Mauricio Briceno authored Oct 23, 2024 and Johan Alfvén committed Nov 14, 2024

- Add support for constant propagation via separate graphir pass set
- Add RewriteFunctions attribute to select traversal direction
- Constant propagation pass traversal direction is forward
- Fixed test/util CreateGraph to not include constant tensors as Graph
  IO
- This change introduces support for TOSA::LOGICAL_LEFT_SHIFT only

Change-Id: I7d63f3c6b11c715fc76ec37f79e85e6a75f6aa87
Signed-off-by: Mauricio Briceno <mauricio.briceno@arm.com>

 - A MEAN op with IFM rank two and reduce in C dimension caused an
output diff
 - The reason was that the intermediate tensor for calculating the sum
had the wrong shape because the reduceAxis shape had been padded to 4D
 - The fix is to use the original reduceAxis shape when calculating the
shape for the sum tensor

Change-Id: I144adfe07f697fecba6f7237e8b216295654f8ae
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

 - Fixed a problem when the final scheduler cost contained wrong info
about Weight buffer when it should not. This caused the tensor
allocation to fail due to over allocation
 - The reason is that after optimizing the cascades, the final call to
ProposeScheduleBuffering decided that it would not need any buffering.
 - However, due to some missing code the weight buffering cost was not
cleared so memory size went over the limit
 - Added code to clear the weight buffering cost when it is not needed

Change-Id: I92db5c6199c99ca1ae4a21b62dfd17f2cae6ac9b
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

Removes contribution from previously buffered weights for an op to the
buffering limit when proposing new weight buffering for the same
op.

Change-Id: If3ca6f6e359c29bc69f26c994503375cd353fba9
Signed-off-by: William Isaksson <william.isaksson@arm.com>

Mauricio Briceno authored Nov 09, 2024 and Johan Alfvén committed Nov 11, 2024

- This ensures produced artifacts are built with the same flags
- For example valgrind can now parse debug symbols properly

Change-Id: I798c0048890f546d2c2ec6a964ffa0f052de0840
Signed-off-by: Mauricio Briceno <mauricio.briceno@arm.com>

Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Change-Id: Ia61c19749958ea271385d5efcdc3682428303ab5

Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Change-Id: Ie1504e5eba860785cfdf9fd09f508bd9c81e9026

Alexander Bengtsson authored Oct 22, 2024 and Alexander Bengtsson committed Nov 11, 2024

- Support TOSA TransposeConv2D with kernel strides 1 or 2
  by adding support for output-padding.
- Simplify DecomposeTransposeConv2D and breakout TFLite-specific padding
  to the TFLite reader.
- Refactor InitializeSlice into TensorSlice.Initialize
- Strides larger than 2 will require further decomposition MLBEDSW-9761

Change-Id: I54143461a5bc677ca7eefd91b9005dbdc7b924ec
Signed-off-by: Alexander Bengtsson <Alexander.Bengtsson@arm.com>

 - Update ConvertToInterpolatingLUT16 to use float to match reference
and avoid rounding issues
 - Update Exp int8 and int16 to use float to match reference

Change-Id: Ia0b6912d665d243b2dab12b002c418d04e2e124d
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

Fixed Transpose attributes not always initialized for supported
Transpose operations.

Change-Id: I205d4920f8da16d9e48c07946e21c66d564e888e
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>

 - Update LUT table generation for RSQRT int16 to use
float to match reference and avoid rounding issues

Change-Id: I1d9af2a2f682dccadc918736f9e40a2e3e1c4986
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

 - Softmax int16 should use double precision in ElementwiseMulScale
in order to match reference and avoid rounding errors
 - Update ElementwiseMulScale to also support double, default
behavior is float

Change-Id: I207f95c273c66d99d994df2aa9cde144a01f80fe
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

The ReshapeOptions attribute is optional for RESHAPE operations.

Change-Id: I607282bdf3871dd364c96a762674378920268725
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>

 - Lowering of softmax can only be done as long as the product of
IFM width and height is within max allowed size for tensor dimension

Change-Id: Ifeacd3a8a8a9d7b1922a32b0475256d93a16fc8d
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

 - Update scale calculations to match reference for FULLY_CONNECTED
 - Refactor duplicate code and move RescalePerChannel to ethos_u_scaling

Change-Id: Ib0fc05fcbdcb124bbe5cba39c253db1f8f084c63
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

- Remove StartChaining() and replace with ClearChainingRegisters()

Change-Id: Ib2904ef6eb93f378699cab221e1c80542c46652d
Signed-off-by: Alexander Bengtsson <Alexander.Bengtsson@arm.com>

Added support for per channel scaling for Ethos-U55/65.

Change-Id: I9d1962ec6061b149abb5ccebaba53fec5cd37333
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>

 - Operators with float input or output is not supported and should be
set to OpType::Passthrough. This does not happen since the architecture
constraint class is missing the float constraint
 - Update constraint class to reject datatype float
 - Use passthrough options in writer since passthrough op's do not
have attribute data

Change-Id: I7f2a8e18155a1c4c41761f4e8ff01ee9c8761298
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

Bjorn Davidsson authored Nov 05, 2024 and Johan Alfvén committed Nov 06, 2024

Check array length before reading pads in
TFLiteGraphOptimiser::ConvertPad

Change-Id: I0237f3466eb038e4ab39ec4b1e5ee23d2a78eec0
Signed-off-by: Björn Davidsson <bjoern.davidsson@arm.com>

* Use DecomposeLargeAxis, but with untransposed OFM shape (ie. the
  IFM shape). Transpose the OFM offset and shape in
  DecomposeLargeAxis.
* Don't pad shapes in RearrangeTranspose because then they won't
  match the perm attribute.
* At the end of decomposition, there can now be multiple ops for
  each swap, so we must look for all ops using the first and last
  tensor when adjusting the tensor and quantization.

Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Change-Id: I3bf3579c247faef42ddd83ae09b803a4de35910d

Rename OpType::DepthwiseConv2DBias to OpType::DepthwiseConv2D.

Change-Id: I6ab44d50339933689d77586f551ed773b6f97102
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>

Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Change-Id: I430d3d6cbc973ea48cb0cf2601f48972815d424c

Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Change-Id: Ib887c2700d0be5d28875ad252d18a564464a4646

Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Change-Id: I72db4a56a31c6887623d687529b0628f83e98a20

Set depth multiplier in GraphBuilder.
Add depth multiplier decomposition in DecomposeDepthwiseConv2D.

Change-Id: Ic1408dadba6e9587bff725308d17b5778d184be4
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>

Add Ethos-U85 to the accelerator configuration documentation

Change-Id: Iad1bb5f6d183730f6c268f143530674c1625a01d
Signed-off-by: Rickard Bolin <rickard.bolin@arm.com>

William Isaksson authored Oct 22, 2024 and Fredrik Svedberg committed Oct 28, 2024

Currrent behaviour gives up on double buffering tensors if the pre-buffering slice happens to be the minimal one. This patch makes sure that we also try to reduce the bufferingDepth before we give up in such a situation.

Change-Id: I6f52ea7b74d18863668c533f0d37bd9b6d05deb8
Signed-off-by: William Isaksson <william.isaksson@arm.com>

Johan Gunnarsson authored Oct 21, 2024 and Rickard Bolin committed Oct 22, 2024

The consumers and producers are managed automatically by the
SchedulerOperation class.

Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Change-Id: I68ec801cbabbcf1dfc9e134a8f3339e7cad39118

 - Batched Conv2D with kernel 1x1 can be optimized in the same way
as FullyConnected
 - Add condition to RewriteFullyConnected to detect the Conv2D case
 - Moved condition to scheduler when to buffer for FullyConnect
 - Extend CanSubdivide with optype FullyConnected
 - Set AxisOrder in TOSA reader for Conv2D

Change-Id: I8ecc9c5bb8700a00e18d54cb8cc209ab70d01f45
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

Original rounding mode of operation was replaced in
GraphIrOptimiser::RewriteDepthwise.

Change-Id: Ifd3d0f37af802bddf2b730eb509a2def29159bf5
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>

Bjorn Davidsson authored Oct 18, 2024 and Fredrik Svedberg committed Oct 21, 2024

Add limited support for 4-bit data to the BufferView class,
and update the code for repacking 4-bit constant weight data
as 8-bit.

Change-Id: I015b059348868bf21272c99c7b425dcf3f2c1a1d
Signed-off-by: Björn Davidsson <bjoern.davidsson@arm.com>

Changed order of input and output rescale fusing since the extra
checks for input fusing on binary elementwise operations expects
fusing to be done in this order.

Change-Id: I5d796383de90a1477240bc45a94f5c2f71616530
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>

Johan Gunnarsson authored Oct 18, 2024 and Johan Alfvén committed Oct 21, 2024

* Add support for reshaping shapes that are used to represent
  offsets. They may have zeros in them, which were clamped to ones
  when reshaping.
* Extend unit tests.

Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Change-Id: Ib0a0796d36a7e6ff2d3cd1af97441dce07b53907

Alexander Bengtsson authored Oct 17, 2024 and Alexander Bengtsson committed Oct 18, 2024

- Handle ArgMax-specific datatype mappings by introducing EthosU85NpuOp::ArgMax.

Change-Id: I0de682a2e28d2d70a67107973c26fc60bba0a50d
Signed-off-by: Alexander Bengtsson <Alexander.Bengtsson@arm.com>

 - Replace all OpType::Conv2DBias with OpType::Conv2D
since there is no different handling in the code base
 - Remove OpType::Conv2DBias

Change-Id: Id444491c87b2a1ea3a9a6c5066b5406531cc025a
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

Johan Gunnarsson authored Oct 17, 2024 and Johan Alfvén committed Oct 18, 2024

* Fix asserts.
* Add unit tests.

Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Change-Id: Ibdcb9a157fca8e32ed59e03469aca2d14d5f220d

Alexander Bengtsson authored Oct 07, 2024 and Alexander Bengtsson committed Oct 18, 2024

Change-Id: I6c977a1b810de949e51204fa5f75782f3cbb8fb6
Signed-off-by: Alexander Bengtsson <Alexander.Bengtsson@arm.com>

 - An output diff occurred because decomposition was not done for
convolutions when batch was greater than 1
 - Add check to CanRunOnHardware checking for batch for conv
 - Use SliceShape instead of full shape in DecomposeConv2D and
DecomposeDepthwiseConv2D

Change-Id: Ic932ad329327beea75df8e76b48d737e53aa3cb0
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

- Assumed padValues were int32, but they can also be int64

Change-Id: I359fdf6f9c4fbb23926669c9f9207da44af18fa3
Signed-off-by: Rickard Bolin <rickard.bolin@arm.com>

Johan Gunnarsson authored Oct 08, 2024 and Johan Alfvén committed Oct 17, 2024

* Remove transpose constraint for Ethos-U85.
* Add decomposition of transpose operations.
* Refactor the ReshapeTo3D functions.

Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Change-Id: I4bf7a69b65b71312c34eed4ae16d7d899383d134