{
FlowInstruction *call;
int builtin;
- Value *def[2];
bld.setPosition(i, false);
- def[0] = bld.mkMovToReg(0, i->getSrc(0))->getDef(0);
- def[1] = bld.mkMovToReg(1, i->getSrc(1))->getDef(0);
+
+ // Generate movs to the input regs for the call we want to generate
+ for (int s = 0; i->srcExists(s); ++s) {
+ Instruction *ld = i->getSrc(s)->getInsn();
+ assert(ld->getSrc(0) != NULL);
+ // check if we are moving an immediate, propagate it in that case
+ if (!ld || ld->fixed || (ld->op != OP_LOAD && ld->op != OP_MOV) ||
+ !(ld->src(0).getFile() == FILE_IMMEDIATE))
+ bld.mkMovToReg(s, i->getSrc(s));
+ else {
+ bld.mkMovToReg(s, ld->getSrc(0));
+ // Clear the src, to make code elimination possible here before we
+ // delete the instruction i later
+ i->setSrc(s, NULL);
+ if (ld->isDead())
+ delete_Instruction(prog, ld);
+ }
+ }
+
switch (i->dType) {
case TYPE_U32: builtin = NVC0_BUILTIN_DIV_U32; break;
case TYPE_S32: builtin = NVC0_BUILTIN_DIV_S32; break;
return;
}
call = bld.mkFlow(OP_CALL, NULL, CC_ALWAYS, NULL);
- bld.mkMov(i->getDef(0), def[(i->op == OP_DIV) ? 0 : 1]);
+ bld.mkMovFromReg(i->getDef(0), i->op == OP_DIV ? 0 : 1);
bld.mkClobber(FILE_GPR, (i->op == OP_DIV) ? 0xe : 0xd, 2);
bld.mkClobber(FILE_PREDICATE, (i->dType == TYPE_S32) ? 0xf : 0x3, 0);
i->ftz = true;
}
+void
+NVC0LegalizeSSA::handleTEXLOD(TexInstruction *i)
+{
+ if (i->tex.levelZero)
+ return;
+
+ ImmediateValue lod;
+
+ // The LOD argument comes right after the coordinates (before depth bias,
+ // offsets, etc).
+ int arg = i->tex.target.getArgCount();
+
+ // SM30+ stores the indirect handle as a separate arg, which comes before
+ // the LOD.
+ if (prog->getTarget()->getChipset() >= NVISA_GK104_CHIPSET &&
+ i->tex.rIndirectSrc >= 0)
+ arg++;
+ // SM20 stores indirect handle combined with array coordinate
+ if (prog->getTarget()->getChipset() < NVISA_GK104_CHIPSET &&
+ !i->tex.target.isArray() &&
+ i->tex.rIndirectSrc >= 0)
+ arg++;
+
+ if (!i->src(arg).getImmediate(lod) || !lod.isInteger(0))
+ return;
+
+ if (i->op == OP_TXL)
+ i->op = OP_TEX;
+ i->tex.levelZero = true;
+ i->moveSources(arg + 1, -1);
+}
+
+void
+NVC0LegalizeSSA::handleShift(Instruction *lo)
+{
+ Value *shift = lo->getSrc(1);
+ Value *dst64 = lo->getDef(0);
+ Value *src[2], *dst[2];
+ operation op = lo->op;
+
+ bld.setPosition(lo, false);
+
+ bld.mkSplit(src, 4, lo->getSrc(0));
+
+ // SM30 and prior don't have the fancy new SHF.L/R ops. So the logic has to
+ // be completely emulated. For SM35+, we can use the more directed SHF
+ // operations.
+ if (prog->getTarget()->getChipset() < NVISA_GK20A_CHIPSET) {
+ // The strategy here is to handle shifts >= 32 and less than 32 as
+ // separate parts.
+ //
+ // For SHL:
+ // If the shift is <= 32, then
+ // (HI,LO) << x = (HI << x | (LO >> (32 - x)), LO << x)
+ // If the shift is > 32, then
+ // (HI,LO) << x = (LO << (x - 32), 0)
+ //
+ // For SHR:
+ // If the shift is <= 32, then
+ // (HI,LO) >> x = (HI >> x, (HI << (32 - x)) | LO >> x)
+ // If the shift is > 32, then
+ // (HI,LO) >> x = (0, HI >> (x - 32))
+ //
+ // Note that on NVIDIA hardware, a shift > 32 yields a 0 value, which we
+ // can use to our advantage. Also note the structural similarities
+ // between the right/left cases. The main difference is swapping hi/lo
+ // on input and output.
+
+ Value *x32_minus_shift, *pred, *hi1, *hi2;
+ DataType type = isSignedIntType(lo->dType) ? TYPE_S32 : TYPE_U32;
+ operation antiop = op == OP_SHR ? OP_SHL : OP_SHR;
+ if (op == OP_SHR)
+ std::swap(src[0], src[1]);
+ bld.mkOp2(OP_ADD, TYPE_U32, (x32_minus_shift = bld.getSSA()), shift, bld.mkImm(0x20))
+ ->src(0).mod = Modifier(NV50_IR_MOD_NEG);
+ bld.mkCmp(OP_SET, CC_LE, TYPE_U8, (pred = bld.getSSA(1, FILE_PREDICATE)),
+ TYPE_U32, shift, bld.mkImm(32));
+ // Compute HI (shift <= 32)
+ bld.mkOp2(OP_OR, TYPE_U32, (hi1 = bld.getSSA()),
+ bld.mkOp2v(op, TYPE_U32, bld.getSSA(), src[1], shift),
+ bld.mkOp2v(antiop, TYPE_U32, bld.getSSA(), src[0], x32_minus_shift))
+ ->setPredicate(CC_P, pred);
+ // Compute LO (all shift values)
+ bld.mkOp2(op, type, (dst[0] = bld.getSSA()), src[0], shift);
+ // Compute HI (shift > 32)
+ bld.mkOp2(op, type, (hi2 = bld.getSSA()), src[0],
+ bld.mkOp1v(OP_NEG, TYPE_S32, bld.getSSA(), x32_minus_shift))
+ ->setPredicate(CC_NOT_P, pred);
+ bld.mkOp2(OP_UNION, TYPE_U32, (dst[1] = bld.getSSA()), hi1, hi2);
+ if (op == OP_SHR)
+ std::swap(dst[0], dst[1]);
+ bld.mkOp2(OP_MERGE, TYPE_U64, dst64, dst[0], dst[1]);
