Question

HLSL新手问题：

我正在尝试移植以下MATLAB代码以在显卡上工作。

function diff_im = anisodiff2D(im, num_iter, delta_t, kappa, option)
im = double(im);

% PDE (partial differential equation) initial condition.
diff_im = im;

% Center pixel distances.
dx = 1;
dy = 1;
dd = sqrt(2);

% 2D convolution masks - finite differences.
hN = [0 1 0; 0 -1 0; 0 0 0];
hS = [0 0 0; 0 -1 0; 0 1 0];
hE = [0 0 0; 0 -1 1; 0 0 0];
hW = [0 0 0; 1 -1 0; 0 0 0];
hNE = [0 0 1; 0 -1 0; 0 0 0];
hSE = [0 0 0; 0 -1 0; 0 0 1];
hSW = [0 0 0; 0 -1 0; 1 0 0];
hNW = [1 0 0; 0 -1 0; 0 0 0];

% Anisotropic diffusion.
for t = 1:num_iter

        % Finite differences. [imfilter(.,.,'conv') can be replaced by conv2(.,.,'same')]
        nablaN = imfilter(diff_im,hN,'conv');
        nablaS = imfilter(diff_im,hS,'conv');   
        nablaW = imfilter(diff_im,hW,'conv');
        nablaE = imfilter(diff_im,hE,'conv');   
        nablaNE = imfilter(diff_im,hNE,'conv');
        nablaSE = imfilter(diff_im,hSE,'conv');   
        nablaSW = imfilter(diff_im,hSW,'conv');
        nablaNW = imfilter(diff_im,hNW,'conv'); 

        % Diffusion function.
        if option == 1
            cN = exp(-(nablaN/kappa).^2);
            cS = exp(-(nablaS/kappa).^2);
            cW = exp(-(nablaW/kappa).^2);
            cE = exp(-(nablaE/kappa).^2);
            cNE = exp(-(nablaNE/kappa).^2);
            cSE = exp(-(nablaSE/kappa).^2);
            cSW = exp(-(nablaSW/kappa).^2);
            cNW = exp(-(nablaNW/kappa).^2);
        elseif option == 2
            cN = 1./(1 + (nablaN/kappa).^2);
            cS = 1./(1 + (nablaS/kappa).^2);
            cW = 1./(1 + (nablaW/kappa).^2);
            cE = 1./(1 + (nablaE/kappa).^2);
            cNE = 1./(1 + (nablaNE/kappa).^2);
            cSE = 1./(1 + (nablaSE/kappa).^2);
            cSW = 1./(1 + (nablaSW/kappa).^2);
            cNW = 1./(1 + (nablaNW/kappa).^2);
        end

        % Discrete PDE solution.
        diff_im = diff_im + ...
                  delta_t*(...
                  (1/(dy^2))*cN.*nablaN + (1/(dy^2))*cS.*nablaS + ...
                  (1/(dx^2))*cW.*nablaW + (1/(dx^2))*cE.*nablaE + ...
                  (1/(dd^2))*cNE.*nablaNE + (1/(dd^2))*cSE.*nablaSE + ...
                  (1/(dd^2))*cSW.*nablaSW + (1/(dd^2))*cNW.*nablaNW );

        % Iteration warning.
        fprintf('\rIteration %d\n',t);
end

在我通过一次通过的那一刻：

texture2D Input0;
sampler2D Input0Sampler = sampler_state
{
    Texture = <Input0>;
    MinFilter = Point;
    MagFilter = Point;
    MipFilter = Point;
    AddressU = Clamp;
    AddressV = Clamp;
};

struct VertexShaderInput
{
    float4 Position : POSITION0;
    float2 TextureCoordinate : TEXCOORD0;
};

struct VertexShaderOutput
{
    float4 Position : POSITION0;
    float2 TextureCoordinate : TEXCOORD0;
};

struct PixelShaderOutput
{
    // TODO: Optionally add/remove output indices to match GPUProcessor.numOutputs
    float4 Index0 : COLOR0;
};

// input texture dimensions
static float w = 1920 - 8;
static float h = 1080 - 8;

static const float2 pixel = float2(1.0 / w, 1.0 / h);
static const float2 halfPixel = float2(pixel.x / 2, pixel.y / 2);

static const float3x3 hN =
{
    0,  1, 0,
    0, -1, 0,
    0,  0, 0
};
static const float3x3 hS =
{
    0,  0, 0,
    0, -1, 0,
    0,  1, 0
};
static const float3x3 hE =
{
    0,  0, 0,
    0, -1, 1,
    0,  0, 0
};
static const float3x3 hW =
{
    0,  0, 0,
    1, -1, 0,
    0,  0, 0
};
static const float3x3 hNE =
{
    0,  0, 1,
    0, -1, 0,
    0,  0, 0
};
static const float3x3 hSE =
{
    0,  0, 0,
    0, -1, 0,
    0,  0, 1
};
static const float3x3 hSW =
{
    0,  0, 0,
    0, -1, 0,
    1,  0, 0
};
static const float3x3 hNW =
{
    1,  0, 0,
    0, -1, 0,
    0,  0, 0
};

VertexShaderOutput VertexShaderFunction(VertexShaderInput vsInput)
{
    //VertexShaderOutput output;

    //output.Position = vsInput.Position;
    //output.TextureCoordinate = vsInput.TextureCoordinate;

    VertexShaderOutput output;
    vsInput.Position.x =  vsInput.Position.x - 2*halfPixel.x;
    vsInput.Position.y =  vsInput.Position.y + 2*halfPixel.y;
    output.Position = vsInput.Position;
    output.TextureCoordinate = vsInput.TextureCoordinate ;
    return output;

