texturing.cpp

Android 一些工具

    context_t const* c = mBuilderContext.c;
    const needs_t& needs = mBuilderContext.needs;
    int Rctx = mBuilderContext.Rctx;

    // We don't have a way to spill registers automatically
    // spill depth and AA regs, when we know we may have to.
    // build the spill list...
    uint32_t spill_list = 0;
    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.linear) {
            // we may run out of register if we have linear filtering
            // at 1 or 4 bytes / pixel on any texture unit.
            if (tmu.format.size == 1) {
                // if depth and AA enabled, we'll run out of 1 register
                if (parts.z.reg > 0 && parts.covPtr.reg > 0)
                    spill_list |= 1<<parts.covPtr.reg;
            }
            if (tmu.format.size == 4) {
                // if depth or AA enabled, we'll run out of 1 or 2 registers
                if (parts.z.reg > 0)
                    spill_list |= 1<<parts.z.reg;
                if (parts.covPtr.reg > 0)   
                    spill_list |= 1<<parts.covPtr.reg;
            }
        }
    }

    Spill spill(registerFile(), *this, spill_list);

    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;

        pointer_t& txPtr = parts.coords[i].ptr;
        pixel_t& texel = parts.texel[i];
            
        // repeat...
        if ((tmu.swrap == GGL_NEEDS_WRAP_11) &&
            (tmu.twrap == GGL_NEEDS_WRAP_11))
        { // 1:1 textures
            comment("fetch texel");
            texel.setTo(regs.obtain(), &tmu.format);
            load(txPtr, texel, WRITE_BACK);
        } else {
            Scratch scratches(registerFile());
            reg_t& s = parts.coords[i].s;
            reg_t& t = parts.coords[i].t;
            if ((mOptLevel&1)==0) {
                comment("reload s/t (multitexture or linear filtering)");
                s.reg = scratches.obtain();
                t.reg = scratches.obtain();
                CONTEXT_LOAD(s.reg, generated_vars.texture[i].spill[0]);
                CONTEXT_LOAD(t.reg, generated_vars.texture[i].spill[1]);
            }

            comment("compute repeat/clamp");
            int u       = scratches.obtain();
            int v       = scratches.obtain();
            int width   = scratches.obtain();
            int height  = scratches.obtain();
            int U = 0;
            int V = 0;

            CONTEXT_LOAD(width,  generated_vars.texture[i].width);
            CONTEXT_LOAD(height, generated_vars.texture[i].height);

            int FRAC_BITS = 0;
            if (tmu.linear) {
                // linear interpolation
                if (tmu.format.size == 1) {
                    // for 8-bits textures, we can afford
                    // 7 bits of fractional precision at no
                    // additional cost (we can't do 8 bits
                    // because filter8 uses signed 16 bits muls)
                    FRAC_BITS = 7;
                } else if (tmu.format.size == 2) {
                    // filter16() is internally limited to 4 bits, so:
                    // FRAC_BITS=2 generates less instructions,
                    // FRAC_BITS=3,4,5 creates unpleasant artifacts,
                    // FRAC_BITS=6+ looks good
                    FRAC_BITS = 6;
                } else if (tmu.format.size == 4) {
                    // filter32() is internally limited to 8 bits, so:
                    // FRAC_BITS=4 looks good
                    // FRAC_BITS=5+ looks better, but generates 3 extra ipp
                    FRAC_BITS = 6;
                } else {
                    // for all other cases we use 4 bits.
                    FRAC_BITS = 4;
                }
            }
            wrapping(u, s.reg, width,  tmu.swrap, FRAC_BITS);
            wrapping(v, t.reg, height, tmu.twrap, FRAC_BITS);

            if (tmu.linear) {
                comment("compute linear filtering offsets");
                // pixel size scale
                const int shift = 31 - gglClz(tmu.format.size);
                U = scratches.obtain();
                V = scratches.obtain();

                // sample the texel center
                SUB(AL, 0, u, u, imm(1<<(FRAC_BITS-1)));
                SUB(AL, 0, v, v, imm(1<<(FRAC_BITS-1)));

                // get the fractionnal part of U,V
                AND(AL, 0, U, u, imm((1<<FRAC_BITS)-1));
                AND(AL, 0, V, v, imm((1<<FRAC_BITS)-1));

                // compute width-1 and height-1
                SUB(AL, 0, width,  width,  imm(1));
                SUB(AL, 0, height, height, imm(1));

                // get the integer part of U,V and clamp/wrap
                // and compute offset to the next texel
                if (tmu.swrap == GGL_NEEDS_WRAP_REPEAT) {
                    // u has already been REPEATed
                    MOV(AL, 1, u, reg_imm(u, ASR, FRAC_BITS));
                    MOV(MI, 0, u, width);                    
                    CMP(AL, u, width);
                    MOV(LT, 0, width, imm(1 << shift));
                    if (shift)
                        MOV(GE, 0, width, reg_imm(width, LSL, shift));
                    RSB(GE, 0, width, width, imm(0));
                } else {
                    // u has not been CLAMPed yet
                    // algorithm:
                    // if ((u>>4) >= width)
                    //      u = width<<4
                    //      width = 0
                    // else
                    //      width = 1<<shift
                    // u = u>>4; // get integer part
                    // if (u<0)
                    //      u = 0
                    //      width = 0
                    // generated_vars.rt = width
                    
