llglsandbox.cpp

	for (U32 i = 0; i < count; ++i)
	{
		//allocate render targets and textures
		if (!dest[i].allocate(res, res, GL_RGBA, false, false, LLTexUnit::TT_TEXTURE, true))
		{
			LL_WARNS() << "Failed to allocate render target." << LL_ENDL;
			// abandon the benchmark test
			delete[] pixels;
			return -1.f;
		}
		dest[i].bindTarget();
		dest[i].clear();
		dest[i].flush();

		if (!texHolder.bind(i))
		{
			// can use a dummy value mDummyTexUnit = new LLTexUnit(-1);
			LL_WARNS() << "Failed to bind tex unit." << LL_ENDL;
			// abandon the benchmark test
			delete[] pixels;
			return -1.f;
		}
		LLImageGL::setManualImage(GL_TEXTURE_2D, 0, GL_RGBA, res,res,GL_RGBA, GL_UNSIGNED_BYTE, pixels);
	}

    delete [] pixels;

	//make a dummy triangle to draw with
	LLPointer<LLVertexBuffer> buff = new LLVertexBuffer(LLVertexBuffer::MAP_VERTEX | LLVertexBuffer::MAP_TEXCOORD0, GL_STATIC_DRAW_ARB);

	if (!buff->allocateBuffer(3, 0, true))
	{
		LL_WARNS() << "Failed to allocate buffer during benchmark." << LL_ENDL;
		// abandon the benchmark test
		return -1.f;
	}

	LLStrider<LLVector3> v;
	LLStrider<LLVector2> tc;

	if (! buff->getVertexStrider(v))
	{
		LL_WARNS() << "GL LLVertexBuffer::getVertexStrider() returned false, "
				   << "buff->getMappedData() is"
				   << (buff->getMappedData()? " not" : "")
				   << " NULL" << LL_ENDL;
		// abandon the benchmark test
		return -1.f;
	}

	// generate dummy triangle
	v[0].set(-1, 1, 0);
	v[1].set(-1, -3, 0);
	v[2].set(3, 1, 0);

	buff->flush();

	// ensure matched pair of bind() and unbind() calls
	class ShaderBinder
	{
	public:
		ShaderBinder(LLGLSLShader& shader):
			mShader(shader)
		{
			mShader.bind();
		}
		~ShaderBinder()
		{
			mShader.unbind();
		}

	private:
		LLGLSLShader& mShader;
	};
	ShaderBinder binder(gBenchmarkProgram);

	buff->setBuffer(LLVertexBuffer::MAP_VERTEX);
	glFinish();

	for (S32 c = -1; c < samples; ++c)
	{
		LLTimer timer;
		timer.start();

		for (U32 i = 0; i < count; ++i)
		{
			dest[i].bindTarget();
			texHolder.bind(i);
			buff->drawArrays(LLRender::TRIANGLES, 0, 3);
			dest[i].flush();
		}

		//wait for current batch of copies to finish
		glFinish();

		F32 time = timer.getElapsedTimeF32();

		if (c >= 0) // <-- ignore the first sample as it tends to be artificially slow
		{ 
			//store result in gigabytes per second
			F32 gb = (F32) ((F64) (res*res*8*count))/(1000000000);
			F32 gbps = gb/time;
			results.push_back(gbps);
		}
	}

	std::sort(results.begin(), results.end());

	F32 gbps = results[results.size()/2];

	LL_INFOS() << "Memory bandwidth is " << llformat("%.3f", gbps) << "GB/sec according to CPU timers" << LL_ENDL;
  
#if LL_DARWIN
    if (gbps > 512.f)
    { 
        LL_WARNS() << "Memory bandwidth is improbably high and likely incorrect; discarding result." << LL_ENDL;
        //OSX is probably lying, discard result
        return -1.f;
    }
#endif

	F32 ms = gBenchmarkProgram.mTimeElapsed/1000000.f;
	F32 seconds = ms/1000.f;

	F64 samples_drawn = res*res*count*samples;
	F32 samples_sec = (samples_drawn/1000000000.0)/seconds;
	gbps = samples_sec*8;

	LL_INFOS() << "Memory bandwidth is " << llformat("%.3f", gbps) << "GB/sec according to ARB_timer_query" << LL_ENDL;

	return gbps;
}