Memory Copy Performance Measure (memcpy, ID3D11DeviceContext::CopyResource)

memcpy_perf_measure.cpp

#include <Windows.h>
#include <d3d11.h>
#include <atlbase.h>
#pragma comment (lib, "d3d11.lib")

#include <iostream>
#include <thread>
#include <chrono>
#include <memory>
#include <vector>
#include <cstdint>
#include <cassert>
#include <algorithm>
#include <execution>

constexpr double MEASURE_SECONDS = 10;
constexpr double GIGABYTE_MAKER = 1 / 1073741824.0;

class performance
{
public:
	performance () : _copy_bytes (0), _running (false) { }
	virtual ~performance () noexcept
	{
		_run.join ();
	}

public:
	uint64_t copy_bytes () const noexcept { return _copy_bytes; }
	std::chrono::duration<double> proceed_time () const noexcept
	{
		return std::chrono::high_resolution_clock::now () - _started;
	}
	bool is_running () const noexcept { return _running; }

protected:
	virtual size_t do_measure () noexcept = 0;

public:
	void run () noexcept
	{
		_running = true;
		_started = std::chrono::high_resolution_clock::now ();

		_run = std::thread ([this]()
			{
				_copy_bytes = 0;

				do
				{
					_copy_bytes += this->do_measure ();
					std::this_thread::yield ();
				}
				while (_running);
			}
		);
	}
	void stop () { _running = false; }

private:
	uint64_t _copy_bytes;
	std::chrono::steady_clock::time_point _started;

	std::thread _run;
	bool _running;
};

class memcpy_performance : public performance
{
private:
	const size_t BUFFER_SIZE = 1024 * 1024 * 16; //< 16MB

public:
	memcpy_performance ()
	{
		_dest.resize (BUFFER_SIZE);
		_src.resize (BUFFER_SIZE);

		for (unsigned int i = 0; i < std::thread::hardware_concurrency (); ++i)
			_temp.push_back (i);
	}

protected:
	virtual size_t do_measure () noexcept override
	{
		std::for_each (std::execution::par_unseq, _temp.begin (), _temp.end (), [this](unsigned int i)
			{
				memcpy (_dest.data (), _src.data (), BUFFER_SIZE);
			});
		return BUFFER_SIZE * _temp.size ();
	}

private:
	std::vector<uint8_t> _dest, _src;
	std::vector<unsigned int> _temp;
};

class D3D11CopyResourceRAM2VRAM_performance : public performance
{
public:
	D3D11CopyResourceRAM2VRAM_performance (size_t size = 4096, DXGI_FORMAT format = DXGI_FORMAT_R8G8B8A8_UNORM)
	{
		HRESULT hr = D3D11CreateDevice (nullptr, D3D_DRIVER_TYPE_HARDWARE, nullptr, 0, nullptr, 0, D3D11_SDK_VERSION, &_d3dDevice, nullptr, &_immediateContext);

		assert (SUCCEEDED (hr));

		memset (&_texDesc, 0, sizeof (D3D11_TEXTURE2D_DESC));
		_texDesc.Width = _texDesc.Height = size;
		_texDesc.ArraySize = 1;
		_texDesc.MipLevels = 1;
		_texDesc.Format = format;
		_texDesc.SampleDesc.Count = 1;

		_texDesc.Usage = D3D11_USAGE_DEFAULT;
		_texDesc.CPUAccessFlags = 0;
		_texDesc.BindFlags = D3D11_BIND_SHADER_RESOURCE;

		hr = _d3dDevice->CreateTexture2D (&_texDesc, nullptr, &_dest);
		assert (SUCCEEDED (hr));

		_texDesc.Usage = D3D11_USAGE_STAGING;
		_texDesc.CPUAccessFlags = D3D11_CPU_ACCESS_READ | D3D11_CPU_ACCESS_WRITE;
		_texDesc.BindFlags = 0;

		hr = _d3dDevice->CreateTexture2D (&_texDesc, nullptr, &_src);
		assert (SUCCEEDED (hr));

		_totalSize = size * size * (format == DXGI_FORMAT_R8G8B8A8_UNORM ? 4 : 16);
	}

protected:
	virtual size_t do_measure () noexcept override
	{
		_immediateContext->CopyResource (_dest, _src);
		_immediateContext->Flush ();

		return _totalSize;
	}

private:
	D3D11_TEXTURE2D_DESC _texDesc;

	CComPtr<ID3D11Device> _d3dDevice;
	CComPtr<ID3D11DeviceContext> _immediateContext;

