crossxtex/DirectXTex/BCDirectCompute.cpp

//-------------------------------------------------------------------------------------
// BCDirectCompute.cpp
//  
// Direct3D 11 Compute Shader BC Compressor
//
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
//-------------------------------------------------------------------------------------

#include "DirectXTexP.h"

#include "BCDirectCompute.h"

#if defined(_DEBUG) || defined(PROFILE)
#pragma comment(lib,"dxguid.lib")
#endif

using namespace DirectX;
using Microsoft::WRL::ComPtr;

namespace
{
    #include "Shaders\Compiled\BC7Encode_EncodeBlockCS.inc"
    #include "Shaders\Compiled\BC7Encode_TryMode02CS.inc"
    #include "Shaders\Compiled\BC7Encode_TryMode137CS.inc"
    #include "Shaders\Compiled\BC7Encode_TryMode456CS.inc"
    #include "Shaders\Compiled\BC6HEncode_EncodeBlockCS.inc"
    #include "Shaders\Compiled\BC6HEncode_TryModeG10CS.inc"
    #include "Shaders\Compiled\BC6HEncode_TryModeLE10CS.inc"

    struct BufferBC6HBC7
    {
        UINT color[4];
    };

    struct ConstantsBC6HBC7
    {
        UINT    tex_width;
        UINT    num_block_x;
        UINT    format;
        UINT    mode_id;
        UINT    start_block_id;
        UINT    num_total_blocks;
        float   alpha_weight;
        UINT    reserved;
    };

    static_assert(sizeof(ConstantsBC6HBC7) == sizeof(UINT) * 8, "Constant buffer size mismatch");

    inline void RunComputeShader(ID3D11DeviceContext* pContext,
        ID3D11ComputeShader* shader,
        ID3D11ShaderResourceView** pSRVs,
        UINT srvCount,
        ID3D11Buffer* pCB,
        ID3D11UnorderedAccessView* pUAV,
        UINT X)
    {
        // Force UAV to nullptr before setting SRV since we are swapping buffers
        ID3D11UnorderedAccessView* nullUAV = nullptr;
        pContext->CSSetUnorderedAccessViews(0, 1, &nullUAV, nullptr);

        pContext->CSSetShader(shader, nullptr, 0);
        pContext->CSSetShaderResources(0, srvCount, pSRVs);
        pContext->CSSetUnorderedAccessViews(0, 1, &pUAV, nullptr);
        pContext->CSSetConstantBuffers(0, 1, &pCB);
        pContext->Dispatch(X, 1, 1);
    }

    inline void ResetContext(ID3D11DeviceContext* pContext)
    {
        ID3D11UnorderedAccessView* nullUAV = nullptr;
        pContext->CSSetUnorderedAccessViews(0, 1, &nullUAV, nullptr);

        ID3D11ShaderResourceView* nullSRV[3] = { nullptr, nullptr, nullptr };
        pContext->CSSetShaderResources(0, 3, nullSRV);

        ID3D11Buffer* nullBuffer[1] = { nullptr };
        pContext->CSSetConstantBuffers(0, 1, nullBuffer);
    }
};

GPUCompressBC::GPUCompressBC() noexcept :
    m_bcformat(DXGI_FORMAT_UNKNOWN),
    m_srcformat(DXGI_FORMAT_UNKNOWN),
    m_alphaWeight(1.f),
    m_bc7_mode02(false),
    m_bc7_mode137(false),
    m_width(0),
    m_height(0)
{
}


//-------------------------------------------------------------------------------------
_Use_decl_annotations_
HRESULT GPUCompressBC::Initialize(ID3D11Device* pDevice)
{
    if (!pDevice)
        return E_INVALIDARG;

