From ab74b8e3549838c0c480555134f5451949bac59f Mon Sep 17 00:00:00 2001 From: Stephane Marchesin Date: Sun, 28 Sep 2008 18:33:23 +0200 Subject: Gallivm: make it compile again, add some opcodes. --- src/gallium/auxiliary/draw/draw_vs_llvm.c | 1 + src/gallium/auxiliary/gallivm/gallivm_cpu.cpp | 1 + src/gallium/auxiliary/gallivm/instructions.cpp | 1210 ++++++++++++++---------- src/gallium/auxiliary/gallivm/instructions.h | 26 +- src/gallium/auxiliary/gallivm/tgsitollvm.cpp | 60 +- 5 files changed, 792 insertions(+), 506 deletions(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/draw/draw_vs_llvm.c b/src/gallium/auxiliary/draw/draw_vs_llvm.c index 2ce30b9a02..727977bc3a 100644 --- a/src/gallium/auxiliary/draw/draw_vs_llvm.c +++ b/src/gallium/auxiliary/draw/draw_vs_llvm.c @@ -32,6 +32,7 @@ * Brian Paul */ +#include "util/u_memory.h" #include "pipe/p_shader_tokens.h" #include "draw_private.h" #include "draw_context.h" diff --git a/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp b/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp index e64bfb1c6c..3a4a41e544 100644 --- a/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp +++ b/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp @@ -46,6 +46,7 @@ #include "tgsi/tgsi_dump.h" #include "util/u_memory.h" +#include "util/u_math.h" #include #include diff --git a/src/gallium/auxiliary/gallivm/instructions.cpp b/src/gallium/auxiliary/gallivm/instructions.cpp index a82dc30306..5fdfe09d18 100644 --- a/src/gallium/auxiliary/gallivm/instructions.cpp +++ b/src/gallium/auxiliary/gallivm/instructions.cpp @@ -83,6 +83,7 @@ Instructions::Instructions(llvm::Module *mod, llvm::Function *func, llvm::BasicB m_llvmPow = 0; m_llvmFloor = 0; m_llvmFlog = 0; + m_llvmFexp = 0; m_llvmLit = 0; m_fmtPtr = 0; @@ -92,194 +93,247 @@ Instructions::Instructions(llvm::Module *mod, llvm::Function *func, llvm::BasicB m_mod = ParseBitcodeFile(buffer); } -llvm::Value * Instructions::add(llvm::Value *in1, llvm::Value *in2) +llvm::BasicBlock * Instructions::currentBlock() const { - return m_builder.CreateAdd(in1, in2, name("add")); + return m_builder.GetInsertBlock(); } -llvm::Value * Instructions::madd(llvm::Value *in1, llvm::Value *in2, - llvm::Value *in3) +llvm::Value * Instructions::abs(llvm::Value *in) { - Value *mulRes = mul(in1, in2); - return add(mulRes, in3); + std::vector vec = extractVector(in); + Value *xabs = callFAbs(vec[0]); + Value *yabs = callFAbs(vec[1]); + Value *zabs = callFAbs(vec[2]); + Value *wabs = callFAbs(vec[3]); + return vectorFromVals(xabs, yabs, zabs, wabs); } - -llvm::Value * Instructions::mul(llvm::Value *in1, llvm::Value *in2) + +llvm::Value * Instructions::add(llvm::Value *in1, llvm::Value *in2) { - return m_builder.CreateMul(in1, in2, name("mul")); + return m_builder.CreateAdd(in1, in2, name("add")); } -const char * Instructions::name(const char *prefix) +llvm::Value * Instructions::arl(llvm::Value *in) { - ++m_idx; - snprintf(m_name, 32, "%s%d", prefix, m_idx); - return m_name; + return floor(in); } -llvm::Value * Instructions::dp3(llvm::Value *in1, llvm::Value *in2) +void Instructions::beginLoop() { - Value *mulRes = mul(in1, in2); - Value *x = m_builder.CreateExtractElement(mulRes, - m_storage->constantInt(0), - name("extractx")); - Value *y = m_builder.CreateExtractElement(mulRes, - m_storage->constantInt(1), - name("extracty")); - Value *z = m_builder.CreateExtractElement(mulRes, - m_storage->constantInt(2), - name("extractz")); - Value *xy = m_builder.CreateAdd(x, y,name("xy")); - Value *dot3 = m_builder.CreateAdd(xy, z, name("dot3")); - return vectorFromVals(dot3, dot3, dot3, dot3); + BasicBlock *begin = BasicBlock::Create(name("loop"), m_func,0); + BasicBlock *end = BasicBlock::Create(name("endloop"), m_func,0); + + m_builder.CreateBr(begin); + Loop loop; + loop.begin = begin; + loop.end = end; + m_builder.SetInsertPoint(begin); + m_loopStack.push(loop); } -llvm::Value *Instructions::callFSqrt(llvm::Value *val) +void Instructions::bgnSub(unsigned label) { - if (!m_llvmFSqrt) { - // predeclare the intrinsic - std::vector fsqrtArgs; - fsqrtArgs.push_back(Type::FloatTy); - PAListPtr fsqrtPal; - FunctionType* fsqrtType = FunctionType::get( - /*Result=*/Type::FloatTy, - /*Params=*/fsqrtArgs, - /*isVarArg=*/false); - m_llvmFSqrt = Function::Create( - /*Type=*/fsqrtType, - /*Linkage=*/GlobalValue::ExternalLinkage, - /*Name=*/"llvm.sqrt.f32", m_mod); - m_llvmFSqrt->setCallingConv(CallingConv::C); - m_llvmFSqrt->setParamAttrs(fsqrtPal); - } - CallInst *call = m_builder.CreateCall(m_llvmFSqrt, val, - name("sqrt")); - call->setCallingConv(CallingConv::C); - call->setTailCall(false); - return call; + llvm::Function *func = findFunction(label); + + Function::arg_iterator args = func->arg_begin(); + Value *ptr_INPUT = args++; + ptr_INPUT->setName("INPUT"); + m_storage->pushArguments(ptr_INPUT); + + llvm::BasicBlock *entry = BasicBlock::Create("entry", func, 0); + + m_func = func; + m_builder.SetInsertPoint(entry); } -llvm::Value * Instructions::rsq(llvm::Value *in1) +void Instructions::brk() { - Value *x = m_builder.CreateExtractElement(in1, - m_storage->constantInt(0), - name("extractx")); - Value *abs = callFAbs(x); - Value *sqrt = callFSqrt(abs); + assert(!m_loopStack.empty()); + BasicBlock *unr = BasicBlock::Create(name("unreachable"), m_func,0); + m_builder.CreateBr(m_loopStack.top().end); + m_builder.SetInsertPoint(unr); +} - Value *rsqrt = m_builder.CreateFDiv(ConstantFP::get(APFloat(1.f)), - sqrt, - name("rsqrt")); - return vectorFromVals(rsqrt, rsqrt, rsqrt, rsqrt); +void Instructions::cal(int label, llvm::Value *input) +{ + std::vector params; + params.push_back(input); + llvm::Function *func = findFunction(label); + + m_builder.CreateCall(func, params.begin(), params.end()); } -llvm::Value * Instructions::vectorFromVals(llvm::Value *x, llvm::Value *y, - llvm::Value *z, llvm::Value *w) +llvm::Value * Instructions::clamp(llvm::Value *in1) { - Constant *const_vec = Constant::getNullValue(m_floatVecType); - Value *res = m_builder.CreateInsertElement(const_vec, x, - m_storage->constantInt(0), - name("vecx")); - res = m_builder.CreateInsertElement(res, y, m_storage->constantInt(1), - name("vecxy")); - res = m_builder.CreateInsertElement(res, z, m_storage->constantInt(2), - name("vecxyz")); - if (w) - res = m_builder.CreateInsertElement(res, w, m_storage->constantInt(3), - name("vecxyzw")); - return res; + // FIXME } -llvm::Value *Instructions::callFAbs(llvm::Value *val) +llvm::Value * Instructions::cmp(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3) { - if (!m_llvmFAbs) { - // predeclare the intrinsic - std::vector fabsArgs; - fabsArgs.push_back(Type::FloatTy); - PAListPtr fabsPal; - FunctionType* fabsType = FunctionType::get( - /*Result=*/Type::FloatTy, - /*Params=*/fabsArgs, - /*isVarArg=*/false); - m_llvmFAbs = Function::Create( - /*Type=*/fabsType, - /*Linkage=*/GlobalValue::ExternalLinkage, - /*Name=*/"fabs", m_mod); - m_llvmFAbs->setCallingConv(CallingConv::C); - m_llvmFAbs->setParamAttrs(fabsPal); - } - CallInst *call = m_builder.CreateCall(m_llvmFAbs, val, - name("fabs")); - call->setCallingConv(CallingConv::C); + llvm::Function *func = m_mod->getFunction("cmp"); + assert(func); + + std::vector params; + params.push_back(in1); + params.push_back(in2); + params.push_back(in3); + CallInst *call = m_builder.CreateCall(func, params.begin(), params.end(), name("cmpres")); call->setTailCall(false); return call; } -llvm::Value * Instructions::lit(llvm::Value *in) +llvm::Value * Instructions::cnd(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3) { - if (!m_llvmLit) { - m_llvmLit = m_mod->getFunction("lit"); - } - CallInst *call = m_builder.CreateCall(m_llvmLit, in, name("litres")); - call->setCallingConv(CallingConv::C); - call->setTailCall(false); - return call; + std::vector vec1 = extractVector(in1); + std::vector vec2 = extractVector(in2); + std::vector vec3 = extractVector(in3); + Constant *half = ConstantFP::get(APFloat(0.5f)); + + Value *xcmp = m_builder.CreateFCmpOGT(vec1[0], half, name("xcmp")); + Value *selx = m_builder.CreateSelect(xcmp, vec2[0], vec3[0], + name("selx")); + + Value *ycmp = m_builder.CreateFCmpOGT(vec1[1], half, name("ycmp")); + Value *sely = m_builder.CreateSelect(ycmp, vec2[1], vec3[1], + name("sely")); + + Value *zcmp = m_builder.CreateFCmpOGT(vec1[2], half, name("zcmp")); + Value *selz = m_builder.CreateSelect(zcmp, vec2[2], vec3[2], + name("selz")); + + Value *wcmp = m_builder.CreateFCmpOGT(vec1[3], half, name("wcmp")); + Value *selw = m_builder.CreateSelect(wcmp, vec2[3], vec3[3], + name("selw")); + + return vectorFromVals(selx, sely, selz, selw); } -llvm::Value * Instructions::sub(llvm::Value *in1, llvm::Value *in2) +llvm::Value * Instructions::cnd0(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3) { - Value *res = m_builder.CreateSub(in1, in2, name("sub")); - return res; + std::vector vec1 = extractVector(in1); + std::vector vec2 = extractVector(in2); + std::vector vec3 = extractVector(in3); + Constant *zero = Constant::getNullValue(Type::FloatTy); + + Value *xcmp = m_builder.CreateFCmpOGE(vec1[0], zero, name("xcmp")); + Value *selx = m_builder.CreateSelect(xcmp, vec2[0], vec3[0], + name("selx")); + + Value *ycmp = m_builder.CreateFCmpOGE(vec1[1], zero, name("ycmp")); + Value *sely = m_builder.CreateSelect(ycmp, vec2[1], vec3[1], + name("sely")); + + Value *zcmp = m_builder.CreateFCmpOGE(vec1[2], zero, name("zcmp")); + Value *selz = m_builder.CreateSelect(zcmp, vec2[2], vec3[2], + name("selz")); + + Value *wcmp = m_builder.CreateFCmpOGE(vec1[3], zero, name("wcmp")); + Value *selw = m_builder.CreateSelect(wcmp, vec2[3], vec3[3], + name("selw")); + + return vectorFromVals(selx, sely, selz, selw); } -llvm::Value * Instructions::callPow(llvm::Value *val1, llvm::Value *val2) +llvm::Value * Instructions::cos(llvm::Value *in) { - if (!m_llvmPow) { - // predeclare the intrinsic - std::vector powArgs; - powArgs.push_back(Type::FloatTy); - powArgs.push_back(Type::FloatTy); - PAListPtr powPal; - FunctionType* powType = FunctionType::get( - /*Result=*/Type::FloatTy, - /*Params=*/powArgs, - /*isVarArg=*/false); - m_llvmPow = Function::Create( - /*Type=*/powType, - /*Linkage=*/GlobalValue::ExternalLinkage, - /*Name=*/"llvm.pow.f32", m_mod); - m_llvmPow->setCallingConv(CallingConv::C); - m_llvmPow->setParamAttrs(powPal); - } - std::vector params; - params.push_back(val1); - params.push_back(val2); - CallInst *call = m_builder.CreateCall(m_llvmPow, params.begin(), params.end(), - name("pow")); - call->setCallingConv(CallingConv::C); +#if 0 + llvm::Function *func = m_mod->getFunction("vcos"); + assert(func); + + CallInst *call = m_builder.CreateCall(func, in, name("cosres")); call->setTailCall(false); return call; +#else + std::vector elems = extractVector(in); + Function *func = m_mod->getFunction("cosf"); + assert(func); + CallInst *cos = m_builder.CreateCall(func, elems[0], name("cosres")); + cos->setCallingConv(CallingConv::C); + cos->setTailCall(true); + return vectorFromVals(cos, cos, cos, cos); +#endif } -llvm::Value * Instructions::pow(llvm::Value *in1, llvm::Value *in2) +llvm::Value * Instructions::cross(llvm::Value *in1, llvm::Value *in2) { Value *x1 = m_builder.CreateExtractElement(in1, m_storage->constantInt(0), name("x1")); + Value *y1 = m_builder.CreateExtractElement(in1, + m_storage->constantInt(1), + name("y1")); + Value *z1 = m_builder.CreateExtractElement(in1, + m_storage->constantInt(2), + name("z1")); + Value *x2 = m_builder.CreateExtractElement(in2, m_storage->constantInt(0), name("x2")); - llvm::Value *val = callPow(x1, x2); - return vectorFromVals(val, val, val, val); + Value *y2 = m_builder.CreateExtractElement(in2, + m_storage->constantInt(1), + name("y2")); + Value *z2 = m_builder.CreateExtractElement(in2, + m_storage->constantInt(2), + name("z2")); + Value *y1z2 = mul(y1, z2); + Value *z1y2 = mul(z1, y2); + + Value *z1x2 = mul(z1, x2); + Value *x1z2 = mul(x1, z2); + + Value *x1y2 = mul(x1, y2); + Value *y1x2 = mul(y1, x2); + + return vectorFromVals(sub(y1z2, z1y2), sub(z1x2, x1z2), sub(x1y2, y1x2)); } -llvm::Value * Instructions::rcp(llvm::Value *in1) +llvm::Value * Instructions::ddx(llvm::Value *in) { - Value *x1 = m_builder.CreateExtractElement(in1, - m_storage->constantInt(0), - name("x1")); - Value *res = m_builder.CreateFDiv(ConstantFP::get(APFloat(1.f)), - x1, name("rcp")); - return vectorFromVals(res, res, res, res); + // FIXME +} + +llvm::Value * Instructions::ddy(llvm::Value *in) +{ + // FIXME +} + +llvm::Value * Instructions::div(llvm::Value *in1, llvm::Value *in2) +{ + return m_builder.CreateFDiv(in1, in2, name("div")); +} + +llvm::Value * Instructions::dot2add(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3) +{ + Value *mulRes = mul(in1, in2); + Value *x = m_builder.CreateExtractElement(mulRes, + m_storage->constantInt(0), + name("extractx")); + Value *y = m_builder.CreateExtractElement(mulRes, + m_storage->constantInt(1), + name("extracty")); + Value *z = m_builder.CreateExtractElement(in3, + m_storage->constantInt(2), + name("extractz")); + Value *xy = m_builder.CreateAdd(x, y,name("xy")); + Value *dot2add = m_builder.CreateAdd(xy, z, name("dot2add")); + return vectorFromVals(dot2add, dot2add, dot2add, dot2add); +} + +llvm::Value * Instructions::dp3(llvm::Value *in1, llvm::Value *in2) +{ + Value *mulRes = mul(in1, in2); + Value *x = m_builder.CreateExtractElement(mulRes, + m_storage->constantInt(0), + name("extractx")); + Value *y = m_builder.CreateExtractElement(mulRes, + m_storage->constantInt(1), + name("extracty")); + Value *z = m_builder.CreateExtractElement(mulRes, + m_storage->constantInt(2), + name("extractz")); + Value *xy = m_builder.CreateAdd(x, y,name("xy")); + Value *dot3 = m_builder.CreateAdd(xy, z, name("dot3")); + return vectorFromVals(dot3, dot3, dot3, dot3); } llvm::Value * Instructions::dp4(llvm::Value *in1, llvm::Value *in2) @@ -302,23 +356,70 @@ llvm::Value * Instructions::dph(llvm::Value *in1, llvm::Value *in2) return vectorFromVals(dph, dph, dph, dph); } -llvm::Value * Instructions::dst(llvm::Value *in1, llvm::Value *in2) +llvm::Value * Instructions::dst(llvm::Value *in1, llvm::Value *in2) +{ + Value *y1 = m_builder.CreateExtractElement(in1, + m_storage->constantInt(1), + name("y1")); + Value *z = m_builder.CreateExtractElement(in1, + m_storage->constantInt(2), + name("z")); + Value *y2 = m_builder.CreateExtractElement(in2, + m_storage->constantInt(1), + name("y2")); + Value *w = m_builder.CreateExtractElement(in2, + m_storage->constantInt(3), + name("w")); + Value *ry = m_builder.CreateMul(y1, y2, name("tyuy")); + return vectorFromVals(ConstantFP::get(APFloat(1.f)), + ry, z, w); +} + +void Instructions::elseop() +{ + assert(!m_ifStack.empty()); + BasicBlock *ifend = BasicBlock::Create(name("ifend"), m_func,0); + m_builder.CreateBr(ifend); + m_builder.SetInsertPoint(m_ifStack.top()); + currentBlock()->setName(name("ifelse")); + m_ifStack.pop(); + m_ifStack.push(ifend); +} + +void Instructions::endif() +{ + assert(!m_ifStack.empty()); + m_builder.CreateBr(m_ifStack.top()); + m_builder.SetInsertPoint(m_ifStack.top()); + m_ifStack.pop(); +} + +void Instructions::endLoop() +{ + assert(!m_loopStack.empty()); + Loop loop = m_loopStack.top(); + m_builder.CreateBr(loop.begin); + loop.end->moveAfter(currentBlock()); + m_builder.SetInsertPoint(loop.end); + m_loopStack.pop(); +} + +void Instructions::end() +{ + m_builder.CreateRetVoid(); +} + +void Instructions::endSub() +{ + m_func = 0; + m_builder.SetInsertPoint(0); +} + +llvm::Value * Instructions::exp(llvm::Value *in) { - Value *y1 = m_builder.CreateExtractElement(in1, - m_storage->constantInt(1), - name("y1")); - Value *z = m_builder.CreateExtractElement(in1, - m_storage->constantInt(2), - name("z")); - Value *y2 = m_builder.CreateExtractElement(in2, - m_storage->constantInt(1), - name("y2")); - Value *w = m_builder.CreateExtractElement(in2, - m_storage->constantInt(3), - name("w")); - Value *ry = m_builder.CreateMul(y1, y2, name("tyuy")); - return vectorFromVals(ConstantFP::get(APFloat(1.f)), - ry, z, w); + std::vector vec = extractVector(in); + return vectorFromVals(callFExp(vec[0]), callFExp(vec[1]), + callFExp(vec[2]), callFExp(vec[3])); } llvm::Value * Instructions::ex2(llvm::Value *in) @@ -330,31 +431,6 @@ llvm::Value * Instructions::ex2(llvm::Value *in) return vectorFromVals(val, val, val, val); } -llvm::Value * Instructions::callFloor(llvm::Value *val) -{ - if (!m_llvmFloor) { - // predeclare the intrinsic - std::vector floorArgs; - floorArgs.push_back(Type::FloatTy); - PAListPtr floorPal; - FunctionType* floorType = FunctionType::get( - /*Result=*/Type::FloatTy, - /*Params=*/floorArgs, - /*isVarArg=*/false); - m_llvmFloor = Function::Create( - /*Type=*/floorType, - /*Linkage=*/GlobalValue::ExternalLinkage, - /*Name=*/"floorf", m_mod); - m_llvmFloor->setCallingConv(CallingConv::C); - m_llvmFloor->setParamAttrs(floorPal); - } - CallInst *call = m_builder.CreateCall(m_llvmFloor, val, - name("floorf")); - call->setCallingConv(CallingConv::C); - call->setTailCall(false); - return call; -} - llvm::Value * Instructions::floor(llvm::Value *in) { std::vector vec = extractVector(in); @@ -362,42 +438,52 @@ llvm::Value * Instructions::floor(llvm::Value *in) callFloor(vec[2]), callFloor(vec[3])); } -llvm::Value * Instructions::arl(llvm::Value *in) -{ - return floor(in); -} - llvm::Value * Instructions::frc(llvm::Value *in) { llvm::Value *flr = floor(in); return sub(in, flr); } -llvm::Value * Instructions::callFLog(llvm::Value *val) +void Instructions::ifop(llvm::Value *in) { - if (!m_llvmFlog) { - // predeclare the intrinsic - std::vector flogArgs; - flogArgs.push_back(Type::FloatTy); - PAListPtr flogPal; - FunctionType* flogType = FunctionType::get( - /*Result=*/Type::FloatTy, - /*Params=*/flogArgs, - /*isVarArg=*/false); - m_llvmFlog = Function::Create( - /*Type=*/flogType, - /*Linkage=*/GlobalValue::ExternalLinkage, - /*Name=*/"logf", m_mod); - m_llvmFlog->setCallingConv(CallingConv::C); - m_llvmFlog->setParamAttrs(flogPal); - } - CallInst *call = m_builder.CreateCall(m_llvmFlog, val, - name("logf")); - call->setCallingConv(CallingConv::C); + BasicBlock *ifthen = BasicBlock::Create(name("ifthen"), m_func,0); + BasicBlock *ifend = BasicBlock::Create(name("ifthenend"), m_func,0); + + //BasicBlock *yblock = new BasicBlock(name("yblock"), m_func,0); + //BasicBlock *zblock = new BasicBlock(name("zblock"), m_func,0); + //BasicBlock *wblock = new BasicBlock(name("wblock"), m_func,0); + + Constant *float0 = Constant::getNullValue(Type::FloatTy); + + Value *x = m_builder.CreateExtractElement(in, m_storage->constantInt(0), + name("extractx")); + Value *xcmp = m_builder.CreateFCmpUNE(x, float0, name("xcmp")); + m_builder.CreateCondBr(xcmp, ifthen, ifend); + //m_builder.SetInsertPoint(yblock); + + m_builder.SetInsertPoint(ifthen); + m_ifStack.push(ifend); +} + +llvm::Value * Instructions::kil(llvm::Value *in) +{ + llvm::Function *func = m_mod->getFunction("kil"); + assert(func); + + CallInst *call = m_builder.CreateCall(func, in, name("kilpres")); call->setTailCall(false); return call; } +llvm::Value * Instructions::lerp(llvm::Value *in1, llvm::Value *in2, + llvm::Value *in3) +{ + llvm::Value *m = mul(in1, in2); + llvm::Value *vec1 = constVector(1.f, 1.f, 1.f, 1.f); + llvm::Value *s = sub(vec1, in1); + return add(m, mul(s, in3)); +} + llvm::Value * Instructions::lg2(llvm::Value *in) { std::vector vec = extractVector(in); @@ -407,120 +493,192 @@ llvm::Value * Instructions::lg2(llvm::Value *in) callFLog(vec[2]), callFLog(vec[3])), const_vec); } -llvm::Value * Instructions::min(llvm::Value *in1, llvm::Value *in2) +llvm::Value * Instructions::lit(llvm::Value *in) +{ + if (!m_llvmLit) { + m_llvmLit = m_mod->getFunction("lit"); + } + CallInst *call = m_builder.CreateCall(m_llvmLit, in, name("litres")); + call->setCallingConv(CallingConv::C); + call->setTailCall(false); + return call; +} + +llvm::Value * Instructions::log(llvm::Value *in) +{ + std::vector vec = extractVector(in); + return vectorFromVals(callFLog(vec[0]), callFLog(vec[1]), + callFLog(vec[2]), callFLog(vec[3])); +} + +llvm::Value * Instructions::madd(llvm::Value *in1, llvm::Value *in2, + llvm::Value *in3) +{ + Value *mulRes = mul(in1, in2); + return add(mulRes, in3); +} + +llvm::Value * Instructions::max(llvm::Value *in1, llvm::Value *in2) { std::vector vec1 = extractVector(in1); std::vector vec2 = extractVector(in2); - Value *xcmp = m_builder.CreateFCmpOLT(vec1[0], vec2[0], name("xcmp")); + Value *xcmp = m_builder.CreateFCmpOGT(vec1[0], vec2[0], + name("xcmp")); Value *selx = m_builder.CreateSelect(xcmp, vec1[0], vec2[0], name("selx")); - Value *ycmp = m_builder.CreateFCmpOLT(vec1[1], vec2[1], name("ycmp")); + Value *ycmp = m_builder.CreateFCmpOGT(vec1[1], vec2[1], + name("ycmp")); Value *sely = m_builder.CreateSelect(ycmp, vec1[1], vec2[1], name("sely")); - Value *zcmp = m_builder.CreateFCmpOLT(vec1[2], vec2[2], name("zcmp")); + Value *zcmp = m_builder.CreateFCmpOGT(vec1[2], vec2[2], + name("zcmp")); Value *selz = m_builder.CreateSelect(zcmp, vec1[2], vec2[2], name("selz")); - Value *wcmp = m_builder.CreateFCmpOLT(vec1[3], vec2[3], name("wcmp")); + Value *wcmp = m_builder.CreateFCmpOGT(vec1[3], vec2[3], + name("wcmp")); Value *selw = m_builder.CreateSelect(wcmp, vec1[3], vec2[3], name("selw")); return vectorFromVals(selx, sely, selz, selw); } -llvm::Value * Instructions::max(llvm::Value *in1, llvm::Value *in2) +llvm::Value * Instructions::min(llvm::Value *in1, llvm::Value *in2) { std::vector vec1 = extractVector(in1); std::vector vec2 = extractVector(in2); - Value *xcmp = m_builder.CreateFCmpOGT(vec1[0], vec2[0], - name("xcmp")); + Value *xcmp = m_builder.CreateFCmpOLT(vec1[0], vec2[0], name("xcmp")); Value *selx = m_builder.CreateSelect(xcmp, vec1[0], vec2[0], name("selx")); - Value *ycmp = m_builder.CreateFCmpOGT(vec1[1], vec2[1], - name("ycmp")); + Value *ycmp = m_builder.CreateFCmpOLT(vec1[1], vec2[1], name("ycmp")); Value *sely = m_builder.CreateSelect(ycmp, vec1[1], vec2[1], name("sely")); - Value *zcmp = m_builder.CreateFCmpOGT(vec1[2], vec2[2], - name("zcmp")); + Value *zcmp = m_builder.CreateFCmpOLT(vec1[2], vec2[2], name("zcmp")); Value *selz = m_builder.CreateSelect(zcmp, vec1[2], vec2[2], name("selz")); - Value *wcmp = m_builder.CreateFCmpOGT(vec1[3], vec2[3], - name("wcmp")); + Value *wcmp = m_builder.CreateFCmpOLT(vec1[3], vec2[3], name("wcmp")); Value *selw = m_builder.CreateSelect(wcmp, vec1[3], vec2[3], name("selw")); return vectorFromVals(selx, sely, selz, selw); } -void Instructions::printVector(llvm::Value *val) +llvm::Value * Instructions::mul(llvm::Value *in1, llvm::Value *in2) { - static const char *frmt = "Vector is [%f, %f, %f, %f]\x0A"; + return m_builder.CreateMul(in1, in2, name("mul")); +} - if (!m_fmtPtr) { - Constant *format = ConstantArray::get(frmt, true); - ArrayType *arrayTy = ArrayType::get(IntegerType::get(8), strlen(frmt) + 1); - GlobalVariable* globalFormat = new GlobalVariable( - /*Type=*/arrayTy, - /*isConstant=*/true, - /*Linkage=*/GlobalValue::InternalLinkage, - /*Initializer=*/0, // has initializer, specified below - /*Name=*/name(".str"), - m_mod); - globalFormat->setInitializer(format); +llvm::Value * Instructions::neg(llvm::Value *in) +{ + Value *neg = m_builder.CreateNeg(in, name("neg")); + return neg; +} - Constant* const_int0 = Constant::getNullValue(IntegerType::get(32)); - std::vector const_ptr_21_indices; - const_ptr_21_indices.push_back(const_int0); - const_ptr_21_indices.push_back(const_int0); - m_fmtPtr = ConstantExpr::getGetElementPtr(globalFormat, - &const_ptr_21_indices[0], const_ptr_21_indices.size()); - } +llvm::Value * Instructions::pow(llvm::Value *in1, llvm::Value *in2) +{ + Value *x1 = m_builder.CreateExtractElement(in1, + m_storage->constantInt(0), + name("x1")); + Value *x2 = m_builder.CreateExtractElement(in2, + m_storage->constantInt(0), + name("x2")); + llvm::Value *val = callPow(x1, x2); + return vectorFromVals(val, val, val, val); +} - Function *func_printf = m_mod->getFunction("printf"); - if (!func_printf) - func_printf = declarePrintf(); - assert(func_printf); - std::vector vec = extractVector(val); - Value *dx = m_builder.CreateFPExt(vec[0], Type::DoubleTy, name("dx")); - Value *dy = m_builder.CreateFPExt(vec[1], Type::DoubleTy, name("dy")); - Value *dz = m_builder.CreateFPExt(vec[2], Type::DoubleTy, name("dz")); - Value *dw = m_builder.CreateFPExt(vec[3], Type::DoubleTy, name("dw")); - std::vector params; - params.push_back(m_fmtPtr); - params.push_back(dx); - params.push_back(dy); - params.push_back(dz); - params.push_back(dw); - CallInst *call = m_builder.CreateCall(func_printf, params.begin(), params.end(), - name("printf")); - call->setCallingConv(CallingConv::C); - call->setTailCall(true); +llvm::Value * Instructions::rcp(llvm::Value *in1) +{ + Value *x1 = m_builder.CreateExtractElement(in1, + m_storage->constantInt(0), + name("x1")); + Value *res = m_builder.CreateFDiv(ConstantFP::get(APFloat(1.f)), + x1, name("rcp")); + return vectorFromVals(res, res, res, res); +} + +llvm::Value * Instructions::rsq(llvm::Value *in1) +{ + Value *x = m_builder.CreateExtractElement(in1, + m_storage->constantInt(0), + name("extractx")); + Value *abs = callFAbs(x); + Value *sqrt = callFSqrt(abs); + + Value *rsqrt = m_builder.CreateFDiv(ConstantFP::get(APFloat(1.f)), + sqrt, + name("rsqrt")); + return vectorFromVals(rsqrt, rsqrt, rsqrt, rsqrt); +} + +llvm::Value * Instructions::scs(llvm::Value *in) +{ + llvm::Function *func = m_mod->getFunction("scs"); + assert(func); + + CallInst *call = m_builder.CreateCall(func, in, name("scsres")); + call->setTailCall(false); + return call; +} + +llvm::Value * Instructions::seq(llvm::Value *in1, llvm::Value *in2) +{ + Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f)); + Constant *const0f = Constant::getNullValue(Type::FloatTy); + + std::vector vec1 = extractVector(in1); + std::vector vec2 = extractVector(in2); + + Value *xcmp = m_builder.CreateFCmpOEQ(vec1[0], vec2[0], name("xcmp")); + Value *x = m_builder.CreateSelect(xcmp, const1f, const0f, name("xsel")); + + Value *ycmp = m_builder.CreateFCmpOEQ(vec1[1], vec2[1], name("ycmp")); + Value *y = m_builder.CreateSelect(ycmp, const1f, const0f, name("ysel")); + + Value *zcmp = m_builder.CreateFCmpOEQ(vec1[2], vec2[2], name("zcmp")); + Value *z = m_builder.CreateSelect(zcmp, const1f, const0f, name("zsel")); + + Value *wcmp = m_builder.CreateFCmpOEQ(vec1[3], vec2[3], name("wcmp")); + Value *w = m_builder.CreateSelect(wcmp, const1f, const0f, name("wsel")); + + return vectorFromVals(x, y, z, w); } -llvm::Function * Instructions::declarePrintf() +llvm::Value * Instructions::sfl(llvm::Value *in1, llvm::Value *in2) { - std::vector args; - PAListPtr params; - FunctionType* funcTy = FunctionType::get( - /*Result=*/IntegerType::get(32), - /*Params=*/args, - /*isVarArg=*/true); - Function* func_printf = Function::Create( - /*Type=*/funcTy, - /*Linkage=*/GlobalValue::ExternalLinkage, - /*Name=*/"printf", m_mod); - func_printf->setCallingConv(CallingConv::C); - func_printf->setParamAttrs(params); - return func_printf; + Constant *const0f = Constant::getNullValue(Type::FloatTy); + + return vectorFromVals(const0f, const0f, const0f, const0f); } +llvm::Value * Instructions::sge(llvm::Value *in1, llvm::Value *in2) +{ + Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f)); + Constant *const0f = Constant::getNullValue(Type::FloatTy); + + std::vector vec1 = extractVector(in1); + std::vector vec2 = extractVector(in2); + + Value *xcmp = m_builder.CreateFCmpOGE(vec1[0], vec2[0], name("xcmp")); + Value *x = m_builder.CreateSelect(xcmp, const1f, const0f, name("xsel")); + + Value *ycmp = m_builder.CreateFCmpOGE(vec1[1], vec2[1], name("ycmp")); + Value *y = m_builder.CreateSelect(ycmp, const1f, const0f, name("ysel")); + + Value *zcmp = m_builder.CreateFCmpOGE(vec1[2], vec2[2], name("zcmp")); + Value *z = m_builder.CreateSelect(zcmp, const1f, const0f, name("zsel")); + + Value *wcmp = m_builder.CreateFCmpOGE(vec1[3], vec2[3], name("wcmp")); + Value *w = m_builder.CreateSelect(wcmp, const1f, const0f, name("wsel")); + + return vectorFromVals(x, y, z, w); +} llvm::Value * Instructions::sgt(llvm::Value *in1, llvm::Value *in2) { @@ -543,7 +701,18 @@ llvm::Value * Instructions::sgt(llvm::Value *in1, llvm::Value *in2) return vectorFromVals(x, y, z, w); } -llvm::Value * Instructions::sge(llvm::Value *in1, llvm::Value *in2) + +llvm::Value * Instructions::sin(llvm::Value *in) +{ + llvm::Function *func = m_mod->getFunction("vsin"); + assert(func); + + CallInst *call = m_builder.CreateCall(func, in, name("sinres")); + call->setTailCall(false); + return call; +} + +llvm::Value * Instructions::sle(llvm::Value *in1, llvm::Value *in2) { Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f)); Constant *const0f = Constant::getNullValue(Type::FloatTy); @@ -551,22 +720,21 @@ llvm::Value * Instructions::sge(llvm::Value *in1, llvm::Value *in2) std::vector vec1 = extractVector(in1); std::vector vec2 = extractVector(in2); - Value *xcmp = m_builder.CreateFCmpOGE(vec1[0], vec2[0], name("xcmp")); + Value *xcmp = m_builder.CreateFCmpOLE(vec1[0], vec2[0], name("xcmp")); Value *x = m_builder.CreateSelect(xcmp, const1f, const0f, name("xsel")); - Value *ycmp = m_builder.CreateFCmpOGE(vec1[1], vec2[1], name("ycmp")); + Value *ycmp = m_builder.CreateFCmpOLE(vec1[1], vec2[1], name("ycmp")); Value *y = m_builder.CreateSelect(ycmp, const1f, const0f, name("ysel")); - Value *zcmp = m_builder.CreateFCmpOGE(vec1[2], vec2[2], name("zcmp")); + Value *zcmp = m_builder.CreateFCmpOLE(vec1[2], vec2[2], name("zcmp")); Value *z = m_builder.CreateSelect(zcmp, const1f, const0f, name("zsel")); - Value *wcmp = m_builder.CreateFCmpOGE(vec1[3], vec2[3], name("wcmp")); + Value *wcmp = m_builder.CreateFCmpOLE(vec1[3], vec2[3], name("wcmp")); Value *w = m_builder.CreateSelect(wcmp, const1f, const0f, name("wsel")); return vectorFromVals(x, y, z, w); } - llvm::Value * Instructions::slt(llvm::Value *in1, llvm::Value *in2) { Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f)); @@ -590,169 +758,331 @@ llvm::Value * Instructions::slt(llvm::Value *in1, llvm::Value *in2) return vectorFromVals(x, y, z, w); } -llvm::Value * Instructions::cross(llvm::Value *in1, llvm::Value *in2) +llvm::Value * Instructions::sne(llvm::Value *in1, llvm::Value *in2) { - Value *x1 = m_builder.CreateExtractElement(in1, - m_storage->constantInt(0), - name("x1")); - Value *y1 = m_builder.CreateExtractElement(in1, - m_storage->constantInt(1), - name("y1")); - Value *z1 = m_builder.CreateExtractElement(in1, - m_storage->constantInt(2), - name("z1")); + Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f)); + Constant *const0f = Constant::getNullValue(Type::FloatTy); - Value *x2 = m_builder.CreateExtractElement(in2, - m_storage->constantInt(0), - name("x2")); - Value *y2 = m_builder.CreateExtractElement(in2, - m_storage->constantInt(1), - name("y2")); - Value *z2 = m_builder.CreateExtractElement(in2, - m_storage->constantInt(2), - name("z2")); - Value *y1z2 = mul(y1, z2); - Value *z1y2 = mul(z1, y2); + std::vector vec1 = extractVector(in1); + std::vector vec2 = extractVector(in2); - Value *z1x2 = mul(z1, x2); - Value *x1z2 = mul(x1, z2); + Value *xcmp = m_builder.CreateFCmpONE(vec1[0], vec2[0], name("xcmp")); + Value *x = m_builder.CreateSelect(xcmp, const1f, const0f, name("xsel")); - Value *x1y2 = mul(x1, y2); - Value *y1x2 = mul(y1, x2); + Value *ycmp = m_builder.CreateFCmpONE(vec1[1], vec2[1], name("ycmp")); + Value *y = m_builder.CreateSelect(ycmp, const1f, const0f, name("ysel")); - return vectorFromVals(sub(y1z2, z1y2), sub(z1x2, x1z2), sub(x1y2, y1x2)); + Value *zcmp = m_builder.CreateFCmpONE(vec1[2], vec2[2], name("zcmp")); + Value *z = m_builder.CreateSelect(zcmp, const1f, const0f, name("zsel")); + + Value *wcmp = m_builder.CreateFCmpONE(vec1[3], vec2[3], name("wcmp")); + Value *w = m_builder.CreateSelect(wcmp, const1f, const0f, name("wsel")); + + return vectorFromVals(x, y, z, w); } +llvm::Value * Instructions::str(llvm::Value *in1, llvm::Value *in2) +{ + Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f)); -llvm::Value * Instructions::abs(llvm::Value *in) + return vectorFromVals(const1f, const1f, const1f, const1f); +} + +llvm::Value * Instructions::sub(llvm::Value *in1, llvm::Value *in2) +{ + Value *res = m_builder.CreateSub(in1, in2, name("sub")); + return res; +} + +llvm::Value * Instructions::trunc(llvm::Value *in) { std::vector vec = extractVector(in); - Value *xabs = callFAbs(vec[0]); - Value *yabs = callFAbs(vec[1]); - Value *zabs = callFAbs(vec[2]); - Value *wabs = callFAbs(vec[3]); - return vectorFromVals(xabs, yabs, zabs, wabs); + Value *icastx = m_builder.CreateFPToSI(vec[0], IntegerType::get(32), + name("ftoix")); + Value *icasty = m_builder.CreateFPToSI(vec[1], IntegerType::get(32), + name("ftoiy")); + Value *icastz = m_builder.CreateFPToSI(vec[2], IntegerType::get(32), + name("ftoiz")); + Value *icastw = m_builder.CreateFPToSI(vec[3], IntegerType::get(32), + name("ftoiw")); + Value *fx = m_builder.CreateSIToFP(icastx, Type::FloatTy, + name("fx")); + Value *fy = m_builder.CreateSIToFP(icasty, Type::FloatTy, + name("fy")); + Value *fz = m_builder.CreateSIToFP(icastz, Type::FloatTy, + name("fz")); + Value *fw = m_builder.CreateSIToFP(icastw, Type::FloatTy, + name("fw")); + return vectorFromVals(fx, fy, fz, fw); } -void Instructions::ifop(llvm::Value *in) +llvm::Value * Instructions::x2d(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3) { - BasicBlock *ifthen = BasicBlock::Create(name("ifthen"), m_func,0); - BasicBlock *ifend = BasicBlock::Create(name("ifthenend"), m_func,0); + std::vector vec1 = extractVector(in1); + std::vector vec2 = extractVector(in2); + std::vector vec3 = extractVector(in3); - //BasicBlock *yblock = new BasicBlock(name("yblock"), m_func,0); - //BasicBlock *zblock = new BasicBlock(name("zblock"), m_func,0); - //BasicBlock *wblock = new BasicBlock(name("wblock"), m_func,0); + Value *x2x3 = m_builder.CreateMul( vec2[0], vec3[0], name("x2x3")); + Value *y2y3 = m_builder.CreateMul( vec2[1], vec3[1], name("y2y3")); + Value *x1px2x3 = m_builder.CreateAdd (vec1[0], x2x3, name("x1 + x2x3")); + Value *x1px2x3py2y3 = m_builder.CreateAdd (x1px2x3, y2y3, name("x1 + x2x3 + y2y3")); - Constant *float0 = Constant::getNullValue(Type::FloatTy); + Value *x2z3 = m_builder.CreateMul( vec2[0], vec3[2], name("x2z3")); + Value *y2w3 = m_builder.CreateMul( vec2[1], vec3[3], name("y2w3")); + Value *y1px2z3 = m_builder.CreateAdd (vec1[1], x2z3, name("y1 + x2z3")); + Value *y1px2z3py2w3 = m_builder.CreateAdd (y1px2z3, y2w3, name("y1 + x2z3 + y2w3")); - Value *x = m_builder.CreateExtractElement(in, m_storage->constantInt(0), - name("extractx")); - Value *xcmp = m_builder.CreateFCmpUNE(x, float0, name("xcmp")); - m_builder.CreateCondBr(xcmp, ifthen, ifend); - //m_builder.SetInsertPoint(yblock); + return vectorFromVals(x1px2x3py2y3, y1px2z3py2w3, x1px2x3py2y3, y1px2z3py2w3); +} - m_builder.SetInsertPoint(ifthen); - m_ifStack.push(ifend); +void Instructions::printVector(llvm::Value *val) +{ + static const char *frmt = "Vector is [%f, %f, %f, %f]\x0A"; + + if (!m_fmtPtr) { + Constant *format = ConstantArray::get(frmt, true); + ArrayType *arrayTy = ArrayType::get(IntegerType::get(8), strlen(frmt) + 1); + GlobalVariable* globalFormat = new GlobalVariable( + /*Type=*/arrayTy, + /*isConstant=*/true, + /*Linkage=*/GlobalValue::InternalLinkage, + /*Initializer=*/0, // has initializer, specified below + /*Name=*/name(".str"), + m_mod); + globalFormat->setInitializer(format); + + Constant* const_int0 = Constant::getNullValue(IntegerType::get(32)); + std::vector const_ptr_21_indices; + const_ptr_21_indices.push_back(const_int0); + const_ptr_21_indices.push_back(const_int0); + m_fmtPtr = ConstantExpr::getGetElementPtr(globalFormat, + &const_ptr_21_indices[0], const_ptr_21_indices.size()); + } + + Function *func_printf = m_mod->getFunction("printf"); + if (!func_printf) + func_printf = declarePrintf(); + assert(func_printf); + std::vector vec = extractVector(val); + Value *dx = m_builder.CreateFPExt(vec[0], Type::DoubleTy, name("dx")); + Value *dy = m_builder.CreateFPExt(vec[1], Type::DoubleTy, name("dy")); + Value *dz = m_builder.CreateFPExt(vec[2], Type::DoubleTy, name("dz")); + Value *dw = m_builder.CreateFPExt(vec[3], Type::DoubleTy, name("dw")); + std::vector params; + params.push_back(m_fmtPtr); + params.push_back(dx); + params.push_back(dy); + params.push_back(dz); + params.push_back(dw); + CallInst *call = m_builder.CreateCall(func_printf, params.begin(), params.end(), + name("printf")); + call->setCallingConv(CallingConv::C); + call->setTailCall(true); } -llvm::BasicBlock * Instructions::currentBlock() const +const char * Instructions::name(const char *prefix) { - return m_builder.GetInsertBlock(); + ++m_idx; + snprintf(m_name, 32, "%s%d", prefix, m_idx); + return m_name; } -void Instructions::elseop() +llvm::Value *Instructions::callFAbs(llvm::Value *val) { - assert(!m_ifStack.empty()); - BasicBlock *ifend = BasicBlock::Create(name("ifend"), m_func,0); - m_builder.CreateBr(ifend); - m_builder.SetInsertPoint(m_ifStack.top()); - currentBlock()->setName(name("ifelse")); - m_ifStack.pop(); - m_ifStack.push(ifend); + if (!m_llvmFAbs) { + // predeclare the intrinsic + std::vector fabsArgs; + fabsArgs.push_back(Type::FloatTy); + PAListPtr fabsPal; + FunctionType* fabsType = FunctionType::get( + /*Result=*/Type::FloatTy, + /*Params=*/fabsArgs, + /*isVarArg=*/false); + m_llvmFAbs = Function::Create( + /*Type=*/fabsType, + /*Linkage=*/GlobalValue::ExternalLinkage, + /*Name=*/"fabs", m_mod); + m_llvmFAbs->setCallingConv(CallingConv::C); + m_llvmFAbs->setParamAttrs(fabsPal); + } + CallInst *call = m_builder.CreateCall(m_llvmFAbs, val, + name("fabs")); + call->setCallingConv(CallingConv::C); + call->setTailCall(false); + return call; } -void Instructions::endif() +llvm::Value * Instructions::callFExp(llvm::Value *val) { - assert(!m_ifStack.empty()); - m_builder.CreateBr(m_ifStack.top()); - m_builder.SetInsertPoint(m_ifStack.top()); - m_ifStack.pop(); + if (!m_llvmFexp) { + // predeclare the intrinsic + std::vector fexpArgs; + fexpArgs.push_back(Type::FloatTy); + PAListPtr fexpPal; + FunctionType* fexpType = FunctionType::get( + /*Result=*/Type::FloatTy, + /*Params=*/fexpArgs, + /*isVarArg=*/false); + m_llvmFexp = Function::Create( + /*Type=*/fexpType, + /*Linkage=*/GlobalValue::ExternalLinkage, + /*Name=*/"expf", m_mod); + m_llvmFexp->setCallingConv(CallingConv::C); + m_llvmFexp->setParamAttrs(fexpPal); + } + CallInst *call = m_builder.CreateCall(m_llvmFexp, val, + name("expf")); + call->setCallingConv(CallingConv::C); + call->setTailCall(false); + return call; } -llvm::Value * Instructions::lerp(llvm::Value *in1, llvm::Value *in2, - llvm::Value *in3) +llvm::Value * Instructions::callFLog(llvm::Value *val) { - llvm::Value *m = mul(in1, in2); - llvm::Value *vec1 = constVector(1.f, 1.f, 1.f, 1.f); - llvm::Value *s = sub(vec1, in1); - return add(m, mul(s, in3)); + if (!m_llvmFlog) { + // predeclare the intrinsic + std::vector flogArgs; + flogArgs.push_back(Type::FloatTy); + PAListPtr flogPal; + FunctionType* flogType = FunctionType::get( + /*Result=*/Type::FloatTy, + /*Params=*/flogArgs, + /*isVarArg=*/false); + m_llvmFlog = Function::Create( + /*Type=*/flogType, + /*Linkage=*/GlobalValue::ExternalLinkage, + /*Name=*/"logf", m_mod); + m_llvmFlog->setCallingConv(CallingConv::C); + m_llvmFlog->setParamAttrs(flogPal); + } + CallInst *call = m_builder.CreateCall(m_llvmFlog, val, + name("logf")); + call->setCallingConv(CallingConv::C); + call->setTailCall(false); + return call; } -void Instructions::beginLoop() +llvm::Value * Instructions::callFloor(llvm::Value *val) { - BasicBlock *begin = BasicBlock::Create(name("loop"), m_func,0); - BasicBlock *end = BasicBlock::Create(name("endloop"), m_func,0); - - m_builder.CreateBr(begin); - Loop loop; - loop.begin = begin; - loop.end = end; - m_builder.SetInsertPoint(begin); - m_loopStack.push(loop); + if (!m_llvmFloor) { + // predeclare the intrinsic + std::vector floorArgs; + floorArgs.push_back(Type::FloatTy); + PAListPtr floorPal; + FunctionType* floorType = FunctionType::get( + /*Result=*/Type::FloatTy, + /*Params=*/floorArgs, + /*isVarArg=*/false); + m_llvmFloor = Function::Create( + /*Type=*/floorType, + /*Linkage=*/GlobalValue::ExternalLinkage, + /*Name=*/"floorf", m_mod); + m_llvmFloor->setCallingConv(CallingConv::C); + m_llvmFloor->setParamAttrs(floorPal); + } + CallInst *call = m_builder.CreateCall(m_llvmFloor, val, + name("floorf")); + call->setCallingConv(CallingConv::C); + call->setTailCall(false); + return call; } -void Instructions::endLoop() +llvm::Value *Instructions::callFSqrt(llvm::Value *val) { - assert(!m_loopStack.empty()); - Loop loop = m_loopStack.top(); - m_builder.CreateBr(loop.begin); - loop.end->moveAfter(currentBlock()); - m_builder.SetInsertPoint(loop.end); - m_loopStack.pop(); + if (!m_llvmFSqrt) { + // predeclare the intrinsic + std::vector fsqrtArgs; + fsqrtArgs.push_back(Type::FloatTy); + PAListPtr fsqrtPal; + FunctionType* fsqrtType = FunctionType::get( + /*Result=*/Type::FloatTy, + /*Params=*/fsqrtArgs, + /*isVarArg=*/false); + m_llvmFSqrt = Function::Create( + /*Type=*/fsqrtType, + /*Linkage=*/GlobalValue::ExternalLinkage, + /*Name=*/"llvm.sqrt.f32", m_mod); + m_llvmFSqrt->setCallingConv(CallingConv::C); + m_llvmFSqrt->setParamAttrs(fsqrtPal); + } + CallInst *call = m_builder.CreateCall(m_llvmFSqrt, val, + name("sqrt")); + call->setCallingConv(CallingConv::C); + call->setTailCall(false); + return call; } -void Instructions::brk() +llvm::Value * Instructions::callPow(llvm::Value *val1, llvm::Value *val2) { - assert(!m_loopStack.empty()); - BasicBlock *unr = BasicBlock::Create(name("unreachable"), m_func,0); - m_builder.CreateBr(m_loopStack.top().end); - m_builder.SetInsertPoint(unr); + if (!m_llvmPow) { + // predeclare the intrinsic + std::vector powArgs; + powArgs.push_back(Type::FloatTy); + powArgs.push_back(Type::FloatTy); + PAListPtr powPal; + FunctionType* powType = FunctionType::get( + /*Result=*/Type::FloatTy, + /*Params=*/powArgs, + /*isVarArg=*/false); + m_llvmPow = Function::Create( + /*Type=*/powType, + /*Linkage=*/GlobalValue::ExternalLinkage, + /*Name=*/"llvm.pow.f32", m_mod); + m_llvmPow->setCallingConv(CallingConv::C); + m_llvmPow->setParamAttrs(powPal); + } + std::vector params; + params.push_back(val1); + params.push_back(val2); + CallInst *call = m_builder.CreateCall(m_llvmPow, params.begin(), params.end(), + name("pow")); + call->setCallingConv(CallingConv::C); + call->setTailCall(false); + return call; } -llvm::Value * Instructions::trunc(llvm::Value *in) +llvm::Value * Instructions::vectorFromVals(llvm::Value *x, llvm::Value *y, + llvm::Value *z, llvm::Value *w) { - std::vector vec = extractVector(in); - Value *icastx = m_builder.CreateFPToSI(vec[0], IntegerType::get(32), - name("ftoix")); - Value *icasty = m_builder.CreateFPToSI(vec[1], IntegerType::get(32), - name("ftoiy")); - Value *icastz = m_builder.CreateFPToSI(vec[2], IntegerType::get(32), - name("ftoiz")); - Value *icastw = m_builder.CreateFPToSI(vec[3], IntegerType::get(32), - name("ftoiw")); - Value *fx = m_builder.CreateSIToFP(icastx, Type::FloatTy, - name("fx")); - Value *fy = m_builder.CreateSIToFP(icasty, Type::FloatTy, - name("fy")); - Value *fz = m_builder.CreateSIToFP(icastz, Type::FloatTy, - name("fz")); - Value *fw = m_builder.CreateSIToFP(icastw, Type::FloatTy, - name("fw")); - return vectorFromVals(fx, fy, fz, fw); + Constant *const_vec = Constant::getNullValue(m_floatVecType); + Value *res = m_builder.CreateInsertElement(const_vec, x, + m_storage->constantInt(0), + name("vecx")); + res = m_builder.CreateInsertElement(res, y, m_storage->constantInt(1), + name("vecxy")); + res = m_builder.CreateInsertElement(res, z, m_storage->constantInt(2), + name("vecxyz")); + if (w) + res = m_builder.CreateInsertElement(res, w, m_storage->constantInt(3), + name("vecxyzw")); + return res; } -void Instructions::end() +llvm::Value * Instructions::constVector(float x, float y, float z, float w) { - m_builder.CreateRetVoid(); + std::vector vec(4); + vec[0] = ConstantFP::get(APFloat(x)); + vec[1] = ConstantFP::get(APFloat(y)); + vec[2] = ConstantFP::get(APFloat(z)); + vec[3] = ConstantFP::get(APFloat(w)); + return ConstantVector::get(m_floatVecType, vec); } -void Instructions::cal(int label, llvm::Value *input) +llvm::Function * Instructions::declarePrintf() { - std::vector params; - params.push_back(input); - llvm::Function *func = findFunction(label); - - m_builder.CreateCall(func, params.begin(), params.end()); + std::vector args; + PAListPtr params; + FunctionType* funcTy = FunctionType::get( + /*Result=*/IntegerType::get(32), + /*Params=*/args, + /*isVarArg=*/true); + Function* func_printf = Function::Create( + /*Type=*/funcTy, + /*Linkage=*/GlobalValue::ExternalLinkage, + /*Name=*/"printf", m_mod); + func_printf->setCallingConv(CallingConv::C); + func_printf->setParamAttrs(params); + return func_printf; } llvm::Function * Instructions::declareFunc(int label) @@ -778,27 +1108,6 @@ llvm::Function * Instructions::declareFunc(int label) return func; } -void Instructions::bgnSub(unsigned label) -{ - llvm::Function *func = findFunction(label); - - Function::arg_iterator args = func->arg_begin(); - Value *ptr_INPUT = args++; - ptr_INPUT->setName("INPUT"); - m_storage->pushArguments(ptr_INPUT); - - llvm::BasicBlock *entry = BasicBlock::Create("entry", func, 0); - - m_func = func; - m_builder.SetInsertPoint(entry); -} - -void Instructions::endSub() -{ - m_func = 0; - m_builder.SetInsertPoint(0); -} - llvm::Function * Instructions::findFunction(int label) { llvm::Function *func = m_functions[label]; @@ -809,17 +1118,6 @@ llvm::Function * Instructions::findFunction(int label) return func; } -llvm::Value * Instructions::constVector(float x, float y, float z, float w) -{ - std::vector vec(4); - vec[0] = ConstantFP::get(APFloat(x)); - vec[1] = ConstantFP::get(APFloat(y)); - vec[2] = ConstantFP::get(APFloat(z)); - vec[3] = ConstantFP::get(APFloat(w)); - return ConstantVector::get(m_floatVecType, vec); -} - - std::vector Instructions::extractVector(llvm::Value *vec) { std::vector elems(4); @@ -834,69 +1132,7 @@ std::vector Instructions::extractVector(llvm::Value *vec) return elems; } -llvm::Value * Instructions::cmp(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3) -{ - llvm::Function *func = m_mod->getFunction("cmp"); - assert(func); - - std::vector params; - params.push_back(in1); - params.push_back(in2); - params.push_back(in3); - CallInst *call = m_builder.CreateCall(func, params.begin(), params.end(), name("cmpres")); - call->setTailCall(false); - return call; -} - -llvm::Value * Instructions::cos(llvm::Value *in) -{ -#if 0 - llvm::Function *func = m_mod->getFunction("vcos"); - assert(func); - - CallInst *call = m_builder.CreateCall(func, in, name("cosres")); - call->setTailCall(false); - return call; -#else - std::vector elems = extractVector(in); - Function *func = m_mod->getFunction("cosf"); - assert(func); - CallInst *cos = m_builder.CreateCall(func, elems[0], name("cosres")); - cos->setCallingConv(CallingConv::C); - cos->setTailCall(true); - return vectorFromVals(cos, cos, cos, cos); -#endif -} - -llvm::Value * Instructions::scs(llvm::Value *in) -{ - llvm::Function *func = m_mod->getFunction("scs"); - assert(func); - - CallInst *call = m_builder.CreateCall(func, in, name("scsres")); - call->setTailCall(false); - return call; -} - -llvm::Value * Instructions::kil(llvm::Value *in) -{ - llvm::Function *func = m_mod->getFunction("kil"); - assert(func); - - CallInst *call = m_builder.CreateCall(func, in, name("kilpres")); - call->setTailCall(false); - return call; -} - -llvm::Value * Instructions::sin(llvm::Value *in) -{ - llvm::Function *func = m_mod->getFunction("vsin"); - assert(func); - CallInst *call = m_builder.CreateCall(func, in, name("sinres")); - call->setTailCall(false); - return call; -} #endif //MESA_LLVM diff --git a/src/gallium/auxiliary/gallivm/instructions.h b/src/gallium/auxiliary/gallivm/instructions.h index d286ce80c7..8df30f62c8 100644 --- a/src/gallium/auxiliary/gallivm/instructions.h +++ b/src/gallium/auxiliary/gallivm/instructions.h @@ -57,15 +57,22 @@ public: llvm::BasicBlock *currentBlock() const; llvm::Value *abs(llvm::Value *in1); - llvm::Value *arl(llvm::Value *in1); llvm::Value *add(llvm::Value *in1, llvm::Value *in2); + llvm::Value *arl(llvm::Value *in1); void beginLoop(); void bgnSub(unsigned); void brk(); void cal(int label, llvm::Value *input); + llvm::Value *clamp(llvm::Value *in); llvm::Value *cmp(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3); + llvm::Value *cnd(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3); + llvm::Value *cnd0(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3); llvm::Value *cos(llvm::Value *in); llvm::Value *cross(llvm::Value *in1, llvm::Value *in2); + llvm::Value *ddx(llvm::Value *in); + llvm::Value *ddy(llvm::Value *in); + llvm::Value *div(llvm::Value *in1, llvm::Value *in2); + llvm::Value *dot2add(llvm::Value *in, llvm::Value *in2, llvm::Value *in3); llvm::Value *dp3(llvm::Value *in1, llvm::Value *in2); llvm::Value *dp4(llvm::Value *in1, llvm::Value *in2); llvm::Value *dph(llvm::Value *in1, llvm::Value *in2); @@ -75,6 +82,7 @@ public: void endLoop(); void end(); void endSub(); + llvm::Value *exp(llvm::Value *in); llvm::Value *ex2(llvm::Value *in); llvm::Value *floor(llvm::Value *in); llvm::Value *frc(llvm::Value *in); @@ -82,32 +90,41 @@ public: llvm::Value *kil(llvm::Value *in); llvm::Value *lerp(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3); - llvm::Value *lit(llvm::Value *in); llvm::Value *lg2(llvm::Value *in); + llvm::Value *lit(llvm::Value *in); + llvm::Value *log(llvm::Value *in); llvm::Value *madd(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3); - llvm::Value *min(llvm::Value *in1, llvm::Value *in2); llvm::Value *max(llvm::Value *in1, llvm::Value *in2); + llvm::Value *min(llvm::Value *in1, llvm::Value *in2); llvm::Value *mul(llvm::Value *in1, llvm::Value *in2); + llvm::Value *neg(llvm::Value *in); llvm::Value *pow(llvm::Value *in1, llvm::Value *in2); llvm::Value *rcp(llvm::Value *in); llvm::Value *rsq(llvm::Value *in); llvm::Value *scs(llvm::Value *in); + llvm::Value *seq(llvm::Value *in1, llvm::Value *in2); + llvm::Value *sfl(llvm::Value *in1, llvm::Value *in2); llvm::Value *sge(llvm::Value *in1, llvm::Value *in2); llvm::Value *sgt(llvm::Value *in1, llvm::Value *in2); llvm::Value *sin(llvm::Value *in); + llvm::Value *sle(llvm::Value *in1, llvm::Value *in2); llvm::Value *slt(llvm::Value *in1, llvm::Value *in2); + llvm::Value *sne(llvm::Value *in1, llvm::Value *in2); + llvm::Value *str(llvm::Value *in1, llvm::Value *in2); llvm::Value *sub(llvm::Value *in1, llvm::Value *in2); llvm::Value *trunc(llvm::Value *in); + llvm::Value *x2d(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3); void printVector(llvm::Value *val); private: const char *name(const char *prefix); llvm::Value *callFAbs(llvm::Value *val); + llvm::Value *callFExp(llvm::Value *val); + llvm::Value *callFLog(llvm::Value *val); llvm::Value *callFloor(llvm::Value *val); llvm::Value *callFSqrt(llvm::Value *val); - llvm::Value *callFLog(llvm::Value *val); llvm::Value *callPow(llvm::Value *val1, llvm::Value *val2); llvm::Value *vectorFromVals(llvm::Value *x, llvm::Value *y, @@ -135,6 +152,7 @@ private: llvm::Function *m_llvmPow; llvm::Function *m_llvmFloor; llvm::Function *m_llvmFlog; + llvm::Function *m_llvmFexp; llvm::Function *m_llvmLit; llvm::Constant *m_fmtPtr; diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp index cc1516a45e..398fbd67bd 100644 --- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp +++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp @@ -286,9 +286,13 @@ translate_instruction(llvm::Module *module, out = instr->rsq(inputs[0]); } break; - case TGSI_OPCODE_EXP: + case TGSI_OPCODE_EXP: { + out = instr->exp(inputs[0]); + } break; - case TGSI_OPCODE_LOG: + case TGSI_OPCODE_LOG: { + out = instr->log(inputs[0]); + } break; case TGSI_OPCODE_MUL: { out = instr->mul(inputs[0], inputs[1]); @@ -338,21 +342,31 @@ translate_instruction(llvm::Module *module, out = instr->lerp(inputs[0], inputs[1], inputs[2]); } break; - case TGSI_OPCODE_CND: + case TGSI_OPCODE_CND: { + out = instr->cnd(inputs[0], inputs[1], inputs[2]); + } break; - case TGSI_OPCODE_CND0: + case TGSI_OPCODE_CND0: { + out = instr->cnd0(inputs[0], inputs[1], inputs[2]); + } break; - case TGSI_OPCODE_DOT2ADD: + case TGSI_OPCODE_DOT2ADD: { + out = instr->dot2add(inputs[0], inputs[1], inputs[2]); + } break; case TGSI_OPCODE_INDEX: break; - case TGSI_OPCODE_NEGATE: + case TGSI_OPCODE_NEGATE: { + out = instr->neg(inputs[0]); + } break; case TGSI_OPCODE_FRAC: { out = instr->frc(inputs[0]); } break; - case TGSI_OPCODE_CLAMP: + case TGSI_OPCODE_CLAMP: { + out = instr->clamp(inputs[0]); + } break; case TGSI_OPCODE_FLOOR: { out = instr->floor(inputs[0]); @@ -392,9 +406,13 @@ translate_instruction(llvm::Module *module, out = instr->cos(inputs[0]); } break; - case TGSI_OPCODE_DDX: + case TGSI_OPCODE_DDX: { + out = instr->ddx(inputs[0]); + } break; - case TGSI_OPCODE_DDY: + case TGSI_OPCODE_DDY: { + out = instr->ddy(inputs[0]); + } break; case TGSI_OPCODE_KILP: break; @@ -408,9 +426,13 @@ translate_instruction(llvm::Module *module, break; case TGSI_OPCODE_RFL: break; - case TGSI_OPCODE_SEQ: + case TGSI_OPCODE_SEQ: { + out = instr->seq(inputs[0], inputs[1]); + } break; - case TGSI_OPCODE_SFL: + case TGSI_OPCODE_SFL: { + out = instr->sfl(inputs[0], inputs[1]); + } break; case TGSI_OPCODE_SGT: { out = instr->sgt(inputs[0], inputs[1]); @@ -420,11 +442,17 @@ translate_instruction(llvm::Module *module, out = instr->sin(inputs[0]); } break; - case TGSI_OPCODE_SLE: + case TGSI_OPCODE_SLE: { + out = instr->sle(inputs[0], inputs[1]); + } break; - case TGSI_OPCODE_SNE: + case TGSI_OPCODE_SNE: { + out = instr->sne(inputs[0], inputs[1]); + } break; - case TGSI_OPCODE_STR: + case TGSI_OPCODE_STR: { + out = instr->str(inputs[0], inputs[1]); + } break; case TGSI_OPCODE_TEX: break; @@ -438,7 +466,9 @@ translate_instruction(llvm::Module *module, break; case TGSI_OPCODE_UP4UB: break; - case TGSI_OPCODE_X2D: + case TGSI_OPCODE_X2D: { + out = instr->x2d(inputs[0], inputs[1], inputs[2]); + } break; case TGSI_OPCODE_ARA: break; -- cgit v1.2.3 From a0a06cbc5b26d7530bd5066f09efe3c1f980d35d Mon Sep 17 00:00:00 2001 From: Stephane Marchesin Date: Sun, 28 Sep 2008 19:48:26 +0200 Subject: Gallivm: more instructions. --- src/gallium/auxiliary/gallivm/instructions.cpp | 61 ++++++++++++++++++++++++-- src/gallium/auxiliary/gallivm/instructions.h | 5 +++ src/gallium/auxiliary/gallivm/tgsitollvm.cpp | 15 ++++--- 3 files changed, 73 insertions(+), 8 deletions(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/gallivm/instructions.cpp b/src/gallium/auxiliary/gallivm/instructions.cpp index 5fdfe09d18..3eaf9aacf6 100644 --- a/src/gallium/auxiliary/gallivm/instructions.cpp +++ b/src/gallium/auxiliary/gallivm/instructions.cpp @@ -163,9 +163,18 @@ void Instructions::cal(int label, llvm::Value *input) m_builder.CreateCall(func, params.begin(), params.end()); } +llvm::Value * Instructions::ceil(llvm::Value *in) +{ + std::vector vec = extractVector(in); + return vectorFromVals(callCeil(vec[0]), callCeil(vec[1]), + callCeil(vec[2]), callCeil(vec[3])); +} + llvm::Value * Instructions::clamp(llvm::Value *in1) { - // FIXME + llvm::Value *zero = constVector(0.0f, 0.0f, 0.0f, 0.0f); + llvm::Value *one = constVector(1.0f, 1.0f, 1.0f, 1.0f); + return min( max(zero, in1), one); } llvm::Value * Instructions::cmp(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3) @@ -289,12 +298,14 @@ llvm::Value * Instructions::cross(llvm::Value *in1, llvm::Value *in2) llvm::Value * Instructions::ddx(llvm::Value *in) { - // FIXME + // FIXME + assert(0); } llvm::Value * Instructions::ddy(llvm::Value *in) { - // FIXME + // FIXME + assert(0); } llvm::Value * Instructions::div(llvm::Value *in1, llvm::Value *in2) @@ -319,6 +330,19 @@ llvm::Value * Instructions::dot2add(llvm::Value *in1, llvm::Value *in2, llvm::Va return vectorFromVals(dot2add, dot2add, dot2add, dot2add); } +llvm::Value * Instructions::dp2(llvm::Value *in1, llvm::Value *in2) +{ + Value *mulRes = mul(in1, in2); + Value *x = m_builder.CreateExtractElement(mulRes, + m_storage->constantInt(0), + name("extractx")); + Value *y = m_builder.CreateExtractElement(mulRes, + m_storage->constantInt(1), + name("extracty")); + Value *xy = m_builder.CreateAdd(x, y,name("xy")); + return vectorFromVals(xy, xy, xy, xy); +} + llvm::Value * Instructions::dp3(llvm::Value *in1, llvm::Value *in2) { Value *mulRes = mul(in1, in2); @@ -581,6 +605,12 @@ llvm::Value * Instructions::neg(llvm::Value *in) return neg; } +llvm::Value * Instructions::nrm(llvm::Value *in) +{ + llvm::Value *v = rsq(in); + return mul(v, in); +} + llvm::Value * Instructions::pow(llvm::Value *in1, llvm::Value *in2) { Value *x1 = m_builder.CreateExtractElement(in1, @@ -887,6 +917,31 @@ const char * Instructions::name(const char *prefix) return m_name; } +llvm::Value * Instructions::callCeil(llvm::Value *val) +{ + if (!m_llvmCeil) { + // predeclare the intrinsic + std::vector ceilArgs; + ceilArgs.push_back(Type::FloatTy); + PAListPtr ceilPal; + FunctionType* ceilType = FunctionType::get( + /*Result=*/Type::FloatTy, + /*Params=*/ceilArgs, + /*isVarArg=*/false); + m_llvmCeil = Function::Create( + /*Type=*/ceilType, + /*Linkage=*/GlobalValue::ExternalLinkage, + /*Name=*/"ceilf", m_mod); + m_llvmCeil->setCallingConv(CallingConv::C); + m_llvmCeil->setParamAttrs(ceilPal); + } + CallInst *call = m_builder.CreateCall(m_llvmCeil, val, + name("ceilf")); + call->setCallingConv(CallingConv::C); + call->setTailCall(false); + return call; +} + llvm::Value *Instructions::callFAbs(llvm::Value *val) { if (!m_llvmFAbs) { diff --git a/src/gallium/auxiliary/gallivm/instructions.h b/src/gallium/auxiliary/gallivm/instructions.h index 8df30f62c8..c3b28e9746 100644 --- a/src/gallium/auxiliary/gallivm/instructions.h +++ b/src/gallium/auxiliary/gallivm/instructions.h @@ -63,6 +63,7 @@ public: void bgnSub(unsigned); void brk(); void cal(int label, llvm::Value *input); + llvm::Value *ceil(llvm::Value *in); llvm::Value *clamp(llvm::Value *in); llvm::Value *cmp(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3); llvm::Value *cnd(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3); @@ -73,6 +74,7 @@ public: llvm::Value *ddy(llvm::Value *in); llvm::Value *div(llvm::Value *in1, llvm::Value *in2); llvm::Value *dot2add(llvm::Value *in, llvm::Value *in2, llvm::Value *in3); + llvm::Value *dp2(llvm::Value *in1, llvm::Value *in2); llvm::Value *dp3(llvm::Value *in1, llvm::Value *in2); llvm::Value *dp4(llvm::Value *in1, llvm::Value *in2); llvm::Value *dph(llvm::Value *in1, llvm::Value *in2); @@ -99,6 +101,7 @@ public: llvm::Value *min(llvm::Value *in1, llvm::Value *in2); llvm::Value *mul(llvm::Value *in1, llvm::Value *in2); llvm::Value *neg(llvm::Value *in); + llvm::Value *nrm(llvm::Value *in); llvm::Value *pow(llvm::Value *in1, llvm::Value *in2); llvm::Value *rcp(llvm::Value *in); llvm::Value *rsq(llvm::Value *in); @@ -120,6 +123,7 @@ public: private: const char *name(const char *prefix); + llvm::Value *callCeil(llvm::Value *val); llvm::Value *callFAbs(llvm::Value *val); llvm::Value *callFExp(llvm::Value *val); llvm::Value *callFLog(llvm::Value *val); @@ -147,6 +151,7 @@ private: llvm::VectorType *m_floatVecType; + llvm::Function *m_llvmCeil; llvm::Function *m_llvmFSqrt; llvm::Function *m_llvmFAbs; llvm::Function *m_llvmPow; diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp index 398fbd67bd..fdfbb76c16 100644 --- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp +++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp @@ -498,11 +498,18 @@ translate_instruction(llvm::Module *module, break; case TGSI_OPCODE_TXB: break; - case TGSI_OPCODE_NRM: + case TGSI_OPCODE_NRM4: + case TGSI_OPCODE_NRM: { + out = instr->nrm(inputs[0]); + } break; - case TGSI_OPCODE_DIV: + case TGSI_OPCODE_DIV: { + out = instr->div(inputs[0], inputs[1]); + } break; - case TGSI_OPCODE_DP2: + case TGSI_OPCODE_DP2: { + out = instr->dp2(inputs[0], inputs[1]); + } break; case TGSI_OPCODE_TXL: break; @@ -620,8 +627,6 @@ translate_instruction(llvm::Module *module, break; case TGSI_OPCODE_M3X2: break; - case TGSI_OPCODE_NRM4: - break; case TGSI_OPCODE_CALLNZ: break; case TGSI_OPCODE_IFC: -- cgit v1.2.3 From 7379d0ef8f533b0aa760cd21b219223602002a56 Mon Sep 17 00:00:00 2001 From: Stephane Marchesin Date: Sun, 28 Sep 2008 23:18:55 +0200 Subject: Gallivm: fix off-by-one. --- src/gallium/auxiliary/gallivm/instructionssoa.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.cpp b/src/gallium/auxiliary/gallivm/instructionssoa.cpp index efddc04e81..9a3ed9f538 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.cpp +++ b/src/gallium/auxiliary/gallivm/instructionssoa.cpp @@ -259,7 +259,7 @@ void InstructionsSoa::createBuiltins() { MemoryBuffer *buffer = MemoryBuffer::getMemBuffer( (const char*)&soabuiltins_data[0], - (const char*)&soabuiltins_data[Elements(soabuiltins_data)]); + (const char*)&soabuiltins_data[Elements(soabuiltins_data)-1]); m_builtins = ParseBitcodeFile(buffer); std::cout<<"Builtins created at "< Date: Tue, 30 Sep 2008 20:50:49 +0200 Subject: Gallivm: port to llvm 2.4. --- configs/linux-llvm | 1 + src/gallium/auxiliary/gallivm/gallivm_builtins.cpp | 252 ++++++++++----------- src/gallium/auxiliary/gallivm/instructions.cpp | 36 +-- src/gallium/auxiliary/gallivm/instructions.h | 2 +- src/gallium/auxiliary/gallivm/instructionssoa.cpp | 6 +- src/gallium/auxiliary/gallivm/instructionssoa.h | 2 +- 6 files changed, 150 insertions(+), 149 deletions(-) (limited to 'src/gallium/auxiliary') diff --git a/configs/linux-llvm b/configs/linux-llvm index 3b32db34d8..489cfd0546 100644 --- a/configs/linux-llvm +++ b/configs/linux-llvm @@ -31,4 +31,5 @@ else LLVM_CXXFLAGS= endif +LD = g++ GL_LIB_DEPS = $(LLVM_LDFLAGS) $(LLVM_LIBS) $(EXTRA_LIB_PATH) -lX11 -lXext -lm -lpthread -lstdc++ diff --git a/src/gallium/auxiliary/gallivm/gallivm_builtins.cpp b/src/gallium/auxiliary/gallivm/gallivm_builtins.cpp index 0fc5c4ec5c..fcc5c05794 100644 --- a/src/gallium/auxiliary/gallivm/gallivm_builtins.cpp +++ b/src/gallium/auxiliary/gallivm/gallivm_builtins.cpp @@ -1,140 +1,140 @@ static const unsigned char llvm_builtins_data[] = { -0x42,0x43,0xc0,0xde,0x21,0x0c,0x00,0x00,0x29,0x02,0x00,0x00,0x01,0x10,0x00,0x00, +0x42,0x43,0xc0,0xde,0x21,0x0c,0x00,0x00,0x27,0x02,0x00,0x00,0x01,0x10,0x00,0x00, 0x10,0x00,0x00,0x00,0x07,0x81,0x23,0x91,0x41,0xc8,0x04,0x49,0x06,0x10,0x32,0x39, 0x92,0x01,0x84,0x0c,0x25,0x05,0x08,0x19,0x1e,0x04,0x8b,0x62,0x80,0x14,0x45,0x02, 0x42,0x92,0x0b,0x42,0xa4,0x10,0x32,0x14,0x38,0x08,0x18,0x49,0x0a,0x32,0x44,0x24, 0x48,0x0a,0x90,0x21,0x23,0x44,0x72,0x80,0x8c,0x14,0x21,0x86,0x0a,0x8a,0x0a,0x64, -0x0c,0x1f,0x00,0x00,0x49,0x18,0x00,0x00,0x02,0x00,0x00,0x00,0x0b,0x04,0x00,0x0c, -0x00,0x00,0x00,0x00,0x51,0x20,0x00,0x00,0x12,0x00,0x00,0x00,0x32,0x22,0x48,0x09, -0x20,0x65,0x82,0x84,0x00,0x26,0x45,0x48,0x05,0x09,0x26,0x45,0xc6,0x05,0x42,0x52, -0x26,0x08,0xae,0x19,0x80,0x61,0x04,0x02,0x98,0x23,0x00,0x83,0x29,0x80,0x21,0x00, -0xb2,0x73,0x04,0x01,0x51,0x8a,0xf4,0x08,0x92,0xa4,0x39,0x47,0x80,0x50,0x2b,0x03, -0x00,0xa0,0x08,0x21,0x5c,0x46,0x2b,0x44,0x08,0x21,0xd4,0x40,0x14,0x01,0x80,0x11, -0x80,0x22,0x88,0x00,0x13,0xa2,0x74,0xb0,0x03,0x3c,0xb0,0x83,0x36,0x80,0x87,0x71, -0x68,0x03,0x76,0x48,0x07,0x77,0xa8,0x07,0x7c,0x68,0x83,0x73,0x70,0x87,0x7a,0xd8, -0x70,0x0f,0xe5,0xd0,0x06,0xf0,0xa0,0x07,0x73,0x20,0x07,0x7a,0x30,0x07,0x72,0xa0, -0x07,0x73,0x20,0x07,0x6d,0x90,0x0e,0x71,0xa0,0x07,0x78,0xa0,0x07,0x78,0xd0,0x06, -0xe9,0x80,0x07,0x7a,0x80,0x07,0x7a,0x80,0x07,0x6d,0x90,0x0e,0x71,0x60,0x07,0x7a, -0x10,0x07,0x76,0xa0,0x07,0x71,0x60,0x07,0x6d,0x90,0x0e,0x73,0x20,0x07,0x7a,0x30, -0x07,0x72,0xa0,0x07,0x73,0x20,0x07,0x6d,0x90,0x0e,0x76,0x40,0x07,0x7a,0x30,0x07, -0x72,0xa0,0x07,0x76,0x40,0x07,0x6d,0x60,0x0e,0x73,0x20,0x07,0x7a,0x30,0x07,0x72, -0xa0,0x07,0x73,0x20,0x07,0x6d,0x60,0x0e,0x76,0x40,0x07,0x7a,0x30,0x07,0x72,0xa0, -0x07,0x76,0x40,0x07,0x6d,0x60,0x0f,0x76,0x40,0x07,0x7a,0x60,0x07,0x74,0xa0,0x07, -0x76,0x40,0x07,0x6d,0x60,0x0f,0x71,0x20,0x07,0x78,0xa0,0x07,0x71,0x20,0x07,0x78, -0xa0,0x07,0x71,0x20,0x07,0x78,0xd0,0x06,0xe1,0x00,0x07,0x7a,0x00,0x07,0x7a,0x60, -0x07,0x74,0xd0,0x06,0xe6,0x80,0x07,0x70,0xa0,0x07,0x71,0x20,0x07,0x78,0xa0,0x07, -0x71,0x20,0x07,0x78,0xa0,0xf3,0x40,0x88,0x04,0x32,0x32,0x02,0x04,0x20,0x76,0x46, -0xfc,0x6c,0x48,0x92,0x00,0x40,0x00,0x00,0x00,0x00,0x0c,0x49,0x12,0x20,0x00,0x00, -0x00,0x00,0x80,0x21,0x89,0x02,0x00,0x01,0x00,0x00,0x00,0x30,0x24,0x59,0x00,0x20, -0x08,0x00,0x00,0x00,0x86,0x24,0x0a,0x00,0x04,0x00,0x00,0x00,0xc0,0x90,0x84,0x01, -0x02,0x00,0x00,0x00,0x00,0x18,0x92,0x1c,0x40,0x00,0x00,0x00,0x00,0x00,0x43,0x12, -0x05,0x00,0x02,0x00,0x00,0x00,0x60,0x48,0x72,0x00,0x01,0x00,0x00,0x00,0x00,0x0c, -0x49,0x14,0x00,0x08,0x00,0x00,0x00,0x80,0x21,0x49,0x01,0x00,0x41,0x00,0x00,0x00, -0x90,0x05,0x02,0x00,0x10,0x00,0x00,0x00,0x32,0x1e,0x98,0x10,0x19,0x11,0x4c,0x90, +0x0c,0x1f,0x00,0x00,0x49,0x18,0x00,0x00,0x03,0x00,0x00,0x00,0x0b,0x84,0xff,0xff, +0xff,0xff,0x1f,0xc0,0x00,0x00,0x00,0x00,0x51,0x20,0x00,0x00,0x12,0x00,0x00,0x00, +0x32,0x22,0x48,0x09,0x20,0x65,0x82,0x84,0x00,0x26,0x45,0x48,0x05,0x09,0x26,0x45, +0xc6,0x05,0x42,0x52,0x26,0x08,0xae,0x19,0x80,0x61,0x04,0x02,0x98,0x23,0x00,0x83, +0x29,0x80,0x21,0x00,0xb2,0x73,0x04,0x01,0x51,0x8a,0xf4,0x08,0x92,0xa4,0x39,0x47, +0x80,0x50,0x2b,0x03,0x00,0xa0,0x08,0x21,0x5c,0x46,0x2b,0x44,0x08,0x21,0xd4,0x40, +0x14,0x01,0x80,0x11,0x80,0x22,0x88,0x00,0x13,0x30,0x7c,0xc0,0x03,0x3b,0xf8,0x05, +0x3b,0xa0,0x83,0x36,0xa8,0x07,0x77,0x58,0x07,0x77,0x78,0x87,0x7b,0x70,0x87,0x36, +0x60,0x87,0x74,0x70,0x87,0x7a,0xc0,0x87,0x36,0x38,0x07,0x77,0xa8,0x87,0x0d,0xf7, +0x50,0x0e,0x6d,0x00,0x0f,0x7a,0x60,0x07,0x74,0xa0,0x07,0x76,0x40,0x07,0x7a,0x60, +0x07,0x74,0xd0,0x06,0xe9,0x10,0x07,0x7a,0x80,0x07,0x7a,0x80,0x07,0x6d,0x90,0x0e, +0x78,0xa0,0x07,0x78,0xa0,0x07,0x78,0xd0,0x06,0xe9,0x10,0x07,0x76,0xa0,0x07,0x71, +0x60,0x07,0x7a,0x10,0x07,0x76,0xd0,0x06,0xe9,0x30,0x07,0x72,0xa0,0x07,0x73,0x20, +0x07,0x7a,0x30,0x07,0x72,0xd0,0x06,0xe9,0x60,0x07,0x74,0xa0,0x07,0x76,0x40,0x07, +0x7a,0x60,0x07,0x74,0xd0,0x06,0xe6,0x30,0x07,0x72,0xa0,0x07,0x73,0x20,0x07,0x7a, +0x30,0x07,0x72,0xd0,0x06,0xe6,0x60,0x07,0x74,0xa0,0x07,0x76,0x40,0x07,0x7a,0x60, +0x07,0x74,0xd0,0x06,0xf6,0x60,0x07,0x74,0xa0,0x07,0x76,0x40,0x07,0x7a,0x60,0x07, +0x74,0xd0,0x06,0xf6,0x10,0x07,0x72,0x80,0x07,0x7a,0x10,0x07,0x72,0x80,0x07,0x7a, +0x10,0x07,0x72,0x80,0x07,0x6d,0x10,0x0e,0x70,0xa0,0x07,0x70,0xa0,0x07,0x76,0x40, +0x07,0x6d,0x60,0x0e,0x78,0x00,0x07,0x7a,0x10,0x07,0x72,0x80,0x07,0x7a,0x10,0x07, +0x72,0x80,0x07,0x3a,0x0f,0x84,0x48,0x20,0x23,0x24,0x40,0x00,0x62,0x67,0x88,0x9f, +0x19,0x92,0x24,0x00,0x10,0x04,0x00,0x00,0x00,0x43,0x92,0x04,0x08,0x00,0x00,0x00, +0x00,0x60,0x48,0xa2,0x00,0x40,0x10,0x00,0x00,0x00,0x0c,0x49,0x16,0x00,0x08,0x02, +0x00,0x00,0x80,0x21,0x89,0x02,0x00,0x41,0x00,0x00,0x00,0x30,0x24,0x61,0x80,0x00, +0x00,0x00,0x00,0x00,0x86,0x24,0x07,0x10,0x00,0x00,0x00,0x00,0xc0,0x90,0x44,0x01, +0x80,0x20,0x00,0x00,0x00,0x18,0x92,0x1c,0x40,0x00,0x00,0x00,0x00,0x00,0x43,0x12, +0x05,0x00,0x82,0x00,0x00,0x00,0x60,0x48,0x52,0x00,0x40,0x10,0x00,0x00,0x00,0x64, +0x81,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x32,0x1e,0x98,0x10,0x19,0x11,0x4c,0x90, 