                    CMP(AL, width, reg_imm(u, ASR, FRAC_BITS));
                    MOV(LE, 0, u, reg_imm(width, LSL, FRAC_BITS));
                    MOV(LE, 0, width, imm(0));
                    MOV(GT, 0, width, imm(1 << shift));
                    MOV(AL, 1, u, reg_imm(u, ASR, FRAC_BITS));
                    MOV(MI, 0, u, imm(0));
                    MOV(MI, 0, width, imm(0));
                }
                CONTEXT_STORE(width, generated_vars.rt);

                const int stride = width;
                CONTEXT_LOAD(stride, generated_vars.texture[i].stride);
                if (tmu.twrap == GGL_NEEDS_WRAP_REPEAT) {
                    // v has already been REPEATed
                    MOV(AL, 1, v, reg_imm(v, ASR, FRAC_BITS));
                    MOV(MI, 0, v, height);
                    CMP(AL, v, height);
                    MOV(LT, 0, height, imm(1 << shift));
                    if (shift)
                        MOV(GE, 0, height, reg_imm(height, LSL, shift));
                    RSB(GE, 0, height, height, imm(0));
                    MUL(AL, 0, height, stride, height);
                } else {
                    // u has not been CLAMPed yet
                    CMP(AL, height, reg_imm(v, ASR, FRAC_BITS));
                    MOV(LE, 0, v, reg_imm(height, LSL, FRAC_BITS));
                    MOV(LE, 0, height, imm(0));
                    if (shift) {
                        MOV(GT, 0, height, reg_imm(stride, LSL, shift));
                    } else {
                        MOV(GT, 0, height, stride);
                    }
                    MOV(AL, 1, v, reg_imm(v, ASR, FRAC_BITS));
                    MOV(MI, 0, v, imm(0));
                    MOV(MI, 0, height, imm(0));
                }
                CONTEXT_STORE(height, generated_vars.lb);
            }
    
            scratches.recycle(width);
            scratches.recycle(height);

            // iterate texture coordinates...
            comment("iterate s,t");
            int dsdx = scratches.obtain();
            int dtdx = scratches.obtain();
            CONTEXT_LOAD(dsdx, generated_vars.texture[i].dsdx);
            CONTEXT_LOAD(dtdx, generated_vars.texture[i].dtdx);
            ADD(AL, 0, s.reg, s.reg, dsdx);
            ADD(AL, 0, t.reg, t.reg, dtdx);
            if ((mOptLevel&1)==0) {
                CONTEXT_STORE(s.reg, generated_vars.texture[i].spill[0]);
                CONTEXT_STORE(t.reg, generated_vars.texture[i].spill[1]);
                scratches.recycle(s.reg);
                scratches.recycle(t.reg);
            }
            scratches.recycle(dsdx);
            scratches.recycle(dtdx);

            // merge base & offset...
            comment("merge base & offset");
            texel.setTo(regs.obtain(), &tmu.format);
            txPtr.setTo(texel.reg, tmu.bits);
            int stride = scratches.obtain();
            CONTEXT_LOAD(stride,    generated_vars.texture[i].stride);
            CONTEXT_LOAD(txPtr.reg, generated_vars.texture[i].data);
            SMLABB(AL, u, v, stride, u);    // u+v*stride 
            base_offset(txPtr, txPtr, u);

            // load texel
            if (!tmu.linear) {
                comment("fetch texel");
                load(txPtr, texel, 0);
            } else {
                // recycle registers we don't need anymore
                scratches.recycle(u);
                scratches.recycle(v);
                scratches.recycle(stride);

                comment("fetch texel, bilinear");
                switch (tmu.format.size) {
                case 1:  filter8(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
                case 2: filter16(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
                case 3: filter24(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
                case 4: filter32(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
                }
            }            
        }
    }
}

void GGLAssembler::build_iterate_texture_coordinates(
    const fragment_parts_t& parts)
{
    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 textures
            const pointer_t& txPtr = parts.coords[i].ptr;
            ADD(AL, 0, txPtr.reg, txPtr.reg, imm(txPtr.size>>3));
        } else {
            Scratch scratches(registerFile());
            int s = parts.coords[i].s.reg;
            int t = parts.coords[i].t.reg;
            if ((mOptLevel&1)==0) {
                s = scratches.obtain();
                t = scratches.obtain();
                CONTEXT_LOAD(s, generated_vars.texture[i].spill[0]);
                CONTEXT_LOAD(t, generated_vars.texture[i].spill[1]);
            }
            int dsdx = scratches.obtain();
            int dtdx = scratches.obtain();
            CONTEXT_LOAD(dsdx, generated_vars.texture[i].dsdx);
            CONTEXT_LOAD(dtdx, generated_vars.texture[i].dtdx);
            ADD(AL, 0, s, s, dsdx);
            ADD(AL, 0, t, t, dtdx);
            if ((mOptLevel&1)==0) {
                CONTEXT_STORE(s, generated_vars.texture[i].spill[0]);
                CONTEXT_STORE(t, generated_vars.texture[i].spill[1]);
            }
        }
    }
}

