Fix export_to_cl_image issue in the fp16 GeMM implementation
- The issue affects Fp16 GeMM on Arm® Mali™-G78
- The issue was caused by a missing fallback implementation for the
case when export_to_cl_image cannot be used
- The new implementation fixes this issues and make the GeMM
implementation for M=1 also faster (4-5% on various networks with fully
connected at the end of the model)
- This patch also enables the H0=0 case in the GeMM examples
Resolves COMPMID-5812, COMPMID-5688, and COMPMID-6147
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Change-Id: Ib7b355ae25337962598dd2ba21665b1a6b48686f
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/514664
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/+/9526
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/examples/gemm_tuner/cl_gemmlowp_reshaped.cpp b/examples/gemm_tuner/cl_gemmlowp_reshaped.cpp
index 115ce6b..9cf9c9f 100644
--- a/examples/gemm_tuner/cl_gemmlowp_reshaped.cpp
+++ b/examples/gemm_tuner/cl_gemmlowp_reshaped.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2020-2021 Arm Limited.
+ * Copyright (c) 2020-2021, 2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -240,6 +240,11 @@
rhs_info.transpose = configs.transpose_rhs;
rhs_info.export_to_cl_image = false; // CL image not supported for quantized cases yet
+ if(rhs_info.h0 == 0)
+ {
+ rhs_info.h0 = std::max(static_cast<unsigned int>(params.N) / rhs_info.n0, 1U);
+ }
+
lhs_reshaped.allocator()->init(TensorInfo(compute_lhs_reshaped_shape(*lhs.info(), lhs_info), 1, params.data_type));
rhs_reshaped.allocator()->init(TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type));
lhs_reshaped.info()->set_quantization_info(q_info);