0x8c,0x09,0x26,0x47,0xc6,0x04,0x43,0x8a,0x8a,0x59,0x8b,0x43,0x50,0xd2,0x09,0x02, 0x81,0xd2,0x73,0x50,0xc9,0x0c,0x2a,0x99,0x41,0x25,0x33,0xa8,0x64,0x56,0x28,0x66, 0x2d,0x0e,0x41,0xcf,0x2a,0x15,0x04,0x4a,0xcf,0x41,0x25,0x33,0xa8,0x64,0x06,0x95, 0xcc,0xa0,0x92,0x59,0x01,0x00,0x00,0x00,0x53,0x82,0x26,0x0c,0x04,0x00,0x00,0x00, 0x22,0x00,0x00,0x00,0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x05,0x00,0x00,0x00, 0x04,0xc6,0x08,0x40,0x10,0x04,0xe1,0x70,0x18,0x23,0x00,0x41,0x10,0x84,0xc3,0x60, -0x04,0x00,0x00,0x00,0x93,0x0c,0xce,0x43,0x4c,0x31,0x3c,0x8e,0x34,0xc9,0x30,0x41, -0xc2,0x14,0x03,0x34,0x51,0x93,0x0c,0x4d,0x44,0x4c,0x31,0x44,0x8d,0x35,0x56,0x01, -0x04,0xc3,0x55,0x21,0x16,0x0e,0x04,0x00,0x0f,0x00,0x00,0x00,0x46,0x41,0x08,0xcc, -0x73,0x9b,0x05,0x21,0x30,0xcf,0x6e,0x18,0x84,0x00,0x2c,0x8b,0x35,0x04,0x80,0x39, -0x04,0x81,0x5d,0x20,0x80,0x0f,0x0c,0x43,0xe4,0xd3,0x36,0x81,0x04,0x3e,0x30,0x0c, -0x91,0x4f,0x5b,0x05,0x12,0xf8,0xc0,0x30,0x44,0x7e,0x7d,0x00,0x05,0xd1,0x4c,0x11, -0x66,0x12,0x83,0xc0,0x3c,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00, -0x2a,0x00,0x00,0x00,0x13,0x04,0x43,0x2c,0x10,0x00,0x00,0x00,0x08,0x00,0x00,0x00, +0x04,0x00,0x00,0x00,0xc3,0x0d,0xce,0x43,0x4c,0x37,0x3c,0x8e,0x34,0xdc,0x30,0x41, +0xc2,0x74,0x03,0x34,0x51,0xc3,0x0d,0x4d,0x44,0x4c,0x37,0x44,0x8d,0x35,0x56,0x01, +0x04,0xc3,0x55,0x21,0x16,0x0e,0x04,0x00,0x0f,0x00,0x00,0x00,0xd6,0x10,0x00,0xe6, +0x10,0x04,0x76,0x81,0x00,0x3e,0x30,0x0c,0x91,0x4f,0x1b,0x05,0x21,0x30,0x8f,0x6d, +0x13,0x48,0xe0,0x03,0xc3,0x10,0xf9,0xb4,0x55,0x20,0x81,0x0f,0x0c,0x43,0xe4,0xd7, +0x66,0x41,0x08,0xcc,0xa3,0x1f,0x40,0x41,0x34,0x53,0x84,0x99,0xc4,0x20,0x30,0x8f, +0x61,0x10,0x02,0xb0,0x2c,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00, +0x27,0x00,0x00,0x00,0x13,0x04,0x43,0x2c,0x10,0x00,0x00,0x00,0x08,0x00,0x00,0x00, 0x24,0x8a,0xa0,0x0c,0x46,0x00,0x4a,0x80,0xc2,0x1c,0x84,0x55,0x55,0xd6,0x1c,0x84, 0x45,0x51,0x16,0x81,0x19,0x80,0x11,0x80,0x31,0x02,0x10,0x04,0x41,0xfc,0x03,0x00, -0x63,0x08,0x0d,0x34,0xc9,0x70,0x55,0xc2,0x2c,0x43,0x20,0x60,0x73,0x0c,0xd3,0x15, +0x63,0x08,0x0d,0x34,0xdc,0x70,0x55,0xc2,0x2c,0x43,0x20,0x60,0x73,0x0c,0xd3,0x15, 0x8d,0x21,0x34,0xd1,0x18,0x42,0xf3,0x8c,0x55,0x00,0x81,0xa0,0x6d,0x73,0x0c,0x19, -0xe7,0x60,0x87,0x52,0x38,0x10,0x00,0x00,0x13,0x00,0x00,0x00,0x17,0x60,0x20,0xc5, -0x74,0x10,0x8d,0x65,0x14,0x13,0xf3,0xd4,0xb4,0x6d,0x14,0x13,0xf3,0xd4,0xb8,0x69, -0x14,0x13,0xf3,0xd4,0xb6,0x75,0x14,0x13,0xf3,0xd4,0xba,0x35,0x0c,0x13,0xf3,0x9c, -0x80,0xe4,0x36,0x48,0x81,0x10,0xc3,0x4a,0x4c,0x54,0xd4,0x6c,0x8b,0x23,0x28,0x76, -0x41,0x4c,0xcc,0xa3,0x1b,0x07,0x21,0x00,0xcb,0x72,0x00,0x05,0xd1,0x4c,0x11,0x66, -0x18,0x83,0xc0,0x3c,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00, -0x81,0x00,0x00,0x00,0x13,0x04,0x4d,0x2c,0x10,0x00,0x00,0x00,0x04,0x00,0x00,0x00, -0x24,0xca,0x60,0x04,0xa0,0x04,0x8a,0x80,0xc2,0x0c,0x00,0x91,0x11,0x00,0x00,0x00, -0x63,0x08,0x4d,0x64,0x16,0xc1,0x49,0x86,0xab,0x22,0x66,0x19,0x02,0x01,0x1b,0x43, -0x70,0xa2,0x59,0x82,0x61,0x0c,0xe1,0x89,0x66,0x09,0x86,0x81,0x0a,0x20,0x0b,0x34, -0x61,0x8e,0x81,0xda,0xa2,0x31,0x84,0x46,0xb2,0x8e,0xe0,0x24,0x83,0x57,0x11,0xb3, -0x0c,0x44,0xf1,0x8d,0x21,0x38,0xd2,0x2c,0x81,0x31,0x86,0xf0,0x48,0xb3,0x04,0xc6, -0x40,0x05,0x00,0x06,0x44,0x18,0x14,0x73,0x0c,0x9c,0x18,0x48,0x63,0x08,0xcd,0x64, -0x64,0x40,0x70,0x92,0xa1,0x0c,0x2a,0x62,0x96,0xe1,0x40,0xcc,0x60,0x0c,0xc1,0x99, -0x66,0x09,0x92,0x31,0x84,0x67,0x9a,0x25,0x48,0x06,0x2a,0x80,0x33,0x38,0xd0,0x00, -0x99,0x63,0x18,0x83,0x34,0x98,0xc6,0x10,0x1a,0xc8,0xd6,0x80,0xe0,0x24,0x03,0x1b, -0x54,0xc4,0x2c,0x83,0xb2,0xb4,0xc1,0x18,0x82,0x03,0xcd,0x12,0x30,0x63,0x08,0x0f, -0x34,0x4b,0xc0,0x0c,0x54,0x00,0x6e,0xa0,0xbc,0xc1,0x32,0xc7,0xa0,0x06,0x70,0x00, -0x61,0x1c,0x84,0x03,0x01,0x00,0x00,0x00,0x4e,0x00,0x00,0x00,0x76,0x52,0x4c,0xcc, -0x73,0xd3,0x24,0x05,0x64,0xec,0xcd,0x8d,0xcc,0xe5,0x87,0x46,0xc6,0x50,0x8a,0x89, -0x79,0xee,0xdb,0x54,0x8a,0x89,0x79,0xee,0xdd,0x1a,0x88,0x89,0x79,0x68,0x73,0x20, -0x26,0xe6,0xa9,0xed,0x81,0x98,0x98,0xc7,0x36,0x0b,0x62,0x62,0x9e,0xdb,0x32,0x88, -0x89,0x79,0x72,0xd3,0x20,0x26,0xe6,0xd9,0x8d,0x83,0x98,0x98,0xa7,0xb7,0x95,0x62, -0x62,0x9e,0xbb,0x27,0x2d,0x20,0x63,0x6f,0x6e,0x64,0x2e,0x3a,0x34,0x35,0x56,0x62, -0x08,0x4e,0x53,0xd9,0xba,0xb5,0x14,0x02,0xf3,0xe0,0xf5,0x25,0x2c,0x82,0xd3,0x0c, -0xbe,0xe0,0x34,0xd3,0x8d,0x9b,0x88,0x21,0x38,0xcd,0x60,0xd7,0x24,0x01,0x63,0xec, -0xcd,0x8d,0xcc,0x45,0x87,0x44,0x80,0x8c,0xbd,0xb9,0x91,0xb9,0xfc,0xc4,0xd0,0x90, -0x02,0x8c,0xb1,0x37,0x37,0x32,0x97,0x1f,0x73,0x29,0x26,0xe6,0xc1,0x71,0x7b,0x29, -0x26,0xe6,0xc1,0x77,0xfb,0x28,0x04,0xe6,0xa9,0x6f,0x52,0x01,0x32,0xf6,0xe6,0x46, -0xe6,0xa2,0x13,0x73,0x63,0x18,0x83,0xc0,0x3c,0xb6,0x41,0x08,0x4e,0x33,0x58,0x47, -0x31,0x31,0x4f,0x5d,0x1f,0xc3,0x22,0x38,0xcd,0xe0,0x0b,0x4e,0x33,0xe1,0xbc,0xa5, -0x18,0x82,0xd3,0x0c,0x77,0x6e,0x20,0xc5,0xc4,0x3c,0xb5,0x4e,0x3a,0x40,0xc6,0xde, -0xdc,0xc8,0x5c,0x7e,0x64,0x70,0x2c,0xa4,0x98,0x98,0xa7,0xee,0xed,0x82,0x10,0x9c, -0xa6,0xba,0x81,0x44,0x70,0x9a,0xc1,0x17,0x9c,0x66,0x32,0x93,0x42,0x60,0x1e,0x7b, -0xb7,0x98,0x62,0x62,0x9e,0xbc,0x36,0x16,0x43,0x70,0x9a,0x0a,0xa7,0x6d,0xa4,0x98, -0x98,0xc7,0xbe,0x8d,0xa4,0x98,0x98,0xc7,0xce,0x0d,0xc6,0x10,0x9c,0x66,0xc0,0x7b, -0x12,0x02,0x32,0xf6,0xe6,0x46,0xe6,0xa2,0x33,0x13,0x73,0x06,0x8b,0xe0,0x34,0x83, -0x2f,0x38,0xcd,0x64,0xd3,0xe6,0x61,0x08,0x4e,0x53,0xd5,0xf6,0x01,0x14,0x44,0x33, -0x45,0x18,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,0x4a,0x00,0x00,0x00, -0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x07,0x00,0x00,0x00,0x24,0xca,0x60,0x04, -0xa0,0x04,0x8a,0x80,0xc2,0x0c,0x00,0xb9,0x61,0x0c,0x04,0x10,0x1e,0xe1,0x19,0xc6, -0x40,0x02,0xe1,0x11,0x1e,0x00,0x00,0x00,0x63,0x08,0xcd,0x63,0x15,0xc1,0x31,0x84, -0x06,0xb2,0x8b,0xe0,0x18,0x42,0x13,0x59,0x46,0x70,0x0c,0xa1,0x71,0x6c,0x23,0x38, -0x16,0x02,0x04,0xc7,0x64,0x61,0x1a,0x37,0x16,0x01,0x04,0x48,0x35,0xc7,0x20,0x79, -0xcf,0x58,0x04,0x10,0x20,0xd5,0x1c,0xc3,0x07,0x06,0xd0,0x58,0x04,0x10,0x20,0xd5, -0x1c,0x43,0x18,0x88,0x41,0x34,0x16,0x01,0x04,0x48,0x35,0xc7,0x30,0x06,0x64,0xe0, -0x98,0x47,0xd0,0xc0,0x80,0xa0,0x89,0x01,0x41,0x23,0x03,0x82,0x63,0x21,0x40,0x70, -0x50,0x66,0x70,0x06,0x68,0x90,0x06,0x58,0x06,0xe1,0x40,0x00,0x25,0x00,0x00,0x00, -0x56,0x52,0x4c,0xcc,0x73,0xd3,0x56,0x41,0x4c,0xcc,0x53,0xdb,0x05,0x31,0x31,0xcf, -0x6d,0x19,0xc4,0xc4,0x3c,0xba,0x6d,0x10,0x13,0xf3,0xf4,0xd6,0x41,0x08,0xc0,0xb2, -0x18,0x46,0x21,0x38,0x4d,0x85,0x9b,0x46,0x21,0x38,0x4d,0xb5,0x9b,0x8a,0x21,0x00, -0xcb,0x82,0xdf,0x66,0x62,0x08,0x4e,0x53,0xdd,0xb7,0x9d,0x18,0x82,0xd3,0x54,0xb7, -0x6e,0x28,0x86,0xe0,0x34,0xd5,0xdd,0xdb,0x47,0x31,0x31,0x4f,0x9d,0x9b,0x87,0x21, -0x00,0xcb,0x52,0xdf,0x06,0x62,0x08,0xc0,0xb2,0xd4,0xbc,0x59,0x10,0x82,0xd3,0x54, -0x96,0x62,0x08,0x4e,0x53,0xe1,0xb6,0x85,0x14,0x13,0xf3,0xd8,0xb4,0x8d,0x14,0x13, -0xf3,0xd8,0xb9,0x89,0x18,0x02,0xb0,0x2c,0xf6,0x6d,0x24,0x86,0x00,0x2c,0x8b,0xcd, -0x1b,0x87,0x21,0x38,0x4d,0x55,0xd3,0xd6,0x30,0x54,0xc0,0x72,0x00,0x05,0xd1,0x4c, -0x11,0x06,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,0x19,0x00,0x00,0x00, -0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00,0x24,0x4a,0x60,0x04, -0x80,0xc2,0x0c,0x00,0x00,0x00,0x00,0x00,0x63,0x08,0xcd,0x33,0x16,0x01,0x04,0x48, -0x34,0xc7,0x00,0x49,0xcf,0x58,0x04,0x10,0x28,0xd1,0x1c,0xc3,0x44,0x39,0x58,0x85, -0x03,0x01,0x00,0x00,0x0a,0x00,0x00,0x00,0x16,0x41,0x4c,0xcc,0x63,0xdb,0x04,0x31, -0x31,0x4f,0x6e,0x0d,0x43,0x05,0x2c,0x07,0x50,0x10,0xcd,0x14,0x61,0x56,0x41,0x4c, -0xcc,0xd3,0x1b,0x45,0x21,0x00,0xcb,0xb2,0x9b,0x04,0x21,0x00,0xcb,0x02,0x00,0x00, -0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,0x1b,0x00,0x00,0x00,0x13,0x04,0x41,0x2c, -0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00,0x24,0xca,0x60,0x04,0xa0,0x04,0x8a,0x80, -0xc2,0x0c,0x00,0x00,0x63,0x08,0xcd,0x33,0x16,0x01,0x04,0xca,0x34,0xc7,0x20,0x51, -0xcf,0x1c,0x43,0x45,0x41,0x73,0x0c,0x16,0x15,0xcd,0x31,0x5c,0x94,0x83,0x58,0x38, -0x10,0x00,0x00,0x00,0x0b,0x00,0x00,0x00,0x86,0x51,0x4c,0xcc,0x53,0xe7,0x76,0x51, -0x4c,0xcc,0x53,0xdb,0x36,0x41,0x4c,0xcc,0x63,0x5b,0x05,0x31,0x31,0x8f,0x6e,0x0d, -0x43,0x05,0x2c,0x66,0x41,0x4c,0xcc,0xd3,0x1f,0x40,0x41,0x34,0x53,0x84,0x19,0x05, -0x21,0x00,0xcb,0x02,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,0x2f,0x00,0x00,0x00, -0x13,0x04,0x45,0x2c,0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00,0x24,0xca,0xa0,0x04, -0x46,0x00,0x8a,0x80,0xc0,0x08,0x00,0x00,0x63,0x08,0x0d,0x34,0xc9,0x30,0x49,0xc4, -0x2c,0x03,0x11,0x50,0x63,0x08,0xcd,0x33,0xc9,0x50,0x49,0xc4,0x2c,0x03,0x21,0x58, -0x63,0x08,0x4d,0x34,0xc9,0x70,0x49,0xc4,0x2c,0x03,0x31,0x60,0x63,0x08,0x8d,0x33, -0xc9,0x90,0x49,0x84,0x69,0x22,0x70,0xc3,0x27,0x1c,0x08,0x00,0x1a,0x00,0x00,0x00, -0x96,0x51,0x4c,0xcc,0x53,0xdf,0x66,0x41,0x08,0xcc,0x83,0xdb,0x04,0x31,0x31,0x4f, -0x6d,0x15,0xc4,0xc4,0x3c,0xb7,0x61,0x10,0x02,0xf3,0xf0,0x47,0x20,0xb9,0x0d,0x52, -0x20,0xc4,0xb0,0x12,0x13,0x15,0x35,0xdb,0xe2,0x08,0x8a,0x5d,0x10,0x13,0xf3,0xec, -0x37,0x90,0x2c,0x4e,0xf4,0x47,0x87,0x54,0xd7,0x17,0x70,0x2c,0x4e,0xf4,0x47,0x87, -0x74,0x02,0xc8,0xe2,0x44,0x7f,0x74,0x48,0xb9,0x69,0x14,0x02,0xf3,0xd4,0xb8,0x6d, -0x18,0x11,0x31,0x55,0xc0,0x62,0x0d,0x43,0x05,0x2c,0x07,0x50,0x10,0xcd,0x14,0x61, -0x46,0x31,0x08,0xcc,0x03,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x71,0x20,0x00,0x00, -0x12,0x00,0x00,0x00,0x66,0x40,0x54,0x82,0x23,0x59,0xc2,0x20,0x09,0x92,0x1d,0x18, -0x4f,0x84,0x34,0x53,0x61,0x03,0xc4,0xe3,0x58,0x85,0x05,0x14,0xbe,0x34,0x45,0xb5, -0x21,0x10,0x82,0x23,0x15,0x46,0x30,0x2c,0xc8,0x64,0x02,0x06,0xf0,0x3c,0x91,0x73, -0x19,0x00,0xe1,0x4b,0x53,0x64,0x0a,0x84,0x84,0x34,0x85,0x31,0x10,0x0a,0xb2,0x3c, -0x56,0x30,0x08,0xcc,0x63,0x0b,0x44,0x25