NVC0LoweringPass::NVC0LoweringPass(Program *prog) : targ(prog->getTarget())
{
bld.setProgram(prog);
- gMemBase = NULL;
}
bool
i->tex.rIndirectSrc = 0;
i->tex.sIndirectSrc = -1;
}
+ // Move the indirect reference to right after the coords
+ else if (i->tex.rIndirectSrc >= 0 && chipset >= NVISA_GM107_CHIPSET) {
+ Value *hnd = i->getIndirectR();
+
+ i->setIndirectR(NULL);
+ i->moveSources(arg, 1);
+ i->setSrc(arg, hnd);
+ i->tex.rIndirectSrc = 0;
+ i->tex.sIndirectSrc = -1;
+ }
} else
// (nvc0) generate and move the tsc/tic/array source to the front
if (i->tex.target.isArray() || i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
for (n = 0; n < i->tex.useOffsets; n++) {
for (c = 0; c < 2; ++c) {
if ((n % 2) == 0 && c == 0)
- offs[n / 2] = i->offset[n][c].get();
+ bld.mkMov(offs[n / 2] = bld.getScratch(), i->offset[n][c].get());
else
bld.mkOp3(OP_INSBF, TYPE_U32,
offs[n / 2],
bld.mkCvt(OP_CVT, TYPE_F32, typedDst[i], TYPE_F16, typedDst[i]);
}
}
+
+ if (format->bgra) {
+ std::swap(typedDst[0], typedDst[2]);
+ }
}
void
convertSurfaceFormat(su);
if (su->op == OP_SUREDB || su->op == OP_SUREDP) {
- Value *pred = su->getSrc(2);
- CondCode cc = CC_NOT_P;
- if (su->getPredicate()) {
- pred = bld.getScratch(1, FILE_PREDICATE);
- cc = su->cc;
- if (cc == CC_NOT_P) {
- bld.mkOp2(OP_OR, TYPE_U8, pred, su->getPredicate(), su->getSrc(2));
- } else {
- bld.mkOp2(OP_AND, TYPE_U8, pred, su->getPredicate(), su->getSrc(2));
- pred->getInsn()->src(1).mod = Modifier(NV50_IR_MOD_NOT);
- }
- }
+ assert(su->getPredicate());
+ Value *pred =
+ bld.mkOp2v(OP_OR, TYPE_U8, bld.getScratch(1, FILE_PREDICATE),
+ su->getPredicate(), su->getSrc(2));
+
Instruction *red = bld.mkOp(OP_ATOM, su->dType, bld.getSSA());
red->subOp = su->subOp;
- if (!gMemBase)
- gMemBase = bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, TYPE_U32, 0);
- red->setSrc(0, gMemBase);
+ red->setSrc(0, bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, TYPE_U32, 0));
red->setSrc(1, su->getSrc(3));
if (su->subOp == NV50_IR_SUBOP_ATOM_CAS)
red->setSrc(2, su->getSrc(4));
// performed
Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));
- assert(cc == CC_NOT_P);
- red->setPredicate(cc, pred);
+ assert(su->cc == CC_NOT_P);
+ red->setPredicate(su->cc, pred);
mov->setPredicate(CC_P, pred);
bld.mkOp2(OP_UNION, TYPE_U32, su->getDef(0),
}
}
+void
+NVC0LoweringPass::processSurfaceCoordsGM107(TexInstruction *su)
+{
+ const int slot = su->tex.r;
+ const int dim = su->tex.target.getDim();
+ const int arg = dim + (su->tex.target.isArray() || su->tex.target.isCube());
+ Value *ind = su->getIndirectR();
+ int pos = 0;
+
+ bld.setPosition(su, false);
+
+ // add texture handle
+ switch (su->op) {
+ case OP_SUSTP:
+ pos = 4;
+ break;
+ case OP_SUREDP:
+ pos = (su->subOp == NV50_IR_SUBOP_ATOM_CAS) ? 2 : 1;
+ break;
+ default:
+ assert(pos == 0);
+ break;
+ }
