texturing.cpp

Android 一些工具

/* libs/pixelflinger/codeflinger/texturing.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.
*/

#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 {

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

// iterators are initialized like this:
// (intToFixedCenter(x) * dx)>>16 + x0
// ((x<<16 + 0x8000) * dx)>>16 + x0
// ((x<<16)*dx + (0x8000*dx))>>16 + x0
// ( (x*dx) + dx>>1 ) + x0
// (x*dx) + (dx>>1 + x0)

void GGLAssembler::init_iterated_color(fragment_parts_t& parts, const reg_t& x)
{
    context_t const* c = mBuilderContext.c;
    const needs_t& needs = mBuilderContext.needs;

    if (mSmooth) {
        // NOTE: we could take this case in the mDithering + !mSmooth case,
        // but this would use up to 4 more registers for the color components
        // for only a little added quality.
        // Currently, this causes the system to run out of registers in
        // some case (see issue #719496)

        comment("compute initial iterated color (smooth and/or dither case)");

        parts.iterated_packed = 0;
        parts.packed = 0;

        // 0x1: color component
        // 0x2: iterators
        const int optReload = mOptLevel >> 1;
        if (optReload >= 3)         parts.reload = 0; // reload nothing
        else if (optReload == 2)    parts.reload = 2; // reload iterators
        else if (optReload == 1)    parts.reload = 1; // reload colors
        else if (optReload <= 0)    parts.reload = 3; // reload both

        if (!mSmooth) {
            // we're not smoothing (just dithering), we never have to 
            // reload the iterators
            parts.reload &= ~2;
        }

        Scratch scratches(registerFile());
        const int t0 = (parts.reload & 1) ? scratches.obtain() : 0;
        const int t1 = (parts.reload & 2) ? scratches.obtain() : 0;
        for (int i=0 ; i<4 ; i++) {
            if (!mInfo[i].iterated)
                continue;            
            
            // this component exists in the destination and is not replaced
            // by a texture unit.
            const int c = (parts.reload & 1) ? t0 : obtainReg();              
            if (i==0) CONTEXT_LOAD(c, iterators.ydady);
            if (i==1) CONTEXT_LOAD(c, iterators.ydrdy);
            if (i==2) CONTEXT_LOAD(c, iterators.ydgdy);
            if (i==3) CONTEXT_LOAD(c, iterators.ydbdy);
            parts.argb[i].reg = c;

            if (mInfo[i].smooth) {
                parts.argb_dx[i].reg = (parts.reload & 2) ? t1 : obtainReg();
                const int dvdx = parts.argb_dx[i].reg;
                CONTEXT_LOAD(dvdx, generated_vars.argb[i].dx);
                MLA(AL, 0, c, x.reg, dvdx, c);
                
                // adjust the color iterator to make sure it won't overflow
                if (!mAA) {
                    // this is not needed when we're using anti-aliasing
                    // because we will (have to) clamp the components
                    // anyway.
                    int end = scratches.obtain();
                    MOV(AL, 0, end, reg_imm(parts.count.reg, LSR, 16));
                    MLA(AL, 1, end, dvdx, end, c);
                    SUB(MI, 0, c, c, end);
                    BIC(AL, 0, c, c, reg_imm(c, ASR, 31)); 
                    scratches.recycle(end);
                }
            }
            
            if (parts.reload & 1) {
                CONTEXT_STORE(c, generated_vars.argb[i].c);
            }
        }
    } else {
        // We're not smoothed, so we can 
        // just use a packed version of the color and extract the
        // components as needed (or not at all if we don't blend)

        // figure out if we need the iterated color
        int load = 0;
        for (int i=0 ; i<4 ; i++) {
            component_info_t& info = mInfo[i];
            if ((info.inDest || info.needed) && !info.replaced)
                load |= 1;
        }
        
        parts.iterated_packed = 1;
        parts.packed = (!mTextureMachine.mask && !mBlending
                && !mFog && !mDithering);
        parts.reload = 0;
        if (load || parts.packed) {
            if (mBlending || mDithering || mInfo[GGLFormat::ALPHA].needed) {
                comment("load initial iterated color (8888 packed)");
                parts.iterated.setTo(obtainReg(),
                        &(c->formats[GGL_PIXEL_FORMAT_RGBA_8888]));
                CONTEXT_LOAD(parts.iterated.reg, packed8888);
            } else {
                comment("load initial iterated color (dest format packed)");

                parts.iterated.setTo(obtainReg(), &mCbFormat);

                // pre-mask the iterated color
                const int bits = parts.iterated.size();
                const uint32_t size = ((bits>=32) ? 0 : (1LU << bits)) - 1;
                uint32_t mask = 0;
                if (mMasking) {
                    for (int i=0 ; i<4 ; i++) {
                        const int component_mask = 1<<i;
                        const int h = parts.iterated.format.c[i].h;
                        const int l = parts.iterated.format.c[i].l;
                        if (h && (!(mMasking & component_mask))) {
                            mask |= ((1<<(h-l))-1) << l;
                        }
                    }
                }

                if (mMasking && ((mask & size)==0)) {
                    // none of the components are present in the mask
                } else {
                    CONTEXT_LOAD(parts.iterated.reg, packed);
                    if (mCbFormat.size == 1) {
                        AND(AL, 0, parts.iterated.reg,
                                parts.iterated.reg, imm(0xFF));
                    } else if (mCbFormat.size == 2) {
                        MOV(AL, 0, parts.iterated.reg,
                                reg_imm(parts.iterated.reg, LSR, 16));
                    }
                }

                // pre-mask the iterated color
                if (mMasking) {
                    build_and_immediate(parts.iterated.reg, parts.iterated.reg,
                            mask, bits);
                }
            }
        }
    }
}

void GGLAssembler::build_iterated_color(
        component_t& fragment,
        const fragment_parts_t& parts,
        int component,
        Scratch& regs)
{
    fragment.setTo( regs.obtain(), 0, 32, CORRUPTIBLE); 

    if (!mInfo[component].iterated)
        return;

    if (parts.iterated_packed) {
        // iterated colors are packed, extract the one we need
        extract(fragment, parts.iterated, component);
    } else {
        fragment.h = GGL_COLOR_BITS;
        fragment.l = GGL_COLOR_BITS - 8;
        fragment.flags |= CLEAR_LO;
        // iterated colors are held in their own register,
        // (smooth and/or dithering case)
        if (parts.reload==3) {
            // this implies mSmooth
            Scratch scratches(registerFile());
            int dx = scratches.obtain();
            CONTEXT_LOAD(fragment.reg, generated_vars.argb[component].c);
            CONTEXT_LOAD(dx, generated_vars.argb[component].dx);
            ADD(AL, 0, dx, fragment.reg, dx);
            CONTEXT_STORE(dx, generated_vars.argb[component].c);
        } else if (parts.reload & 1) {
            CONTEXT_LOAD(fragment.reg, generated_vars.argb[component].c);
        } else {
            // we don't reload, so simply rename the register and mark as
            // non CORRUPTIBLE so that the texture env or blending code
            // won't modify this (renamed) register
            regs.recycle(fragment.reg);
            fragment.reg = parts.argb[component].reg;
            fragment.flags &= ~CORRUPTIBLE;
        }
        if (mInfo[component].smooth && mAA) {
            // when using smooth shading AND anti-aliasing, we need to clamp
            // the iterators because there is always an extra pixel on the
            // edges, which most of the time will cause an overflow
            // (since technically its outside of the domain).
            BIC(AL, 0, fragment.reg, fragment.reg,
                    reg_imm(fragment.reg, ASR, 31));
            component_sat(fragment);
        }
    }
}

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

void GGLAssembler::decodeLogicOpNeeds(const needs_t& needs)
{
    // gather some informations about the components we need to process...
    const int opcode = GGL_READ_NEEDS(LOGIC_OP, needs.n) | GGL_CLEAR;
    switch(opcode) {
    case GGL_COPY:
        mLogicOp = 0;
        break;
    case GGL_CLEAR:
    case GGL_SET:
        mLogicOp = LOGIC_OP;
        break;
    case GGL_AND:
    case GGL_AND_REVERSE:
    case GGL_AND_INVERTED:
    case GGL_XOR:
    case GGL_OR:
    case GGL_NOR:
    case GGL_EQUIV:
    case GGL_OR_REVERSE:
    case GGL_OR_INVERTED:
    case GGL_NAND:
        mLogicOp = LOGIC_OP|LOGIC_OP_SRC|LOGIC_OP_DST;
        break;
    case GGL_NOOP:
    case GGL_INVERT:
        mLogicOp = LOGIC_OP|LOGIC_OP_DST;
        break;        
    case GGL_COPY_INVERTED:
        mLogicOp = LOGIC_OP|LOGIC_OP_SRC;
        break;
    };        
}

void GGLAssembler::decodeTMUNeeds(const needs_t& needs, context_t const* c)
{
    uint8_t replaced=0;
    mTextureMachine.mask = 0;
    mTextureMachine.activeUnits = 0;
    for (int i=GGL_TEXTURE_UNIT_COUNT-1 ; i>=0 ; i--) {
        texture_unit_t& tmu = mTextureMachine.tmu[i];
        if (replaced == 0xF) {
            // all components are replaced, skip this TMU.
            tmu.format_idx = 0;
            tmu.mask = 0;
            tmu.replaced = replaced;
            continue;
        }
        tmu.format_idx = GGL_READ_NEEDS(T_FORMAT, needs.t[i]);
        tmu.format = c->formats[tmu.format_idx];
        tmu.bits = tmu.format.size*8;
        tmu.swrap = GGL_READ_NEEDS(T_S_WRAP, needs.t[i]);
        tmu.twrap = GGL_READ_NEEDS(T_T_WRAP, needs.t[i]);
        tmu.env = ggl_needs_to_env(GGL_READ_NEEDS(T_ENV, needs.t[i]));
        tmu.pot = GGL_READ_NEEDS(T_POT, needs.t[i]);
        tmu.linear = GGL_READ_NEEDS(T_LINEAR, needs.t[i])
                && tmu.format.size!=3; // XXX: only 8, 16 and 32 modes for now

        // 5551 linear filtering is not supported
        if (tmu.format_idx == GGL_PIXEL_FORMAT_RGBA_5551)
            tmu.linear = 0;
        
        tmu.mask = 0;
        tmu.replaced = replaced;

        if (tmu.format_idx) {
            mTextureMachine.activeUnits++;
            if (tmu.format.c[0].h)    tmu.mask |= 0x1;
            if (tmu.format.c[1].h)    tmu.mask |= 0x2;
            if (tmu.format.c[2].h)    tmu.mask |= 0x4;
            if (tmu.format.c[3].h)    tmu.mask |= 0x8;
            if (tmu.env == GGL_REPLACE) {
                replaced |= tmu.mask;
            } else if (tmu.env == GGL_DECAL) {
                if (!tmu.format.c[GGLFormat::ALPHA].h) {
                    // if we don't have alpha, decal does nothing
                    tmu.mask = 0;
                } else {
                    // decal always ignores At
                    tmu.mask &= ~(1<<GGLFormat::ALPHA);
                }
            }
        }
        mTextureMachine.mask |= tmu.mask;
        //printf("%d: mask=%08lx, replaced=%08lx\n",
        //    i, int(tmu.mask), int(tmu.replaced));
    }
    mTextureMachine.replaced = replaced;
    mTextureMachine.directTexture = 0;
    //printf("replaced=%08lx\n", mTextureMachine.replaced);
}


void GGLAssembler::init_textures(
        tex_coord_t* coords,
        const reg_t& x, const reg_t& y)
{
    context_t const* c = mBuilderContext.c;
    const needs_t& needs = mBuilderContext.needs;
    int Rctx = mBuilderContext.Rctx;
    int Rx = x.reg;
    int Ry = y.reg;

    if (mTextureMachine.mask) {
        comment("compute texture coordinates");
    }

