Merge branch 'hotfix/v01.02.00.02'
IPU1 eary boot needs to set EVE clock to OPP High
- Previously, IPU1 code will read DSP OPP mode first, then program
EVE frequency accordingly. If DSP is OPP High, then EVE is put
into OPP High, otherwise, OPP Normal.
- Now IPU1 gets early boot, before DSP frequency is programmed,
which happens at the end of u-boot, just before starting Linux kernel.
- Need to put EVE frequency to OPP High
- MCT-1234
- Previously, IPU1 code will read DSP OPP mode first, then program
EVE frequency accordingly. If DSP is OPP High, then EVE is put
into OPP High, otherwise, OPP Normal.
- Now IPU1 gets early boot, before DSP frequency is programmed,
which happens at the end of u-boot, just before starting Linux kernel.
- Need to put EVE frequency to OPP High
- MCT-1234
Start hotfix v01.02.00.02
Merge branch 'hotfix/v01.02.00.01'
Fix IPU1 l3_noc message during remoteproc
- Happens in CoreSDK filesystem
- Recommended fix is to configure and use watchdog timers in IPU1 firmware
- Fix is tested and confirmed by CoreSDK team
- MCT-1232
- Happens in CoreSDK filesystem
- Recommended fix is to configure and use watchdog timers in IPU1 firmware
- Fix is tested and confirmed by CoreSDK team
- MCT-1232
Start hotfix v01.02.00.01
Merge branch 'release/v01.02.00'
Update manifest for OpenCL v01.02.00
- SRAS00006792
- SRAS00006792
Add v01.02.00.x into release notes
- Clean up documentation build warnings
- Clean up documentation build warnings
Fix for clGetKernelArgInfo incorrect behavior for CL_KERNEL_ARG_TYPE_QUALIFIER
According to clarification from Khronos in a previous pull request:
https://github.com/KhronosGroup/OpenCL-CTS/pull/141, the
CL_KERNEL_ARG_TYPE_QUALIFIER field for const values must always
return CL_KERNEL_ARG_TYPE_NONE even if the const keyword is specified
for the argument.
Return CL_KERNEL_ARG_TYPE_CONST only for pointers and not const values.
(MCT-1230)
According to clarification from Khronos in a previous pull request:
https://github.com/KhronosGroup/OpenCL-CTS/pull/141, the
CL_KERNEL_ARG_TYPE_QUALIFIER field for const values must always
return CL_KERNEL_ARG_TYPE_NONE even if the const keyword is specified
for the argument.
Return CL_KERNEL_ARG_TYPE_CONST only for pointers and not const values.
(MCT-1230)
Handle CL_MEM_HOST_* flags during sub buffer creation
(MCT-1229)
(MCT-1229)
Prevent IPU from sleeping before attaching to EVEs
- MCT-1228
- MCT-1228
OpenCL 1.2 separation of program compilation and linking
- Implemention of clCompileProgram() and clLinkProgram() APIs
- Added support to program loadBinary for compiled objects and libraries
- Added two examples
- vecadd_compile_link: Demonstrate compile and link separation
- vecadd_compile_link_loadbinary: Demonstrate loadBinary with
compile and link separation
- Added support for clProgramBuildInfo attribute CL_PROGRAM_BINARY_TYPE
- Use of OpenCL 1.2 version number, new readme for release, update
example documentation
(MCT-890, MCT-891, MCT-892)
- Implemention of clCompileProgram() and clLinkProgram() APIs
- Added support to program loadBinary for compiled objects and libraries
- Added two examples
- vecadd_compile_link: Demonstrate compile and link separation
- vecadd_compile_link_loadbinary: Demonstrate loadBinary with
compile and link separation
- Added support for clProgramBuildInfo attribute CL_PROGRAM_BINARY_TYPE
- Use of OpenCL 1.2 version number, new readme for release, update
example documentation
(MCT-890, MCT-891, MCT-892)
Fix carveout block allocation in resource table
- Carveout blocks in the resource table could be allocated in two different
ways. One is CMA (with reserved memory marked as "reusable"), which will
use 1MB alignment. The other is DMA pool (with reserved memory marked as
"no-map"), which will round up block size and alignment to power of 2.
In DMA pool case, if not careful, carveout blocks could end up not able
to be allocated in the available memory, while CMA succeeds.
