VaryingPacking.cpp

Branch :

Show log
Commit
Author : Olli Etuaho
Date : 2018-03-20 14:44:53
Hash : 94bbed1e
Message : Collect static use information during parsing We now collect metadata for variables in the symbol table. The metadata is stored in a map using the variable unique id as a key, so we can store the variables themselves as constexpr while still having dynamic metadata. For now we collect whether a variable is statically read or written. This can be used to more accurately determine whether a variable is statically used, but can also enable more optimizations in the future, such as pruning variables that are never read or folding variables that are never written after initialization. The collection is done during parsing, so that nothing is pruned from the AST before the static use is recorded. Static writes are flagged in ParseContext::checkCanBeLValue, as that function is already called for all variables that are written. Static reads are flagged whenever there's an operation that requires a variable to be read. This includes: * Unary and binary math ops * Comma ops * Ternary ops * Assignments * Returning the variable * Passing the variable as an in or inout argument to a function * Using the variable as a constructor argument * Using the variable as an if statement condition * Using the variable as a loop condition or expression * Using the variable as an index * Using the variable as a switch statement init expression In case there are statements that simply refer to a variable without doing operations on it, the variable is being treated as statically read. Examples of such statements: my_var; my_arr[2]; These are a bit of a corner case, but it makes sense to treat them as static use for validation purposes. Collecting correct static use information costs us a bit of compiler performance, but the regression is on the order of just a few percent in the compiler perf tests. BUG=angleproject:2262 TEST=angle_unittests, angle_end2end_tests Change-Id: Ib0d7add7e4a7d11bffeb2a4861eeea982c562234 Reviewed-on: https://chromium-review.googlesource.com/977964 Reviewed-by: Corentin Wallez <cwallez@chromium.org> Commit-Queue: Olli Etuaho <oetuaho@nvidia.com>

src/libANGLE/VaryingPacking.cpp

//
// Copyright 2015 The ANGLE Project Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
//
// VaryingPacking:
//   Class which describes a mapping from varyings to registers, according
//   to the spec, or using custom packing algorithms. We also keep a register
//   allocation list for the D3D renderer.
//

#include "libANGLE/VaryingPacking.h"

#include "common/utilities.h"
#include "libANGLE/Program.h"
#include "libANGLE/Shader.h"

namespace gl
{

namespace
{

// true if varying x has a higher priority in packing than y
bool ComparePackedVarying(const PackedVarying &x, const PackedVarying &y)
{
    // If the PackedVarying 'x' or 'y' to be compared is an array element, this clones an equivalent
    // non-array shader variable 'vx' or 'vy' for actual comparison instead.
    sh::ShaderVariable vx, vy;
    const sh::ShaderVariable *px, *py;
    if (x.isArrayElement())
    {
        vx           = *x.varying;
        vx.arraySizes.clear();
        px           = &vx;
    }
    else
    {
        px = x.varying;
    }

    if (y.isArrayElement())
    {
        vy           = *y.varying;
        vy.arraySizes.clear();
        py           = &vy;
    }
    else
    {
        py = y.varying;
    }

    return gl::CompareShaderVar(*px, *py);
}

}  // anonymous namespace

// Implementation of VaryingPacking
VaryingPacking::VaryingPacking(GLuint maxVaryingVectors, PackMode packMode)
    : mRegisterMap(maxVaryingVectors), mPackMode(packMode)
{
}

VaryingPacking::~VaryingPacking() = default;

// Packs varyings into generic varying registers, using the algorithm from
// See [OpenGL ES Shading Language 1.00 rev. 17] appendix A section 7 page 111
// Also [OpenGL ES Shading Language 3.00 rev. 4] Section 11 page 119
// Returns false if unsuccessful.
bool VaryingPacking::packVarying(const PackedVarying &packedVarying)
{
    const sh::ShaderVariable &varying = *packedVarying.varying;

    // "Non - square matrices of type matCxR consume the same space as a square matrix of type matN
    // where N is the greater of C and R."
    // Here we are a bit more conservative and allow packing non-square matrices more tightly.
    // Make sure we use transposed matrix types to count registers correctly.
    ASSERT(!varying.isStruct());
    GLenum transposedType       = gl::TransposeMatrixType(varying.type);
    unsigned int varyingRows    = gl::VariableRowCount(transposedType);
    unsigned int varyingColumns = gl::VariableColumnCount(transposedType);

