Add errata for the performance regressions on OpenCL
- Experimented performance regressions for some networks on OpenCL when
using Arm(R) Mali(TM) DDK r8p0
- Performance regression in Convolution Layer OpenCL backend on
Mali(TM) G77, when QSYMM8_PER_CHANNEL is used as weights' data type
Change-Id: Ieb0ddde26545cf9f03e0598f3492f3e731aa3398
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5617
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
diff --git a/docs/user_guide/errata.dox b/docs/user_guide/errata.dox
index cca2a39..635358b 100644
--- a/docs/user_guide/errata.dox
+++ b/docs/user_guide/errata.dox
@@ -30,6 +30,12 @@
@section S7_1_errata Errata
+- 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
- OSs Affected: Linux
@@ -38,6 +44,15 @@
- 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
+ - OSs Affected: All
+ - Conditions:
+ - Mali™ G77
+ - Convolution Layer in use
+ - 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
- Conditions: