gglassembler.cpp

Android 一些工具

/* libs/pixelflinger/codeflinger/GGLAssembler.cpp
**
** Copyright 2006, The Android Open Source Project
**
** Licensed under the Apache License, Version 2.0 (the "License"); 
** you may not use this file except in compliance with the License. 
** You may obtain a copy of the License at 
**
**     http://www.apache.org/licenses/LICENSE-2.0 
**
** Unless required by applicable law or agreed to in writing, software 
** distributed under the License is distributed on an "AS IS" BASIS, 
** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 
** See the License for the specific language governing permissions and 
** limitations under the License.
*/

#define LOG_TAG "GGLAssembler"

#include <assert.h>
#include <stdint.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/types.h>
#include <cutils/log.h>

#include "codeflinger/GGLAssembler.h"

namespace android {

// ----------------------------------------------------------------------------

GGLAssembler::GGLAssembler(ARMAssemblerInterface* target)
    : ARMAssemblerProxy(target), RegisterAllocator(), mOptLevel(7)
{
}

GGLAssembler::~GGLAssembler()
{
}

void GGLAssembler::prolog()
{
    ARMAssemblerProxy::prolog();
}

void GGLAssembler::epilog(uint32_t touched)
{
    ARMAssemblerProxy::epilog(touched);
}

void GGLAssembler::reset(int opt_level)
{
    ARMAssemblerProxy::reset();
    RegisterAllocator::reset();
    mOptLevel = opt_level;
}

// ---------------------------------------------------------------------------

int GGLAssembler::scanline(const needs_t& needs, context_t const* c)
{
    int err = 0;
    int opt_level = mOptLevel;
    while (opt_level >= 0) {
        reset(opt_level);
        err = scanline_core(needs, c);
        if (err == 0)
            break;
        opt_level--;
    }
    
    // XXX: in theory, pcForLabel is not valid before generate()
    uint32_t* fragment_start_pc = pcForLabel("fragment_loop");
    uint32_t* fragment_end_pc = pcForLabel("epilog");
    const int per_fragment_ops = int(fragment_end_pc - fragment_start_pc);
    
    // build a name for our pipeline
    char name[64];    
    sprintf(name,
            "scanline__%08X:%08X_%08X_%08X [%3d ipp]",
            needs.p, needs.n, needs.t[0], needs.t[1], per_fragment_ops);

    if (err) {
        LOGE("Error while generating ""%s""\n", name);
        disassemble(name);
        return -1;
    }

    return generate(name);
}

int GGLAssembler::scanline_core(const needs_t& needs, context_t const* c)
{
    int64_t duration = ggl_system_time();

    mBlendFactorCached = 0;
    mBlending = 0;
    mMasking = 0;
    mAA        = GGL_READ_NEEDS(P_AA, needs.p);
    mDithering = GGL_READ_NEEDS(P_DITHER, needs.p);
    mAlphaTest = GGL_READ_NEEDS(P_ALPHA_TEST, needs.p) + GGL_NEVER;
    mDepthTest = GGL_READ_NEEDS(P_DEPTH_TEST, needs.p) + GGL_NEVER;
    mFog       = GGL_READ_NEEDS(P_FOG, needs.p) != 0;
    mSmooth    = GGL_READ_NEEDS(SHADE, needs.n) != 0;
    mBuilderContext.needs = needs;
    mBuilderContext.c = c;
    mBuilderContext.Rctx = reserveReg(R0); // context always in R0
    mCbFormat = c->formats[ GGL_READ_NEEDS(CB_FORMAT, needs.n) ];

    // ------------------------------------------------------------------------

    decodeLogicOpNeeds(needs);

    decodeTMUNeeds(needs, c);

    mBlendSrc  = ggl_needs_to_blendfactor(GGL_READ_NEEDS(BLEND_SRC, needs.n));
    mBlendDst  = ggl_needs_to_blendfactor(GGL_READ_NEEDS(BLEND_DST, needs.n));
    mBlendSrcA = ggl_needs_to_blendfactor(GGL_READ_NEEDS(BLEND_SRCA, needs.n));
    mBlendDstA = ggl_needs_to_blendfactor(GGL_READ_NEEDS(BLEND_DSTA, needs.n));

    if (!mCbFormat.c[GGLFormat::ALPHA].h) {
        if ((mBlendSrc == GGL_ONE_MINUS_DST_ALPHA) ||
            (mBlendSrc == GGL_DST_ALPHA)) {
            mBlendSrc = GGL_ONE;
        }
        if ((mBlendSrcA == GGL_ONE_MINUS_DST_ALPHA) ||
            (mBlendSrcA == GGL_DST_ALPHA)) {
            mBlendSrcA = GGL_ONE;
        }
        if ((mBlendDst == GGL_ONE_MINUS_DST_ALPHA) ||
            (mBlendDst == GGL_DST_ALPHA)) {
            mBlendDst = GGL_ONE;
        }
        if ((mBlendDstA == GGL_ONE_MINUS_DST_ALPHA) ||
            (mBlendDstA == GGL_DST_ALPHA)) {
            mBlendDstA = GGL_ONE;
        }
    }

