Age | Commit message (Collapse) | Author |
|
The new version introduces the following major changes:
* Change public interface to simplify and standardize the user experience
- Use the term "Workload" uniformly
- Simplify operator interface to be a set of static methods:
validate_op(), create_op()
* Separate the kernel writing into its own component (template_writer).
This is to allow the co-development of GpuKernelWriter, and to allow
easy replacement once GpuKernelWriter is mature.
* Optimize the core fusion algorithm used by the component graph. The
details can be found in GpuKernelComponentGraph::fuse()
* Use Gpu instead of Cl prefixes for most of the Workload interfaces
(except for runtime and kernel components, which have to be language specific)
This allows the potential extension to other Gpu langauges in the
future.
* Refactor runtime memory interface so that auxiliary tensor handling
is separate from the user tensor passing. This is because the former
is less stable and may require extension in the future.
* Hide source code object from the user as it is not required at the
moment
* Deprecate the old prototype entirely by disabling it in SCons build
Resolves COMPMID-5510, COMPMID-5512, COMPMID-5513
Change-Id: If69d2362856f2de4503546b7b6cf48a525cf3079
Signed-off-by: SiCong Li <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8406
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Resolves COMPMID-5634
Change-Id: I075de70d509d0c4430b4bcf3f218384e237a3a56
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/453708
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8473
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
|
|
* Use the same rounding function for the left-over part with the
vectorized part.
Resolves: COMPMID-5640
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I07450b2a43390b77539b78cd5d3e6772bdc38548
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8520
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
|
|
* Resolved MLCE-946
Change-Id: I7a2b8d068bbc810602dd0959e8c99d87a323152f
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8513
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
|
|
Use rational approximation with optimised coefficients
to calculate tanh_f16. Method is ~2.5x faster than previous
and has lower relative and absolute error.
This will fix https://github.com/ARM-software/ComputeLibrary/issues/1002
Credit to George Steed for suggesting use of vcageq instead of min+max
Signed-off-by: Jonathan Deakin <jonathan.deakin@arm.com>
Change-Id: Id70da3aab666c68b0d798266a837b59c00937bf7
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8480
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
* Resolves COMPMID-5654
Change-Id: I2654f5178b4400abf333a9b7ef5c9a239ce4ef73
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8507
Reviewed-by: Ramy Elgammal <ramy.elgammal@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
|
|
* Use the same implementation as other layers.
Resolves: COMPMID-5108
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I5a50259b398b71ca1f61b5ee8daa539bf8263fac
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8501
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
|
|
- Adds tests for batched matrix multiplication
- Bugfix for issue : 3d tensors input tensors with offsets in GemmLowp results in mismatches
Resolves : [COMPMID-5507]
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: I68e036fbca642c1841dd4321033045aadc8f5636
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/461298
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8482
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
* Do not change the tensor info after configure stage.
* By fixing this, the 1D optimization for activation layer can be
applied to all data types and tensor layout.
Resolves: COMPMID-5644
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I557f9bb84e5e456c28d6b423584887d7a3648ad4
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8470
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Mohmun02 <MohammedSuhail.Munshi@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
|
|
* Resolves MLCE-942
Change-Id: I4e61e3712c7bc7e711b1fa280d7c3e73d08a57e3
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8407
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
|
|
- Use a 1d execution window to improve memory access pattern.
Resolves: [COMPMID-5465]
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: Ida30669ffa06eb002ca43a6edf15e25a6eaad2f6
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8344
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Resolves: COMPMID-5494
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: I8f512745855b8ca21181a9ab21323bfff6aeb866
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/458884
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8391
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Partially resolves: COMPMID-5642
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: Icbea9de88d9c55494e782a08bb17aeeb2a9e592d
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8385
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
|
|
* Use the window instead of the tensor shape to determine the
number of elements in the x-dimension.
* Remove the LUT implementation in 32-bit build.
Resolves: COMPMID-5641
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I0a79aa38d8f6a105ad01785bd94571f5a2ecb348
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8380
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
|
|
Resolves: COMPMID-5600
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I5196d1639c48d0b8a116d47ed1d6c7334dc8f41e
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8374
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
|
|
Resolves: COMPMID-5462
Change-Id: I2c7151c8faf4016cc33592fff04d492d7cbc8fd6
Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8366
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Resolves COMPMID-5632
Change-Id: I2bdbe69a610ca2510fbd74d5d412842679299762
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8365
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
|
|
- Does not Pass RESULT_MULTIPLIER, RESULT_SHIFT When PER_CHANNEL_QUANTIZATION is defined.
