intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-19 06:24:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

Change interface to pass Interface Descriptor Data as pointer

Browse Source 
Change-Id: I0f33109b800a7607206954bb1e5cb0826290e6f3
This commit is contained in:
Zdanowicz, Zbigniew
2018-05-11 13:33:16 +02:00
committed by sys_ocldev
parent 2e06df06ff
commit 33fee15711
5 changed files with 31 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
runtime/command_queue/gpgpu_walker.inl
View File

@@ -595,7 +595,8 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchWalker(
localWorkSizes,
offsetInterfaceDescriptorTable,
interfaceDescriptorIndex,
preemptionMode);
preemptionMode,
nullptr);
if (&dispatchInfo == &*multiDispatchInfo.begin()) {
// If hwTimeStampAlloc is passed (not nullptr), then we know that profiling is enabled
@@ -761,7 +762,8 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
localWorkSizes,
offsetInterfaceDescriptorTable,
interfaceDescriptorIndex,
preemptionMode);
preemptionMode,
nullptr);
// Implement enabling special WA DisableLSQCROPERFforOCL if needed
GpgpuWalkerHelper<GfxFamily>::applyWADisableLSQCROPERFforOCL(commandStream, scheduler, true);

11
runtime/helpers/kernel_commands.h
View File

@@ -41,8 +41,9 @@ struct MultiDispatchInfo;
template <typename GfxFamily>
struct KernelCommandsHelper : public PerThreadDataHelper {
typedef typename GfxFamily::BINDING_TABLE_STATE BINDING_TABLE_STATE;
typedef typename GfxFamily::RENDER_SURFACE_STATE RENDER_SURFACE_STATE;
using BINDING_TABLE_STATE = typename GfxFamily::BINDING_TABLE_STATE;
using RENDER_SURFACE_STATE = typename GfxFamily::RENDER_SURFACE_STATE;
using INTERFACE_DESCRIPTOR_DATA = typename GfxFamily::INTERFACE_DESCRIPTOR_DATA;
static uint32_t computeSlmValues(uint32_t valueIn);
@@ -58,7 +59,8 @@ struct KernelCommandsHelper : public PerThreadDataHelper {
uint32_t threadsPerThreadGroup,
uint32_t sizeSlm,
bool barrierEnable,
PreemptionMode preemptionMode);
PreemptionMode preemptionMode,
INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor);
static void sendMediaStateFlush(
LinearStream &commandStream,
@@ -100,7 +102,8 @@ struct KernelCommandsHelper : public PerThreadDataHelper {
const size_t localWorkSize[3],
const uint64_t offsetInterfaceDescriptorTable,
const uint32_t interfaceDescriptorIndex,
PreemptionMode preemptionMode);
PreemptionMode preemptionMode,
INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor);
static size_t getSizeRequiredCS();
static bool isPipeControlWArequired();

17
runtime/helpers/kernel_commands.inl
View File

@@ -144,9 +144,9 @@ size_t KernelCommandsHelper<GfxFamily>::sendInterfaceDescriptorData(
uint32_t threadsPerThreadGroup,
uint32_t sizeSlm,
bool barrierEnable,
PreemptionMode preemptionMode) {
typedef typename GfxFamily::SAMPLER_STATE SAMPLER_STATE;
typedef typename GfxFamily::INTERFACE_DESCRIPTOR_DATA INTERFACE_DESCRIPTOR_DATA;
PreemptionMode preemptionMode,
INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor) {
using SAMPLER_STATE = typename GfxFamily::SAMPLER_STATE;
// Allocate some memory for the interface descriptor
auto pInterfaceDescriptor = static_cast<INTERFACE_DESCRIPTOR_DATA *>(ptrOffset(indirectHeap.getCpuBase(), (size_t)offsetInterfaceDescriptor));
@@ -305,11 +305,9 @@ size_t KernelCommandsHelper<GfxFamily>::sendIndirectState(
const size_t localWorkSize[3],
const uint64_t offsetInterfaceDescriptorTable,
const uint32_t interfaceDescriptorIndex,
PreemptionMode preemptionMode) {
typedef typename GfxFamily::INTERFACE_DESCRIPTOR_DATA INTERFACE_DESCRIPTOR_DATA;
typedef typename GfxFamily::RENDER_SURFACE_STATE RENDER_SURFACE_STATE;
typedef typename GfxFamily::SAMPLER_STATE SAMPLER_STATE;
PreemptionMode preemptionMode,
INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor) {
using SAMPLER_STATE = typename GfxFamily::SAMPLER_STATE;
DEBUG_BREAK_IF(simd != 8 && simd != 16 && simd != 32);
@@ -395,7 +393,8 @@ size_t KernelCommandsHelper<GfxFamily>::sendIndirectState(
threadsPerThreadGroup,
kernel.slmTotalSize,
!!patchInfo.executionEnvironment->HasBarriers,
preemptionMode);
preemptionMode,
inlineInterfaceDescriptor);
if (DebugManager.flags.AddPatchInfoCommentsForAUBDump.get()) {
PatchInfoData patchInfoData(kernelStartOffset, 0, PatchInfoAllocationType::InstructionHeap, dsh.getGraphicsAllocation()->getGpuAddress(), offsetInterfaceDescriptor, PatchInfoAllocationType::DynamicStateHeap);