+ delete_Instruction(prog, lo);
+ return;
+ }
+
+ Instruction *hi = new_Instruction(func, op, TYPE_U32);
+ lo->bb->insertAfter(lo, hi);
+
+ hi->sType = lo->sType;
+ lo->dType = TYPE_U32;
+
+ hi->setDef(0, (dst[1] = bld.getSSA()));
+ if (lo->op == OP_SHR)
+ hi->subOp |= NV50_IR_SUBOP_SHIFT_HIGH;
+ lo->setDef(0, (dst[0] = bld.getSSA()));
+
+ bld.setPosition(hi, true);
+
+ if (lo->op == OP_SHL)
+ std::swap(hi, lo);
+
+ hi->setSrc(0, new_ImmediateValue(prog, 0u));
+ hi->setSrc(1, shift);
+ hi->setSrc(2, lo->op == OP_SHL ? src[0] : src[1]);
+
+ lo->setSrc(0, src[0]);
+ lo->setSrc(1, shift);
+ lo->setSrc(2, src[1]);
+
+ bld.mkOp2(OP_MERGE, TYPE_U64, dst64, dst[0], dst[1]);
+}
+
+void
+NVC0LegalizeSSA::handleSET(CmpInstruction *cmp)
+{
+ DataType hTy = cmp->sType == TYPE_S64 ? TYPE_S32 : TYPE_U32;
+ Value *carry;
+ Value *src0[2], *src1[2];
+ bld.setPosition(cmp, false);
+
+ bld.mkSplit(src0, 4, cmp->getSrc(0));
+ bld.mkSplit(src1, 4, cmp->getSrc(1));
+ bld.mkOp2(OP_SUB, hTy, NULL, src0[0], src1[0])
+ ->setFlagsDef(0, (carry = bld.getSSA(1, FILE_FLAGS)));
+ cmp->setFlagsSrc(cmp->srcCount(), carry);
+ cmp->setSrc(0, src0[1]);
+ cmp->setSrc(1, src1[1]);
+ cmp->sType = hTy;
+}
+
bool
NVC0LegalizeSSA::visit(Function *fn)
{
Instruction *next;
for (Instruction *i = bb->getEntry(); i; i = next) {
next = i->next;
- if (i->sType == TYPE_F32) {
- if (prog->getType() != Program::TYPE_COMPUTE)
- handleFTZ(i);
- continue;
- }
+
+ if (i->sType == TYPE_F32 && prog->getType() != Program::TYPE_COMPUTE)
+ handleFTZ(i);
+
switch (i->op) {
case OP_DIV:
case OP_MOD:
- handleDIV(i);
+ if (i->sType != TYPE_F32)
+ handleDIV(i);
break;
case OP_RCP:
case OP_RSQ:
if (i->dType == TYPE_F64)
handleRCPRSQ(i);
break;
+ case OP_TXL:
+ case OP_TXF:
+ handleTEXLOD(i->asTex());
+ break;
+ case OP_SHR:
+ case OP_SHL:
+ if (typeSizeof(i->sType) == 8)
+ handleShift(i);
+ break;
+ case OP_SET:
+ case OP_SET_AND:
+ case OP_SET_OR:
+ case OP_SET_XOR:
+ if (typeSizeof(i->sType) == 8 && i->sType != TYPE_F64)
+ handleSET(i->asCmp());
+ break;
default:
break;
}
: rZero(NULL),
carry(NULL),
pOne(NULL),
- needTexBar(prog->getTarget()->getChipset() >= 0xe0)
+ needTexBar(prog->getTarget()->getChipset() >= 0xe0 &&
+ prog->getTarget()->getChipset() < 0x110)
{
}
for (int s = 0; i->srcExists(s); ++s) {
if (s == 2 && i->op == OP_SUCLAMP)
continue;
+ if (s == 1 && i->op == OP_SHLADD)
+ continue;
ImmediateValue *imm = i->getSrc(s)->asImm();
if (imm) {
if (i->op == OP_SELP && s == 2) {
bb->remove(bb->getEntry());
}
+// replaces instructions which would end up as f2f or i2i with faster
+// alternatives:
+// - fabs(a) -> fadd(0, abs a)
+// - fneg(a) -> fadd(neg 0, neg a)
+// - ineg(a) -> iadd(0, neg a)
+// - fneg(abs a) -> fadd(neg 0, neg abs a)
+// - sat(a) -> sat add(0, a)
+void
+NVC0LegalizePostRA::replaceCvt(Instruction *cvt)
+{
+ if (!isFloatType(cvt->sType) && typeSizeof(cvt->sType) != 4)
+ return;
+ if (cvt->sType != cvt->dType)
+ return;
+ // we could make it work, but in this case we have optimizations disabled
+ // and we don't really care either way.
+ if (cvt->src(0).getFile() != FILE_GPR &&
+ cvt->src(0).getFile() != FILE_MEMORY_CONST)
+ return;
+
+ Modifier mod0, mod1;
+
+ switch (cvt->op) {
+ case OP_ABS:
+ if (cvt->src(0).mod)
+ return;
+ if (!isFloatType(cvt->sType))
+ return;
+ mod0 = 0;
+ mod1 = NV50_IR_MOD_ABS;
+ break;
+ case OP_NEG:
+ if (!isFloatType(cvt->sType) && cvt->src(0).mod)
+ return;
+ if (isFloatType(cvt->sType) &&
+ (cvt->src(0).mod && cvt->src(0).mod != Modifier(NV50_IR_MOD_ABS)))
+ return;
+
+ mod0 = isFloatType(cvt->sType) ? NV50_IR_MOD_NEG : 0;
+ mod1 = cvt->src(0).mod == Modifier(NV50_IR_MOD_ABS) ?
+ NV50_IR_MOD_NEG_ABS : NV50_IR_MOD_NEG;
+ break;
+ case OP_SAT:
+ if (!isFloatType(cvt->sType) && cvt->src(0).mod.abs())
+ return;
+ mod0 = 0;
+ mod1 = cvt->src(0).mod;
+ cvt->saturate = true;
+ break;
+ default:
+ return;
+ }
+
+ cvt->op = OP_ADD;
+ cvt->moveSources(0, 1);
+ cvt->setSrc(0, rZero);
+ cvt->src(0).mod = mod0;
+ cvt->src(1).mod = mod1;
+}
+
bool
NVC0LegalizePostRA::visit(BasicBlock *bb)
{
} else
if (i->op == OP_LOAD && i->subOp == NV50_IR_SUBOP_LDC_IS) {
int offset = i->src(0).get()->reg.data.offset;
- if (abs(offset) > 0x10000)
+ if (abs(offset) >= 0x10000)
i->src(0).get()->reg.fileIndex += offset >> 16;
i->src(0).get()->reg.data.offset = (int)(short)offset;
} else {
// TODO: Move this to before register allocation for operations that
// need the $c register !
