- Aug 21, 2024
-
-
No longer rounds up to the ofm block to improve performance estimates. Change-Id: I47bff892ee3f71623c83aea3796e13a978cc09b8 Signed-off-by:
William Isaksson <william.isaksson@arm.com>
-
Johan Alfvén authored
- Extend heuristic with a bandwidth ratio to solve use cases where it is beneficial to do buffering Change-Id: I32121f73c523e3eae60e623a93aaf3255eab50c2 Signed-off-by:
Johan Alfven <johan.alfven@arm.com>
-
- Aug 20, 2024
-
-
Johan Gunnarsson authored
* Add support for TOSA NEGATE with fewer than 4 dimensions. * Add asserts to ensure that both IFMs have same data type where that's required. * Add helper to create const tensors with a certain value. * Update some other places to not trigger the new asserts. Signed-off-by:
Johan Gunnarsson <johan.gunnarsson@arm.com> Change-Id: I8f572d3e754c3521e739bf146aa0a00358c758ec
-
- Aug 19, 2024
-
-
Change-Id: I5ad75ba62971d19f760408cbcfa0b0eac1967c67 Signed-off-by:
Alexander Bengtsson <Alexander.Bengtsson@arm.com>
-
Johan Gunnarsson authored
* Move all CONCAT processing to graphir_optimiser. * Convert TFLite PACK into TOSA CONCAT. * Drop OpType::ConcatTFLite. Use OpType::Concat instead. Signed-off-by:
Johan Gunnarsson <johan.gunnarsson@arm.com> Change-Id: I757b0945f9ad28d0b1e3e6e18bcff9fcbd93398d
-
Johan Gunnarsson authored
* Add support for TOSA CONCAT fewer than 4 dimensions. * Add some missing RecordOptimisation. * Fix handling of constant bool tensors. * Fix TOSA validation. Signed-off-by:
Johan Gunnarsson <johan.gunnarsson@arm.com> Change-Id: I577f6abf31ef28c1ca128bf2bf3b007b86bec767
-
- Aug 16, 2024
-
-
Adds function for calculating per channel bws and all required inputs. Currently only used for m2m cycle estimation and weight transfer cycle estimation for weight decoder selection. Change-Id: I5de9678f09b5d3cfa7bb97d07fd780b72777f397 Signed-off-by:
William Isaksson <william.isaksson@arm.com>
-
- Fix implicit narrowing assert in python-bindings - Fix unused parameter warning in graphir_optimiser Change-Id: Iaaa2c1109d4e813d41f0b95a8e7569999a600d1b Signed-off-by:
Alexander Bengtsson <Alexander.Bengtsson@arm.com>
-
Johan Alfvén authored
- Fixed a case where consecutive ops have the same weights and it is beneficial to share the same weight buffering. - Added detection of this use case in the weight buffering algorithm so it will force a full ofm depth slice and hence the same weight buffer can be reused and only one dma transer is needed. Change-Id: I132f8bdc8d2c692541351e84441ad6c47a5c477f Signed-off-by:
Johan Alfven <johan.alfven@arm.com>
-
- Aug 15, 2024
-
-
Alexander Bengtsson authored
- Remove Small 4x4 IFM constraint to allow larger blocks for Matmul operations. - Consider OFM-coverage instead of IFM-fetch in relativeCost calculations. Change-Id: I5aee67f35bd979cc77328afb22b254dbdd0cabdd Signed-off-by:
Alexander Bengtsson <Alexander.Bengtsson@arm.com>
-
- Aug 13, 2024
-
-
Johan Alfvén authored
- Add simple heuristic when to enable weight buffering since dram configurations does not always benefit from it - Solution to be refined based on architecture and memory configuration Change-Id: I501c180dd846e6886520e0a2d25c2425d7df8592 Signed-off-by:
Johan Alfven <johan.alfven@arm.com>
-
Change-Id: I08f236996768794907d0fc44b09c29daa2509556 Signed-off-by:
Alexander Bengtsson <Alexander.Bengtsson@arm.com>
-
Alexander Bengtsson authored