	CComPtr<ID3D11Texture2D> _dest, _src;
	size_t _totalSize;
};

class D3D11CopyResourceVRAM2VRAM_performance : public performance
{
public:
	D3D11CopyResourceVRAM2VRAM_performance (size_t size = 4096, DXGI_FORMAT format = DXGI_FORMAT_R8G8B8A8_UNORM)
	{
		HRESULT hr = D3D11CreateDevice (nullptr, D3D_DRIVER_TYPE_HARDWARE, nullptr, 0, nullptr, 0, D3D11_SDK_VERSION, &_d3dDevice, nullptr, &_immediateContext);

		assert (SUCCEEDED (hr));

		memset (&_texDesc, 0, sizeof (D3D11_TEXTURE2D_DESC));
		_texDesc.Width = _texDesc.Height = size;
		_texDesc.ArraySize = 1;
		_texDesc.MipLevels = 1;
		_texDesc.Format = format;
		_texDesc.SampleDesc.Count = 1;

		_texDesc.Usage = D3D11_USAGE_DEFAULT;
		_texDesc.CPUAccessFlags = 0;
		_texDesc.BindFlags = D3D11_BIND_SHADER_RESOURCE;

		hr = _d3dDevice->CreateTexture2D (&_texDesc, nullptr, &_dest);
		assert (SUCCEEDED (hr));

		hr = _d3dDevice->CreateTexture2D (&_texDesc, nullptr, &_src);
		assert (SUCCEEDED (hr));

		_totalSize = size * size * (format == DXGI_FORMAT_R8G8B8A8_UNORM ? 4 : 16);
	}

protected:
	virtual size_t do_measure () noexcept override
	{
		_immediateContext->CopyResource (_dest, _src);
		_immediateContext->Flush ();

		return _totalSize;
	}

private:
	D3D11_TEXTURE2D_DESC _texDesc;

	CComPtr<ID3D11Device> _d3dDevice;
	CComPtr<ID3D11DeviceContext> _immediateContext;

	CComPtr<ID3D11Texture2D> _dest, _src;
	size_t _totalSize;
};

void measure (const char * testname, performance* perf)
{
	printf ("==== %s Performance Measure ====\n", testname);
	std::shared_ptr<performance> _measure (perf);
	_measure->run ();
	while (_measure->is_running ())
	{
		if (_measure->proceed_time ().count () >= MEASURE_SECONDS)
			_measure->stop ();

		printf ("\r%3.3lfs... %lfGB/s... Total Copied: %lfGB",
			_measure->proceed_time ().count (),
			(_measure->copy_bytes () / _measure->proceed_time ().count ()) * GIGABYTE_MAKER,
			_measure->copy_bytes () * GIGABYTE_MAKER);

		std::this_thread::yield ();
	}
	printf ("\r%3.3lfs... %lfGB/s... Total Copied: %lfGB",
		_measure->proceed_time ().count (), 
		(_measure->copy_bytes () / _measure->proceed_time ().count ()) * GIGABYTE_MAKER,
		_measure->copy_bytes ()* GIGABYTE_MAKER);
	putchar ('\n');
}

int main (int argc, char* argv[])
{
	measure (u8"CPU memcpy", new memcpy_performance ());
	measure (u8"Direct3D Texture2D Copy Resource RAM to VRAM(4096 * 4096 * R8G8B8A8)", new D3D11CopyResourceRAM2VRAM_performance (4096, DXGI_FORMAT_R8G8B8A8_UNORM));
	measure (u8"Direct3D Texture2D Copy Resource RAM to VRAM(4096 * 4096 * RGBAF)", new D3D11CopyResourceRAM2VRAM_performance (4096, DXGI_FORMAT_R32G32B32A32_FLOAT));
	measure (u8"Direct3D Texture2D Copy Resource VRAM to VRAM(4096 * 4096 * R8G8B8A8)", new D3D11CopyResourceVRAM2VRAM_performance (4096, DXGI_FORMAT_R8G8B8A8_UNORM));
	measure (u8"Direct3D Texture2D Copy Resource VRAM to VRAM(4096 * 4096 * RGBAF)", new D3D11CopyResourceVRAM2VRAM_performance (4096, DXGI_FORMAT_R32G32B32A32_FLOAT));

	return 0;
}

CPU: AMD Ryzen 5 1600X (No Overclocked)
RAM: DDR4 2133MHz 8GBx2 (No Overclocked)
M/B: ASRock B350M Pro4
GPU: Zotac NVIDIA GeForce GTX 970

Result:

