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>

6 files changed, 17 insertions, 16 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
index e85dd59425..290fe87230 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
@@ -293,14 +293,14 @@ GemmImplementation<float, float>::with_estimate(
 {
     GemmMethod::GEMM_HYBRID,
     "a64_smallK_hybrid_fp32_mla_8x4",
-    [](const GemmArgs &args) { return args._Ksize <= 8 && (args._Nsize % 4)==0 && !args._indirect_input; },
+    [](const GemmArgs &args) { return args._Ksize <= 8 && (args._Nsize % 4)==0 && !args._indirect_input && !args._accumulate; },
     nullptr,
     [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_fp32_mla_8x4, float, float>(args); }
 },
 {
     GemmMethod::GEMM_HYBRID,
     "a64_smallK_hybrid_fp32_mla_6x4",
-    [](const GemmArgs &args) { return (args._Ksize > 8 && args._Ksize <= 16) && (args._Nsize % 4)==0 && !args._indirect_input; },
+    [](const GemmArgs &args) { return (args._Ksize > 8 && args._Ksize <= 16) && (args._Nsize % 4)==0 && !args._indirect_input && !args._accumulate; },
     nullptr,
     [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_fp32_mla_6x4, float, float>(args); }
 },
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp b/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp
index 89c2d5a23e..0cc4d4f3d9 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp
@@ -530,7 +530,7 @@ public:
                                  (m_end - m_start), (nmax - n0), kern_k, b_panel, this->_ldb, out_arg,
                                  (this->_bias && first_pass) ? this->_bias + (multi * this->_bias_multi_stride) + n0 : nullptr,
                                  last_pass ? _args._act : Activation(),
-                                 !first_pass,
+                                 !first_pass || _args._accumulate,
                                  // Quantization parameters
                                  _os, _col_bias+(multi * _args._Nsize), n0);
                 } else if (_convolver) {
@@ -563,7 +563,7 @@ public:
                                  (m_end - m_start), (nmax - n0), kern_k, b_panel, this->_ldb, out_arg,
                                  (this->_bias && first_pass) ? this->_bias + (multi * this->_bias_multi_stride) + n0 : nullptr,
                                  last_pass ? _args._act : Activation(),
-                                 !first_pass,
+                                 !first_pass || _args._accumulate,
                                  // Quantization parameters
                                  _os, _col_bias+(multi * _args._Nsize), n0);
                 } else {
@@ -579,7 +579,7 @@ public:
                                  (m_end - m_start), (nmax - n0), kern_k, b_panel, this->_ldb, out_arg,
                                  (this->_bias && first_pass) ? this->_bias + (multi * this->_bias_multi_stride) + n0 : nullptr,
                                  last_pass ? _args._act : Activation(),
-                                 !first_pass,
+                                 !first_pass || _args._accumulate,
                                  // Quantization parameters
                                  _os, _col_bias+(multi * _args._Nsize), n0);
                 }
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
index fd20e53f60..0dc0d55b27 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020, 2022-2023 Arm Limited.
+ * Copyright (c) 2017-2020, 2022-2024 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -128,14 +128,14 @@ GemmImplementation<int8_t, int32_t>::with_estimate(
 {
     GemmMethod::GEMM_HYBRID,
     "a64_smallK_hybrid_s8s32_dot_8x4",
-    [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; },
+    [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input && !args._accumulate; },
     [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_s8s32_dot_8x4, int8_t, int32_t>(args); }
 },
 {
     GemmMethod::GEMM_HYBRID,
     "a64_smallK_hybrid_s8s32_dot_6x4",
-    [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; },
+    [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input && !args._accumulate; },
     [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_s8s32_dot_6x4, int8_t, int32_t>(args); }
 },
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp b/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
index 4f732f7d94..d8b464584a 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
@@ -350,6 +350,7 @@ class GemmInterleaved : public GemmCommon<To, Tr> {
     const bool _thread_columns;
 
     const Activation _act;
+    const bool _accumulate;
 