    // Special pack mode for D3D9. Each varying takes a full register, no sharing.
    // TODO(jmadill): Implement more sophisticated component packing in D3D9.
    if (mPackMode == PackMode::ANGLE_NON_CONFORMANT_D3D9)
    {
        varyingColumns = 4;
    }

    // "Variables of type mat2 occupies 2 complete rows."
    // For non-WebGL contexts, we allow mat2 to occupy only two columns per row.
    else if (mPackMode == PackMode::WEBGL_STRICT && varying.type == GL_FLOAT_MAT2)
    {
        varyingColumns = 4;
    }

    // "Arrays of size N are assumed to take N times the size of the base type"
    // GLSL ES 3.10 section 4.3.6: Output variables cannot be arrays of arrays or arrays of
    // structures, so we may use getBasicTypeElementCount().
    const unsigned int elementCount = varying.getBasicTypeElementCount();
    varyingRows *= (packedVarying.isArrayElement() ? 1 : elementCount);

    unsigned int maxVaryingVectors = static_cast<unsigned int>(mRegisterMap.size());

    // Fail if we are packing a single over-large varying.
    if (varyingRows > maxVaryingVectors)
    {
        return false;
    }

    // "For 2, 3 and 4 component variables packing is started using the 1st column of the 1st row.
    // Variables are then allocated to successive rows, aligning them to the 1st column."
    if (varyingColumns >= 2 && varyingColumns <= 4)
    {
        for (unsigned int row = 0; row <= maxVaryingVectors - varyingRows; ++row)
        {
            if (isFree(row, 0, varyingRows, varyingColumns))
            {
                insert(row, 0, packedVarying);
                return true;
            }
        }

        // "For 2 component variables, when there are no spare rows, the strategy is switched to
        // using the highest numbered row and the lowest numbered column where the variable will
        // fit."
        if (varyingColumns == 2)
        {
            for (unsigned int r = maxVaryingVectors - varyingRows + 1; r-- >= 1;)
            {
                if (isFree(r, 2, varyingRows, 2))
                {
                    insert(r, 2, packedVarying);
                    return true;
                }
            }
        }

        return false;
    }

    // "1 component variables have their own packing rule. They are packed in order of size, largest
    // first. Each variable is placed in the column that leaves the least amount of space in the
    // column and aligned to the lowest available rows within that column."
    ASSERT(varyingColumns == 1);
    unsigned int contiguousSpace[4]     = {0};
    unsigned int bestContiguousSpace[4] = {0};
    unsigned int totalSpace[4]          = {0};

    for (unsigned int row = 0; row < maxVaryingVectors; ++row)
    {
        for (unsigned int column = 0; column < 4; ++column)
        {
            if (mRegisterMap[row][column])
            {
                contiguousSpace[column] = 0;
            }
            else
            {
                contiguousSpace[column]++;
                totalSpace[column]++;

                if (contiguousSpace[column] > bestContiguousSpace[column])
                {
                    bestContiguousSpace[column] = contiguousSpace[column];
                }
            }
        }
    }

    unsigned int bestColumn = 0;
    for (unsigned int column = 1; column < 4; ++column)
    {
        if (bestContiguousSpace[column] >= varyingRows &&
            (bestContiguousSpace[bestColumn] < varyingRows ||
             totalSpace[column] < totalSpace[bestColumn]))
        {
            bestColumn = column;
        }
    }

    if (bestContiguousSpace[bestColumn] >= varyingRows)
    {
        for (unsigned int row = 0; row < maxVaryingVectors; row++)
        {
            if (isFree(row, bestColumn, varyingRows, 1))
            {
                for (unsigned int arrayIndex = 0; arrayIndex < varyingRows; ++arrayIndex)
                {
                    // If varyingRows > 1, it must be an array.
                    PackedVaryingRegister registerInfo;
                    registerInfo.packedVarying     = &packedVarying;
                    registerInfo.registerRow       = row + arrayIndex;
                    registerInfo.registerColumn    = bestColumn;
                    registerInfo.varyingArrayIndex =
                        (packedVarying.isArrayElement() ? packedVarying.arrayIndex : arrayIndex);
                    registerInfo.varyingRowIndex   = 0;
                    // Do not record register info for builtins.
                    // TODO(jmadill): Clean this up.
                    if (!packedVarying.varying->isBuiltIn())
                    {
                        mRegisterList.push_back(registerInfo);
                    }
                    mRegisterMap[row + arrayIndex][bestColumn] = true;
                }
                break;
            }
        }
        return true;
    }