- if (typeSizeof(i->dType) == 8) {
+ if (typeSizeof(i->sType) == 8 || typeSizeof(i->dType) == 8) {
Instruction *hi;
hi = BuildUtil::split64BitOpPostRA(func, i, rZero, carry);
if (hi)
next = hi;
}
+ if (i->op == OP_SAT || i->op == OP_NEG || i->op == OP_ABS)
+ replaceCvt(i);
+
if (i->op != OP_MOV && i->op != OP_PFETCH)
replaceZero(i);
}
NVC0LoweringPass::NVC0LoweringPass(Program *prog) : targ(prog->getTarget())
{
bld.setProgram(prog);
- gMemBase = NULL;
}
bool
{
uint8_t b = prog->driver->io.auxCBSlot;
uint32_t off = prog->driver->io.texBindBase + slot * 4;
+
+ if (ptr)
+ ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(2));
+
return bld.
mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
}
if (i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
// XXX this ignores tsc, and assumes a 1:1 mapping
assert(i->tex.rIndirectSrc >= 0);
- Value *hnd = loadTexHandle(
- bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
- i->getIndirectR(), bld.mkImm(2)),
- i->tex.r);
- i->tex.r = 0xff;
- i->tex.s = 0x1f;
- i->setIndirectR(hnd);
+ if (!i->tex.bindless) {
+ Value *hnd = loadTexHandle(i->getIndirectR(), i->tex.r);
+ i->tex.r = 0xff;
+ i->tex.s = 0x1f;
+ i->setIndirectR(hnd);
+ }
i->setIndirectS(NULL);
} else if (i->tex.r == i->tex.s || i->op == OP_TXF) {
- i->tex.r += prog->driver->io.texBindBase / 4;
+ if (i->tex.r == 0xffff)
+ i->tex.r = prog->driver->io.fbtexBindBase / 4;
+ else
+ i->tex.r += prog->driver->io.texBindBase / 4;
i->tex.s = 0; // only a single cX[] value possible here
} else {
Value *hnd = bld.getScratch();
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) {
Value *ticRel = i->getIndirectR();
Value *tscRel = i->getIndirectS();
+ if (i->tex.r == 0xffff) {
+ i->tex.r = 0x20;
+ i->tex.s = 0x10;
+ }
+
if (ticRel) {
i->setSrc(i->tex.rIndirectSrc, NULL);
if (i->tex.r)
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],
if (chipset >= NVISA_GM107_CHIPSET)
s += dim;
if (i->tex.target.isArray()) {
- bld.mkOp3(OP_INSBF, TYPE_U32, i->getSrc(s),
+ Value *offset = bld.getScratch();
+ bld.mkOp3(OP_INSBF, TYPE_U32, offset,
bld.loadImm(NULL, imm), bld.mkImm(0xc10),
i->getSrc(s));
+ i->setSrc(s, offset);
} else {
i->moveSources(s, 1);
i->setSrc(s, bld.loadImm(NULL, imm << 16));
bool
NVC0LoweringPass::handleManualTXD(TexInstruction *i)
{
- static const uint8_t qOps[4][2] =
- {
- { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(MOV2, MOV2, ADD, ADD) }, // l0
- { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(MOV2, MOV2, ADD, ADD) }, // l1
- { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l2
- { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l3
- };
+ // Always done from the l0 perspective. This is the way that NVIDIA's
+ // driver does it, and doing it from the "current" lane's perpsective
+ // doesn't seem to always work for reasons that aren't altogether clear,
+ // even in frag shaders.
+ //
+ // Note that we must move not only the coordinates into lane0, but also all
+ // ancillary arguments, like array indices and depth compare as they may
+ // differ between lanes. Offsets for TXD are supposed to be uniform, so we
+ // leave them alone.
+ static const uint8_t qOps[2] =
+ { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(MOV2, MOV2, ADD, ADD) };
+
Value *def[4][4];
- Value *crd[3];
+ Value *crd[3], *arr[2], *shadow;
Instruction *tex;
Value *zero = bld.loadImm(bld.getSSA(), 0);
int l, c;
// indirect are both in the leading arg, while for Kepler, array and
// indirect are separate (and both precede the coordinates). Maxwell is
// handled in a separate function.
- unsigned array;
+ int array;
if (targ->getChipset() < NVISA_GK104_CHIPSET)
array = i->tex.target.isArray() || i->tex.rIndirectSrc >= 0;
else
for (c = 0; c < dim; ++c)
crd[c] = bld.getScratch();
+ for (c = 0; c < array; ++c)
+ arr[c] = bld.getScratch();
+ shadow = bld.getScratch();
- bld.mkOp(OP_QUADON, TYPE_NONE, NULL);
for (l = 0; l < 4; ++l) {
Value *src[3], *val;
- // mov coordinates from lane l to all lanes
+
+ bld.mkOp(OP_QUADON, TYPE_NONE, NULL);
+ // we're using the texture result from lane 0 in all cases, so make sure
+ // that lane 0 is pointing at the proper array index, indirect value,
+ // and depth compare.
+ if (l != 0) {
+ for (c = 0; c < array; ++c)
+ bld.mkQuadop(0x00, arr[c], l, i->getSrc(c), zero);
+ if (i->tex.target.isShadow()) {
+ // The next argument after coords is the depth compare
+ bld.mkQuadop(0x00, shadow, l, i->getSrc(array + dim), zero);
+ }
+ }
+ // mov position coordinates from lane l to all lanes
for (c = 0; c < dim; ++c)
bld.mkQuadop(0x00, crd[c], l, i->getSrc(c + array), zero);
// add dPdx from lane l to lanes dx
for (c = 0; c < dim; ++c)
- bld.mkQuadop(qOps[l][0], crd[c], l, i->dPdx[c].get(), crd[c]);
+ bld.mkQuadop(qOps[0], crd[c], l, i->dPdx[c].get(), crd[c]);
// add dPdy from lane l to lanes dy
for (c = 0; c < dim; ++c)
- bld.mkQuadop(qOps[l][1], crd[c], l, i->dPdy[c].get(), crd[c]);
+ bld.mkQuadop(qOps[1], crd[c], l, i->dPdy[c].get(), crd[c]);
// normalize cube coordinates
if (i->tex.target.isCube()) {
for (c = 0; c < 3; ++c)
}
// texture
bld.insert(tex = cloneForward(func, i));
+ if (l != 0) {
+ for (c = 0; c < array; ++c)
+ tex->setSrc(c, arr[c]);
+ if (i->tex.target.isShadow())
+ tex->setSrc(array + dim, shadow);
+ }
for (c = 0; c < dim; ++c)
tex->setSrc(c + array, src[c]);
+ // broadcast results from lane 0 to all lanes so that the moves *into*
+ // the target lane pick up the proper value.