14
unit_tests/helpers/kernel_commands_tests.cpp
View File

@@ -103,7 +103,7 @@ HWTEST_F(KernelCommandsTest, programInterfaceDescriptorDataResourceUsage) {
size_t crossThreadDataSize = kernel->getCrossThreadDataSize();
KernelCommandsHelper<FamilyType>::sendInterfaceDescriptorData(
indirectHeap, 0, 0, crossThreadDataSize, 64, 0, 0, 0, 1, 0 * KB, false, pDevice->getPreemptionMode());
indirectHeap, 0, 0, crossThreadDataSize, 64, 0, 0, 0, 1, 0 * KB, false, pDevice->getPreemptionMode(), nullptr);
auto usedIndirectHeapAfter = indirectHeap.getUsed();
EXPECT_EQ(sizeof(INTERFACE_DESCRIPTOR_DATA), usedIndirectHeapAfter - usedIndirectHeapBefore);
@@ -327,7 +327,8 @@ HWTEST_F(KernelCommandsTest, sendIndirectStateResourceUsage) {
localWorkSizes,
IDToffset,
0,
pDevice->getPreemptionMode());
pDevice->getPreemptionMode(),
nullptr);
// It's okay these are EXPECT_GE as they're only going to be used for
// estimation purposes to avoid OOM.
@@ -404,7 +405,8 @@ HWTEST_F(KernelCommandsTest, usedBindingTableStatePointer) {
localWorkSizes,
0,
0,
pDevice->getPreemptionMode());
pDevice->getPreemptionMode(),
nullptr);
EXPECT_EQ(0x00000000u, *(&bindingTableStatesPointers[0]));
EXPECT_EQ(0x00000040u, *(&bindingTableStatesPointers[1]));
@@ -556,7 +558,8 @@ HWTEST_F(KernelCommandsTest, usedBindingTableStatePointersForGlobalAndConstantAn
localWorkSizes,
0,
0,
pDevice->getPreemptionMode());
pDevice->getPreemptionMode(),
nullptr);
bti = reinterpret_cast<typename FamilyType::BINDING_TABLE_STATE *>(reinterpret_cast<unsigned char *>(ssh.getCpuBase()) + localSshOffset + btiOffset);
for (uint32_t i = 0; i < numSurfaces; ++i) {
@@ -792,7 +795,8 @@ HWTEST_F(KernelCommandsTest, GivenKernelWithSamplersWhenIndirectStateIsProgramme
localWorkSizes,
interfaceDescriptorTableOffset,
0,
pDevice->getPreemptionMode());
pDevice->getPreemptionMode(),
nullptr);
bool isMemorySame = memcmp(borderColorPointer, mockDsh, borderColorSize) == 0;
EXPECT_TRUE(isMemorySame);

5
unit_tests/kernel/kernel_slm_tests.cpp
View File

@@ -109,8 +109,9 @@ HWTEST_P(KernelSLMAndBarrierTest, test_SLMProgramming) {
0,
1,
kernelInfo.workloadInfo.slmStaticSize,
!!executionEnvironment.HasBarriers,
pDevice->getPreemptionMode()); // Barriers Enabled
!!executionEnvironment.HasBarriers, // Barriers Enabled
pDevice->getPreemptionMode(),
nullptr);
// add the heap base + offset
uint32_t *pIdData = (uint32_t *)indirectHeap.getCpuBase() + offsetInterfaceDescriptorData;