    return false;
}

bool VaryingPacking::isFree(unsigned int registerRow,
                            unsigned int registerColumn,
                            unsigned int varyingRows,
                            unsigned int varyingColumns) const
{
    for (unsigned int row = 0; row < varyingRows; ++row)
    {
        ASSERT(registerRow + row < mRegisterMap.size());
        for (unsigned int column = 0; column < varyingColumns; ++column)
        {
            ASSERT(registerColumn + column < 4);
            if (mRegisterMap[registerRow + row][registerColumn + column])
            {
                return false;
            }
        }
    }

    return true;
}

void VaryingPacking::insert(unsigned int registerRow,
                            unsigned int registerColumn,
                            const PackedVarying &packedVarying)
{
    unsigned int varyingRows    = 0;
    unsigned int varyingColumns = 0;

    const auto &varying = *packedVarying.varying;
    ASSERT(!varying.isStruct());
    GLenum transposedType = gl::TransposeMatrixType(varying.type);
    varyingRows           = gl::VariableRowCount(transposedType);
    varyingColumns        = gl::VariableColumnCount(transposedType);

    PackedVaryingRegister registerInfo;
    registerInfo.packedVarying  = &packedVarying;
    registerInfo.registerColumn = registerColumn;

    // GLSL ES 3.10 section 4.3.6: Output variables cannot be arrays of arrays or arrays of
    // structures, so we may use getBasicTypeElementCount().
    const unsigned int arrayElementCount = varying.getBasicTypeElementCount();
    for (unsigned int arrayElement = 0; arrayElement < arrayElementCount; ++arrayElement)
    {
        if (packedVarying.isArrayElement() && arrayElement != packedVarying.arrayIndex)
        {
            continue;
        }
        for (unsigned int varyingRow = 0; varyingRow < varyingRows; ++varyingRow)
        {
            registerInfo.registerRow     = registerRow + (arrayElement * varyingRows) + varyingRow;
            registerInfo.varyingRowIndex = varyingRow;
            registerInfo.varyingArrayIndex = arrayElement;
            // Do not record register info for builtins.
            // TODO(jmadill): Clean this up.
            if (!packedVarying.varying->isBuiltIn())
            {
                mRegisterList.push_back(registerInfo);
            }

            for (unsigned int columnIndex = 0; columnIndex < varyingColumns; ++columnIndex)
            {
                mRegisterMap[registerInfo.registerRow][registerColumn + columnIndex] = true;
            }
        }
    }
}

bool VaryingPacking::collectAndPackUserVaryings(gl::InfoLog &infoLog,
                                                const ProgramMergedVaryings &mergedVaryings,
                                                const std::vector<std::string> &tfVaryings)
{
    std::set<std::string> uniqueFullNames;
    mPackedVaryings.clear();

    for (const auto &ref : mergedVaryings)
    {
        const sh::Varying *input  = ref.second.vertex;
        const sh::Varying *output = ref.second.fragment;

        // Only pack statically used varyings that have a matched input or output, plus special
        // builtins. Note that we pack all statically used user-defined varyings even if they are
        // not active. GLES specs are a bit vague on whether it's allowed to only pack active
        // varyings, though GLES 3.1 spec section 11.1.2.1 says that "device-dependent
        // optimizations" may be used to make vertex shader outputs fit.
        if ((input && output && output->staticUse) ||
            (input && input->isBuiltIn() && input->active) ||
            (output && output->isBuiltIn() && output->active))
        {
            const sh::Varying *varying = output ? output : input;

            // Don't count gl_Position. Also don't count gl_PointSize for D3D9.
            if (varying->name != "gl_Position" &&
                !(varying->name == "gl_PointSize" &&
                  mPackMode == PackMode::ANGLE_NON_CONFORMANT_D3D9))
            {
                // Will get the vertex shader interpolation by default.
                auto interpolation = ref.second.get()->interpolation;