+ if (l != 0)
+ for (c = 0; i->defExists(c); ++c)
+ bld.mkQuadop(0x00, tex->getDef(c), 0, tex->getDef(c), zero);
+ bld.mkOp(OP_QUADPOP, TYPE_NONE, NULL);
+
// save results
for (c = 0; i->defExists(c); ++c) {
Instruction *mov;
mov->lanes = 1 << l;
}
}
- bld.mkOp(OP_QUADPOP, TYPE_NONE, NULL);
for (c = 0; i->defExists(c); ++c) {
Instruction *u = bld.mkOp(OP_UNION, TYPE_U32, i->getDef(c));
txq->moveSources(0, 1);
txq->setSrc(0, src);
} else {
- Value *hnd = loadTexHandle(
- bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
- txq->getIndirectR(), bld.mkImm(2)),
- txq->tex.r);
+ Value *hnd = loadTexHandle(txq->getIndirectR(), txq->tex.r);
txq->tex.r = 0xff;
txq->tex.s = 0x1f;
mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off + 8), ptr);
}
-inline Value *
-NVC0LoweringPass::loadSuInfo32(Value *ptr, uint32_t off)
-{
- return loadResInfo32(ptr, off, prog->driver->io.suInfoBase);
-}
-
-inline Value *
-NVC0LoweringPass::loadSuInfo64(Value *ptr, uint32_t off)
-{
- return loadResInfo64(ptr, off, prog->driver->io.suInfoBase);
-}
-
-inline Value *
-NVC0LoweringPass::loadSuLength32(Value *ptr, uint32_t off)
-{
- return loadResLength32(ptr, off, prog->driver->io.suInfoBase);
-}
-
-inline Value *
-NVC0LoweringPass::loadBufInfo32(Value *ptr, uint32_t off)
-{
- return loadResInfo32(ptr, off, prog->driver->io.bufInfoBase);
-}
-
inline Value *
NVC0LoweringPass::loadBufInfo64(Value *ptr, uint32_t off)
{
return loadResLength32(ptr, off, prog->driver->io.bufInfoBase);
}
-inline Value *
-NVC0LoweringPass::loadUboInfo32(Value *ptr, uint32_t off)
-{
- return loadResInfo32(ptr, off, prog->driver->io.uboInfoBase);
-}
-
inline Value *
NVC0LoweringPass::loadUboInfo64(Value *ptr, uint32_t off)
{
mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
}
-/* On nvc0, surface info is obtained via the surface binding points passed
- * to the SULD/SUST instructions.
- * On nve4, surface info is stored in c[] and is used by various special
- * instructions, e.g. for clamping coordiantes or generating an address.
- * They couldn't just have added an equivalent to TIC now, couldn't they ?
- */
-#define NVE4_SU_INFO_ADDR 0x00
-#define NVE4_SU_INFO_FMT 0x04
-#define NVE4_SU_INFO_DIM_X 0x08
-#define NVE4_SU_INFO_PITCH 0x0c
-#define NVE4_SU_INFO_DIM_Y 0x10
-#define NVE4_SU_INFO_ARRAY 0x14
-#define NVE4_SU_INFO_DIM_Z 0x18
-#define NVE4_SU_INFO_UNK1C 0x1c
-#define NVE4_SU_INFO_WIDTH 0x20
-#define NVE4_SU_INFO_HEIGHT 0x24
-#define NVE4_SU_INFO_DEPTH 0x28
-#define NVE4_SU_INFO_TARGET 0x2c
-#define NVE4_SU_INFO_BSIZE 0x30
-#define NVE4_SU_INFO_RAW_X 0x34
-#define NVE4_SU_INFO_MS_X 0x38
-#define NVE4_SU_INFO_MS_Y 0x3c
-
-#define NVE4_SU_INFO__STRIDE 0x40
-
-#define NVE4_SU_INFO_DIM(i) (0x08 + (i) * 8)
-#define NVE4_SU_INFO_SIZE(i) (0x20 + (i) * 4)
-#define NVE4_SU_INFO_MS(i) (0x38 + (i) * 4)
+inline Value *
+NVC0LoweringPass::loadSuInfo32(Value *ptr, int slot, uint32_t off, bool bindless)
+{
+ uint32_t base = slot * NVC0_SU_INFO__STRIDE;
+
+ if (ptr) {
+ ptr = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(slot));
+ if (bindless)
+ ptr = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(511));
+ else
+ ptr = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(7));
+ ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(6));
+ base = 0;
+ }
+ off += base;
+
+ return loadResInfo32(ptr, off, bindless ? prog->driver->io.bindlessBase :
+ prog->driver->io.suInfoBase);
+}
+
+Value *
+NVC0LoweringPass::loadMsAdjInfo32(TexInstruction::Target target, uint32_t index, int slot, Value *ind, bool bindless)
+{
+ if (!bindless || targ->getChipset() < NVISA_GM107_CHIPSET)
+ return loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(index), bindless);
+
+ assert(bindless);
+
+ Value *samples = bld.getSSA();
+ // this shouldn't be lowered because it's being inserted before the current instruction
+ TexInstruction *tex = new_TexInstruction(func, OP_TXQ);
+ tex->tex.target = target;
+ tex->tex.query = TXQ_TYPE;
+ tex->tex.mask = 0x4;
+ tex->tex.r = 0xff;
+ tex->tex.s = 0x1f;
+ tex->tex.rIndirectSrc = 0;
+ tex->setDef(0, samples);
+ tex->setSrc(0, ind);
+ tex->setSrc(1, bld.loadImm(NULL, 0));
+ bld.insert(tex);
+
+ // doesn't work with sample counts other than 1/2/4/8 but they aren't supported
+ switch (index) {
+ case 0: {
+ Value *tmp = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), samples, bld.mkImm(2));
+ return bld.mkOp2v(OP_SHR, TYPE_U32, bld.getSSA(), tmp, bld.mkImm(2));
+ }
+ case 1: {
+ Value *tmp = bld.mkCmp(OP_SET, CC_GT, TYPE_U32, bld.getSSA(), TYPE_U32, samples, bld.mkImm(2))->getDef(0);
+ return bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), tmp, bld.mkImm(1));
+ }
+ default: {
+ assert(false);
+ return NULL;
+ }
+ }
+}
static inline uint16_t getSuClampSubOp(const TexInstruction *su, int c)
{
int dim = suq->tex.target.getDim();