     const int _maxthreads;
     int _nthreads;
@@ -680,7 +681,7 @@ public:
                       _Ksections(args._Ksections), _Ktotal(get_ktotal(args)),
                       _rounded_Ksize(roundup(_Ksize, strategy::k_unroll())),
                       _nbatches(args._nbatches), _nmulti(args._nmulti), _thread_columns(is_thread_columns(args)),
-                      _act(args._act), _maxthreads(args._maxthreads), _nthreads(args._maxthreads),
+                      _act(args._act), _accumulate(args._accumulate), _maxthreads(args._maxthreads), _nthreads(args._maxthreads),
                       _k_block(get_k_block_size(args)), _x_block(get_x_block_size(args)), _Mround(roundup(args._Msize, strategy::out_height())),
                       _os(os) { }
 
@@ -690,7 +691,7 @@ public:
                       _Ksections(args._Ksections), _Ktotal(get_ktotal(args)),
                       _rounded_Ksize(roundup(_Ksize, strategy::k_unroll())),
                       _nbatches(args._nbatches), _nmulti(args._nmulti), _thread_columns(is_thread_columns(args)),
-                      _act(args._act), _maxthreads(args._maxthreads), _nthreads(args._maxthreads),
+                      _act(args._act), _accumulate(args._accumulate),  _maxthreads(args._maxthreads), _nthreads(args._maxthreads),
                       _k_block(get_k_block_size(args)), _x_block(get_x_block_size(args)), _Mround(roundup(args._Msize, strategy::out_height())),
                       _os() { }
 
@@ -823,7 +824,7 @@ public:
                             // Only do bias on the first pass
                             ((first_pass && this->_bias) ? this->_bias + (multi * this->_bias_multi_stride) : nullptr),
                             // Only do activation on the last pass, and accumulation on any non-first pass.
-                            (last_pass ? _act : Activation()), !first_pass,
+                            (last_pass ? _act : Activation()), (!first_pass || _accumulate),
                             // Pass in quantization parameters for requantizing kernels (others will ignore)
                             _os, col_bias + (multi * _Nsize),
                             // Accumulation buffer
@@ -971,7 +972,7 @@ public:
                             // Only do bias on the first pass
                             ((first_pass && this->_bias) ? this->_bias + (current.multi() * this->_bias_multi_stride) : nullptr),
                             // Only do activation on the last pass, and accumulation on any non-first pass.
-                            (last_pass ? _act : Activation()), !first_pass,
+                            (last_pass ? _act : Activation()), (!first_pass || _accumulate),
                             // Pass in quantization parameters for requantizing kernels (others will ignore)
                             _os, col_bias + (current.multi() * _Nsize),
                             // Accumulation buffer
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
index af5cfbbf2b..dfacb687a8 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020, 2022-2023 Arm Limited.
+ * Copyright (c) 2017-2020, 2022-2024 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -94,14 +94,14 @@ GemmImplementation<uint8_t, uint32_t>::with_estimate(
 {
     GemmMethod::GEMM_HYBRID,
     "a64_smallK_hybrid_u8u32_dot_8x4",
-    [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; },
+    [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input && !args._accumulate; },
     [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_u8u32_dot_8x4, uint8_t, uint32_t>(args); }
 },
 {
     GemmMethod::GEMM_HYBRID,
     "a64_smallK_hybrid_u8u32_dot_6x4",
-    [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; },
+    [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input && !args._accumulate; },
     [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_u8u32_dot_6x4, uint8_t, uint32_t>(args); }
 },
diff --git a/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp b/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
index 92c884ce18..dbada36052 100644
--- a/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
@@ -180,7 +180,7 @@ public:
                                  this->_Cptr + (multi * this->_C_multi_stride) + n,
                                  (nmax - n), (kmax-k0),
                                  this->_bias ? this->_bias + (multi * this->_bias_multi_stride) + n : nullptr,
-                                 _args._act, (k0 != 0),
+                                 _args._act, (k0 != 0) || _args._accumulate,
                                  _os, col_bias, n + (_args._Nsize * multi));
                 }
             }