Resolves COMPMID-5499
Signed-off-by: Adnan AlSinan <adnan.alsinan@arm.com>
Change-Id: Ie433eaf83c003a6d5ccfeb89eb2783528dc2b48e
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8316
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Resolves COMPMID-5631
Change-Id: I37d1d0d043f8d44d782d2225091af607ad131b58
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8364
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
|
|
* For DDK version 30 and higher, force the CL compiler to use
64 registers for NHWC direct convolution.
Resolves: COMPMID-5508
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I7d9ecc3b5a4eceaff44542cd26f6f05e30ab2c1f
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8351
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
|
|
Resolves COMPMID-5601
Signed-off-by: Adnan AlSinan <adnan.alsinan@arm.com>
Change-Id: I1baf92c4751d784d017c0b2f7de1fc09e42ce69c
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8309
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
|
|
* Resolves ARMCL-588
Change-Id: I0ae29e001fecbcf31b40fc51f8b25107456626de
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8337
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
|
|
* Use fixed-point arithmetic where possible.
* Various optimization for the FP32-based implementation.
This implementation is kept as the fall-back solution
in case of unrealistic quantization parameters that exceed
the range of fixed-point solution.
Resolves: COMPMID-5458
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I221d2d3801ecaae4fe0b7cf6ae8ef00ca3743665
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8317
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
* Resolves MLCE-924
Change-Id: I3cc3d30893c2ee0865eacafdc1d9ba3d5b876d32
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8326
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Change-Id: Id37b59adbc8c4cbe218d1652aeb02a0b4ce42c66
Signed-off-by: Jonathan Deakin <jonathan.deakin@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8256
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
SVE merges for interleaved kernels were not guarding bias reads with the
correct predicates, leading to overreads and crashes in some cases. Fix to
use the appropriate predicate.
Resolves: COMPMID-5627
Change-Id: Ib049531c4a3bea56e90623b7b9f0d6a7ab4db2c8
Signed-off-by: David Mansell <David.Mansell@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8315
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
|
|
* Resolves COMPMID-5599
Change-Id: I4c1df48eda289c82ca567f6808fccd0b09065223
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8302
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
multiplication with variable input tensors
Resolves: COMPMID-5506
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: I8345a3b7a83ef46f9ec7a77197cc65c933ec9ac6
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8239
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Resolves COMPMID-5533
Change-Id: Ice3d9469c7486a700c58fb61fc692b13f368d202
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8148
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Resolves: COMPMID-5580
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: Ia731560c23a6ab2e0ead5a857fbabb9cbc25154c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/452428
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8268
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
|
|
This patch introduces several performance optimizations regarding the Bilinear Scale operator with REPLICATE Border mode. Changes apply only to NHWC.
This patch
- Reduces the memory footprint by disabling precomputation of indices and weights when they're not used
- Rewrites the kernels for QASYMM8/QASYMM8_SIGNED/U8(Uint8)
- Adds S8(Int8) Bilinear Scale for Border mode REPLICATE
- Removes Bilinear Scale SVE kernels for Quantized and Integer types and adjust the heuristics to choose the Neon™ implementation
- Adds new test cases where the input and output of the Bilinear Scale operator have different quantization scale and offset
Resolves: COMPMID-5453, COMPMID-5454
Change-Id: I3d251e76e0c6978fd5a0a1795ec62ab536bec93c
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8250
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
* Also fix a bug in mul_U8_U8_U8.
Resolves: COMPMID-5460
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: Ie1edafeae7aaad91164caeeb04661a8974a7fc1b
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8244
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
The M0 and N0 were incorrectly set for the case of broadcasting when the
elementwise component is non-root.
This is because we previously always use rhs tensor to derive the load
M0, N0. But for non-root components, the addend/divisor tensor can be
in the lhs or rhs. Thus this would fail in case the addend/divisor is in
the lhs.
- Also fixes broken Dynamic Fusion test
Resolves COMPMID-5482
Signed-off-by: SiCong Li <sicong.li@arm.com>
Change-Id: I37f27ffa392781387db15739b1666f1dad28c554
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/445890
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Mohammed Suhail Munshi <mohammedsuhail.munshi@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8111
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
OpenCL implementation uses built in erf.
NEON implementation requires new vectorized erf.
Uses the following approximation:
erf(x) = 1 - 1 / (1 + a1x + a2x^2 + a3x^3 + a4x^4)^4
a1 = 0.278393, a2 = 0.230389, a3 = 0.000972, a4 = 0.078108
From https://en.wikipedia.org/wiki/Error_function#Numerical_approximations
Signed-off-by: Murray Kornelsen <murray.kornelsen@mail.mcgill.ca>
Change-Id: I2d3964b2c26a4334166b17135f9104bc6324fad2
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7921
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Pablo Marquez Tello <pablo.tello@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Implements LayerNorm for qasymm8 tensors.
Uses uint8x16 loads and stores.
Summation is performed in integer arithmetic (vpaddl)
Normalization is performed in float32 before requantizing back to int8.