int arg = dim + (suq->tex.target.isArray() || suq->tex.target.isCube());
Value *ind = suq->getIndirectR();
- uint32_t base;
+ int slot = suq->tex.r;
int c, d;
- if (ind) {
- ind = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(),
- ind, bld.mkImm(suq->tex.r));
- ind = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(),
- ind, bld.mkImm(7));
- ind = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
- ind, bld.mkImm(6));
- base = 0;
- } else {
- base = suq->tex.r * NVE4_SU_INFO__STRIDE;
- }
-
for (c = 0, d = 0; c < 3; ++c, mask >>= 1) {
if (c >= arg || !(mask & 1))
continue;
int offset;
if (c == 1 && suq->tex.target == TEX_TARGET_1D_ARRAY) {
- offset = NVE4_SU_INFO_SIZE(2);
+ offset = NVC0_SU_INFO_SIZE(2);
} else {
- offset = NVE4_SU_INFO_SIZE(c);
+ offset = NVC0_SU_INFO_SIZE(c);
}
- bld.mkMov(suq->getDef(d++), loadSuInfo32(ind, base + offset));
+ bld.mkMov(suq->getDef(d++), loadSuInfo32(ind, slot, offset, suq->tex.bindless));
if (c == 2 && suq->tex.target.isCube())
bld.mkOp2(OP_DIV, TYPE_U32, suq->getDef(d - 1), suq->getDef(d - 1),
bld.loadImm(NULL, 6));
if (mask & 1) {
if (suq->tex.target.isMS()) {
- Value *ms_x = loadSuInfo32(ind, base + NVE4_SU_INFO_MS(0));
- Value *ms_y = loadSuInfo32(ind, base + NVE4_SU_INFO_MS(1));
+ Value *ms_x = loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(0), suq->tex.bindless);
+ Value *ms_y = loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(1), suq->tex.bindless);
Value *ms = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(), ms_x, ms_y);
bld.mkOp2(OP_SHL, TYPE_U32, suq->getDef(d++), bld.loadImm(NULL, 1), ms);
} else {
void
NVC0LoweringPass::adjustCoordinatesMS(TexInstruction *tex)
{
- uint16_t base;
const int arg = tex->tex.target.getArgCount();
+ int slot = tex->tex.r;
if (tex->tex.target == TEX_TARGET_2D_MS)
tex->tex.target = TEX_TARGET_2D;
Value *tx = bld.getSSA(), *ty = bld.getSSA(), *ts = bld.getSSA();
Value *ind = tex->getIndirectR();
- if (ind) {
- ind = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(),
- ind, bld.mkImm(tex->tex.r));
- ind = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(),
- ind, bld.mkImm(7));
- ind = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
- ind, bld.mkImm(6));
- base = 0;
- } else {
- base = tex->tex.r * NVE4_SU_INFO__STRIDE;
- }
-
- Value *ms_x = loadSuInfo32(ind, base + NVE4_SU_INFO_MS(0));
- Value *ms_y = loadSuInfo32(ind, base + NVE4_SU_INFO_MS(1));
+ Value *ms_x = loadMsAdjInfo32(tex->tex.target, 0, slot, ind, tex->tex.bindless);
+ Value *ms_y = loadMsAdjInfo32(tex->tex.target, 1, slot, ind, tex->tex.bindless);
bld.mkOp2(OP_SHL, TYPE_U32, tx, x, ms_x);
bld.mkOp2(OP_SHL, TYPE_U32, ty, y, ms_y);
const bool atom = su->op == OP_SUREDB || su->op == OP_SUREDP;
const bool raw =
su->op == OP_SULDB || su->op == OP_SUSTB || su->op == OP_SUREDB;
- const int idx = su->tex.r;
+ 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());
- const uint16_t base = idx * NVE4_SU_INFO__STRIDE;
int c;
Value *zero = bld.mkImm(0);
Value *p1 = NULL;
Value *src[3];
Value *bf, *eau, *off;
Value *addr, *pred;
- Value *ind = NULL;
+ Value *ind = su->getIndirectR();
off = bld.getScratch(4);
bf = bld.getScratch(4);
adjustCoordinatesMS(su);
- if (su->tex.rIndirectSrc >= 0) {
- ind = su->getIndirectR();
- if (su->tex.r > 0) {
- ind = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), ind,
- bld.loadImm(NULL, su->tex.r));
- }
- ind = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ind, bld.mkImm(7));
- ind = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), ind, bld.mkImm(6));
- }
-
// calculate clamped coordinates
for (c = 0; c < arg; ++c) {
int dimc = c;
src[c] = bld.getScratch();
if (c == 0 && raw)
- v = loadSuInfo32(ind, base + NVE4_SU_INFO_RAW_X);
+ v = loadSuInfo32(ind, slot, NVC0_SU_INFO_RAW_X, su->tex.bindless);
else
- v = loadSuInfo32(ind, base + NVE4_SU_INFO_DIM(dimc));
+ v = loadSuInfo32(ind, slot, NVC0_SU_INFO_DIM(dimc), su->tex.bindless);
bld.mkOp3(OP_SUCLAMP, TYPE_S32, src[c], su->getSrc(c), v, zero)
->subOp = getSuClampSubOp(su, dimc);
}
bld.mkOp2(OP_AND, TYPE_U32, off, src[0], bld.loadImm(NULL, 0xffff));
} else
if (dim == 3) {
- v = loadSuInfo32(ind, base + NVE4_SU_INFO_UNK1C);
+ v = loadSuInfo32(ind, slot, NVC0_SU_INFO_UNK1C, su->tex.bindless);
bld.mkOp3(OP_MADSP, TYPE_U32, off, src[2], v, src[1])
->subOp = NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
- v = loadSuInfo32(ind, base + NVE4_SU_INFO_PITCH);
+ v = loadSuInfo32(ind, slot, NVC0_SU_INFO_PITCH, su->tex.bindless);
bld.mkOp3(OP_MADSP, TYPE_U32, off, off, v, src[0])
->subOp = NV50_IR_SUBOP_MADSP(0,2,8); // u32 u16l u16l
} else {
assert(dim == 2);
- v = loadSuInfo32(ind, base + NVE4_SU_INFO_PITCH);
+ v = loadSuInfo32(ind, slot, NVC0_SU_INFO_PITCH, su->tex.bindless);
bld.mkOp3(OP_MADSP, TYPE_U32, off, src[1], v, src[0])
->subOp = (su->tex.target.isArray() || su->tex.target.isCube()) ?