- This commit addresses the possibe carveout block allocation failure
in DMA pool case. In particular,
- AM57 DSP: 6MB carveout block is split into 4MB and 2MB.
- Without this fix, DMA pool allocation actually will fail on DSP2
due to its smaller 8MB reserved memory, with its first 1MB being
allocated to VRING/IPC.
- AM57 IPU1: slight carveout block size adjustment.
- Take this opportunity to make code/data blocks tighter, so that
unused memory can be returned to Linux in "reusable" allocation.
- MCT-1225, LCPD-17021, LCPD-17331
- Carveout blocks in the resource table could be allocated in two different
ways. One is CMA (with reserved memory marked as "reusable"), which will
use 1MB alignment. The other is DMA pool (with reserved memory marked as
"no-map"), which will round up block size and alignment to power of 2.
In DMA pool case, if not careful, carveout blocks could end up not able
to be allocated in the available memory, while CMA succeeds.
- This commit addresses the possibe carveout block allocation failure
in DMA pool case. In particular,
- AM57 DSP: 6MB carveout block is split into 4MB and 2MB.
- Without this fix, DMA pool allocation actually will fail on DSP2
due to its smaller 8MB reserved memory, with its first 1MB being
allocated to VRING/IPC.
- AM57 IPU1: slight carveout block size adjustment.
- Take this opportunity to make code/data blocks tighter, so that
unused memory can be returned to Linux in "reusable" allocation.
- MCT-1225, LCPD-17021, LCPD-17331
Merge tag 'v01.01.19.02' into develop
OpenCL 1.1.19.02 for PSDK 6.1
OpenCL 1.1.19.02 for PSDK 6.1
Merge branch 'hotfix/v01.01.19.02'
Start hotfix v01.01.19.02
- to use updated TIDL library build
- to use updated TIDL library build
Merge tag 'v01.01.19.01' into develop
OpenCL v01.01.19.01 for Processor SDK 6.1
OpenCL v01.01.19.01 for Processor SDK 6.1
Merge branch 'hotfix/v01.01.19.01'
OpenCL v01.01.19.01 for Processor SDK 6.1
OpenCL v01.01.19.01 for Processor SDK 6.1
OpenCL 1.1.19.01 for new TIDL library
- new TIDL library with MobileNetV2 support
- PLSDK-2956
- new TIDL library with MobileNetV2 support
- PLSDK-2956
Fix platforms example for AM57 RTOS and K2x
- C++ Context throw exception if queried device type does not exist.
Check if devices exist first before getting devices.
- Fix AM57 RTOS compilation warnings
- MCT-1218, MCT-1220
- C++ Context throw exception if queried device type does not exist.
Check if devices exist first before getting devices.
- Fix AM57 RTOS compilation warnings
- MCT-1218, MCT-1220
Fix platforms example for AM57 RTOS and K2x
- C++ Context throw exception if queried device type does not exist.
Check if devices exist first before getting devices.
- Fix AM57 RTOS compilation warnings
- MCT-1218, MCT-1220
- C++ Context throw exception if queried device type does not exist.
Check if devices exist first before getting devices.
- Fix AM57 RTOS compilation warnings
- MCT-1218, MCT-1220
Merge tag 'v01.01.19.00' into develop
OpenCL 1.1.19.00 for Processor 6.1
OpenCL 1.1.19.00 for Processor 6.1
Merge branch 'release/v01.01.19.00'
OpenCL 1.1.19.00 for Processor SDK 6.1
OpenCL 1.1.19.00 for Processor SDK 6.1
Update platform examples to query all devices
- including the CUSTOM devices
- Update TIDL library build to yocto build
- MCT-1220
- including the CUSTOM devices
- Update TIDL library build to yocto build
- MCT-1220
Fix warnings in persistent examples
(MCT-1218)
(MCT-1218)
Add fix for CL_DEVICE_TYPE_ALL in Platform getDevices
(MCT-1220)
(MCT-1220)
Bugfix for platform getDevices and readme doc update
- When querying for the number of DSP devices from the platform, EVE
devices were also being included in the count
- Fix logic to not include EVE/CUSTOM devices when querying for
ACCELERATOR|CUSTOM devices
- Update readme document index
(MCT-1220)
- When querying for the number of DSP devices from the platform, EVE
devices were also being included in the count
- Fix logic to not include EVE/CUSTOM devices when querying for
ACCELERATOR|CUSTOM devices
- Update readme document index
(MCT-1220)
Update manifest for version 1.1.19.00
SRAS#:SRAS00006550
SRAS#:SRAS00006550
Build against the latest TIDL library
- MCT-1217
- MCT-1217
Update build using the latest PSDK 6.1 nightly
Set release version to v01.01.19 and OpenCL version to 1.1
Remove race condition in AM57 monitor makefiles
- The obj directory creation during the monitor build process was
subject to a potential race condition since the dependency chain did
not guarantee it.