    // init texture coordinates for each tmu
    const int cb_format_idx = GGL_READ_NEEDS(CB_FORMAT, needs.n);
    const bool multiTexture = mTextureMachine.activeUnits > 1;
    for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT; i++) {
        const texture_unit_t& tmu = mTextureMachine.tmu[i];
        if (tmu.format_idx == 0)
            continue;
        if ((tmu.swrap == GGL_NEEDS_WRAP_11) &&
            (tmu.twrap == GGL_NEEDS_WRAP_11)) 
        {
            // 1:1 texture
            pointer_t& txPtr = coords[i].ptr;
            txPtr.setTo(obtainReg(), tmu.bits);
            CONTEXT_LOAD(txPtr.reg, state.texture[i].iterators.ydsdy);
            ADD(AL, 0, Rx, Rx, reg_imm(txPtr.reg, ASR, 16));    // x += (s>>16)
            CONTEXT_LOAD(txPtr.reg, state.texture[i].iterators.ydtdy);
            ADD(AL, 0, Ry, Ry, reg_imm(txPtr.reg, ASR, 16));    // y += (t>>16)
            // merge base & offset
            CONTEXT_LOAD(txPtr.reg, generated_vars.texture[i].stride);
            SMLABB(AL, Rx, Ry, txPtr.reg, Rx);               // x+y*stride
            CONTEXT_LOAD(txPtr.reg, generated_vars.texture[i].data);
            base_offset(txPtr, txPtr, Rx);
        } else {
            Scratch scratches(registerFile());
            reg_t& s = coords[i].s;
            reg_t& t = coords[i].t;
            // s = (x * dsdx)>>16 + ydsdy
            // s = (x * dsdx)>>16 + (y*dsdy)>>16 + s0
            // t = (x * dtdx)>>16 + ydtdy
            // t = (x * dtdx)>>16 + (y*dtdy)>>16 + t0
            s.setTo(obtainReg());
            t.setTo(obtainReg());
            const int need_w = GGL_READ_NEEDS(W, needs.n);
            if (need_w) {
                CONTEXT_LOAD(s.reg, state.texture[i].iterators.ydsdy);
                CONTEXT_LOAD(t.reg, state.texture[i].iterators.ydtdy);
            } else {
                int ydsdy = scratches.obtain();
                int ydtdy = scratches.obtain();
                CONTEXT_LOAD(s.reg, generated_vars.texture[i].dsdx);
                CONTEXT_LOAD(ydsdy, state.texture[i].iterators.ydsdy);
                CONTEXT_LOAD(t.reg, generated_vars.texture[i].dtdx);
                CONTEXT_LOAD(ydtdy, state.texture[i].iterators.ydtdy);
                MLA(AL, 0, s.reg, Rx, s.reg, ydsdy);
                MLA(AL, 0, t.reg, Rx, t.reg, ydtdy);
            }
            
            if ((mOptLevel&1)==0) {
                CONTEXT_STORE(s.reg, generated_vars.texture[i].spill[0]);
                CONTEXT_STORE(t.reg, generated_vars.texture[i].spill[1]);
                recycleReg(s.reg);
                recycleReg(t.reg);
            }
        }

        // direct texture?
        if (!multiTexture && !mBlending && !mDithering && !mFog && 
            cb_format_idx == tmu.format_idx && !tmu.linear &&
            mTextureMachine.replaced == tmu.mask) 
        {
                mTextureMachine.directTexture = i + 1; 
        }
    }
}

void GGLAssembler::build_textures(  fragment_parts_t& parts,
                                    Scratch& regs)
{