void GGLAssembler::build_blending(
component_t& temp, // incomming fragment / output
const pixel_t& pixel, // framebuffer
int component,
Scratch& regs)
{
if (!mInfo[component].blend)
return;
int fs = component==GGLFormat::ALPHA ? mBlendSrcA : mBlendSrc;
int fd = component==GGLFormat::ALPHA ? mBlendDstA : mBlendDst;
if (fs==GGL_SRC_ALPHA_SATURATE && component==GGLFormat::ALPHA)
fs = GGL_ONE;
const int blending = blending_codes(fs, fd);
if (!temp.size()) {
// here, blending will produce something which doesn't depend on
// that component (eg: GL_ZERO:GL_*), so the register has not been
// allocated yet. Will never be used as a source.
temp = component_t(regs.obtain(), CORRUPTIBLE);
}
// we are doing real blending...
// fb: extracted dst
// fragment: extracted src
// temp: component_t(fragment) and result
// scoped register allocator
Scratch scratches(registerFile());
comment("blending");
// we can optimize these cases a bit...
// (1) saturation is not needed
// (2) we can use only one multiply instead of 2
// (3) we can reduce the register pressure
// R = S*f + D*(1-f) = (S-D)*f + D
// R = S*(1-f) + D*f = (D-S)*f + S
const bool same_factor_opt1 =
(fs==GGL_DST_COLOR && fd==GGL_ONE_MINUS_DST_COLOR) ||
(fs==GGL_SRC_COLOR && fd==GGL_ONE_MINUS_SRC_COLOR) ||
(fs==GGL_DST_ALPHA && fd==GGL_ONE_MINUS_DST_ALPHA) ||
(fs==GGL_SRC_ALPHA && fd==GGL_ONE_MINUS_SRC_ALPHA);
const bool same_factor_opt2 =
(fs==GGL_ONE_MINUS_DST_COLOR && fd==GGL_DST_COLOR) ||
(fs==GGL_ONE_MINUS_SRC_COLOR && fd==GGL_SRC_COLOR) ||
(fs==GGL_ONE_MINUS_DST_ALPHA && fd==GGL_DST_ALPHA) ||
(fs==GGL_ONE_MINUS_SRC_ALPHA && fd==GGL_SRC_ALPHA);
// XXX: we could also optimize these cases:
// R = S*f + D*f = (S+D)*f
// R = S*(1-f) + D*(1-f) = (S+D)*(1-f)
// R = S*D + D*S = 2*S*D
// see if we need to extract 'component' from the destination (fb)
integer_t fb;
if (blending & (BLEND_DST|FACTOR_DST)) {
fb.setTo(scratches.obtain(), 32);
extract(fb, pixel, component);
if (mDithering) {
// XXX: maybe what we should do instead, is simply
// expand fb -or- fragment to the larger of the two
if (fb.size() < temp.size()) {
// for now we expand 'fb' to min(fragment, 8)
int new_size = temp.size() < 8 ? temp.size() : 8;
expand(fb, fb, new_size);
}
}
}
// convert input fragment to integer_t
if (temp.l && (temp.flags & CORRUPTIBLE)) {
MOV(AL, 0, temp.reg, reg_imm(temp.reg, LSR, temp.l));
temp.h -= temp.l;
temp.l = 0;
}
integer_t fragment(temp.reg, temp.size(), temp.flags);
// if not done yet, convert input fragment to integer_t
if (temp.l) {
// here we know temp is not CORRUPTIBLE
fragment.reg = scratches.obtain();
MOV(AL, 0, fragment.reg, reg_imm(temp.reg, LSR, temp.l));
fragment.flags |= CORRUPTIBLE;
}
if (!(temp.flags & CORRUPTIBLE)) {
// temp is not corruptible, but since it's the destination it
// will be modified, so we need to allocate a new register.
temp.reg = regs.obtain();
temp.flags &= ~CORRUPTIBLE;
fragment.flags &= ~CORRUPTIBLE;
}
if ((blending & BLEND_SRC) && !same_factor_opt1) {
// source (fragment) is needed for the blending stage
// so it's not CORRUPTIBLE (unless we're doing same_factor_opt1)
fragment.flags &= ~CORRUPTIBLE;
}
if (same_factor_opt1) {
// R = S*f + D*(1-f) = (S-D)*f + D
integer_t factor;
build_blend_factor(factor, fs,
component, pixel, fragment, fb, scratches);
// fb is always corruptible from this point
fb.flags |= CORRUPTIBLE;
build_blendFOneMinusF(temp, factor, fragment, fb);
} else if (same_factor_opt2) {
// R = S*(1-f) + D*f = (D-S)*f + S
integer_t factor;
// fb is always corrruptible here
fb.flags |= CORRUPTIBLE;
build_blend_factor(factor, fd,
component, pixel, fragment, fb, scratches);
build_blendOneMinusFF(temp, factor, fragment, fb);
} else {
integer_t src_factor;
integer_t dst_factor;
// if destination (fb) is not needed for the blending stage,
// then it can be marked as CORRUPTIBLE
if (!(blending & BLEND_DST)) {
fb.flags |= CORRUPTIBLE;
}
// XXX: try to mark some registers as CORRUPTIBLE
// in most case we could make those corruptible
// when we're processing the last component
// but not always, for instance
// when fragment is constant and not reloaded
// when fb is needed for logic-ops or masking
// when a register is aliased (for instance with mAlphaSource)
// blend away...
if (fs==GGL_ZERO) {
if (fd==GGL_ZERO) { // R = 0
// already taken care of
} else if (fd==GGL_ONE) { // R = D
// already taken care of
} else { // R = D*fd
// compute fd
build_blend_factor(dst_factor, fd,
component, pixel, fragment, fb, scratches);
mul_factor(temp, fb, dst_factor);
}
} else if (fs==GGL_ONE) {
if (fd==GGL_ZERO) { // R = S
// NOP, taken care of
} else if (fd==GGL_ONE) { // R = S + D
component_add(temp, fb, fragment); // args order matters
component_sat(temp);
} else { // R = S + D*fd
// compute fd
build_blend_factor(dst_factor, fd,
component, pixel, fragment, fb, scratches);
mul_factor_add(temp, fb, dst_factor, component_t(fragment));
component_sat(temp);
}
} else {
// compute fs
build_blend_factor(src_factor, fs,
component, pixel, fragment, fb, scratches);
if (fd==GGL_ZERO) { // R = S*fs
mul_factor(temp, fragment, src_factor);
} else if (fd==GGL_ONE) { // R = S*fs + D
mul_factor_add(temp, fragment, src_factor, component_t(fb));
component_sat(temp);
} else { // R = S*fs + D*fd
mul_factor(temp, fragment, src_factor);
if (scratches.isUsed(src_factor.reg))
scratches.recycle(src_factor.reg);
// compute fd
build_blend_factor(dst_factor, fd,
component, pixel, fragment, fb, scratches);
mul_factor_add(temp, fb, dst_factor, temp);
if (!same_factor_opt1 && !same_factor_opt2) {
component_sat(temp);
}
}
}
}
// now we can be corrupted (it's the dest)
temp.flags |= CORRUPTIBLE;
}
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());
}