- Fix the dependency chain such that the obj directory is always
available before an obj file is built
(MCT-1219)
- The obj directory creation during the monitor build process was
subject to a potential race condition since the dependency chain did
not guarantee it.
- Fix the dependency chain such that the obj directory is always
available before an obj file is built
(MCT-1219)
Remove compile time warnings for OpenCL examples
- Use -Wall when compiling OpenCL examples
- Fix issues leading to compiler time warnings for examples
(MCT-1218)
- Use -Wall when compiling OpenCL examples
- Fix issues leading to compiler time warnings for examples
(MCT-1218)
OpenCL 1.2 printf with support for vector data types
- Add support for disabling [core %d] addition to printf messages
as per OpenCL 1.2 specification using a new environment variable,
TI_OCL_PRINTF_COREID
- Switch format identifier flag _MFLL to _MFL in _printfi.c to handle
cl_long and cl_ulong vectors.
- Fixed printf +inf to inf
- Fixed printf #X case from 0x0 to 0
- Fix alignment for longlong2 and double2 vector data types
- Use modified va_arg_128bit for double2 and longlong2
- Update llvm-3.6 arm build to use version created with gcc-arm-8.3.0
- Update debug_printf doc to reflect new features
- Added _ltoa function definition since it is not available for
cgt versions 8.3.2 and lower.
(MCT-888)
- Add support for disabling [core %d] addition to printf messages
as per OpenCL 1.2 specification using a new environment variable,
TI_OCL_PRINTF_COREID
- Switch format identifier flag _MFLL to _MFL in _printfi.c to handle
cl_long and cl_ulong vectors.
- Fixed printf +inf to inf
- Fixed printf #X case from 0x0 to 0
- Fix alignment for longlong2 and double2 vector data types
- Use modified va_arg_128bit for double2 and longlong2
- Update llvm-3.6 arm build to use version created with gcc-arm-8.3.0
- Update debug_printf doc to reflect new features
- Added _ltoa function definition since it is not available for
cgt versions 8.3.2 and lower.
(MCT-888)
Update build using the latest PSDK 6.0.0.7 release
Revert "Move EVE runtime memory into IPU1 DTS reserved mem"
This reverts commit 90aa292476f4b9153db6dc6742446876ae06718b.
- Moving EVE firmware runtime memory out of CMEM to IPU memory
does NOT save any memory, because we still cannot get rid of
first 32MB reserved CMEM due to OpenMP requirement.
Revert the memory movement, put EVE firmware runtime memory
back in first 32MB of reserved CMEM.
- MCT-1141
This reverts commit 90aa292476f4b9153db6dc6742446876ae06718b.
- Moving EVE firmware runtime memory out of CMEM to IPU memory
does NOT save any memory, because we still cannot get rid of
first 32MB reserved CMEM due to OpenMP requirement.
Revert the memory movement, put EVE firmware runtime memory
back in first 32MB of reserved CMEM.
- MCT-1141
Added missing bandwidth.h to the edmabw example
(MCT-1190)
(MCT-1190)
Collect dataQ/minValue/maxValue for TIDL trace
- MCT-1201
- MCT-1201
Fix g++ 8.3.0 OpenCL example undefined behavior
- Fix undefined behavior of non-void functions with no return statements,
when using g++ 8.3.0. Change them to void functions.
- MCT-1200
- Fix undefined behavior of non-void functions with no return statements,
when using g++ 8.3.0. Change them to void functions.
- MCT-1200
Fix DSP L2 cache size update builtin functions
- Previous L2 cache size update functions used
- signed integer to represent cache size delta before and after,
- unsigned integer to represent scratch size.
When comparing these two numbers, everything is casted to
unsigned integers for comparison according to C/C++ rules.