    // Check for DirectCompute support
    D3D_FEATURE_LEVEL fl = pDevice->GetFeatureLevel();

    if (fl < D3D_FEATURE_LEVEL_10_0)
    {
        // DirectCompute not supported on Feature Level 9.x hardware
        return HRESULT_FROM_WIN32(ERROR_NOT_SUPPORTED);
    }

    if (fl < D3D_FEATURE_LEVEL_11_0)
    {
        // DirectCompute support on Feature Level 10.x hardware is optional, and this function needs it
        D3D11_FEATURE_DATA_D3D10_X_HARDWARE_OPTIONS hwopts;
        HRESULT hr = pDevice->CheckFeatureSupport(D3D11_FEATURE_D3D10_X_HARDWARE_OPTIONS, &hwopts, sizeof(hwopts));
        if (FAILED(hr))
        {
            memset(&hwopts, 0, sizeof(hwopts));
        }

        if (!hwopts.ComputeShaders_Plus_RawAndStructuredBuffers_Via_Shader_4_x)
        {
            return HRESULT_FROM_WIN32(ERROR_NOT_SUPPORTED);
        }
    }

    // Save a device reference and obtain immediate context
    m_device = pDevice;

    pDevice->GetImmediateContext(m_context.ReleaseAndGetAddressOf());
    assert(m_context);

    //--- Create compute shader library: BC6H -----------------------------------------

    // Modes 11-14
    HRESULT hr = pDevice->CreateComputeShader(BC6HEncode_TryModeG10CS, sizeof(BC6HEncode_TryModeG10CS), nullptr, m_BC6H_tryModeG10CS.ReleaseAndGetAddressOf());
    if (FAILED(hr))
        return hr;

    // Modes 1-10
    hr = pDevice->CreateComputeShader(BC6HEncode_TryModeLE10CS, sizeof(BC6HEncode_TryModeLE10CS), nullptr, m_BC6H_tryModeLE10CS.ReleaseAndGetAddressOf());
    if (FAILED(hr))
        return hr;

    // Encode
    hr = pDevice->CreateComputeShader(BC6HEncode_EncodeBlockCS, sizeof(BC6HEncode_EncodeBlockCS), nullptr, m_BC6H_encodeBlockCS.ReleaseAndGetAddressOf());
    if (FAILED(hr))
        return hr;

    //--- Create compute shader library: BC7 ------------------------------------------

    // Modes 4, 5, 6
    hr = pDevice->CreateComputeShader(BC7Encode_TryMode456CS, sizeof(BC7Encode_TryMode456CS), nullptr, m_BC7_tryMode456CS.ReleaseAndGetAddressOf());
    if (FAILED(hr))
        return hr;

    // Modes 1, 3, 7
    hr = pDevice->CreateComputeShader(BC7Encode_TryMode137CS, sizeof(BC7Encode_TryMode137CS), nullptr, m_BC7_tryMode137CS.ReleaseAndGetAddressOf());
    if (FAILED(hr))
        return hr;

    // Modes 0, 2
    hr = pDevice->CreateComputeShader(BC7Encode_TryMode02CS, sizeof(BC7Encode_TryMode02CS), nullptr, m_BC7_tryMode02CS.ReleaseAndGetAddressOf());
    if (FAILED(hr))
        return hr;

    // Encode
    hr = pDevice->CreateComputeShader(BC7Encode_EncodeBlockCS, sizeof(BC7Encode_EncodeBlockCS), nullptr, m_BC7_encodeBlockCS.ReleaseAndGetAddressOf());
    if (FAILED(hr))
        return hr;

    return S_OK;
}


//-------------------------------------------------------------------------------------
HRESULT GPUCompressBC::Prepare(size_t width, size_t height, uint32_t flags, DXGI_FORMAT format, float alphaWeight)
{
    if (!width || !height || alphaWeight < 0.f)
        return E_INVALIDARG;

    if ((width > UINT32_MAX) || (height > UINT32_MAX))
        return E_INVALIDARG;

    m_width = width;
    m_height = height;

    m_alphaWeight = alphaWeight;

    if (flags & TEX_COMPRESS_BC7_QUICK)
    {
        m_bc7_mode02 = false;
        m_bc7_mode137 = false;
    }
    else
    {
        m_bc7_mode02 = (flags & TEX_COMPRESS_BC7_USE_3SUBSETS) != 0;
        m_bc7_mode137 = true;
    }

    size_t xblocks = std::max<size_t>(1, (width + 3) >> 2);
    size_t yblocks = std::max<size_t>(1, (height + 3) >> 2);
    size_t num_blocks = xblocks * yblocks;

    switch (format)
    {
        // BC6H GPU compressor takes RGBAF32 as input
    case DXGI_FORMAT_BC6H_TYPELESS:
    case DXGI_FORMAT_BC6H_UF16:
    case DXGI_FORMAT_BC6H_SF16:
        m_srcformat = DXGI_FORMAT_R32G32B32A32_FLOAT;
        break;

