Add in place summation to CPU GEMM kernels

Instead of dispatching the sum postop for GEMM kernels to a
separate kernel + add, that requires an extra destination sized
allocation, plus 3 extra load/stores per element,
just do it in the GEMM kernel.

Resolves: ONCPUML-1442

Signed-off-by: Radu Salavat <radu.salavat@arm.com>
Co-authored-by: Milos Puzovic <milos.puzovic@arm.com>
Change-Id: I7a1f2da3300875fa1ac88b705a34390969518077
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11298
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 12 insertions, 1 deletions

diff --git a/src/cpu/operators/CpuGemm.cpp b/src/cpu/operators/CpuGemm.cpp
index e035de0131..905e86c185 100644
--- a/src/cpu/operators/CpuGemm.cpp
+++ b/src/cpu/operators/CpuGemm.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021-2023 Arm Limited.
+ * Copyright (c) 2021-2024 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -53,6 +53,7 @@ cpu::AsmGemmInfo init_assembly_metadata(const GEMMInfo &info)
     asm_info.fast_mode               = info.fast_math();
     asm_info.fixed_format            = info.fixed_format();
     asm_info.weight_format           = info.weight_format();
+    asm_info.accumulate              = info.accumulate();
     asm_info.transpose_b =
         info.pretranspose_B(); // The "pretranspose_B" flag here is not the same as the pretranspose_B_array method. The flag here signals to pretranspose_B_array method if we want to perform additional transpose on B before the pretranspose_B_array method
 
@@ -219,6 +220,16 @@ Status CpuGemm::validate(const ITensorInfo *a,
                          const GEMMInfo    &gemm_info)
 {
     ARM_COMPUTE_UNUSED(alpha);
+    // When using accumulation(in place summation), for now, the only supported values for alpha and beta are 1 respectively 0.
+    // Do the appropriate checks before proceeding.
+    if (gemm_info.accumulate())
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(alpha != 1, "Accumulation is not supported when alpha is different from 1");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+            (beta != 0 && c != nullptr),
+            "Accumulation is not supported when beta is different from 0 with a non-null bias matrix c");
+    }
+
     const bool is_c_bias    = beta == 1 && c != nullptr;
     const bool run_addition = c != nullptr && beta != 0 && beta != 1;
     // Check if we should use the pretransposed_b or original b