However, the intended behavior is to perform signed comparison.
The bug caused these builtin functions not changing L2 cache size.
- Change everything to signed integers for signed comparison.
- Reduce L2 cache size to 64KB in TIDL environment on AM57,
to provide enough L2 scratch sapce for TIDL library use.
- MCT-1198
- Previous L2 cache size update functions used
- signed integer to represent cache size delta before and after,
- unsigned integer to represent scratch size.
When comparing these two numbers, everything is casted to
unsigned integers for comparison according to C/C++ rules.
However, the intended behavior is to perform signed comparison.
The bug caused these builtin functions not changing L2 cache size.
- Change everything to signed integers for signed comparison.
- Reduce L2 cache size to 64KB in TIDL environment on AM57,
to provide enough L2 scratch sapce for TIDL library use.
- MCT-1198
Fix core id search for received message
- Search whole array instead of first k entries, where k is
the number of available compute units, because array is
now indexed by physical core id (used to be logical core id).
- MCT-1195
- Search whole array instead of first k entries, where k is
the number of available compute units, because array is
now indexed by physical core id (used to be logical core id).
- MCT-1195
Bugfix for clEnqueueBarrierWithWaitList and clEnqueueMarkerWithWaitList
- The Event::Barrier and Event::Marker types needed to be checked
alongside Event::WaitForEvents when deciding whether to push
further events on a command queue. This is because both
MarkerEvent and BarrierEvent classes inherit from WaitForEventsEvent
(MCT-1175)
- The Event::Barrier and Event::Marker types needed to be checked
alongside Event::WaitForEvents when deciding whether to push
further events on a command queue. This is because both
MarkerEvent and BarrierEvent classes inherit from WaitForEventsEvent
(MCT-1175)
Bugfix for clSetEventCallback()
- In OpenCL 1.2 clSetEventCallback() can accept CL_COMPLETE,
CL_SUBMITTED and CL_RUNNING as command_exec_callback_type
- Set Event::Running status when an Event is dispatched
(MCT-1174, MCT-1176)
- In OpenCL 1.2 clSetEventCallback() can accept CL_COMPLETE,
CL_SUBMITTED and CL_RUNNING as command_exec_callback_type
- Set Event::Running status when an Event is dispatched
(MCT-1174, MCT-1176)
Merge branch 'release/v01.01.18.00'
Update documentation for v01.01.18.00
- Change default CORESDK_VERSION and TIDL_VERSION
- Changes needed for TIDL-API v01.03.00
- MCT-1135
- Change default CORESDK_VERSION and TIDL_VERSION
- Changes needed for TIDL-API v01.03.00
- MCT-1135
Replace __DATE__ and __TIME__ with git SHA
Replace _DATE_ and _TIME_ with the SHA of the ti-opencl git repo used
for the build. This impacts:
1. clocl version reporting with -v
E.g. 'OpenCL 1.1 TI product version 01.01.18.00 (e7069f4)'
2. OpenCL runtime CL_PLATFORM_VERSION query.
E.g. 'OpenCL 1.1 TI product version 01.01.18.00 (e7069f4)'
The git SHA is a more useful indicator of the source code baseline used
to build OpenCL.
(MCT-1119)
Replace _DATE_ and _TIME_ with the SHA of the ti-opencl git repo used
for the build. This impacts:
1. clocl version reporting with -v
E.g. 'OpenCL 1.1 TI product version 01.01.18.00 (e7069f4)'
2. OpenCL runtime CL_PLATFORM_VERSION query.
E.g. 'OpenCL 1.1 TI product version 01.01.18.00 (e7069f4)'
The git SHA is a more useful indicator of the source code baseline used
to build OpenCL.
(MCT-1119)
Merge tag 'v01.01.17.02' into develop
For CoreSDK 2019 gcc8 build
For CoreSDK 2019 gcc8 build
Merge branch 'hotfix/v01.01.17.02'
Minor fix for g++-arm-8.2.1 compilation
- MCT-1112
- MCT-1112
Create hotfix v01.01.17.02 branch
Minor fix for g++-arm-8.2.1 compilation
- MCT-1112
- MCT-1112
Fix potential double free in HeapsMultiThreadedPolicy
Delete the copy constructor and operator= in HeapsMultiThreadedPolicy
to avoid potential double freeing of freed memory warning from
static analyzers.