        // BC7 GPU compressor takes RGBA32 as input
    case DXGI_FORMAT_BC7_TYPELESS:
    case DXGI_FORMAT_BC7_UNORM:
        m_srcformat = DXGI_FORMAT_R8G8B8A8_UNORM;
        break;

    case DXGI_FORMAT_BC7_UNORM_SRGB:
        m_srcformat = DXGI_FORMAT_R8G8B8A8_UNORM_SRGB;
        break;

    default:
        m_bcformat = m_srcformat = DXGI_FORMAT_UNKNOWN;
        return HRESULT_FROM_WIN32(ERROR_NOT_SUPPORTED);
    }

    m_bcformat = format;

    auto pDevice = m_device.Get();
    if (!pDevice)
        return E_POINTER;

    // Create structured buffers
    uint64_t sizeInBytes = uint64_t(num_blocks) * sizeof(BufferBC6HBC7);
    if (sizeInBytes >= UINT32_MAX)
        return HRESULT_FROM_WIN32(ERROR_ARITHMETIC_OVERFLOW);

    auto bufferSize = static_cast<size_t>(sizeInBytes);

    {
        D3D11_BUFFER_DESC desc = {};
        desc.BindFlags = D3D11_BIND_UNORDERED_ACCESS | D3D11_BIND_SHADER_RESOURCE;
        desc.Usage = D3D11_USAGE_DEFAULT;
        desc.MiscFlags = D3D11_RESOURCE_MISC_BUFFER_STRUCTURED;
        desc.StructureByteStride = sizeof(BufferBC6HBC7);
        desc.ByteWidth = static_cast<UINT>(bufferSize);

        HRESULT hr = pDevice->CreateBuffer(&desc, nullptr, m_output.ReleaseAndGetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }

        hr = pDevice->CreateBuffer(&desc, nullptr, m_err1.ReleaseAndGetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }

        hr = pDevice->CreateBuffer(&desc, nullptr, m_err2.ReleaseAndGetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }
    }

    // Create staging output buffer
    {
        D3D11_BUFFER_DESC desc = {};
        desc.Usage = D3D11_USAGE_STAGING;
        desc.CPUAccessFlags = D3D11_CPU_ACCESS_READ;
        desc.ByteWidth = static_cast<UINT>(bufferSize);

        HRESULT hr = pDevice->CreateBuffer(&desc, nullptr, m_outputCPU.ReleaseAndGetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }
    }

    // Create constant buffer
    {
        D3D11_BUFFER_DESC desc = {};
        desc.BindFlags = D3D11_BIND_CONSTANT_BUFFER;
        desc.Usage = D3D11_USAGE_DYNAMIC;
        desc.CPUAccessFlags = D3D11_CPU_ACCESS_WRITE;
        desc.ByteWidth = sizeof(ConstantsBC6HBC7);

        HRESULT hr = pDevice->CreateBuffer(&desc, nullptr, m_constBuffer.ReleaseAndGetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }
    }

    // Create shader resource views
    {
        D3D11_SHADER_RESOURCE_VIEW_DESC desc = {};
        desc.Buffer.NumElements = static_cast<UINT>(num_blocks);
        desc.ViewDimension = D3D11_SRV_DIMENSION_BUFFER;

        HRESULT hr = pDevice->CreateShaderResourceView(m_err1.Get(), &desc, m_err1SRV.ReleaseAndGetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }

        hr = pDevice->CreateShaderResourceView(m_err2.Get(), &desc, m_err2SRV.ReleaseAndGetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }
    }

