Revert "Allow cpu buffers copy on DG2 in 32bit"

This reverts commit 2055daf.

Signed-off-by: Maciej Plewka <[email protected]>

Author: MaciejPlewka

opencl/source/mem_obj/buffer.cpp

@@ -643,7 +643,7 @@ bool Buffer::isReadWriteOnCpuAllowed(const Device &device) {
         return false;
     }
 
-    if (graphicsAllocation->isAllocatedInLocalMemoryPool() && !graphicsAllocation->isAllocationLockable()) {
+    if (graphicsAllocation->isAllocatedInLocalMemoryPool()) {
         return false;
     }
 

opencl/test/unit_test/command_queue/read_write_buffer_cpu_copy.cpp

@@ -9,7 +9,6 @@
 #include "shared/source/helpers/basic_math.h"
 #include "shared/source/helpers/local_memory_access_modes.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"
-#include "shared/test/common/mocks/mock_gmm.h"
 #include "shared/test/common/test_macros/test.h"
 
 #include "opencl/test/unit_test/command_queue/enqueue_read_buffer_fixture.h"
@@ -332,10 +331,7 @@ TEST(ReadWriteBufferOnCpu, whenLocalMemoryPoolAllocationIsAskedForPreferenceThen
     cl_int retVal = 0;
     std::unique_ptr<Buffer> buffer(Buffer::create(&ctx, CL_MEM_READ_WRITE, MemoryConstants::pageSize, nullptr, retVal));
     ASSERT_NE(nullptr, buffer.get());
-    auto gmm = std::make_unique<MockGmm>(device->getGmmHelper());
-    reinterpret_cast<MemoryAllocation *>(buffer->getGraphicsAllocation(device->getRootDeviceIndex()))->setGmm(gmm.release(), 0);
     reinterpret_cast<MemoryAllocation *>(buffer->getGraphicsAllocation(device->getRootDeviceIndex()))->overrideMemoryPool(MemoryPool::LocalMemory);
-    reinterpret_cast<MemoryAllocation *>(buffer->getGraphicsAllocation(device->getRootDeviceIndex()))->getDefaultGmm()->resourceParams.Flags.Info.NotLockable = 1;
 
     EXPECT_FALSE(buffer->isReadWriteOnCpuAllowed(device->getDevice()));
     EXPECT_FALSE(buffer->isReadWriteOnCpuPreferred(reinterpret_cast<void *>(0x1000), MemoryConstants::pageSize, device->getDevice()));

opencl/test/unit_test/mem_obj/buffer_tests.cpp

@@ -68,9 +68,7 @@ TEST(Buffer, whenBufferAllocatedInLocalMemoryThenCpuCopyIsDisallowed) {
     MockBuffer buffer(allocation);
     UltDeviceFactory factory{1, 0};
     auto &device = *factory.rootDevices[0];
-    auto gmm = std::make_unique<MockGmm>(device.getGmmHelper());
-    allocation.setGmm(gmm.get(), 0);
-    allocation.getDefaultGmm()->resourceParams.Flags.Info.NotLockable = 1;
+
     allocation.memoryPool = MemoryPool::LocalMemory;
     EXPECT_FALSE(buffer.isReadWriteOnCpuAllowed(device));
 

opencl/test/unit_test/xe_hpg_core/dg2/windows/CMakeLists.txt

@@ -6,7 +6,6 @@
 
 set(IGDRCL_SRCS_tests_xe_hpg_core_dg2_windows
     ${CMAKE_CURRENT_SOURCE_DIR}/CMakeLists.txt
-    ${CMAKE_CURRENT_SOURCE_DIR}/wddm_tests_dg2.cpp
 )
 if(WIN32)
   target_sources(igdrcl_tests PRIVATE ${IGDRCL_SRCS_tests_xe_hpg_core_dg2_windows})

opencl/test/unit_test/xe_hpg_core/dg2/windows/wddm_tests_dg2.cpp

@@ -38,11 +38,5 @@ void HwHelperHw<Family>::setExtraAllocationData(AllocationData &allocationData,
             allocationData.storageInfo.isLockable = false;
         }
     }
-
-    if (HwInfoConfig::get(hwInfo.platform.eProductFamily)->isStorageInfoAdjustmentRequired()) {
-        if (properties.allocationType == AllocationType::BUFFER && !properties.flags.preferCompressed && !properties.flags.shareable) {
-            allocationData.storageInfo.isLockable = true;
-        }
-    }
 }
 } // namespace NEO

shared/source/helpers/extra_allocation_data_xehp_and_later.inl

@@ -860,16 +860,19 @@ bool WddmMemoryManager::isCpuCopyRequired(const void *ptr) {
     // function checks what is the delta between reading from cachead memory
     // compare to reading from provided pointer
     // if value is above threshold, it means that pointer is uncached.
-    constexpr auto slownessFactor = 50u;
-    static int64_t meassurmentOverhead = std::numeric_limits<int64_t>::max();
-    static int64_t fastestLocalRead = std::numeric_limits<int64_t>::max();
+    constexpr auto slownessFactor = 5u;
+    static volatile int64_t meassurmentOverhead = std::numeric_limits<int64_t>::max();
+    static volatile int64_t fastestLocalRead = std::numeric_limits<int64_t>::max();
+    static volatile int64_t max = std::numeric_limits<int64_t>::min();
+    static volatile int64_t min = std::numeric_limits<int64_t>::max();
 
     // local variable that we will read for comparison
-    int cacheable = 1;
+    volatile int cacheable = 1;
+    // auto hostPtr = std::make_unique<volatile int[]>(0x13000 / sizeof(int));
     volatile int *localVariablePointer = &cacheable;
-    volatile const int *volatileInputPtr = static_cast<volatile const int *>(ptr);
+    volatile int *volatileInputPtr = (volatile int *)(ptr);
 
-    int64_t timestamp0, timestamp1, localVariableReadDelta, inputPointerReadDelta;
+    volatile int64_t timestamp0, timestamp1, localVariableReadDelta, inputPointerReadDelta;
 
     // compute timing overhead
     _mm_lfence();
@@ -883,7 +886,7 @@ bool WddmMemoryManager::isCpuCopyRequired(const void *ptr) {
     }
 
     // dummy read
-    cacheable = *localVariablePointer;
+    //cacheable = *localVariablePointer;
 
     _mm_lfence();
     timestamp0 = __rdtsc();
@@ -901,7 +904,7 @@ bool WddmMemoryManager::isCpuCopyRequired(const void *ptr) {
         fastestLocalRead = localVariableReadDelta;
     }
     // dummy read
-    cacheable = *volatileInputPtr;
+    //cacheable = *volatileInputPtr;
 
     _mm_lfence();
     timestamp0 = __rdtsc();
@@ -914,6 +917,9 @@ bool WddmMemoryManager::isCpuCopyRequired(const void *ptr) {
     if (inputPointerReadDelta <= 0) {
         inputPointerReadDelta = 1;
     }
+    max = std::max(max, localVariableReadDelta);
+    min = std::min(min, inputPointerReadDelta);
+    // UNRECOVERABLE_IF(max > min);
     return inputPointerReadDelta > slownessFactor * fastestLocalRead;
 }
 

shared/source/os_interface/windows/wddm_memory_manager.cpp

@@ -185,11 +185,3 @@ template <>
 bool HwInfoConfigHw<gfxProduct>::isMatrixMultiplyAccumulateSupported(const HardwareInfo &hwInfo) const {
     return true;
 }
-template <>
-bool HwInfoConfigHw<gfxProduct>::isStorageInfoAdjustmentRequired() const {
-    if constexpr (is32bit) {
-        return true;
-    } else {
-        return false;
-    }
-}

shared/source/xe_hpg_core/dg2/os_agnostic_hw_info_config_dg2.inl

@@ -17,4 +17,3 @@ HWTEST_EXCLUDE_PRODUCT(HwInfoConfigTest, givenHwInfoConfigWhenAskedIfPipeControl
 HWTEST_EXCLUDE_PRODUCT(HwInfoConfigTest, givenHwInfoConfigWhenAskedIfTile64With3DSurfaceOnBCSIsSupportedThenTrueIsReturned, IGFX_DG2);
 HWTEST_EXCLUDE_PRODUCT(HwInfoConfigTest, WhenAllowRenderCompressionIsCalledThenTrueIsReturned, IGFX_DG2);
 HWTEST_EXCLUDE_PRODUCT(HwInfoConfigTest, whenConvertingTimestampsToCsDomainThenNothingIsChanged, IGFX_DG2);
-HWTEST_EXCLUDE_PRODUCT(HwInfoConfigTest, givenHwInfoConfigWhenAskedIfStorageInfoAdjustmentIsRequiredThenFalseIsReturned, IGFX_DG2);

shared/test/unit_test/xe_hpg_core/dg2/excludes_xe_hpg_core_dg2.cpp

@@ -450,12 +450,3 @@ DG2TEST_F(ProductConfigTests, givenDg2G11DeviceIdWhenDifferentRevisionIsPassedTh
         EXPECT_EQ(productConfig, DG2_G11);
     }
 }
-
-DG2TEST_F(HwInfoConfigTestDg2, givenHwInfoConfigWhenAskedIfStorageInfoAdjustmentIsRequiredThenTrueIsReturned) {
-    auto hwInfoConfig = HwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
-    if constexpr (is32bit) {
-        EXPECT_TRUE(hwInfoConfig->isStorageInfoAdjustmentRequired());
-    } else {
-        EXPECT_FALSE(hwInfoConfig->isStorageInfoAdjustmentRequired());
-    }
-}