(MCT-1111)
Delete the copy constructor and operator= in HeapsMultiThreadedPolicy
to avoid potential double freeing of freed memory warning from
static analyzers.
(MCT-1111)
Migrate build to Ubuntu 18.04/g++-7
- Fix broken g++-5 build on old Ubuntu
- MCT-1110
- Fix broken g++-5 build on old Ubuntu
- MCT-1110
Migrate to ubuntu18.04/g++-7
- MCT-1110
- MCT-1110
Fix core_scheduler bug for sub_devices
- Core_scheduler should allocate compute unit from the compute units
being passed in.
- MCT-1107
- Core_scheduler should allocate compute unit from the compute units
being passed in.
- MCT-1107
Create __dsp_speed() builtin function from existing dsp_speed() in the monitor
- Add edmabw example demonstrating use of __dsp_speed()
- Add doc entry for edmabw example and vecadd_subdevice example
(MCT-1190)
- Add edmabw example demonstrating use of __dsp_speed()
- Add doc entry for edmabw example and vecadd_subdevice example
(MCT-1190)
Eliminate dynamic_cast(s) related to DSPDevice
(MCT-1183)
(MCT-1183)
Eliminate duplication across EVE & DSP worker.cpp
(MCT-1192)
(MCT-1192)
Disable custom device for OpenCL RTOS
- Not supported at this moment
- Fix clRetainDevice reference of EVEDevice and CPUDevice
- Fix SYSBIOS CCS debug script
- Sync development build tools versions with the latest PSDK build
- MCT-1154
- Not supported at this moment
- Fix clRetainDevice reference of EVEDevice and CPUDevice
- Fix SYSBIOS CCS debug script
- Sync development build tools versions with the latest PSDK build
- MCT-1154
Removed unused code from the runtime
(MCT-1189)
(MCT-1189)
Handle EVE and CPU devices in the clRetainDevice and clReleaseDevice functions
(MCT-1187)
(MCT-1187)
Fix compilation warnings for OpenCL examples
- Enable use of deprecated 1.1 OpenCL APIs without warnings
- Fix use of Buffer constructor
(MCT-1180)
- Enable use of deprecated 1.1 OpenCL APIs without warnings
- Fix use of Buffer constructor
(MCT-1180)
Fix RTOS dsplib_fft compiling error with new BIOS
- Add a missing header file
- MCT-1186
- Add a missing header file
- MCT-1186
Addition of preprocessor macros and version update
- Add __OPENCL_C_VERSION__
- Update __OPENCL_VERSION__ to CL_VERSION_1_2
- Update version in dsp/eve device
(MCT-1181)
- Add __OPENCL_C_VERSION__
- Update __OPENCL_VERSION__ to CL_VERSION_1_2
- Update version in dsp/eve device
(MCT-1181)
Bugfix for incorrect return value in clRetainDevice() and clReleaseDevice()
(MCT-1182)
(MCT-1182)
Update OpenCL version to 1.2.0
(MCT-1179)
(MCT-1179)
Support for new compute info flags
- CL_DEVICE_IMAGE_MAX_ARRAY_SIZE
- CL_DEVICE_IMAGE_MAX_BUFFER_SIZE
- CL_DEVICE_LINKER_AVAILABLE
- CL_DEVICE_PREFERRED_INTEROP_USER_SYNC
- CL_DEVICE_PRINTF_BUFFER_SIZE
(MCT-1177)
- CL_DEVICE_IMAGE_MAX_ARRAY_SIZE
- CL_DEVICE_IMAGE_MAX_BUFFER_SIZE
- CL_DEVICE_LINKER_AVAILABLE
- CL_DEVICE_PREFERRED_INTEROP_USER_SYNC
- CL_DEVICE_PRINTF_BUFFER_SIZE
(MCT-1177)
Support for new CL_MEM_HOST_* buffer creation flags
- CL_MEM_HOST_WRITE_ONLY
- CL_MEM_HOST_READ_ONLY
- CL_MEM_HOST_NO_ACCESS
(MCT-1173)
- CL_MEM_HOST_WRITE_ONLY
- CL_MEM_HOST_READ_ONLY
- CL_MEM_HOST_NO_ACCESS
(MCT-1173)
clEnqueueFillImage() implementation
- Returns CL_IMAGE_FORMAT_NOT_SUPPORTED
(MCT-1133)
- Returns CL_IMAGE_FORMAT_NOT_SUPPORTED
(MCT-1133)
clGetKernelArgInfo(), clCreateImage() implementation and