    // Create unordered access views
    {
        D3D11_UNORDERED_ACCESS_VIEW_DESC desc = {};
        desc.Buffer.NumElements = static_cast<UINT>(num_blocks);
        desc.ViewDimension = D3D11_UAV_DIMENSION_BUFFER;

        HRESULT hr = pDevice->CreateUnorderedAccessView(m_output.Get(), &desc, m_outputUAV.ReleaseAndGetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }

        hr = pDevice->CreateUnorderedAccessView(m_err1.Get(), &desc, m_err1UAV.ReleaseAndGetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }

        hr = pDevice->CreateUnorderedAccessView(m_err2.Get(), &desc, m_err2UAV.ReleaseAndGetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }
    }

    return S_OK;
}


//-------------------------------------------------------------------------------------
HRESULT GPUCompressBC::Compress(const Image& srcImage, const Image& destImage)
{
    if (!srcImage.pixels || !destImage.pixels)
        return E_INVALIDARG;

    if (srcImage.width != destImage.width
        || srcImage.height != destImage.height
        || srcImage.width != m_width
        || srcImage.height != m_height
        || srcImage.format != m_srcformat
        || destImage.format != m_bcformat)
    {
        return E_UNEXPECTED;
    }

    //--- Create input texture --------------------------------------------------------
    auto pDevice = m_device.Get();
    if (!pDevice)
        return E_POINTER;

    // We need to avoid the hardware doing additional colorspace conversion
    DXGI_FORMAT inputFormat = (m_srcformat == DXGI_FORMAT_R8G8B8A8_UNORM_SRGB) ? DXGI_FORMAT_R8G8B8A8_UNORM : m_srcformat;

    ComPtr<ID3D11Texture2D> sourceTex;
    {
        D3D11_TEXTURE2D_DESC desc = {};
        desc.Width = static_cast<UINT>(srcImage.width);
        desc.Height = static_cast<UINT>(srcImage.height);
        desc.MipLevels = 1;
        desc.ArraySize = 1;
        desc.Format = inputFormat;
        desc.SampleDesc.Count = 1;
        desc.Usage = D3D11_USAGE_DEFAULT;
        desc.BindFlags = D3D11_BIND_SHADER_RESOURCE;

        D3D11_SUBRESOURCE_DATA initData;
        initData.pSysMem = srcImage.pixels;
        initData.SysMemPitch = static_cast<DWORD>(srcImage.rowPitch);
        initData.SysMemSlicePitch = static_cast<DWORD>(srcImage.slicePitch);

        HRESULT hr = pDevice->CreateTexture2D(&desc, &initData, sourceTex.GetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }
    }

    ComPtr<ID3D11ShaderResourceView> sourceSRV;
    {
        D3D11_SHADER_RESOURCE_VIEW_DESC desc = {};
        desc.Texture2D.MipLevels = 1;
        desc.Format = inputFormat;
        desc.ViewDimension = D3D11_SRV_DIMENSION_TEXTURE2D;

        HRESULT hr = pDevice->CreateShaderResourceView(sourceTex.Get(), &desc, sourceSRV.GetAddressOf());
        if (FAILED(hr))
        {
            return hr;
        }
    }

    //--- Compress using DirectCompute ------------------------------------------------
    bool isbc7 = false;
    switch (m_bcformat)
    {
    case DXGI_FORMAT_BC6H_TYPELESS:
    case DXGI_FORMAT_BC6H_UF16:
    case DXGI_FORMAT_BC6H_SF16:
        break;

    case DXGI_FORMAT_BC7_TYPELESS:
    case DXGI_FORMAT_BC7_UNORM:
    case DXGI_FORMAT_BC7_UNORM_SRGB:
        isbc7 = true;
        break;

    default:
        return E_UNEXPECTED;
    }

    const UINT MAX_BLOCK_BATCH = 64;

    auto pContext = m_context.Get();
    if (!pContext)
        return E_UNEXPECTED;

    size_t xblocks = std::max<size_t>(1, (m_width + 3) >> 2);
    size_t yblocks = std::max<size_t>(1, (m_height + 3) >> 2);

    auto num_total_blocks = static_cast<UINT>(xblocks * yblocks);
    UINT num_blocks = num_total_blocks;
    UINT start_block_id = 0;
    while (num_blocks > 0)
    {
        UINT n = std::min<UINT>(num_blocks, MAX_BLOCK_BATCH);
        UINT uThreadGroupCount = n;