- Update mypy in pre-commit to v0.981 Change-Id: I5e7d067584e3ad0a977f70980ac157ac97737d47 Signed-off-by:
Alexander Bengtsson <Alexander.Bengtsson@arm.com>
-
Signed-off-by:
Johan Gunnarsson <johan.gunnarsson@arm.com> Change-Id: I611ee58eb6e60041edcd20513ad79b86d0f4040a
-
Ifm shape used in per block estimate instead of ifm block. Currently this is not correct as we are rounding up to the ifm shape later. Change-Id: Ie2237056e7cde9fae8783eac95c91a46bcb1a9c2 Signed-off-by:
William Isaksson <william.isaksson@arm.com>
-
- Aug 12, 2024
-
-
- Rename RESCALE_ROUND to RESCALE_DOUBLE_ROUND to match TOSA specification Change-Id: Ic7818dcb559615fe71ab40220139a81e117a3496 Signed-off-by:
Alexander Bengtsson <Alexander.Bengtsson@arm.com>
-
Change-Id: Ifacfe1da39c75a9a4ff033c4249cdc50b1f747b0 Signed-off-by:
Alexander Bengtsson <Alexander.Bengtsson@arm.com>
-
Johan Gunnarsson authored
All CONV-like ops use the same validation paths, but FULLY_CONNECTED has different shape sizes. Signed-off-by:
Johan Gunnarsson <johan.gunnarsson@arm.com> Change-Id: Ic9a1008a9151001c0cf192cdf59345a274023acb
-
- Aug 09, 2024
-
-
Johan Gunnarsson authored
* Add support for TOSA CAST with fewer than 4 dimensions. Signed-off-by:
Johan Gunnarsson <johan.gunnarsson@arm.com> Change-Id: I6adb420e94ccfb5581635351e7db25efad91ece3
-
- Fix names of Rescale entries in graphapi OpAttr This will map rescale attributes to the correct attribute struct: rescale_attr_t Change-Id: I6788fde6c1b1813bf877bbc9b95896cbbca1cbd0 Signed-off-by:
Alexander Bengtsson <Alexander.Bengtsson@arm.com>
-
Johan Gunnarsson authored
In the hillclimb bottleneck fix, include all non-direct neighbours to the bottleneck when making the last attempt at swapping live ranges. Signed-off-by:
Johan Gunnarsson <johan.gunnarsson@arm.com> Change-Id: I4f963c03118aff9d296ddb2cd5445d69816af931
-
- Aug 08, 2024
-
-
Changes default perf query shapes to the op slice rather than the op shape. Change-Id: Ib3819df85e5c274efd98a92942a403573cdf00b7 Signed-off-by:
wilisa01 <william.isaksson@arm.com>
-
- Aug 07, 2024
-
-
Johan Alfvén authored
- Fixed output diff in regression tests - MoveSplitSliceToConsumer must run after any other optimisation that can change ifm/ofm shapes - Since graphir_optimiser has functions that can change shapes MoveSplitSliceToConsumer must be run in this step instead of tflite_graph_optimiser Change-Id: Ifa0f42d977e719111099e16d8bc98311e3bb1d62 Signed-off-by:
Johan Alfven <johan.alfven@arm.com>
-
Johan Alfvén authored
- Add missing NONE check in format tens list to prevent compilation error - The check was removed when code was refactored Change-Id: Ib79aff6198df3f0cbf40b40291e9a4befab39653 Signed-off-by:
Johan Alfven <johan.alfven@arm.com>
-
- Aug 06, 2024
-
-
- Convert scale attributes to fractions instead of points - Add FieldTypeId specialization for enums. Change-Id: I84e5cb7a1e5743dccde6b99352f42dc949ad84da Signed-off-by:
Alexander Bengtsson <Alexander.Bengtsson@arm.com>
-
- Aug 05, 2024
-
-
- The issue is that the intrinsic _BitScanReverse64 is not available in WIN32 - The solution is to split the 64-bit value up and use the 32-bit intrinsic instead Change-Id: Ic2093ce9f5d1a300986bd1327571edac650f5274 Signed-off-by:
Tim Hall <tim.hall@arm.com>
-
- Aug 02, 2024
-
-
Erik Waldemarson authored
Change-Id: I2061757c72178c7bebfad243a6b6b08fa9d7afb1 Signed-off-by:
Erik Waldemarson <erik.waldemarson@arm.com>
-
- Aug 01, 2024
