| 8f916d08dc104ecde5b888c2a8da76196b3f1e57 |
|
26-Jan-2018 |
Michael Butler <butlermichael@google.com> |
Elide NOP for fused operations
Test: mm
Test: cts & vts binaries
Change-Id: I89771b9afd79c86a49462f2e7f2a9a4cc07a775e
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| 62cf7fdcc03da02682d47f731094c2032b4fc14a |
|
22-Dec-2017 |
Michael Butler <butlermichael@google.com> |
Adjust Hexagon NN library's output quantized range
The Hexagon NN library chooses its own quantized ranges. To restore the
quantized range provided by the NN API, new operations are added to the
end of a hexagon nn graph to dequantize to float and quantize to set the
NN API's range.
Bug: 67478959
Test: mm
Test: runs on cts and vts binaries
Change-Id: I9cabf4c7c92335aaa0f04bd5e8d2133a6a53f001
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| dd779b6e6891811fb7dc317e72c50ff97e672f8c |
|
22-Dec-2017 |
Michael Butler <butlermichael@google.com> |
Disable floating point operations for HVX service
HVX is only performant when running on quantized values. Further, as an
optimization, the current HVX driver will convert some floating point
tensors into quantized values, perform the operation, and then convert
them back to floating point. This results in a loss in precision causing
some tests to fail.
Bug: 67478959
Test: mm
Test: vts and cts binaries run correctly using HVX
Change-Id: I0bcdc8fd3b0279d55b7992db98bab77262b0ce7e
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| 25d8e8714c120716b2982790818dad502693e2f4 |
|
19-Dec-2017 |
Michael Butler <butlermichael@google.com> |
Integrate with 12/07 drop
Bug: 67478959
Test: mm, test vts and cts binaries
Change-Id: I4245170cf8ff5bb3c017354dbce9930c3e7c8574
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| 9001cd5bf64977f6f61112a37052a03c87ee9550 |
|
08-Dec-2017 |
Michael Butler <butlermichael@google.com> |
Check result of HIDL call
Bug: 67310896
Test: mm, cts & vts binaries
Change-Id: Ib8df59b50b9bcbb63a1da061fc0efe296a846793
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| a6e68df59d64fceb115b91c69d472f0eebbe6400 |
|
04-Dec-2017 |
Michael Butler <butlermichael@google.com> |
Format Hexagon NN Service
Adds .clang-format file to project and formats all Hexagon NN
Service .cpp and .h files.
Bug: 69063955
Test: mm -j40 (to check formatting)
Change-Id: I2bf513e85f6f1c6feb93de029a1366d59dc9bb34
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| 2c6f06f42a661bf71a14a9bc465cc2be494ee329 |
|
02-Dec-2017 |
David Gross <dgross@google.com> |
Merge ""Canonicalize" class RunTimePoolInfo." into oc-mr1-nn-dev
|
| fa6d0ab09af2b9415ae2fe8251bb75aa7f5c19f9 |
|
01-Dec-2017 |
David Gross <dgross@google.com> |
"Canonicalize" class RunTimePoolInfo.
This class manages an important resource (an mmap() memory) region.
Rather than treating the class as a bundle of state with methods to
explicitly manipulate that resource, give it the full C++ concrete
data type treatment: Constructors, destructor, and move assignment, with
all data private. Now the lifetime of the managed resource matches the
lifetime of the class instance that manages the resource: The
resource is freed when the instance is destroyed or assigned to.
Test: nn/runtime/test
Bug: 69685100
Bug: 69929024
Change-Id: I2514ed437e92f4210d241883883a9bffe75f68a1
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| aff6eb8d0fae092ce5f694e1bfb252dd0a99ccbf |
|
28-Oct-2017 |
Michael Butler <butlermichael@google.com> |
Add functionality for boosting and slowing hexagon chip
Bug: 67478959
Test: mm, vts, cts
Change-Id: I1482a033d735d85c4459934856ba95510f2c1aca
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| 0f4cdac298cbf691540154154c223db02425b3d0 |
|
26-Oct-2017 |
Michael Butler <butlermichael@google.com> |
Integrate 10/15 hexagon nnlib drop
Additionally configured service to restart driver after the driver has
crashed.
Bug: 67478959
Test: mm, vts, cts
Change-Id: I4b95292e1fd4424676e049b08953ac00e4ab22f3
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| 7017487d7cc109bf952ce7c5cc8a9135f18c801c |
|
07-Oct-2017 |
Michael Butler <butlermichael@google.com> |
Added implicit padding to NN HVX service
Additionally fixed a bug with activation values and performed tighter
validation checks on padding.
Fixed issue with transpose.
Bug: 67478959
Test: mm, vts, nn/ml/runtime/tests
Change-Id: Ia5cb5ab90ec9f069fd8c62c13f54d20fe2fc5ab4
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| 20a5bcaf2e2776f4f4050a0de00ba9710a242019 |
|
06-Oct-2017 |
Michael Butler <butlermichael@google.com> |
Integrate 10/4 hexagon nnlib drop
Additionally removed debugging and allowed for truly concurrent
executions.
Bug: 67478959
Test: mm, vts
Change-Id: I7d7f94647eee4cc523095c57bc882ee2273a8844
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|
| 82d31136fb09dcd04542655553a9c0455b8f183d |
|
04-Oct-2017 |
Michael Butler <butlermichael@google.com> |
Update to the NeuralNetworks HVX service driver
Adds new operations with the latest drop and addresses the comments in
ag/2999193.
Also accounted for ag/3017213/ and added missing transpose functionality
Bug: 63905942
Bug: 64316595
Test: mm, vts
Change-Id: I12b6444cc44428b7dd2086ca13744e769ca3bbbd
/hardware/qcom/neuralnetworks/hvxservice/1.0/HexagonModel.cpp
|