        {
            D3D11_MAPPED_SUBRESOURCE mapped;
            HRESULT hr = pContext->Map(m_constBuffer.Get(), 0, D3D11_MAP_WRITE_DISCARD, 0, &mapped);
            if (FAILED(hr))
                return hr;

            ConstantsBC6HBC7 param;
            param.tex_width = static_cast<UINT>(srcImage.width);
            param.num_block_x = static_cast<UINT>(xblocks);
            param.format = static_cast<UINT>(m_bcformat);
            param.mode_id = 0;
            param.start_block_id = start_block_id;
            param.num_total_blocks = num_total_blocks;
            param.alpha_weight = m_alphaWeight;
            memcpy(mapped.pData, &param, sizeof(param));

            pContext->Unmap(m_constBuffer.Get(), 0);
        }

        if (isbc7)
        {
            //--- BC7 -----------------------------------------------------------------
            ID3D11ShaderResourceView* pSRVs[] = { sourceSRV.Get(), nullptr };
            RunComputeShader(pContext, m_BC7_tryMode456CS.Get(), pSRVs, 2, m_constBuffer.Get(),
                m_err1UAV.Get(), std::max<UINT>((uThreadGroupCount + 3) / 4, 1));

            if (m_bc7_mode137)
            {
                for (UINT i = 0; i < 3; ++i)
                {
                    static const UINT modes[] = { 1, 3, 7 };

                    // Mode 1: err1 -> err2
                    // Mode 3: err2 -> err1
                    // Mode 7: err1 -> err2
                    {
                        D3D11_MAPPED_SUBRESOURCE mapped;
                        HRESULT hr = pContext->Map(m_constBuffer.Get(), 0, D3D11_MAP_WRITE_DISCARD, 0, &mapped);
                        if (FAILED(hr))
                        {
                            ResetContext(pContext);
                            return hr;
                        }

                        ConstantsBC6HBC7 param;
                        param.tex_width = static_cast<UINT>(srcImage.width);
                        param.num_block_x = static_cast<UINT>(xblocks);
                        param.format = static_cast<UINT>(m_bcformat);
                        param.mode_id = modes[i];
                        param.start_block_id = start_block_id;
                        param.num_total_blocks = num_total_blocks;
                        param.alpha_weight = m_alphaWeight;
                        memcpy(mapped.pData, &param, sizeof(param));
                        pContext->Unmap(m_constBuffer.Get(), 0);
                    }

                    pSRVs[1] = (i & 1) ? m_err2SRV.Get() : m_err1SRV.Get();
                    RunComputeShader(pContext, m_BC7_tryMode137CS.Get(), pSRVs, 2, m_constBuffer.Get(),
                        (i & 1) ? m_err1UAV.Get() : m_err2UAV.Get(), uThreadGroupCount);
                }
            }

            if (m_bc7_mode02)
            {
                // 3 subset modes tend to be used rarely and add significant compression time
                for (UINT i = 0; i < 2; ++i)
                {
                    static const UINT modes[] = { 0, 2 };
                    // Mode 0: err2 -> err1
                    // Mode 2: err1 -> err2
                    {
                        D3D11_MAPPED_SUBRESOURCE mapped;
                        HRESULT hr = pContext->Map(m_constBuffer.Get(), 0, D3D11_MAP_WRITE_DISCARD, 0, &mapped);
                        if (FAILED(hr))
                        {
                            ResetContext(pContext);
                            return hr;
                        }