-
-
- Fix off-by-one error in the GCC static name hash. Signed-off-by:
Philip Hall <philip.hall@arm.com> Change-Id: I1ed1099d87c0afb66afc8fef5f98bbc3e04e7f26
-
- Jul 31, 2024
-
-
Alexander Bengtsson authored
- Add support for TOSA Mul operations by extending ConvertAttributes in GraphIrOptimiser. Change-Id: Ia9c0fc6d8af0445d78042c0b9a0d195029fdfb78 Signed-off-by:
Alexander Bengtsson <Alexander.Bengtsson@arm.com>
-
- Jul 19, 2024
-
-
Bjorn Davidsson authored
When ofm datatype is Bool, we were incorrectly clamping output to [0, 255]. As True is represented as -1, it was clamped to 0. Fixed by having IntegerMin and IntegerMax handle Bool. Change-Id: I14ad89e128ab3339318f9967f8693959e3378652 Signed-off-by:
Björn Davidsson <bjoern.davidsson@arm.com>
-
- Jul 18, 2024
-
-
Updates to latest version, we are now using the same version of FlatBuffers and the FlatBuffers compiler. Change-Id: I29887794e8f338dc3f962c014c59a83b9db4d5f7 Signed-off-by:
wilisa01 <william.isaksson@arm.com>
-
Update to FlatBuffers 24.3.25 required by TensorFlow 2.17. Change-Id: I4c5e4399a8ecc61548bd7d5ee8807c42a955aa4c Signed-off-by:
wilisa01 <william.isaksson@arm.com>
-
Patch to fix mac util being reported as estimated above 100% for depth first convolutions, by accounting for sparsity and improving convolution cycle estimation. Also removes a bug where we divide the cycle count with the amount of wd_cores. Change-Id: Ie565c3524f121578779e21684725b12cfe304f7f Signed-off-by:
wilisa01 <william.isaksson@arm.com>
-
FWD encoder was returning 0 as the unencoded stream size, leading to incorrect estimates, and sometimes a divide-by-zero exception. Also now setting the correct weight format when estimating cycles. Change-Id: I4c82ca0f31b94cd8fe9343457157e01ebdf93c05 Signed-off-by:
Björn Davidsson <bjoern.davidsson@arm.com>
-
- Jul 17, 2024
-
-
Bjorn Davidsson authored
Slice shape is no longer initialized before getting to decomposition, added check. Ofm stride multiplier was removed, restored now. Change-Id: I7854d5bca5918ce35c6401b0efc508d6ad8873be Signed-off-by:
Björn Davidsson <bjoern.davidsson@arm.com>
-
- Jul 16, 2024
-
-
Philip Hall authored
- Feature bit values ended up as an enum rather than bitmasks. This commit fixes the bits so that WeightBuffering is no longer permanently disabled. Signed-off-by:
Philip Hall <philip.hall@arm.com> Change-Id: Ic4ce8fdb63332b220a18d1687e6c4d2ba10e7800
-
- Jul 15, 2024
-
-
Philip Hall authored
- Added scheduler feature control flags in the form of disable bits to control aspects of the scheduling stage. Signed-off-by:
Philip Hall <philip.hall@arm.com> Change-Id: I6f9c53d1c1f41ef37c1a5603901b3e09bdcee747
-
Fredrik Svedberg authored
Fixed support for TOSA DEPTHWISE_CONV2D weight format. Change-Id: I60ec82aa602dac521aa3c1285a7736bc28b5fc86 Signed-off-by:
Fredrik Svedberg <fredrik.svedberg@arm.com>
-
- The internal TOSA loader did not honour the type for CLAMP padding values. This commit converts the padding from the input type before setting the attribute. - Fix incorrect float conversion size validation. Signed-off-by:
Philip Hall <philip.hall@arm.com> Change-Id: I4c0873a066fca72ac259d8f8f531edab906e0ca9
-
- Jul 12, 2024
-
-
Fredrik Svedberg authored
The usage bitmap was never updated (due to operator precedence). Change-Id: Iaf27d997c80be76c3dc5c185a2d5ab860a6ce42a Signed-off-by:
Fredrik Svedberg <fredrik.svedberg@arm.com>
-