    // if we need the framebuffer, read it now
    const int blending =    blending_codes(mBlendSrc, mBlendDst) |
                            blending_codes(mBlendSrcA, mBlendDstA);

    // XXX: handle special cases, destination not modified...
    if ((mBlendSrc==GGL_ZERO) && (mBlendSrcA==GGL_ZERO) &&
        (mBlendDst==GGL_ONE) && (mBlendDstA==GGL_ONE)) {
        // Destination unmodified (beware of logic ops)
    } else if ((mBlendSrc==GGL_ZERO) && (mBlendSrcA==GGL_ZERO) &&
        (mBlendDst==GGL_ZERO) && (mBlendDstA==GGL_ZERO)) {
        // Destination is zero (beware of logic ops)
    }
    
    const int masking = GGL_READ_NEEDS(MASK_ARGB, needs.n);
    for (int i=0 ; i<4 ; i++) {
        const int mask = 1<<i;
        component_info_t& info = mInfo[i];
        int fs = i==GGLFormat::ALPHA ? mBlendSrcA : mBlendSrc;
        int fd = i==GGLFormat::ALPHA ? mBlendDstA : mBlendDst;
        if (fs==GGL_SRC_ALPHA_SATURATE && i==GGLFormat::ALPHA)
            fs = GGL_ONE;
        info.masked =   !!(masking & mask);
        info.inDest =   !info.masked && mCbFormat.c[i].h && 
                        ((mLogicOp & LOGIC_OP_SRC) || (!mLogicOp));
        if (mCbFormat.components >= GGL_LUMINANCE &&
                (i==GGLFormat::GREEN || i==GGLFormat::BLUE)) {
            info.inDest = false;
        }
        info.needed =   (i==GGLFormat::ALPHA) && 
                        (isAlphaSourceNeeded() || mAlphaTest != GGL_ALWAYS);
        info.replaced = !!(mTextureMachine.replaced & mask);
        info.iterated = (!info.replaced && (info.inDest || info.needed)); 
        info.smooth =   mSmooth && info.iterated;
        info.fog =      mFog && info.inDest && (i != GGLFormat::ALPHA);
        info.blend =    (fs != int(GGL_ONE)) || (fd > int(GGL_ZERO));

        mBlending |= (info.blend ? mask : 0);
        mMasking |= (mCbFormat.c[i].h && info.masked) ? mask : 0;
    }


    fragment_parts_t parts;

    // ------------------------------------------------------------------------
    prolog();
    // ------------------------------------------------------------------------

    build_scanline_prolog(parts, needs);

    if (registerFile().status())
        return registerFile().status();

    // ------------------------------------------------------------------------
    label("fragment_loop");
    // ------------------------------------------------------------------------
    {
        Scratch regs(registerFile());

        if (mDithering) {
            // update the dither index.
            MOV(AL, 0, parts.count.reg,
                    reg_imm(parts.count.reg, ROR, GGL_DITHER_ORDER_SHIFT));
            ADD(AL, 0, parts.count.reg, parts.count.reg,
                    imm( 1 << (32 - GGL_DITHER_ORDER_SHIFT)));
            MOV(AL, 0, parts.count.reg,
                    reg_imm(parts.count.reg, ROR, 32 - GGL_DITHER_ORDER_SHIFT));
        }

        // XXX: could we do an early alpha-test here in some cases?
        // It would probaly be used only with smooth-alpha and no texture
        // (or no alpha component in the texture).

        // Early z-test
        if (mAlphaTest==GGL_ALWAYS) {
            build_depth_test(parts, Z_TEST|Z_WRITE);
        } else {
            // we cannot do the z-write here, because
            // it might be killed by the alpha-test later
            build_depth_test(parts, Z_TEST);
        }

        { // texture coordinates
            Scratch scratches(registerFile());

            // texel generation
            build_textures(parts, regs);
        }        

        if ((blending & (FACTOR_DST|BLEND_DST)) || mMasking ||
                (mLogicOp & LOGIC_OP_DST)) {
            // blending / logic_op / masking need the framebuffer
            mDstPixel.setTo(regs.obtain(), &mCbFormat);

            // load the framebuffer pixel
            comment("fetch color-buffer");
            load(parts.cbPtr, mDstPixel);
        }

        if (registerFile().status())
            return registerFile().status();

        pixel_t pixel;
        int directTex = mTextureMachine.directTexture;
        if (directTex | parts.packed) {
            // note: we can't have both here
            // iterated color or direct texture
            pixel = directTex ? parts.texel[directTex-1] : parts.iterated;
            pixel.flags &= ~CORRUPTIBLE;
        } else {
            if (mDithering) {
                const int ctxtReg = mBuilderContext.Rctx;
                const int mask = GGL_DITHER_SIZE-1;
                parts.dither = reg_t(regs.obtain());
                AND(AL, 0, parts.dither.reg, parts.count.reg, imm(mask));
                ADD(AL, 0, parts.dither.reg, parts.dither.reg, ctxtReg);
                LDRB(AL, parts.dither.reg, parts.dither.reg,
                        immed12_pre(GGL_OFFSETOF(ditherMatrix)));
            }
        
            // allocate a register for the resulting pixel
            pixel.setTo(regs.obtain(), &mCbFormat, FIRST);

            build_component(pixel, parts, GGLFormat::ALPHA,    regs);

            if (mAlphaTest!=GGL_ALWAYS) {
                // only handle the z-write part here. We know z-test
                // was successful, as well as alpha-test.
                build_depth_test(parts, Z_WRITE);
            }

            build_component(pixel, parts, GGLFormat::RED,      regs);
            build_component(pixel, parts, GGLFormat::GREEN,    regs);
            build_component(pixel, parts, GGLFormat::BLUE,     regs);

            pixel.flags |= CORRUPTIBLE;
        }

        if (registerFile().status())
            return registerFile().status();
        
        if (pixel.reg == -1) {
            // be defensive here. if we're here it's probably
            // that this whole fragment is a no-op.
            pixel = mDstPixel;
        }
        
        // logic operation
        build_logic_op(pixel, regs);

        // masking
        build_masking(pixel, regs); 

        comment("store");
        store(parts.cbPtr, pixel, WRITE_BACK);
    }

    if (registerFile().status())
        return registerFile().status();

    // update the iterated color...
    if (parts.reload != 3) {
        build_smooth_shade(parts);
    }

    // update iterated z
    build_iterate_z(parts);

    // update iterated fog
    build_iterate_f(parts);

    SUB(AL, S, parts.count.reg, parts.count.reg, imm(1<<16));
    B(PL, "fragment_loop");
    label("epilog");
    epilog(registerFile().touched());

    if ((mAlphaTest!=GGL_ALWAYS) || (mDepthTest!=GGL_ALWAYS)) {
        if (mDepthTest!=GGL_ALWAYS) {
            label("discard_before_textures");
            build_iterate_texture_coordinates(parts);
        }
        label("discard_after_textures");
        build_smooth_shade(parts);
        build_iterate_z(parts);
        build_iterate_f(parts);
        ADD(AL, 0, parts.cbPtr.reg, parts.cbPtr.reg, imm(parts.cbPtr.size>>3));
        SUB(AL, S, parts.count.reg, parts.count.reg, imm(1<<16));
        B(PL, "fragment_loop");
        epilog(registerFile().touched());
    }

    return registerFile().status();
}

// ---------------------------------------------------------------------------

void GGLAssembler::build_scanline_prolog(
    fragment_parts_t& parts, const needs_t& needs)
{
    Scratch scratches(registerFile());
    int Rctx = mBuilderContext.Rctx;

    // compute count
    comment("compute ct (# of pixels to process)");
    parts.count.setTo(obtainReg());
    int Rx = scratches.obtain();    
    int Ry = scratches.obtain();
    CONTEXT_LOAD(Rx, iterators.xl);
    CONTEXT_LOAD(parts.count.reg, iterators.xr);
    CONTEXT_LOAD(Ry, iterators.y);

    // parts.count = iterators.xr - Rx
    SUB(AL, 0, parts.count.reg, parts.count.reg, Rx);
    SUB(AL, 0, parts.count.reg, parts.count.reg, imm(1));

    if (mDithering) {
        // parts.count.reg = 0xNNNNXXDD
        // NNNN = count-1
        // DD   = dither offset
        // XX   = 0xxxxxxx (x = garbage)
        Scratch scratches(registerFile());
        int tx = scratches.obtain();
        int ty = scratches.obtain();
        AND(AL, 0, tx, Rx, imm(GGL_DITHER_MASK));
        AND(AL, 0, ty, Ry, imm(GGL_DITHER_MASK));
        ADD(AL, 0, tx, tx, reg_imm(ty, LSL, GGL_DITHER_ORDER_SHIFT));
        ORR(AL, 0, parts.count.reg, tx, reg_imm(parts.count.reg, LSL, 16));
    } else {
        // parts.count.reg = 0xNNNN0000
        // NNNN = count-1
        MOV(AL, 0, parts.count.reg, reg_imm(parts.count.reg, LSL, 16));
    }

    // compute dst ptr
    comment("compute color-buffer pointer");
    const int cb_bits = mCbFormat.size*8;
    int Rs = scratches.obtain();
    parts.cbPtr.setTo(obtainReg(), cb_bits);
    CONTEXT_LOAD(Rs, state.buffers.color.stride);
    CONTEXT_LOAD(parts.cbPtr.reg, state.buffers.color.data);