clGetProgramInfo() new parameters
- clGetKernelArgInfo() implementation including addition of
"-cl-kernel-arg-info" clocl option
- clCreateImage() implementation which returns CL_INVALID_OPERATION
- Implementation of CL_PROGRAM_NUM_KERNELS and CL_PROGRAM_KERNEL_NAMES
for clGetProgramInfo()
(MCT-1140)
clGetProgramInfo() new parameters
- clGetKernelArgInfo() implementation including addition of
"-cl-kernel-arg-info" clocl option
- clCreateImage() implementation which returns CL_INVALID_OPERATION
- Implementation of CL_PROGRAM_NUM_KERNELS and CL_PROGRAM_KERNEL_NAMES
for clGetProgramInfo()
(MCT-1140)
Implementation of CL_MAP_WRITE_INVALIDATE_REGION
(MCT-896)
(MCT-896)
clEnqueueBarrierWithWaitList implementation
(MCT-1139)
(MCT-1139)
clUnloadPlatformCompiler() implementation
(MCT-894)
(MCT-894)
clEnqueueMigrateMemObjects implementation
(MCT-901)
(MCT-901)
Update three examples to use make_kernel instead of the older KernelFunctor
- simple
- timeout
- abort_exit
(MCT-902)
- simple
- timeout
- abort_exit
(MCT-902)
Adding KernelFunctor support from OpenCL 1.1 cl.hpp header
- Usage through CL_USE_DEPRECATED_OPENCL_1_1_APIS_TI
defined in cl.h
(MCT-899, MCT-903)
- Usage through CL_USE_DEPRECATED_OPENCL_1_1_APIS_TI
defined in cl.h
(MCT-899, MCT-903)
OpenCL 1.2 Headers
(MCT-903, MCT-899)
(MCT-903, MCT-899)
Bugfix for clSetUserEventStatus()
- Input execution_status can also be a negative integer apart from
CL_COMPLETE.
(MCT-1170)
- Input execution_status can also be a negative integer apart from
CL_COMPLETE.
(MCT-1170)
Bugfix - buffer creation flags in vecadd_subdevice example
(MCT-1138)
(MCT-1138)
AM57: DSP custom device, TIDL-API Built-in Kernels
- Add DSP as custom device, in addition to being ACCELERATOR device
- Add TIDL-API Built-in Kernels to OpenCL DSP runtime
- Clean up subdevice/device dependent program handling,
move subdevice/rootdevice logic from api_program into Coal::Program
- MCT-1143, MCT-1154
- Add DSP as custom device, in addition to being ACCELERATOR device
- Add TIDL-API Built-in Kernels to OpenCL DSP runtime
- Clean up subdevice/device dependent program handling,
move subdevice/rootdevice logic from api_program into Coal::Program
- MCT-1143, MCT-1154
Merge tag 'v01.01.18.00' into develop
OpenCL 1.1.18.00 for Processor SDK 5.3
OpenCL 1.1.18.00 for Processor SDK 5.3
Move EVE runtime memory into IPU1 DTS reserved mem
- Part 1 of 2 to move OpenCL runtime memory out of CMEM
- Previously, OpenCL reserves first 32MB of CMEM for OpenCL runtime
(OpenMP DSP memory and EVE memory). This makes CMEM block hard to move.
- New approach for EVE is to allocate OpenCL runtime memory from IPU1's
DTS reserved memory using remoteproc and carveout entries in resource
table. This makes sense because EVE is loaded by IPU1.
- Remove hard coded CMEM information in the OpenCL runtime for EVE.
- MCT-948, MCT-1141
- Part 1 of 2 to move OpenCL runtime memory out of CMEM
- Previously, OpenCL reserves first 32MB of CMEM for OpenCL runtime
(OpenMP DSP memory and EVE memory). This makes CMEM block hard to move.
- New approach for EVE is to allocate OpenCL runtime memory from IPU1's
DTS reserved memory using remoteproc and carveout entries in resource
table. This makes sense because EVE is loaded by IPU1.
- Remove hard coded CMEM information in the OpenCL runtime for EVE.