+ su->setSrc(arg + pos, loadTexHandle(ind, slot + 32));
+
+ // prevent read fault when the image is not actually bound
+ CmpInstruction *pred =
+ bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
+ TYPE_U32, bld.mkImm(0),
+ loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR));
+ if (su->op != OP_SUSTP && su->tex.format) {
+ const TexInstruction::ImgFormatDesc *format = su->tex.format;
+ int blockwidth = format->bits[0] + format->bits[1] +
+ format->bits[2] + format->bits[3];
+
+ assert(format->components != 0);
+ // make sure that the format doesn't mismatch when it's not FMT_NONE
+ bld.mkCmp(OP_SET_OR, CC_NE, TYPE_U32, pred->getDef(0),
+ TYPE_U32, bld.loadImm(NULL, blockwidth / 8),
+ loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE),
+ pred->getDef(0));
+ }
+ su->setPredicate(CC_NOT_P, pred->getDef(0));
+}
+
+void
+NVC0LoweringPass::handleSurfaceOpGM107(TexInstruction *su)
+{
+ processSurfaceCoordsGM107(su);
+
+ if (su->op == OP_SULDP)
+ convertSurfaceFormat(su);
+
+ if (su->op == OP_SUREDP) {
+ Value *def = su->getDef(0);
+
+ su->op = OP_SUREDB;
+ su->setDef(0, bld.getSSA());
+
+ bld.setPosition(su, true);
+
+ // make sure to initialize dst value when the atomic operation is not
+ // performed
+ Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));
+
+ assert(su->cc == CC_NOT_P);
+ mov->setPredicate(CC_P, su->getPredicate());
+
+ bld.mkOp2(OP_UNION, TYPE_U32, def, su->getDef(0), mov->getDef(0));
+ }
+}
+
bool
NVC0LoweringPass::handleWRSV(Instruction *i)
{
case OP_SUSTP:
case OP_SUREDB:
case OP_SUREDP:
- if (targ->getChipset() >= NVISA_GK104_CHIPSET)
+ if (targ->getChipset() >= NVISA_GM107_CHIPSET)
+ handleSurfaceOpGM107(i->asTex());
+ else if (targ->getChipset() >= NVISA_GK104_CHIPSET)
handleSurfaceOpNVE4(i->asTex());
else
handleSurfaceOpNVC0(i->asTex());
/* Kepler+ has a special opcode to compute a new base address to be used
* for indirect loads.
+ *
+ * Maxwell+ has an additional similar requirement for indirect
+ * interpolation ops in frag shaders.
*/
- if (targ->getChipset() >= NVISA_GK104_CHIPSET && !i->perPatch &&
- (i->op == OP_VFETCH || i->op == OP_EXPORT) && i->src(0).isIndirect(0)) {
+ bool doAfetch = false;
+ if (targ->getChipset() >= NVISA_GK104_CHIPSET &&
+ !i->perPatch &&
+ (i->op == OP_VFETCH || i->op == OP_EXPORT) &&
+ i->src(0).isIndirect(0)) {
+ doAfetch = true;
+ }
+ if (targ->getChipset() >= NVISA_GM107_CHIPSET &&
+ (i->op == OP_LINTERP || i->op == OP_PINTERP) &&
+ i->src(0).isIndirect(0)) {
+ doAfetch = true;
+ }
+
+ if (doAfetch) {
+ Value *addr = cloneShallow(func, i->getSrc(0));
Instruction *afetch = bld.mkOp1(OP_AFETCH, TYPE_U32, bld.getSSA(),
- cloneShallow(func, i->getSrc(0)));
+ i->getSrc(0));
afetch->setIndirect(0, 0, i->getIndirect(0, 0));
- i->src(0).get()->reg.data.offset = 0;
+ addr->reg.data.offset = 0;
+ i->setSrc(0, addr);
i->setIndirect(0, 0, afetch->getDef(0));
}