void GGLAssembler::filter8(
        const fragment_parts_t& parts,
        pixel_t& texel, const texture_unit_t& tmu,
        int U, int V, pointer_t& txPtr,
        int FRAC_BITS)
{
    if (tmu.format.components != GGL_ALPHA &&
        tmu.format.components != GGL_LUMINANCE)
    {
        // this is a packed format, and we don't support
        // linear filtering (it's probably RGB 332)
        // Should not happen with OpenGL|ES
        LDRB(AL, texel.reg, txPtr.reg);
        return;
    }

    // ------------------------
    // about ~22 cycles / pixel
    Scratch scratches(registerFile());

    int pixel= scratches.obtain();
    int d    = scratches.obtain();
    int u    = scratches.obtain();
    int k    = scratches.obtain();
    int rt   = scratches.obtain();
    int lb   = scratches.obtain();

    // RB -> U * V

    CONTEXT_LOAD(rt, generated_vars.rt);
    CONTEXT_LOAD(lb, generated_vars.lb);

    int offset = pixel;
    ADD(AL, 0, offset, lb, rt);
    LDRB(AL, pixel, txPtr.reg, reg_scale_pre(offset));
    SMULBB(AL, u, U, V);
    SMULBB(AL, d, pixel, u);
    RSB(AL, 0, k, u, imm(1<<(FRAC_BITS*2)));
    
    // LB -> (1-U) * V
    RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
    LDRB(AL, pixel, txPtr.reg, reg_scale_pre(lb));
    SMULBB(AL, u, U, V);
    SMLABB(AL, d, pixel, u, d);
    SUB(AL, 0, k, k, u);
    
    // LT -> (1-U)*(1-V)
    RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
    LDRB(AL, pixel, txPtr.reg);
    SMULBB(AL, u, U, V);
    SMLABB(AL, d, pixel, u, d);

    // RT -> U*(1-V)
    LDRB(AL, pixel, txPtr.reg, reg_scale_pre(rt));
    SUB(AL, 0, u, k, u);
    SMLABB(AL, texel.reg, pixel, u, d);
    
    for (int i=0 ; i<4 ; i++) {
        if (!texel.format.c[i].h) continue;
        texel.format.c[i].h = FRAC_BITS*2+8;
        texel.format.c[i].l = FRAC_BITS*2; // keeping 8 bits in enough
    }
    texel.format.size = 4;
    texel.format.bitsPerPixel = 32;
    texel.flags |= CLEAR_LO;
}

void GGLAssembler::filter16(
        const fragment_parts_t& parts,
        pixel_t& texel, const texture_unit_t& tmu,
        int U, int V, pointer_t& txPtr,
        int FRAC_BITS)
{    
    // compute the mask
    // XXX: it would be nice if the mask below could be computed
    // automatically.
    uint32_t mask = 0;
    int shift = 0;
    int prec = 0;
    switch (tmu.format_idx) {
        case GGL_PIXEL_FORMAT_RGB_565:
            // source: 00000ggg.ggg00000 | rrrrr000.000bbbbb
            // result: gggggggg.gggrrrrr | rrrrr0bb.bbbbbbbb
            mask = 0x07E0F81F;
            shift = 16;
            prec = 5;
            break;
        case GGL_PIXEL_FORMAT_RGBA_4444:
            // 0000,1111,0000,1111 | 0000,1111,0000,1111
            mask = 0x0F0F0F0F;
            shift = 12;
            prec = 4;
            break;
        case GGL_PIXEL_FORMAT_LA_88:
            // 0000,0000,1111,1111 | 0000,0000,1111,1111
            // AALL -> 00AA | 00LL
            mask = 0x00FF00FF;
            shift = 8;
            prec = 8;
            break;
        default:
            // unsupported format, do something sensical...
            LOGE("Unsupported 16-bits texture format (%d)", tmu.format_idx);
            LDRH(AL, texel.reg, txPtr.reg);
            return;
    }

    const int adjust = FRAC_BITS*2 - prec;
    const int round  = 0;

    // update the texel format
    texel.format.size = 4;
    texel.format.bitsPerPixel = 32;
    texel.flags |= CLEAR_HI|CLEAR_LO;
    for (int i=0 ; i<4 ; i++) {
        if (!texel.format.c[i].h) continue;
        const uint32_t offset = (mask & tmu.format.mask(i)) ? 0 : shift;
        texel.format.c[i].h = tmu.format.c[i].h + offset + prec;
        texel.format.c[i].l = texel.format.c[i].h - (tmu.format.bits(i) + prec);
    }

    // ------------------------
    // about ~40 cycles / pixel
    Scratch scratches(registerFile());

    int pixel= scratches.obtain();
    int d    = scratches.obtain();
    int u    = scratches.obtain();
    int k    = scratches.obtain();

    // RB -> U * V