                        ConstantsBC6HBC7 param;
                        param.tex_width = static_cast<UINT>(srcImage.width);
                        param.num_block_x = static_cast<UINT>(xblocks);
                        param.format = static_cast<UINT>(m_bcformat);
                        param.mode_id = modes[i];
                        param.start_block_id = start_block_id;
                        param.num_total_blocks = num_total_blocks;
                        param.alpha_weight = m_alphaWeight;
                        memcpy(mapped.pData, &param, sizeof(param));
                        pContext->Unmap(m_constBuffer.Get(), 0);
                    }

                    pSRVs[1] = (i & 1) ? m_err1SRV.Get() : m_err2SRV.Get();
                    RunComputeShader(pContext, m_BC7_tryMode02CS.Get(), pSRVs, 2, m_constBuffer.Get(),
                        (i & 1) ? m_err2UAV.Get() : m_err1UAV.Get(), uThreadGroupCount);
                }
            }

            pSRVs[1] = (m_bc7_mode02 || m_bc7_mode137) ? m_err2SRV.Get() : m_err1SRV.Get();
            RunComputeShader(pContext, m_BC7_encodeBlockCS.Get(), pSRVs, 2, m_constBuffer.Get(),
                m_outputUAV.Get(), std::max<UINT>((uThreadGroupCount + 3) / 4, 1));
        }
        else
        {
            //--- BC6H ----------------------------------------------------------------
            ID3D11ShaderResourceView* pSRVs[] = { sourceSRV.Get(), nullptr };
            RunComputeShader(pContext, m_BC6H_tryModeG10CS.Get(), pSRVs, 2, m_constBuffer.Get(),
                m_err1UAV.Get(), std::max<UINT>((uThreadGroupCount + 3) / 4, 1));

            for (UINT i = 0; i < 10; ++i)
            {
                {
                    D3D11_MAPPED_SUBRESOURCE mapped;
                    HRESULT hr = pContext->Map(m_constBuffer.Get(), 0, D3D11_MAP_WRITE_DISCARD, 0, &mapped);
                    if (FAILED(hr))
                    {
                        ResetContext(pContext);
                        return hr;
                    }

                    ConstantsBC6HBC7 param;
                    param.tex_width = static_cast<UINT>(srcImage.width);
                    param.num_block_x = static_cast<UINT>(xblocks);
                    param.format = static_cast<UINT>(m_bcformat);
                    param.mode_id = i;
                    param.start_block_id = start_block_id;
                    param.num_total_blocks = num_total_blocks;
                    memcpy(mapped.pData, &param, sizeof(param));
                    pContext->Unmap(m_constBuffer.Get(), 0);
                }

                pSRVs[1] = (i & 1) ? m_err2SRV.Get() : m_err1SRV.Get();
                RunComputeShader(pContext, m_BC6H_tryModeLE10CS.Get(), pSRVs, 2, m_constBuffer.Get(),
                    (i & 1) ? m_err1UAV.Get() : m_err2UAV.Get(), std::max<UINT>((uThreadGroupCount + 1) / 2, 1));
            }

            pSRVs[1] = m_err1SRV.Get();
            RunComputeShader(pContext, m_BC6H_encodeBlockCS.Get(), pSRVs, 2, m_constBuffer.Get(),
                m_outputUAV.Get(), std::max<UINT>((uThreadGroupCount + 1) / 2, 1));
        }

        start_block_id += n;
        num_blocks -= n;
    }

    ResetContext(pContext);

    //--- Copy output texture back to CPU ---------------------------------------------

    pContext->CopyResource(m_outputCPU.Get(), m_output.Get());

    D3D11_MAPPED_SUBRESOURCE mapped;
    HRESULT hr = pContext->Map(m_outputCPU.Get(), 0, D3D11_MAP_READ, 0, &mapped);
    if (SUCCEEDED(hr))
    {
        auto pSrc = static_cast<const uint8_t *>(mapped.pData);
        uint8_t *pDest = destImage.pixels;

        size_t pitch = xblocks * sizeof(BufferBC6HBC7);

        size_t rows = std::max<size_t>(1, (destImage.height + 3) >> 2);

        for (size_t h = 0; h < rows; ++h)
        {
            memcpy(pDest, pSrc, destImage.rowPitch);

            pSrc += pitch;
            pDest += destImage.rowPitch;
        }

        pContext->Unmap(m_outputCPU.Get(), 0);
    }

    return hr;
}