==== CPU memcpy Performance Measure ====
10.000s... 13.245923GB/s... Total Copied: 132.562500GB
==== Direct3D Texture2D Copy Resource RAM to VRAM(4096 * 4096 * R8G8B8A8) Performance Measure ====
10.000s... 12.181203GB/s... Total Copied: 121.812500GB
==== Direct3D Texture2D Copy Resource RAM to VRAM(4096 * 4096 * RGBAF) Performance Measure ====
10.000s... 13.474920GB/s... Total Copied: 134.750000GB
==== Direct3D Texture2D Copy Resource VRAM to VRAM(4096 * 4096 * R8G8B8A8) Performance Measure ====
10.000s... 127.330644GB/s... Total Copied: 1273.312500GB
==== Direct3D Texture2D Copy Resource VRAM to VRAM(4096 * 4096 * RGBAF) Performance Measure ====
10.000s... 74.749695GB/s... Total Copied: 747.500000GB

CPU: AMD Ryzen 7 3700X (No Overclocked)
RAM: ESSENCORE DDR4 2400MHz 16GBx2 (No Overclocked, RAM Timing XMP)
M/B: MSI B550M Mortar WiFi
GPU: AMD Radeon RX 5700XT Reference Model by Sapphire

Result:

==== CPU memcpy Performance Measure ====
10.005s... 14.043436GB/s... Total Copied: 140.500000GB
==== Direct3D Texture2D Copy Resource RAM to VRAM(4096 * 4096 * R8G8B8A8) Performance Measure ====
10.000s... 22.462423GB/s... Total Copied: 224.625000GB
==== Direct3D Texture2D Copy Resource RAM to VRAM(4096 * 4096 * RGBAF) Performance Measure ====
10.000s... 23.749940GB/s... Total Copied: 237.500000GB
==== Direct3D Texture2D Copy Resource VRAM to VRAM(4096 * 4096 * R8G8B8A8) Performance Measure ====
10.000s... 183.493069GB/s... Total Copied: 1834.937500GB
==== Direct3D Texture2D Copy Resource VRAM to VRAM(4096 * 4096 * RGBAF) Performance Measure ====
10.000s... 187.524475GB/s... Total Copied: 1875.250000GB

Background Informations
PCI-Express 4.0 x16 Maximum Bandwidth: 31.5 GB/s.

CPU: AMD Ryzen 7 3700X (No Overclocked)
RAM: ESSENCORE DDR4 2400MHz 16GBx2 (No Overclocked, RAM Timing XMP)
M/B: MSI B550M Mortar WiFi
GPU: GALAX NVIDIA GeForce RTX 3070 EX OC

Result:

==== CPU memcpy Performance Measure ====
10.006s... 14.916766GB/s... Total Copied: 149.250000GB
==== Direct3D Texture2D Copy Resource RAM to VRAM(4096 * 4096 * R8G8B8A8) Performance Measure ====
10.000s... 22.962413GB/s... Total Copied: 229.625000GB
==== Direct3D Texture2D Copy Resource RAM to VRAM(4096 * 4096 * RGBAF) Performance Measure ====
10.000s... 25.024848GB/s... Total Copied: 250.250000GB
==== Direct3D Texture2D Copy Resource VRAM to VRAM(4096 * 4096 * R8G8B8A8) Performance Measure ====
10.000s... 328.060400GB/s... Total Copied: 3280.625000GB
==== Direct3D Texture2D Copy Resource VRAM to VRAM(4096 * 4096 * RGBAF) Performance Measure ====
10.000s... 174.724219GB/s... Total Copied: 1747.250000GB

CPU: AMD Ryzen 3 2200U
RAM: Crucial DDR4 4GBx2
M/B: Lenovo ideapad 330S-15ARR (81FB)
GPU: AMD Radeon Vega 3 Mobile Graphics

Result:

==== CPU memcpy Performance Measure ====
10.007s... 5.233843GB/s... Total Copied: 52.375000GB
==== Direct3D Texture2D Copy Resource RAM to VRAM(4096 * 4096 * R8G8B8A8) Performance Measure ====
10.000s... 6.624958GB/s... Total Copied: 66.250000GB
==== Direct3D Texture2D Copy Resource RAM to VRAM(4096 * 4096 * RGBAF) Performance Measure ====
10.000s... 7.474972GB/s... Total Copied: 74.750000GB
==== Direct3D Texture2D Copy Resource VRAM to VRAM(4096 * 4096 * R8G8B8A8) Performance Measure ====
10.000s... 10.406215GB/s... Total Copied: 104.062500GB
==== Direct3D Texture2D Copy Resource VRAM to VRAM(4096 * 4096 * RGBAF) Performance Measure ====
10.000s... 8.499961GB/s... Total Copied: 85.000000GB