void GGLAssembler::build_blend_factor(
integer_t& factor, int f, int component,
const pixel_t& dst_pixel,
integer_t& fragment,
integer_t& fb,
Scratch& scratches)
{
integer_t src_alpha(fragment);
// src_factor/dst_factor won't be used after blending,
// so it's fine to mark them as CORRUPTIBLE (if not aliased)
factor.flags |= CORRUPTIBLE;
switch(f) {
case GGL_ONE_MINUS_SRC_ALPHA:
case GGL_SRC_ALPHA:
if (component==GGLFormat::ALPHA && !isAlphaSourceNeeded()) {
// we're processing alpha, so we already have
// src-alpha in fragment, and we need src-alpha just this time.
} else {
// alpha-src will be needed for other components
if (!mBlendFactorCached || mBlendFactorCached==f) {
src_alpha = mAlphaSource;
factor = mAlphaSource;
factor.flags &= ~CORRUPTIBLE;
// we already computed the blend factor before, nothing to do.
if (mBlendFactorCached)
return;
// this is the first time, make sure to compute the blend
// factor properly.
mBlendFactorCached = f;
break;
} else {
// we have a cached alpha blend factor, but we want another one,
// this should really not happen because by construction,
// we cannot have BOTH source and destination
// blend factors use ALPHA *and* ONE_MINUS_ALPHA (because
// the blending stage uses the f/(1-f) optimization
// for completeness, we handle this case though. Since there
// are only 2 choices, this meens we want "the other one"
// (1-factor)
factor = mAlphaSource;
factor.flags &= ~CORRUPTIBLE;
RSB(AL, 0, factor.reg, factor.reg, imm((1<<factor.s)));
mBlendFactorCached = f;
return;
}
}
// fall-through...
case GGL_ONE_MINUS_DST_COLOR:
case GGL_DST_COLOR:
case GGL_ONE_MINUS_SRC_COLOR:
case GGL_SRC_COLOR:
case GGL_ONE_MINUS_DST_ALPHA:
case GGL_DST_ALPHA:
case GGL_SRC_ALPHA_SATURATE:
// help us find out what register we can use for the blend-factor
// CORRUPTIBLE registers are chosen first, or a new one is allocated.
if (fragment.flags & CORRUPTIBLE) {
factor.setTo(fragment.reg, 32, CORRUPTIBLE);
fragment.flags &= ~CORRUPTIBLE;
} else if (fb.flags & CORRUPTIBLE) {
factor.setTo(fb.reg, 32, CORRUPTIBLE);
fb.flags &= ~CORRUPTIBLE;
} else {
factor.setTo(scratches.obtain(), 32, CORRUPTIBLE);
}
break;
}
// XXX: doesn't work if size==1
switch(f) {
case GGL_ONE_MINUS_DST_COLOR:
case GGL_DST_COLOR:
factor.s = fb.s;
ADD(AL, 0, factor.reg, fb.reg, reg_imm(fb.reg, LSR, fb.s-1));
break;
case GGL_ONE_MINUS_SRC_COLOR:
case GGL_SRC_COLOR:
factor.s = fragment.s;
ADD(AL, 0, factor.reg, fragment.reg,
reg_imm(fragment.reg, LSR, fragment.s-1));
break;
case GGL_ONE_MINUS_SRC_ALPHA:
case GGL_SRC_ALPHA:
factor.s = src_alpha.s;
ADD(AL, 0, factor.reg, src_alpha.reg,
reg_imm(src_alpha.reg, LSR, src_alpha.s-1));
break;
case GGL_ONE_MINUS_DST_ALPHA:
case GGL_DST_ALPHA:
// XXX: should be precomputed
extract(factor, dst_pixel, GGLFormat::ALPHA);
ADD(AL, 0, factor.reg, factor.reg,
reg_imm(factor.reg, LSR, factor.s-1));
break;
case GGL_SRC_ALPHA_SATURATE:
// XXX: should be precomputed
// XXX: f = min(As, 1-Ad)
// btw, we're guaranteed that Ad's size is <= 8, because
// it's extracted from the framebuffer
break;
}
switch(f) {
case GGL_ONE_MINUS_DST_COLOR:
case GGL_ONE_MINUS_SRC_COLOR:
case GGL_ONE_MINUS_DST_ALPHA:
case GGL_ONE_MINUS_SRC_ALPHA:
RSB(AL, 0, factor.reg, factor.reg, imm((1<<factor.s)));
}
// don't need more than 8-bits for the blend factor
// and this will prevent overflows in the multiplies later
if (factor.s > 8) {
MOV(AL, 0, factor.reg, reg_imm(factor.reg, LSR, factor.s-8));
factor.s = 8;
}
}
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)
}
int fbComponents = 0;
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;
fbComponents |= mCbFormat.c[i].h ? mask : 0;
}
mAllMasked = (mMasking == fbComponents);
if (mAllMasked) {
mDithering = 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 (registerFile().status())
return registerFile().status();
}
if ((blending & (FACTOR_DST|BLEND_DST)) ||
(mMasking && !mAllMasked) ||
(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));
ADDR_ADD(AL, 0, parts.dither.reg, ctxtReg, parts.dither.reg);
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;
}
if (!mAllMasked) {
// 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);
if (!mAllMasked) {
ADDR_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());
}