- MCT-948, MCT-1141
Add support for ION based shared memory allocator
Update the shared memory allocator to use ION instead of CMEM. The
default allocator is CMEM. Set USE_ION=1 during build to switch to ION.
(MCT-1121)
Update the shared memory allocator to use ION instead of CMEM. The
default allocator is CMEM. Set USE_ION=1 during build to switch to ION.
(MCT-1121)
Refactor - remove dead code from the runtime
host/src/core/dsp/tal
* The pcie and devmem variants are no longer required.
- pcie is for C6678 PCIe device, no longer supported by the OpenCL
runtime.
- devmem is referenced only by mpm based mailbox. This can be removed
since we've switched to MessageQ based mailbox on AM57x and K2x.
host/init - Used by C6678, no longer supported
Removed references to BUILD_C6678 in sources, Makefile(s).
(MCT-1125)
host/src/core/dsp/tal
* The pcie and devmem variants are no longer required.
- pcie is for C6678 PCIe device, no longer supported by the OpenCL
runtime.
- devmem is referenced only by mpm based mailbox. This can be removed
since we've switched to MessageQ based mailbox on AM57x and K2x.
host/init - Used by C6678, no longer supported
Removed references to BUILD_C6678 in sources, Makefile(s).
(MCT-1125)
Merge in TargetAddressSpace
- Fix a bug in POCL: regenerate_kernel_metadata() needs to be called
once for all kernels (not one time for each kernel)
- Skip struct address space as it messes with __ocl_event type
- Merge in LLVMUtils, UVA
- Minor cleanups to minimize difference
- MCT-862
- Fix a bug in POCL: regenerate_kernel_metadata() needs to be called
once for all kernels (not one time for each kernel)
- Skip struct address space as it messes with __ocl_event type
- Merge in LLVMUtils, UVA
- Minor cleanups to minimize difference
- MCT-862
Merge in WorkitemLoops, WorkitemHandler
- Clean up _local_size_x,y,z that TI do not use
- WorkitemLoops: version LLVM user vs use that TI fixes
- MCT-862
- Clean up _local_size_x,y,z that TI do not use
- WorkitemLoops: version LLVM user vs use that TI fixes
- MCT-862
Merge in WorkitemHandlerChooser
- Partial VUA, TargetAddrSpace merge
- Simplify WorkItemAliasAnalysis header file to minimize changes
- MCT-862
- Partial VUA, TargetAddrSpace merge
- Simplify WorkItemAliasAnalysis header file to minimize changes
- MCT-862
Merge in files with POCL v1.2
- Mostly cosmetic, all TI modifications are guarded by TI_POCL
- Remove BarrierBlock class as in POCL v1.2
- Simplify Flatten class header file to minimize changes
- Merge in new Workgroup files from POCL1.2
- MCT-862
- Mostly cosmetic, all TI modifications are guarded by TI_POCL
- Remove BarrierBlock class as in POCL v1.2
- Simplify Flatten class header file to minimize changes
- Merge in new Workgroup files from POCL1.2
- MCT-862
Updated version to 1.1.19.0
Merge tag 'v01.01.17.01' into develop for PSDK 5.2
Merge branch 'hotfix/v01.01.17.01'
clEnqueueMarker incorrectly returns CL_OUT_OF_HOST_MEMORY
A test added into clEnqueueMarker incorrectly returns
CL_OUT_OF_HOST_MEMORY if there are no events in the command queue.
Modify the test to return CL_OUT_OF_HOST_MEMORY iff the event count in
the command queue > 0 and events array returned is null.
(MCT-1092)
A test added into clEnqueueMarker incorrectly returns
CL_OUT_OF_HOST_MEMORY if there are no events in the command queue.
Modify the test to return CL_OUT_OF_HOST_MEMORY iff the event count in
the command queue > 0 and events array returned is null.
(MCT-1092)
Created hotfix branch
Updated version to 1.1.18.0
Merge tag 'v01.01.17.00' into develop
Tag for PSDK 5.2 release
Tag for PSDK 5.2 release
Merge branch 'release/v01.01.17.00'
Documentation - update readme for 1.1.17
(MCT-1090)
(MCT-1090)
Update PSDK and TIDL library dependencies
(MCT-1070)
(MCT-1070)
Updated OpenCL manifest for 1.1.17
SRAS ID: SRAS00005715
SRAS ID: SRAS00005715