NV50_IR_SUBOP_MADSP_SD : NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
if (raw) {
bf = src[0];
} else {
- v = loadSuInfo32(ind, base + NVE4_SU_INFO_FMT);
+ v = loadSuInfo32(ind, slot, NVC0_SU_INFO_FMT, su->tex.bindless);
bld.mkOp3(OP_VSHL, TYPE_U32, bf, src[0], v, zero)
->subOp = NV50_IR_SUBOP_V1(7,6,8|2);
}
case 2:
z = off;
if (!su->tex.target.isArray() && !su->tex.target.isCube()) {
- z = loadSuInfo32(ind, base + NVE4_SU_INFO_UNK1C);
+ z = loadSuInfo32(ind, slot, NVC0_SU_INFO_UNK1C, su->tex.bindless);
subOp = NV50_IR_SUBOP_SUBFM_3D;
}
break;
}
// part 2
- v = loadSuInfo32(ind, base + NVE4_SU_INFO_ADDR);
+ v = loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR, su->tex.bindless);
if (su->tex.target == TEX_TARGET_BUFFER) {
eau = v;
}
// add array layer offset
if (su->tex.target.isArray() || su->tex.target.isCube()) {
- v = loadSuInfo32(ind, base + NVE4_SU_INFO_ARRAY);
+ v = loadSuInfo32(ind, slot, NVC0_SU_INFO_ARRAY, su->tex.bindless);
if (dim == 1)
bld.mkOp3(OP_MADSP, TYPE_U32, eau, src[1], v, eau)
->subOp = NV50_IR_SUBOP_MADSP(4,0,0); // u16 u24 u32
// let's just set it 0 for raw access and hope it works
v = raw ?
- bld.mkImm(0) : loadSuInfo32(ind, base + NVE4_SU_INFO_FMT);
+ bld.mkImm(0) : loadSuInfo32(ind, slot, NVC0_SU_INFO_FMT, su->tex.bindless);
// get rid of old coordinate sources, make space for fmt info and predicate
su->moveSources(arg, 3 - arg);
su->setSrc(0, addr);
su->setSrc(1, v);
su->setSrc(2, pred);
+ su->setIndirectR(NULL);
// prevent read fault when the image is not actually bound
CmpInstruction *pred1 =
bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
TYPE_U32, bld.mkImm(0),
- loadSuInfo32(ind, base + NVE4_SU_INFO_ADDR));
+ loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR, su->tex.bindless));
if (su->op != OP_SUSTP && su->tex.format) {
const TexInstruction::ImgFormatDesc *format = su->tex.format;
assert(format->components != 0);
bld.mkCmp(OP_SET_OR, CC_NE, TYPE_U32, pred1->getDef(0),
TYPE_U32, bld.loadImm(NULL, blockwidth / 8),
- loadSuInfo32(ind, base + NVE4_SU_INFO_BSIZE),
+ loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE, su->tex.bindless),
pred1->getDef(0));
}
su->setPredicate(CC_NOT_P, pred1->getDef(0));
bld.mkCvt(OP_CVT, TYPE_F32, typedDst[i], TYPE_F16, typedDst[i]);
}
}
+
+ if (format->bgra) {
+ std::swap(typedDst[0], typedDst[2]);
+ }
+}
+
+void
+NVC0LoweringPass::insertOOBSurfaceOpResult(TexInstruction *su)
+{
+ if (!su->getPredicate())
+ return;
+
+ bld.setPosition(su, true);
+
+ for (unsigned i = 0; su->defExists(i); ++i) {
+ ValueDef &def = su->def(i);
+
+ Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));
+ assert(su->cc == CC_NOT_P);
+ mov->setPredicate(CC_P, su->getPredicate());
+ Instruction *uni = bld.mkOp2(OP_UNION, TYPE_U32, bld.getSSA(), NULL, mov->getDef(0));
+
+ def.replace(uni->getDef(0), false);
+ uni->setSrc(0, def.get());
+ }
}
void
{
processSurfaceCoordsNVE4(su);
- if (su->op == OP_SULDP)
+ if (su->op == OP_SULDP) {
convertSurfaceFormat(su);
+ insertOOBSurfaceOpResult(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::processSurfaceCoordsNVC0(TexInstruction *su)
{
- const int idx = su->tex.r;
+ 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());
- const uint16_t base = idx * NVE4_SU_INFO__STRIDE;
int c;
Value *zero = bld.mkImm(0);
Value *src[3];
Value *v;
- Value *ind = NULL;
+ Value *ind = su->getIndirectR();
bld.setPosition(su, false);
adjustCoordinatesMS(su);
- if (su->tex.rIndirectSrc >= 0) {
- ind = su->getIndirectR();
- if (su->tex.r > 0) {
- ind = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), ind,
- bld.loadImm(NULL, su->tex.r));
- }
- ind = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ind, bld.mkImm(7));
- ind = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), ind, bld.mkImm(6));
+ if (ind) {
+ Value *ptr;
+ ptr = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), ind, bld.mkImm(su->tex.r));
+ ptr = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(7));
+ su->setIndirectR(ptr);
}
// get surface coordinates
// calculate pixel offset
if (su->op == OP_SULDP || su->op == OP_SUREDP) {
- v = loadSuInfo32(ind, base + NVE4_SU_INFO_BSIZE);
+ v = loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE, su->tex.bindless);
su->setSrc(0, bld.mkOp2v(OP_MUL, TYPE_U32, bld.getSSA(), src[0], v));
}
// add array layer offset
if (su->tex.target.isArray() || su->tex.target.isCube()) {
- v = loadSuInfo32(ind, base + NVE4_SU_INFO_ARRAY);
+ v = loadSuInfo32(ind, slot, NVC0_SU_INFO_ARRAY, su->tex.bindless);
assert(dim > 1);
su->setSrc(2, bld.mkOp2v(OP_MUL, TYPE_U32, bld.getSSA(), src[2], v));
}
CmpInstruction *pred =
bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
TYPE_U32, bld.mkImm(0),
- loadSuInfo32(ind, base + NVE4_SU_INFO_ADDR));
+ loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR, su->tex.bindless));
if (su->op != OP_SUSTP && su->tex.format) {
const TexInstruction::ImgFormatDesc *format = su->tex.format;
int blockwidth = format->bits[0] + format->bits[1] +
// 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, base + NVE4_SU_INFO_BSIZE),
+ loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE, su->tex.bindless),
pred->getDef(0));
}
su->setPredicate(CC_NOT_P, pred->getDef(0));
processSurfaceCoordsNVC0(su);
- if (su->op == OP_SULDP)
+ if (su->op == OP_SULDP) {
convertSurfaceFormat(su);
+ insertOOBSurfaceOpResult(su);
+ }
if (su->op == OP_SUREDB || su->op == OP_SUREDP) {
const int dim = su->tex.target.getDim();
}
}
+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();
+ Value *handle;
+ int pos = 0;
+
+ bld.setPosition(su, false);
+
+ adjustCoordinatesMS(su);
+
+ // 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;
+ }
+ if (su->tex.bindless)
+ handle = ind;
+ else
+ handle = loadTexHandle(ind, slot + 32);
+ su->setSrc(arg + pos, handle);
+
+ // The address check doesn't make sense here. The format check could make
+ // sense but it's a bit of a pain.