                // Note that we lose the vertex shader static use information here. The data for the
                // variable is taken from the fragment shader.
                if (varying->isStruct())
                {
                    ASSERT(!varying->isArray());
                    for (const auto &field : varying->fields)
                    {
                        ASSERT(!field.isStruct() && !field.isArray());
                        mPackedVaryings.push_back(
                            PackedVarying(field, interpolation, varying->name));
                        uniqueFullNames.insert(mPackedVaryings.back().fullName());
                    }
                }
                else
                {
                    mPackedVaryings.push_back(PackedVarying(*varying, interpolation));
                    uniqueFullNames.insert(mPackedVaryings.back().fullName());
                }
                continue;
            }
        }

        // Keep Transform FB varyings in the merged list always.
        if (!input)
        {
            continue;
        }

        for (const std::string &tfVarying : tfVaryings)
        {
            std::vector<unsigned int> subscripts;
            std::string baseName = ParseResourceName(tfVarying, &subscripts);
            size_t subscript     = GL_INVALID_INDEX;
            if (!subscripts.empty())
            {
                subscript = subscripts.back();
            }
            // Already packed for fragment shader.
            if (uniqueFullNames.count(tfVarying) > 0 || uniqueFullNames.count(baseName) > 0)
            {
                continue;
            }
            if (input->isStruct())
            {
                const sh::ShaderVariable *field = FindShaderVarField(*input, tfVarying);
                if (field != nullptr)
                {
                    ASSERT(!field->isStruct() && !field->isArray());
                    mPackedVaryings.emplace_back(*field, input->interpolation, input->name);
                    mPackedVaryings.back().vertexOnly = true;
                    mPackedVaryings.back().arrayIndex = GL_INVALID_INDEX;
                    uniqueFullNames.insert(tfVarying);
                }
            }
            // Array as a whole and array element conflict has already been checked in
            // linkValidateTransformFeedback.
            else if (baseName == input->name)
            {
                // only pack varyings that are not builtins.
                if (tfVarying.compare(0, 3, "gl_") != 0)
                {
                    mPackedVaryings.emplace_back(*input, input->interpolation);
                    mPackedVaryings.back().vertexOnly = true;
                    mPackedVaryings.back().arrayIndex = static_cast<GLuint>(subscript);
                    uniqueFullNames.insert(tfVarying);
                }
                // Continue to match next array element for 'input' if the current match is array
                // element.
                if (subscript == GL_INVALID_INDEX)
                {
                    break;
                }
            }
        }
    }

    std::sort(mPackedVaryings.begin(), mPackedVaryings.end(), ComparePackedVarying);

    return packUserVaryings(infoLog, mPackedVaryings, tfVaryings);
}

// See comment on packVarying.
bool VaryingPacking::packUserVaryings(gl::InfoLog &infoLog,
                                      const std::vector<PackedVarying> &packedVaryings,
                                      const std::vector<std::string> &transformFeedbackVaryings)
{

    // "Variables are packed into the registers one at a time so that they each occupy a contiguous
    // subrectangle. No splitting of variables is permitted."
    for (const PackedVarying &packedVarying : packedVaryings)
    {
        if (!packVarying(packedVarying))
        {
            infoLog << "Could not pack varying " << packedVarying.fullName();

            // TODO(jmadill): Implement more sophisticated component packing in D3D9.
            if (mPackMode == PackMode::ANGLE_NON_CONFORMANT_D3D9)
            {
                infoLog << "Note: Additional non-conformant packing restrictions are enforced on "
                           "D3D9.";
            }

            return false;
        }
    }

    // Sort the packed register list
    std::sort(mRegisterList.begin(), mRegisterList.end());

    // Assign semantic indices
    for (unsigned int semanticIndex = 0;
         semanticIndex < static_cast<unsigned int>(mRegisterList.size()); ++semanticIndex)
    {
        mRegisterList[semanticIndex].semanticIndex = semanticIndex;
    }

    return true;
}

}  // namespace rx

kc3-lang/angle/src/libANGLE/VaryingPacking.cpp

Commit

kc3-lang/angle /src/libANGLE/VaryingPacking.cpp