    //return output;
}

float4 Convolution(VertexShaderOutput input, float3x3 kernel)
{
    //PixelShaderOutput output;
    float4 pixel = float4(0.0f, 0.0f, 0.0f, 0.0f);

    for (int i = -1; i <= 1; ++i)
    {
        for (int j = -1; j <= 1; ++j)
        {
            pixel += kernel[i+1][j+1] * tex2D(Input0Sampler, input.TextureCoordinate + float2(i,j));
        };
    };

    return pixel;
}

PixelShaderOutput PixelShaderFunction(VertexShaderOutput psInput)
{       
    PixelShaderOutput output;
    output.Index0 = tex2D(Input0Sampler, psInput.TextureCoordinate);

    float dx, dy, dd;
    dx = 1; dy = 1; dd = pow(2, 0.5);
    float delta_t = 1/7;


    float4 nablaN = Convolution(psInput, hN);
    float4 nablaS = Convolution(psInput, hS);
    float4 nablaW = Convolution(psInput, hW);
    float4 nablaE = Convolution(psInput, hE);
    float4 nablaNE = Convolution(psInput, hNE);
    float4 nablaSE = Convolution(psInput, hSE);
    float4 nablaSW = Convolution(psInput, hSW);
    float4 nablaNW = Convolution(psInput, hNW);

    float4 cN  = 1 / pow( 1 + (nablaN / 40), 2);
    float4 cS  = 1 / pow( 1 + (nablaS / 40), 2);
    float4 cW  = 1 / pow( 1 + (nablaW / 40), 2);
    float4 cE  = 1 / pow( 1 + (nablaE / 40), 2);
    float4 cNE = 1 / pow( 1 + (nablaNE / 40), 2);
    float4 cSE = 1 / pow( 1 + (nablaSE / 40), 2);
    float4 cSW = 1 / pow( 1 + (nablaSW / 40), 2);
    float4 cNW = 1 / pow( 1 + (nablaNW / 40), 2);   

    output.Index0 += delta_t * 
    (
        mul(cN, nablaN) + mul(cS, nablaS) + mul(cW, nablaW) + mul(cE, nablaE) + (dd*dd)*(mul(cNE, nablaNE) + mul(cSE, nablaSE) + mul(cSW, nablaSW) + mul(cNW, nablaNW))
    );

    return output;

}

technique PeronaMalik
{
    pass pass1
    {
        VertexShader = compile vs_2_0 VertexShaderFunction();
        PixelShader  = compile ps_2_0 PixelShaderFunction();
    }
}

问题是我需要在matlab代码中连续多次应用这种技术。我是否应该使用多个通行证来做到这一点？

编辑

如果我决定使用C＃来控制传球，那么我可以尝试：

byte[] theBytes = TemplateMatch.Bytes;
for (int iters = 0; iters < 3; iters++)
{
    t.SetData<byte>(theBytes);
    GraphicsDevice.SetRenderTarget(renOutput);

    effect.Parameters["Input0"].SetValue(t);
    quad.RenderFullScreenQuad(effect);
    for (int i = 0; i < effect.Techniques.Count; i++)
    {
        for (int j = 0; j < effect.Techniques[i].Passes.Count; j++)
        {
            effect.Techniques[i].Passes[j].Apply();
        }
    }

    GraphicsDevice.SetRenderTarget(null);

    renOutput.GetData<float>(arrayOutput);
    Buffer.BlockCopy(arrayOutput, 0, theBytes, 0, theBytes.Length);
}

但是在第二次迭代中我得到一个错误

在资源上激活时，您可能无法在资源上调用SetData GraphicsDevice的。在调用SetData之前将其从设备中取消设置。

在线t.SetData（theBytes）;

修改

我试过

byte[] theBytes = TemplateMatch.Bytes;
for (int iters = 0; iters < 3; iters++)
{                    
    t.SetData<byte>(theBytes);
    GraphicsDevice.SetRenderTarget(renOutput);

    effect.Parameters["Input0"].SetValue(t);
    quad.RenderFullScreenQuad(effect);
    for (int i = 0; i < effect.Techniques.Count; i++)
    {
        for (int j = 0; j < effect.Techniques[i].Passes.Count; j++)
        {
            effect.Techniques[i].Passes[j].Apply();
        }
    }

    GraphicsDevice.SetRenderTarget(null);

    renOutput.GetData<float>(arrayOutput);
    Buffer.BlockCopy(arrayOutput, 0, theBytes, 0, theBytes.Length);
    GraphicsDevice.Textures[0] = null;
}

这似乎解决了运行时错误，但这给了我相同的未经过滤的图像！

修改

我修改了上面的matlab代码，以便它在freemat中运行。我已经逐步完成了代码，我发现问题与这样的行有关cN = exp（ - （nablaN / kappa）。^ 2）; 在freemat版本中，这些不会评估为零（而我的HLSL版本确实如此）。这让我怀疑这个问题与HLSL的精度问题或我如何在显卡上处理浮点运算有关。

Answer 1

你正在申请通行证但没有画任何东西。

至少你应该有类似的代码：

for (int i = 0; i < effect.Techniques.Count; i++)
{
    for (int j = 0; j < effect.Techniques[i].Passes.Count; j++)
    {
        effect.Techniques[i].Passes[j].Apply();
        quad.RenderFullScreenQuad(effect); 
    }
}

如何在HLSL中使用多个传递？

1 个答案: