diff options
Diffstat (limited to 'docs/user_guide/errata.dox')
-rw-r--r-- | docs/user_guide/errata.dox | 75 |
1 files changed, 60 insertions, 15 deletions
diff --git a/docs/user_guide/errata.dox b/docs/user_guide/errata.dox index 635358bf92..056e45a432 100644 --- a/docs/user_guide/errata.dox +++ b/docs/user_guide/errata.dox @@ -1,5 +1,5 @@ /// -/// Copyright (c) 2019-2021 Arm Limited. +/// Copyright (c) 2019-2023 Arm Limited. /// /// SPDX-License-Identifier: MIT /// @@ -30,22 +30,67 @@ namespace arm_compute @section S7_1_errata Errata -- Experimented performance regressions for some networks on OpenCL when using Arm® Mali™ DDK r8p0 - - Versions Affected: >= v21.05 +- (COMPMID-6493) Crash when running Arm Compute Library compiled for SVE2 on a computer that support SVE only. + - Versions: >= v21.02 && <=v23.08 + - OSs: Linux, Android. + - Conditions: + - Compile the latest Arm Compute Library for SVE2 (arch=armv8.6-a-sve2). + - multi_isa = 0 + - Device with SVE but without SVE2 support. + - Result: + - Crash due to illegal instruction. + - To run SVE only, build with arch="armv8.2-a-sve", arch="armv8.6-a-sve", or with multi_isa=1. + +- (COMPMID-6404) Under certain conditions, CLTile may produce incorrect result. + - Versions: >= v19.02 && < v23.08 + - OSs: Linux, Android. + - Conditions: + - The size of the lowest dimension of the input tensor is greater than 16 bytes. + - The size of the lowest dimension of the input tensor is not a multiple of 16. + - Result: + - Incorrect result is produced. + +- (COMPMID-6271) Under certain conditions, CLArgMinMaxLayer validation tests may fail + - Versions Affected: >= v20.02 && < v23.08 + - OSs Affected: Linux + - Conditions: + - Backend: OpenCL + - Axis == 0 + - Result: + - Sporadic mismatches only on certain devices + +- (COMPMID-5324) Issue identified with direct and depthwise convolutions for certain Arm® Mali™ DDK versions. + - Versions Affected: < v22.08 + - Conditions: + - Arm® Mali™ DDK Versions : >= r23p0 && <= r38p0 + - Mali™ GPUs: Bifrost GPU family with the exception of G71 + - Backend: OpenCL + - Build options Include : "cl-fast-relaxed-math" + - Result: Reduced accuracy issue, while using direct and depthwise convolutions fused with LU_BOUNDED_RELU activation. + +- (COMPMID-5134) An issue has been identified when running the graph_deepspeech_v0_4_1 graph example. + - Versions Affected: >= v21.08 + - Conditions: + - Data type input: F32 + - Backend: OpenCL + - Result: The execution of the graph_deepspeech_v0_4_1 could fail on OpenCL backend for systems with a small RAM. The issue is due to the extra temporary memory required to reshape the network weights + +- (COMPMID-4013) Experimented performance regressions for some networks on OpenCL when using Arm® Mali™ DDK r8p0 + - Versions Affected: v21.05 - OSs Affected: All - Conditions: - Arm® Mali™ DDK r8p0 -- Under certain conditions, CLFullyConnectedLayer quantized tests may fail due to an issue in the test framework. - - Versions Affected: 21.02 +- (COMPMID-5146) Under certain conditions, CLFullyConnectedLayer quantized tests may fail due to an issue in the test framework. + - Versions Affected: v21.02 - OSs Affected: Linux - Conditions: - armv7a architecture - release mode - asserts enabled -- Performance regression in Convolution Layer OpenCL backend on Mali™ G77 when QSYMM8_PER_CHANNEL is used as weights' data type. - - Versions Affected: >= v20.11 +- (COMPMID-4367) Performance regression in Convolution Layer OpenCL backend on Mali™ G77 when QSYMM8_PER_CHANNEL is used as weights' data type. + - Versions Affected: >= v20.11 && < v21.08 - OSs Affected: All - Conditions: - Mali™ G77 @@ -53,34 +98,34 @@ namespace arm_compute - OpenCL backend - Convolution Layer uses QSYMM8_PER_CHANNEL as the data type of its weight -- A wrong test configuration has been found in CLGEMMMatrixMultiplyReshapedOnlyRHS set of tests. - - Versions Affected: >= 20.11 +- (COMPMID-4306) A wrong test configuration has been found in CLGEMMMatrixMultiplyReshapedOnlyRHS set of tests. + - Versions Affected: >= v20.11 && < v21.05 - Conditions: - Data type input: F32/F16 - Fused bounded relu activation with coefficient 'a' being negative -- Under certain conditions, the validation test case 'CL/DirectConvolutionLayer/Float/FP32/RunSmall9x9\@InputShape=32x37x3x4:StrideX=1:StrideY=1:PadX=0:PadY=0:KernelSize=9:NumKernels=1:DataType=F32:ActivationInfo=LU_BOUNDED_RELU:DataLayout=NHWC' may fail. +- (COMPMID-5135) Under certain conditions, the validation test case 'CL/DirectConvolutionLayer/Float/FP32/RunSmall9x9\@InputShape=32x37x3x4:StrideX=1:StrideY=1:PadX=0:PadY=0:KernelSize=9:NumKernels=1:DataType=F32:ActivationInfo=LU_BOUNDED_RELU:DataLayout=NHWC' may fail. - Versions Affected: >= v20.08 - Conditions: - The validation suite has to run in nightly mode and execute 40k+ test cases before the test mentioned above -- Under certain conditions, benchmark examples can hang when OpenCL profiling queues are enabled. +- (COMPMID-5136) Under certain conditions, benchmark examples can hang when OpenCL profiling queues are enabled. - Versions Affected: >= v19.11 - OSs Affected: Linux - Conditions: - Arm® Mali™ DDK r1p0 - r8p0, and - Linux kernel >= 4.4 -- On Android with arm64-v8a/arm64-v8.2-a architecture, Arm® Neon™ validation tests can fail when compiled using Android Ndk +- (COMPMID-5137) On Android with armv8a/armv8.2-a architecture, Arm® Neon™ validation tests can fail when compiled using Android Ndk >= r18b in debug mode (https://github.com/android/ndk/issues/1135). - Versions Affected: >= v19.11 - OSs Affected: Android - Conditions: - - arm64-v8a/arm64-v8.2-a architecture, and + - armv8a/armv8.2-a architecture, and - Compiled using Android NDK >= r18b in debug mode. -- An issue has been identified with CLCast. - - Versions Affected: >= 18.11 +- (COMPMID-4288) An issue has been identified with CLCast. + - Versions Affected: >= v18.11 && < v21.05 - Conditions: - Data type input: F32 - Data type output: All integer types |