From 34336f9378eca4d39913ac0d9ba411a494631ad5 Mon Sep 17 00:00:00 2001
From: Matthew Bentham <Matthew.Bentham@arm.com>
Date: Thu, 27 Apr 2023 12:13:50 +0000
Subject: Make Convert workloads use arm_compute::NECast in CpuAcc backend

NECast can use conversion instructions where they are available
so this should in general be faster.

Signed-off-by: Matthew Bentham <Matthew.Bentham@arm.com>
Change-Id: I3f259e17b280a4f4c36f363965ffbc8ee8c4c29f
---
 .../workloads/NeonConvertFp32ToFp16Workload.cpp    | 58 ++++++++++++++++++----
 1 file changed, 47 insertions(+), 11 deletions(-)

(limited to 'src/backends/neon/workloads/NeonConvertFp32ToFp16Workload.cpp')

diff --git a/src/backends/neon/workloads/NeonConvertFp32ToFp16Workload.cpp b/src/backends/neon/workloads/NeonConvertFp32ToFp16Workload.cpp
index 089716a4b4..017ed9867e 100644
--- a/src/backends/neon/workloads/NeonConvertFp32ToFp16Workload.cpp
+++ b/src/backends/neon/workloads/NeonConvertFp32ToFp16Workload.cpp
@@ -5,6 +5,7 @@
 
 #include "NeonConvertFp32ToFp16Workload.hpp"
 
+#include <arm_compute/runtime/NEON/functions/NECast.h>
 #include <Half.hpp>
 #include <Profiling.hpp>
 
@@ -12,32 +13,67 @@
 
 #include <backendsCommon/WorkloadUtils.hpp>
 
+static constexpr arm_compute::ConvertPolicy g_AclConvertPolicy = arm_compute::ConvertPolicy::SATURATE;
+
 namespace armnn
 {
 
+arm_compute::Status NeonConvertFp32ToFp16WorkloadValidate(const TensorInfo& input, const TensorInfo& output)
+{
+    // Fallback to portable software implementation if Compute Library NECast won't work, so
+    // this method always returns success
+
+    armnn::IgnoreUnused(input);
+    armnn::IgnoreUnused(output);
+    return arm_compute::Status();
+}
+
 NeonConvertFp32ToFp16Workload::NeonConvertFp32ToFp16Workload(const ConvertFp32ToFp16QueueDescriptor& descriptor,
                                                              const WorkloadInfo& info)
     : Float32ToFloat16Workload<ConvertFp32ToFp16QueueDescriptor>(descriptor, info)
 {
     this->m_Data.ValidateInputsOutputs("NeonConvertFp32ToFp16Workload", 1, 1);
-    GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
+
+    arm_compute::ITensor& input  = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    if (arm_compute::NECast::validate(input.info(), output.info(), g_AclConvertPolicy))
+    {
+        // Use NECast if supported (needs hardware support for FP16)
+        m_Cast.reset(new arm_compute::NECast);
+        m_Cast->configure(&input, &output, g_AclConvertPolicy);
+    }
+    else
+    {
+        // Else use software implementation from Half.hpp
+        GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
+    }
 }
 
 void NeonConvertFp32ToFp16Workload::Execute() const
 {
     ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConvertFp32ToFp16Workload_Execute", this->GetGuid());
 
-    auto convertFunc = [](uint8_t* dst, const uint8_t* src, size_t size)
-        {
-            auto input = reinterpret_cast<const float*>(src);
-            auto output = reinterpret_cast<Half*>(dst);
-            size_t numElements = size/2; // 2 bytes per fp16
-            armnnUtils::FloatingPointConverter::ConvertFloat32To16(input, numElements, output);
-        };
-
-    for (const auto& pair : m_TensorHandlePairs)
+    if (m_Cast)
     {
-        CopyTensorContentsGeneric(pair.first, pair.second, convertFunc);
+        // Use NECast if supported and initialised
+        m_Cast->run();
+    }
+    else
+    {
+        // Else use softwre implementabion using Half.hpp
+        auto convertFunc = [](uint8_t* dst, const uint8_t* src, size_t size)
+            {
+                auto input = reinterpret_cast<const float*>(src);
+                auto output = reinterpret_cast<Half*>(dst);
+                size_t numElements = size/2; // 2 bytes per fp16
+                armnnUtils::FloatingPointConverter::ConvertFloat32To16(input, numElements, output);
+            };
+
+        for (const auto& pair : m_TensorHandlePairs)
+        {
+            CopyTensorContentsGeneric(pair.first, pair.second, convertFunc);
+        }
     }
 }
 
-- 
cgit v1.2.1