Add support for non-constant weights and biases in CpuFullyConnected

Changing the approach for specifying that weights and biases tensors are
non-constant by making it a member of TensorInfo rather than an option
of the functions.

Resolves: COMPMID-4222

Change-Id: I96e6f3868f51785c9700a3ef6a1fe7b05747862c
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6162
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 31 insertions, 0 deletions

diff --git a/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp b/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp
index 97893b0672..1dd6286dbf 100644
--- a/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp
+++ b/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp
@@ -206,6 +206,7 @@ private:
     std::vector<TypeInput>           _indirect_pad{};
     arm_gemm::ConvolutionParameters  _cp{};
     experimental::MemoryRequirements _aux_mem{ Count };
+    bool                             _B_pretranspose_required{ false };
 };
 
 template <typename TypeInput, typename TypeOutput, class OutputStage>
@@ -391,6 +392,7 @@ void Fallback<TypeInput, TypeOutput, OutputStage>::configure(const ITensorInfo *
         const size_t       B_pretranspose_size = _gemm_kernel_asm->get_B_pretransposed_array_size();
         _pretranspose_info                     = TensorInfo(TensorShape(B_pretranspose_size), 1, DataType::U8);
         _aux_mem[Pretranspose]                 = MemoryInfo(offset_int_vec(Pretranspose), MemoryLifetime::Persistent, B_pretranspose_size, alignment);
+        _B_pretranspose_required               = true;
     }
 
     // Handle indirect GEMM convolution
@@ -485,6 +487,35 @@ void Fallback<TypeInput, TypeOutput, OutputStage>::run(ITensorPack &tensors)
         in1_ptr        = reinterpret_cast<const TypeInput *>(b->buffer() + b->info()->offset_first_element_in_bytes());
     }
 
+    // If necessary, run pretranspose every time if either weights or biases are non-constant
+    if((b && !b->info()->are_values_constant()) || (c && !c->info()->are_values_constant() && c->info()->data_type() == DataType::S32))
+    {
+        if(c && c->info()->data_type() == DataType::S32)
+        {
+            _gemm_kernel_asm->set_quantized_bias(reinterpret_cast<const int32_t *>(c->buffer() + c->info()->offset_first_element_in_bytes()), 0);
+        }
+
+        // Pretranspose B if required
+        if(_B_pretranspose_required)
+        {
+            const int  ldb            = b->info()->strides_in_bytes().y() / sizeof(TypeInput);
+            const auto b_ptr          = reinterpret_cast<const TypeInput *>(b->buffer() + b->info()->offset_first_element_in_bytes());
+            const int  multi_stride_b = b->info()->strides_in_bytes().z() / sizeof(TypeInput);
+
+            CpuAuxTensorHandler pretranspose(offset_int_vec(Pretranspose), _pretranspose_info, tensors, true);
+            ARM_COMPUTE_ERROR_ON(pretranspose.get()->buffer() == nullptr);
+
+            if(b->info()->are_values_constant())
+            {
+                _gemm_kernel_asm->requantize_bias(pretranspose.get()->buffer(), b_ptr, ldb, multi_stride_b);
+            }
+            else
+            {
+                _gemm_kernel_asm->pretranspose_B_array(pretranspose.get()->buffer(), b_ptr, ldb, multi_stride_b);
+            }
+        }
+    }
+
     const auto scheduling_hint = scheduling_hint_heuristic(_kernel_info.method, d->info()->data_type());
 
     // Set workspace if needed and reset number of threads as buffer manager gets re-created with max_threads