_command_queue_desc_t desc = {
+        ZE_COMMAND_QUEUE_DESC_VERSION_CURRENT,
+        ZE_COMMAND_QUEUE_FLAG_NONE,
+        ZE_COMMAND_QUEUE_MODE_DEFAULT,
+        ZE_COMMAND_QUEUE_PRIORITY_NORMAL,
+        0};
+
+    MockCsrHw2<FamilyType> csr(*neoDevice->getExecutionEnvironment(), 0);
+    csr.initializeTagAllocation();
+    csr.setupContext(*neoDevice->getDefaultEngine().osContext);
+
+    L0::CommandQueue *commandQueue = CommandQueue::create(productFamily,
+                                                          device,
+                                                          &csr,
+                                                          &desc,
+                                                          true);
+    ASSERT_NE(nullptr, commandQueue);
+
+    std::unique_ptr<L0::CommandList> commandList(CommandList::create(productFamily, device, true));
+    auto commandListHandle = commandList->toHandle();
+    auto status = commandQueue->executeCommandLists(1, &commandListHandle, nullptr, false);
+
+    auto globalFence = csr.getGlobalFenceAllocation();
+    if (globalFence) {
+        bool found = false;
+        for (auto alloc : csr.copyOfAllocations) {
+            if (alloc == globalFence) {
+                found = true;
+                break;
+            }
+        }
+        EXPECT_TRUE(found);
+    }
+    EXPECT_EQ(status, ZE_RESULT_SUCCESS);
+    EXPECT_TRUE(csr.programHardwareContextCalled);
+    commandQueue->destroy();
+}
+
 } // namespace ult
 } // namespace L0

opencl/test/unit_test/command_stream/command_stream_receiver_flush_task_1_tests.cpp

@@ -8,11 +8,11 @@
 #include "shared/source/gmm_helper/gmm_helper.h"
 #include "shared/source/os_interface/os_context.h"
 #include "shared/test/unit_test/helpers/debug_manager_state_restore.h"
+#include "shared/test/unit_test/helpers/dispatch_flags_helper.h"
 #include "shared/test/unit_test/helpers/ult_hw_helper.h"
 
 #include "opencl/source/command_queue/gpgpu_walker.h"
 #include "opencl/test/unit_test/fixtures/ult_command_stream_receiver_fixture.h"
-#include "opencl/test/unit_test/helpers/dispatch_flags_helper.h"
 #include "opencl/test/unit_test/helpers/unit_test_helper.h"
 #include "opencl/test/unit_test/mocks/mock_buffer.h"
 #include "opencl/test/unit_test/mocks/mock_command_queue.h"