Signed-off-by: Murray Kornelsen <murray.kornelsen@mail.mcgill.ca>
Change-Id: I2407c8b34717fb47adab98791bd76fb8a3c62f4a
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7922
Comments-Addressed: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Resolves: COMPMID-5531
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: Id1a5f909d64e897cdef2e920240ac5ae244166b4
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8242
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
input tensors
- Add a test for CPU to batched matrix multiplication with variable input tensors
- Disable assembly kernel when using _reshape_b_only_on_first_run flag
Resolves COMPMID-5501
Signed-off-by: Adnan AlSinan <adnan.alsinan@arm.com>
Change-Id: If96b182584617806a9dfe597dbfaf05241b123c2
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8234
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
The heuristic has been tweaked to call direct convolution when we think
it can be faster than gemm-based convolution. The main change is
affecting the selection of the convolution method on the first layer.
In general, the question we should ask for the first convolution layer of
a model is: when the execution time of im2col + gemm < direct?. Since
im2col does not depend on the OFM, it means that when OFM is big enough,
the contribution of im2col is small and the GEMM approach is preferable.
From internal experiments, the OFM threshold is 64.
Resolves COMPMID-5504, COMPMID-5504, COMPMID-5477
Change-Id: If1bd1fa93c185ffa874388e29866244e62ca3494
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8231
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
|
|
This patch removes index and weight pre-computations where it's not used and reduces some calculations inside the inner-most loop of Scale.
Resolves: COMPMID-5452
Change-Id: Ie149b1b76a90a8cb659ada0f97aef78caf69932f
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8220
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
|
|
Resolves COMPMID-5498
Change-Id: I474f3f963257014255d082aab0ccbe3efe5aa067
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8222
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Ramy Elgammal <ramy.elgammal@arm.com>
Reviewed-by: Ramy Elgammal <ramy.elgammal@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
- That would force CpuConv2d::get_convolution_method()
choose GEMM_CONV2D or GEMM methods instead.
Resolves: COMPMID-5531
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: I4dec8772e8c150da003d9a89c1d036057c4d28b0
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8233
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
|
|
The optimization concerns the case where the depth multiplier is > 1.
The depth multiplier for loop has been removed from the OpenCL kernel
and the GWS has been mapped to the output shape. In this way, we can
still perform a tile with N0 columns and improve the performance of
depthwise conv over 80% when depth multiplier is > 1.
Resolves COMPMID-5568
Change-Id: I604e287d4eeb31c54b9cc6c3072a698cd0e3e136
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8184
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
input tensors
Resolves : [COMPMID-5502]
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: Ida001dc597973f9180468737a3e32e5022e6baee
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/450342
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Mohammed Suhail Munshi <mohammedsuhail.munshi@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8224
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Ran into issues with f16 meanstddevnorm. Essentially, with large enough tensors and/or large values in tensors, output becomes all 0.
This is due to the variance computation. In f16, it reaches infinity quite easily, then the division results in 0.
This change modifies the OpenCL and NEON implementations to compute the sum of squares and the variance using f32, while other operations remain f16.
Update: Found that the square operation also benefits from f32, rather than squaring in f16 and accumulating f32.
Signed-off-by: Murray Kornelsen <murray.kornelsen@mail.mcgill.ca>
Change-Id: Ide00afd84ec6d26fec4d53b073e295814f08ba46
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7959
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
From an internal performance evaluation, it seems that Winograd-based
Conv2D offers better performance than alternative methods such as direct
convolution and gemm-based conv already from IFM=8. Before the condition
was for IFM>=16
Resolves COMPMID-5532
Change-Id: I9ff04835d6fd07f5f0abeec9645c9d9cc913b6b7
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8147
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
* Resolves COMPMID-5211
Change-Id: I7cc72662bb1cf52bf112685639d3dbba33d1333f
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7993
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
|
|
repeatedly when mapping multiple children of the same parent.
Signed-off-by: Murray Kornelsen <murray.kornelsen@mail.mcgill.ca>
Change-Id: I7ff554915a37320dfd94f15a6fb01a72a235cf39
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7920
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
Previously, the add_as_1d_array kernels were used on tensors with
non-matching strides which caused the wrong elements to be added. The
fix is to check that the strides are equal when selecting the addition
kernel.
Change-Id: I914ca2b95e5b8ed1875ec5ebe129bdfe2845496b
Signed-off-by: Jonathan Deakin <jonathan.deakin@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8120
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
* Resolves MLCE-799
Change-Id: I3d5b2afbf65d159aa5e645743c1e139110dfc20e
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8093
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
|
|
* Resolves MLCE-903
Change-Id: I39fb3b4b395a37c0f32481830f94d85ec15e205f
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8124
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
|