+ if (su->tex.bindless)
+ return;
+
+ // 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, su->tex.bindless));
+ 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, su->tex.bindless),
+ 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);
+ insertOOBSurfaceOpResult(su);
+ }
+
+ if (su->op == OP_SUREDP) {
+ Value *def = su->getDef(0);
+
+ su->op = OP_SUREDB;
+
+ // There may not be a predicate in the bindless case.
+ if (su->getPredicate()) {
+ 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)
{
assert(prog->getType() != Program::TYPE_FRAGMENT); // INTERP
}
} else if (i->src(0).getFile() == FILE_MEMORY_CONST) {
+ int8_t fileIndex = i->getSrc(0)->reg.fileIndex - 1;
+ Value *ind = i->getIndirect(0, 1);
+
if (targ->getChipset() >= NVISA_GK104_CHIPSET &&
- prog->getType() == Program::TYPE_COMPUTE) {
+ prog->getType() == Program::TYPE_COMPUTE &&
+ (fileIndex >= 6 || ind)) {
// The launch descriptor only allows to set up 8 CBs, but OpenGL
- // requires at least 12 UBOs. To bypass this limitation, we store the
- // addrs into the driver constbuf and we directly load from the global
- // memory.
- int8_t fileIndex = i->getSrc(0)->reg.fileIndex - 1;
- Value *ind = i->getIndirect(0, 1);
-
+ // requires at least 12 UBOs. To bypass this limitation, for constant
+ // buffers 7+, we store the addrs into the driver constbuf and we
+ // directly load from the global memory.
if (ind) {
// Clamp the UBO index when an indirect access is used to avoid
// loading information from the wrong place in the driver cb.
- ind = bld.mkOp2v(OP_MIN, TYPE_U32, ind,
+ // TODO - synchronize the max with the driver.
+ ind = bld.mkOp2v(OP_MIN, TYPE_U32, bld.getSSA(),
bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(),
ind, bld.loadImm(NULL, fileIndex)),
- bld.loadImm(NULL, 12));
+ bld.loadImm(NULL, 13));
+ fileIndex = 0;
}
- if (i->src(0).isIndirect(1)) {
- Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
- Value *ptr = loadUboInfo64(ind, fileIndex * 16);
- Value *length = loadUboLength32(ind, fileIndex * 16);
- Value *pred = new_LValue(func, FILE_PREDICATE);
- if (i->src(0).isIndirect(0)) {
- bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
- bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
- }
- i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
- i->setIndirect(0, 1, NULL);
- i->setIndirect(0, 0, ptr);
- bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
- i->setPredicate(CC_NOT_P, pred);
- if (i->defExists(0)) {
- bld.mkMov(i->getDef(0), bld.mkImm(0));
- }
- } else if (fileIndex >= 0) {
- Value *ptr = loadUboInfo64(ind, fileIndex * 16);
- if (i->src(0).isIndirect(0)) {
- bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
- }
- i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
- i->setIndirect(0, 1, NULL);
- i->setIndirect(0, 0, ptr);
+ Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
+ Value *ptr = loadUboInfo64(ind, fileIndex * 16);
+ Value *length = loadUboLength32(ind, fileIndex * 16);
+ Value *pred = new_LValue(func, FILE_PREDICATE);
+ if (i->src(0).isIndirect(0)) {
+ bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
+ bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
}
+ i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
+ i->setIndirect(0, 1, NULL);
+ i->setIndirect(0, 0, ptr);
+ bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
+ i->setPredicate(CC_NOT_P, pred);
+ Value *zero, *dst = i->getDef(0);
+ i->setDef(0, bld.getSSA());
+
+ bld.setPosition(i, true);
+ bld.mkMov((zero = bld.getSSA()), bld.mkImm(0))
+ ->setPredicate(CC_P, pred);
+ bld.mkOp2(OP_UNION, TYPE_U32, dst, i->getDef(0), zero);
} else if (i->src(0).isIndirect(1)) {
Value *ptr;
if (i->src(0).isIndirect(0))
// TGSI backend may use 4th component of TID,NTID,CTAID,NCTAID
i->op = OP_MOV;
i->setSrc(0, bld.mkImm((sv == SV_NTID || sv == SV_NCTAID) ? 1 : 0));
+ } else
+ if (sv == SV_TID) {
+ // Help CSE combine TID fetches
+ Value *tid = bld.mkOp1v(OP_RDSV, TYPE_U32, bld.getScratch(),
+ bld.mkSysVal(SV_COMBINED_TID, 0));
+ i->op = OP_EXTBF;
+ i->setSrc(0, tid);
+ switch (sym->reg.data.sv.index) {
+ case 0: i->setSrc(1, bld.mkImm(0x1000)); break;
+ case 1: i->setSrc(1, bld.mkImm(0x0a10)); break;
+ case 2: i->setSrc(1, bld.mkImm(0x061a)); break;
+ }
}
if (sv == SV_VERTEX_COUNT) {
bld.setPosition(i, true);
i->setSrc(0, bld.mkImm(sv == SV_GRIDID ? 0 : 1));
return true;
}
+ // Fallthrough
+ case SV_WORK_DIM:
addr += prog->driver->prop.cp.gridInfoBase;
bld.mkLoad(TYPE_U32, i->getDef(0),
bld.mkSymbol(FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
break;
case SV_SAMPLE_POS: {
- Value *off = new_LValue(func, FILE_GPR);
- ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
+ Value *sampleID = bld.getScratch();
+ ld = bld.mkOp1(OP_PIXLD, TYPE_U32, sampleID, bld.mkImm(0));
ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
- bld.mkOp2(OP_SHL, TYPE_U32, off, i->getDef(0), bld.mkImm(3));
- bld.mkLoad(TYPE_F32,
- i->getDef(0),
- bld.mkSymbol(
- FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
- TYPE_U32, prog->driver->io.sampleInfoBase +
- 4 * sym->reg.data.sv.index),
- off);
+ Value *offset = calculateSampleOffset(sampleID);
+
+ assert(prog->driver->prop.fp.readsSampleLocations);
+
+ if (targ->getChipset() >= NVISA_GM200_CHIPSET) {
+ bld.mkLoad(TYPE_F32,
+ i->getDef(0),
+ bld.mkSymbol(
+ FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
+ TYPE_U32, prog->driver->io.sampleInfoBase),
+ offset);
+ bld.mkOp2(OP_EXTBF, TYPE_U32, i->getDef(0), i->getDef(0),
+ bld.mkImm(0x040c + sym->reg.data.sv.index * 16));
+ bld.mkCvt(OP_CVT, TYPE_F32, i->getDef(0), TYPE_U32, i->getDef(0));
+ bld.mkOp2(OP_MUL, TYPE_F32, i->getDef(0), i->getDef(0), bld.mkImm(1.0f / 16.0f));
+ } else {
+ bld.mkLoad(TYPE_F32,
+ i->getDef(0),
+ bld.mkSymbol(
+ FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
+ TYPE_U32, prog->driver->io.sampleInfoBase +
+ 4 * sym->reg.data.sv.index),
+ offset);
+ }
break;
}
case SV_SAMPLE_MASK: {
default:
if (prog->getType() == Program::TYPE_TESSELLATION_EVAL && !i->perPatch)
vtx = bld.mkOp1v(OP_PFETCH, TYPE_U32, bld.getSSA(), bld.mkImm(0));
- ld = bld.mkFetch(i->getDef(0), i->dType,
- FILE_SHADER_INPUT, addr, i->getIndirect(0, 0), vtx);
- ld->perPatch = i->perPatch;
+ if (prog->getType() == Program::TYPE_FRAGMENT) {
+ bld.mkInterp(NV50_IR_INTERP_FLAT, i->getDef(0), addr, NULL);
+ } else {
+ ld = bld.mkFetch(i->getDef(0), i->dType,
+ FILE_SHADER_INPUT, addr, i->getIndirect(0, 0), vtx);
+ ld->perPatch = i->perPatch;
+ }
break;
}
bld.getBB()->remove(i);
bool
NVC0LoweringPass::handleSQRT(Instruction *i)
{
+ if (targ->isOpSupported(OP_SQRT, i->dType))
+ return true;
+
if (i->dType == TYPE_F64) {
Value *pred = bld.getSSA(1, FILE_PREDICATE);
Value *zero = bld.loadImm(NULL, 0.0);
return true;
}
+Value *
+NVC0LoweringPass::calculateSampleOffset(Value *sampleID)
+{
+ Value *offset = bld.getScratch();
+ if (targ->getChipset() >= NVISA_GM200_CHIPSET) {
+ // Sample location offsets (in bytes) are calculated like so:
+ // offset = (SV_POSITION.y % 4 * 2) + (SV_POSITION.x % 2)
+ // offset = offset * 32 + sampleID % 8 * 4;
+ // which is equivalent to:
+ // offset = (SV_POSITION.y & 0x3) << 6 + (SV_POSITION.x & 0x1) << 5;
+ // offset += sampleID << 2
+
+ // The second operand (src1) of the INSBF instructions are like so:
+ // 0xssll where ss is the size and ll is the offset.
+ // so: dest = src2 | (src0 & (1 << ss - 1)) << ll
+
+ // Add sample ID (offset = (sampleID & 0x7) << 2)
+ bld.mkOp3(OP_INSBF, TYPE_U32, offset, sampleID, bld.mkImm(0x0302), bld.mkImm(0x0));
+
+ Symbol *xSym = bld.mkSysVal(SV_POSITION, 0);
+ Symbol *ySym = bld.mkSysVal(SV_POSITION, 1);
+ Value *coord = bld.getScratch();
+
+ // Add X coordinate (offset |= (SV_POSITION.x & 0x1) << 5)
+ bld.mkInterp(NV50_IR_INTERP_LINEAR, coord,
+ targ->getSVAddress(FILE_SHADER_INPUT, xSym), NULL);
+ bld.mkCvt(OP_CVT, TYPE_U32, coord, TYPE_F32, coord)
+ ->rnd = ROUND_ZI;
+ bld.mkOp3(OP_INSBF, TYPE_U32, offset, coord, bld.mkImm(0x0105), offset);
+
+ // Add Y coordinate (offset |= (SV_POSITION.y & 0x3) << 6)
+ bld.mkInterp(NV50_IR_INTERP_LINEAR, coord,
+ targ->getSVAddress(FILE_SHADER_INPUT, ySym), NULL);
+ bld.mkCvt(OP_CVT, TYPE_U32, coord, TYPE_F32, coord)
+ ->rnd = ROUND_ZI;
+ bld.mkOp3(OP_INSBF, TYPE_U32, offset, coord, bld.mkImm(0x0206), offset);
+ } else {
+ bld.mkOp2(OP_SHL, TYPE_U32, offset, sampleID, bld.mkImm(3));
+ }
+ return offset;
+}
+
+// Handle programmable sample locations for GM20x+
+void
+NVC0LoweringPass::handlePIXLD(Instruction *i)
+{
+ if (i->subOp != NV50_IR_SUBOP_PIXLD_OFFSET)
+ return;
+ if (targ->getChipset() < NVISA_GM200_CHIPSET)
+ return;
+
+ assert(prog->driver->prop.fp.readsSampleLocations);
+
+ bld.mkLoad(TYPE_F32,
+ i->getDef(0),
+ bld.mkSymbol(
+ FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
+ TYPE_U32, prog->driver->io.sampleInfoBase),
+ calculateSampleOffset(i->getSrc(0)));
+
+ bld.getBB()->remove(i);
+}
+
// Generate a binary predicate if an instruction is predicated by
// e.g. an f32 value.
void
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());
case OP_BUFQ:
handleBUFQ(i);
break;
+ case OP_PIXLD:
+ handlePIXLD(i);
+ break;
default:
break;
}
/* 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));
}