,0x21,0x0d,0x00,0x00,0x00,0x00,0x00,0x00}; +0xe7,0x60,0x87,0x52,0x38,0x10,0x00,0x00,0x10,0x00,0x00,0x00,0x27,0x50,0x20,0x05, +0xd1,0x0c,0x17,0x60,0x20,0xc5,0x74,0x10,0x8d,0x65,0x14,0x13,0xf3,0xd4,0xb4,0x6d, +0x14,0x13,0xf3,0xd4,0xb8,0x69,0x14,0x13,0xf3,0xd4,0xb6,0x75,0x14,0x13,0xf3,0xd4, +0xba,0x35,0x0c,0x13,0xf3,0xd8,0x05,0x31,0x31,0x8f,0x6e,0x1c,0x84,0x00,0x2c,0xcb, +0x01,0x14,0x44,0x33,0x45,0x98,0x61,0x0c,0x02,0xf3,0x00,0x00,0x00,0x00,0x00,0x00, +0x61,0x20,0x00,0x00,0x81,0x00,0x00,0x00,0x13,0x04,0x4d,0x2c,0x10,0x00,0x00,0x00, +0x04,0x00,0x00,0x00,0x24,0xca,0x60,0x04,0xa0,0x04,0x8a,0x80,0xc2,0x0c,0x00,0x91, +0x11,0x00,0x00,0x00,0x63,0x08,0x4d,0x64,0x16,0xc1,0xe1,0x86,0xab,0x22,0x66,0x19, +0x02,0x01,0x1b,0x43,0x70,0xa2,0x59,0x82,0x61,0x0c,0xe1,0x89,0x66,0x09,0x86,0x81, +0x0a,0x20,0x0b,0x34,0x61,0x8e,0x81,0xda,0xa2,0x31,0x84,0x46,0xb2,0x8e,0xe0,0x70, +0x83,0x57,0x11,0xb3,0x0c,0x44,0xf1,0x8d,0x21,0x38,0xd2,0x2c,0x81,0x31,0x86,0xf0, +0x48,0xb3,0x04,0xc6,0x40,0x05,0x00,0x06,0x44,0x18,0x14,0x73,0x0c,0x9c,0x18,0x48, +0x63,0x08,0xcd,0x64,0x64,0x40,0x70,0xb8,0xa1,0x0c,0x2a,0x62,0x96,0xe1,0x40,0xcc, +0x60,0x0c,0xc1,0x99,0x66,0x09,0x92,0x31,0x84,0x67,0x9a,0x25,0x48,0x06,0x2a,0x80, +0x33,0x38,0xd0,0x00,0x99,0x63,0x18,0x83,0x34,0x98,0xc6,0x10,0x1a,0xc8,0xd6,0x80, +0xe0,0x70,0x03,0x1b,0x54,0xc4,0x2c,0x83,0xb2,0xb4,0xc1,0x18,0x82,0x03,0xcd,0x12, +0x30,0x63,0x08,0x0f,0x34,0x4b,0xc0,0x0c,0x54,0x00,0x6e,0xa0,0xbc,0xc1,0x32,0xc7, +0xa0,0x06,0x70,0x00,0x61,0x1c,0x84,0x03,0x01,0x00,0x00,0x00,0x4e,0x00,0x00,0x00, +0x76,0x52,0x4c,0xcc,0x73,0xd3,0x24,0x05,0x64,0xec,0xcd,0x8d,0xcc,0xe5,0x87,0x46, +0xc6,0x50,0x8a,0x89,0x79,0xee,0xdb,0x54,0x8a,0x89,0x79,0xee,0xdd,0x1a,0x88,0x89, +0x79,0x68,0x73,0x20,0x26,0xe6,0xa9,0xed,0x81,0x98,0x98,0xc7,0x36,0x0b,0x62,0x62, +0x9e,0xdb,0x32,0x88,0x89,0x79,0x72,0xd3,0x20,0x26,0xe6,0xd9,0x8d,0x83,0x98,0x98, +0xa7,0xb7,0x95,0x62,0x62,0x9e,0xbb,0x27,0x2d,0x20,0x63,0x6f,0x6e,0x64,0x2e,0x3a, +0x34,0x35,0x56,0x62,0x08,0x4e,0x53,0xd9,0xba,0xb5,0x14,0x02,0xf3,0xe0,0xf5,0x25, +0x2c,0x82,0xd3,0x0c,0xbe,0xe0,0x34,0xd3,0x8d,0x9b,0x88,0x21,0x38,0xcd,0x60,0xd7, +0x24,0x01,0x63,0xec,0xcd,0x8d,0xcc,0x45,0x87,0x44,0x80,0x8c,0xbd,0xb9,0x91,0xb9, +0xfc,0xc4,0xd0,0x90,0x02,0x8c,0xb1,0x37,0x37,0x32,0x97,0x1f,0x73,0x29,0x26,0xe6, +0xc1,0x71,0x7b,0x29,0x26,0xe6,0xc1,0x77,0xfb,0x28,0x04,0xe6,0xa9,0x6f,0x52,0x01, +0x32,0xf6,0xe6,0x46,0xe6,0xa2,0x13,0x73,0x63,0x18,0x83,0xc0,0x3c,0xb6,0x41,0x08, +0x4e,0x33,0x58,0x47,0x31,0x31,0x4f,0x5d,0x1f,0xc3,0x22,0x38,0xcd,0xe0,0x0b,0x4e, +0x33,0xe1,0xbc,0xa5,0x18,0x82,0xd3,0x0c,0x77,0x6e,0x20,0xc5,0xc4,0x3c,0xb5,0x4e, +0x3a,0x40,0xc6,0xde,0xdc,0xc8,0x5c,0x7e,0x64,0x70,0x2c,0xa4,0x98,0x98,0xa7,0xee, +0x6f,0x20,0x11,0x9c,0x66,0xf0,0x05,0xa7,0x99,0xec,0x82,0x10,0x9c,0xa6,0x32,0x93, +0x42,0x60,0x1e,0x7b,0xb7,0x98,0x62,0x62,0x9e,0xbc,0x36,0x16,0x43,0x70,0x9a,0x0a, +0xa7,0x6d,0xa4,0x98,0x98,0xc7,0xbe,0x8d,0xa4,0x98,0x98,0xc7,0xce,0x0d,0xc6,0x10, +0x9c,0x66,0xc0,0x7b,0x12,0x02,0x32,0xf6,0xe6,0x46,0xe6,0xa2,0x33,0x13,0x73,0x06, +0x8b,0xe0,0x34,0x83,0x2f,0x38,0xcd,0x64,0xd3,0x07,0x50,0x10,0xcd,0x14,0x61,0xe6, +0x61,0x08,0x4e,0x53,0xd5,0x36,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00, +0x4a,0x00,0x00,0x00,0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x07,0x00,0x00,0x00, +0x24,0xca,0x60,0x04,0xa0,0x04,0x8a,0x80,0xc2,0x0c,0x00,0xb9,0x61,0x0c,0x04,0x10, +0x1e,0xe1,0x19,0xc6,0x40,0x02,0xe1,0x11,0x1e,0x00,0x00,0x00,0x63,0x08,0xcd,0x63, +0x15,0xc1,0x31,0x84,0x06,0xb2,0x8b,0xe0,0x18,0x42,0x13,0x59,0x46,0x70,0x0c,0xa1, +0x71,0x6c,0x23,0x38,0x16,0x02,0x04,0xc7,0x64,0x61,0x1a,0x37,0x16,0x01,0x04,0x48, +0x35,0xc7,0x20,0x79,0xcf,0x58,0x04,0x10,0x20,0xd5,0x1c,0xc3,0x07,0x06,0xd0,0x58, +0x04,0x10,0x20,0xd5,0x1c,0x43,0x18,0x88,0x41,0x34,0x16,0x01,0x04,0x48,0x35,0xc7, +0x30,0x06,0x64,0xe0,0x98,0x47,0xd0,0xc0,0x80,0xa0,0x89,0x01,0x41,0x23,0x03,0x82, +0x63,0x21,0x40,0x70,0x50,0x66,0x70,0x06,0x68,0x90,0x06,0x58,0x06,0xe1,0x40,0x00, +0x25,0x00,0x00,0x00,0x56,0x52,0x4c,0xcc,0x73,0xd3,0x56,0x41,0x4c,0xcc,0x53,0xdb, +0x05,0x31,0x31,0xcf,0x6d,0x19,0xc4,0xc4,0x3c,0xba,0x6d,0x10,0x13,0xf3,0xf4,0xd6, +0x41,0x08,0xc0,0xb2,0x18,0x46,0x21,0x38,0x4d,0x85,0x9b,0x46,0x21,0x38,0x4d,0xb5, +0x9b,0x8a,0x21,0x00,0xcb,0x82,0xdf,0x66,0x62,0x08,0x4e,0x53,0xdd,0xb7,0x9d,0x18, +0x82,0xd3,0x54,0xb7,0x6e,0x28,0x86,0xe0,0x34,0xd5,0xdd,0xdb,0x47,0x31,0x31,0x4f, +0x9d,0x9b,0x87,0x21,0x00,0xcb,0x52,0xdf,0x06,0x62,0x08,0xc0,0xb2,0xd4,0xbc,0x59, +0x10,0x82,0xd3,0x54,0x96,0x62,0x08,0x4e,0x53,0xe1,0xb6,0x85,0x14,0x13,0xf3,0xd8, +0xb4,0x8d,0x14,0x13,0xf3,0xd8,0xb9,0x89,0x18,0x02,0xb0,0x2c,0xf6,0x6d,0x24,0x86, +0x00,0x2c,0x8b,0xcd,0x1b,0x87,0x21,0x38,0x4d,0x55,0xd3,0xd6,0x30,0x54,0xc0,0x72, +0x00,0x05,0xd1,0x4c,0x11,0x06,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00, +0x19,0x00,0x00,0x00,0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00, +0x24,0x4a,0x60,0x04,0x80,0xc2,0x0c,0x00,0x00,0x00,0x00,0x00,0x63,0x08,0xcd,0x33, +0x16,0x01,0x04,0x48,0x34,0xc7,0x00,0x49,0xcf,0x58,0x04,0x10,0x28,0xd1,0x1c,0xc3, +0x44,0x39,0x58,0x85,0x03,0x01,0x00,0x00,0x0a,0x00,0x00,0x00,0x26,0x41,0x08,0xc0, +0xb2,0x18,0x45,0x21,0x00,0xcb,0xb2,0x5b,0x04,0x31,0x31,0x8f,0x6d,0x13,0xc4,0xc4, +0x3c,0xb9,0x35,0x0c,0x15,0xb0,0x58,0x05,0x31,0x31,0x4f,0x7f,0x00,0x05,0xd1,0x4c, +0x11,0x06,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,0x1b,0x00,0x00,0x00, +0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00,0x24,0xca,0x60,0x04, +0xa0,0x04,0x8a,0x80,0xc2,0x0c,0x00,0x00,0x63,0x08,0xcd,0x33,0x16,0x01,0x04,0xca, +0x34,0xc7,0x20,0x51,0xcf,0x1c,0x43,0x45,0x41,0x73,0x0c,0x16,0x15,0xcd,0x31,0x5c, +0x94,0x83,0x58,0x38,0x10,0x00,0x00,0x00,0x0b,0x00,0x00,0x00,0x76,0x51,0x4c,0xcc, +0x53,0xdb,0x86,0x51,0x4c,0xcc,0x53,0xe7,0x36,0x41,0x4c,0xcc,0x63,0x5b,0x05,0x31, +0x31,0x8f,0x6e,0x16,0xc4,0xc4,0x3c,0xbd,0x51,0x10,0x02,0xb0,0x2c,0xd6,0x30,0x54, +0xc0,0x72,0x00,0x05,0xd1,0x4c,0x11,0x06,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00, +0x2c,0x00,0x00,0x00,0x13,0x04,0x45,0x2c,0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00, +0x24,0xca,0xa0,0x04,0x46,0x00,0x8a,0x80,0xc0,0x08,0x00,0x00,0x63,0x08,0x0d,0x34, +0xdc,0x30,0x49,0xc4,0x2c,0x03,0x11,0x50,0x63,0x08,0xcd,0x33,0xdc,0x50,0x49,0xc4, +0x2c,0x03,0x21,0x58,0x63,0x08,0x4d,0x34,0xdc,0x70,0x49,0xc4,0x2c,0x03,0x31,0x60, +0x63,0x08,0x8d,0x33,0xdc,0x90,0x49,0x84,0x69,0x22,0x70,0xc3,0x27,0x1c,0x08,0x00, +0x17,0x00,0x00,0x00,0x96,0x51,0x4c,0xcc,0x53,0xdf,0x66,0x41,0x08,0xcc,0x83,0xdb, +0x04,0x31,0x31,0x4f,0x6d,0x15,0xc4,0xc4,0x3c,0xb7,0x61,0x10,0x02,0xf3,0xf0,0x76, +0x41,0x4c,0xcc,0xb3,0x1f,0x81,0x11,0x11,0x13,0x15,0x35,0x37,0x90,0x2c,0x4e,0xf4, +0x47,0x87,0x54,0xd7,0x17,0x70,0x2c,0x4e,0xf4,0x47,0x87,0x74,0x02,0xc8,0xe2,0x44, +0x7f,0x74,0x48,0xb9,0x69,0x14,0x02,0xf3,0xd4,0xb8,0x6d,0x18,0x11,0x31,0x55,0xc0, +0x62,0x0d,0x43,0x05,0x2c,0x07,0x50,0x10,0xcd,0x14,0x61,0x46,0x31,0x08,0xcc,0x03, +0x00,0x00,0x00,0x00,0x71,0x20,0x00,0x00,0x12,0x00,0x00,0x00,0x66,0x40,0x54,0x82, +0x23,0x19,0xc3,0xa0,0x20,0x8b,0x1d,0x18,0x4f,0x84,0x34,0x53,0x61,0x03,0xc4,0xe3, +0x58,0x85,0x05,0x14,0xbe,0x34,0x45,0xb5,0x21,0x10,0x82,0x23,0x15,0x46,0x30,0x2c, +0xc8,0x64,0x02,0x06,0xf0,0x3c,0x91,0x73,0x19,0x00,0xe1,0x4b,0x53,0x64,0x0a,0x84, +0x84,0x34,0x85,0x25,0x0c,0x92,0x20,0x59,0xc1,0x20,0x30,0x8f,0x2d,0x10,0x95,0x84, +0x34,0x00,0x00,0x00,0x00,0x00,0x00,0x00}; diff --git a/src/gallium/auxiliary/gallivm/instructions.cpp b/src/gallium/auxiliary/gallivm/instructions.cpp index 3eaf9aacf6..599975d5ad 100644 --- a/src/gallium/auxiliary/gallivm/instructions.cpp +++ b/src/gallium/auxiliary/gallivm/instructions.cpp @@ -923,7 +923,7 @@ llvm::Value * Instructions::callCeil(llvm::Value *val) // predeclare the intrinsic std::vector ceilArgs; ceilArgs.push_back(Type::FloatTy); - PAListPtr ceilPal; + AttrListPtr ceilPal; FunctionType* ceilType = FunctionType::get( /*Result=*/Type::FloatTy, /*Params=*/ceilArgs, @@ -933,7 +933,7 @@ llvm::Value * Instructions::callCeil(llvm::Value *val) /*Linkage=*/GlobalValue::ExternalLinkage, /*Name=*/"ceilf", m_mod); m_llvmCeil->setCallingConv(CallingConv::C); - m_llvmCeil->setParamAttrs(ceilPal); + m_llvmCeil->setAttributes(ceilPal); } CallInst *call = m_builder.CreateCall(m_llvmCeil, val, name("ceilf")); @@ -948,7 +948,7 @@ llvm::Value *Instructions::callFAbs(llvm::Value *val) // predeclare the intrinsic std::vector fabsArgs; fabsArgs.push_back(Type::FloatTy); - PAListPtr fabsPal; + AttrListPtr fabsPal; FunctionType* fabsType = FunctionType::get( /*Result=*/Type::FloatTy, /*Params=*/fabsArgs, @@ -958,7 +958,7 @@ llvm::Value *Instructions::callFAbs(llvm::Value *val) /*Linkage=*/GlobalValue::ExternalLinkage, /*Name=*/"fabs", m_mod); m_llvmFAbs->setCallingConv(CallingConv::C); - m_llvmFAbs->setParamAttrs(fabsPal); + m_llvmFAbs->setAttributes(fabsPal); } CallInst *call = m_builder.CreateCall(m_llvmFAbs, val, name("fabs")); @@ -973,7 +973,7 @@ llvm::Value * Instructions::callFExp(llvm::Value *val) // predeclare the intrinsic std::vector fexpArgs; fexpArgs.push_back(Type::FloatTy); - PAListPtr fexpPal; + AttrListPtr fexpPal; FunctionType* fexpType = FunctionType::get( /*Result=*/Type::FloatTy, /*Params=*/fexpArgs, @@ -983,7 +983,7 @@ llvm::Value * Instructions::callFExp(llvm::Value *val) /*Linkage=*/GlobalValue::ExternalLinkage, /*Name=*/"expf", m_mod); m_llvmFexp->setCallingConv(CallingConv::C); - m_llvmFexp->setParamAttrs(fexpPal); + m_llvmFexp->setAttributes(fexpPal); } CallInst *call = m_builder.CreateCall(m_llvmFexp, val, name("expf")); @@ -998,7 +998,7 @@ llvm::Value * Instructions::callFLog(llvm::Value *val) // predeclare the intrinsic std::vector flogArgs; flogArgs.push_back(Type::FloatTy); - PAListPtr flogPal; + AttrListPtr flogPal; FunctionType* flogType = FunctionType::get( /*Result=*/Type::FloatTy, /*Params=*/flogArgs, @@ -1008,7 +1008,7 @@ llvm::Value * Instructions::callFLog(llvm::Value *val) /*Linkage=*/GlobalValue::ExternalLinkage, /*Name=*/"logf", m_mod); m_llvmFlog->setCallingConv(CallingConv::C); - m_llvmFlog->setParamAttrs(flogPal); + m_llvmFlog->setAttributes(flogPal); } CallInst *call = m_builder.CreateCall(m_llvmFlog, val, name("logf")); @@ -1023,7 +1023,7 @@ llvm::Value * Instructions::callFloor(llvm::Value *val) // predeclare the intrinsic std::vector floorArgs; floorArgs.push_back(Type::FloatTy); - PAListPtr floorPal; + AttrListPtr floorPal; FunctionType* floorType = FunctionType::get( /*Result=*/Type::FloatTy, /*Params=*/floorArgs, @@ -1033,7 +1033,7 @@ llvm::Value * Instructions::callFloor(llvm::Value *val) /*Linkage=*/GlobalValue::ExternalLinkage, /*Name=*/"floorf", m_mod); m_llvmFloor->setCallingConv(CallingConv::C); - m_llvmFloor->setParamAttrs(floorPal); + m_llvmFloor->setAttributes(floorPal); } CallInst *call = m_builder.CreateCall(m_llvmFloor, val, name("floorf")); @@ -1048,7 +1048,7 @@ llvm::Value *Instructions::callFSqrt(llvm::Value *val) // predeclare the intrinsic std::vector fsqrtArgs; fsqrtArgs.push_back(Type::FloatTy); - PAListPtr fsqrtPal; + AttrListPtr fsqrtPal; FunctionType* fsqrtType = FunctionType::get( /*Result=*/Type::FloatTy, /*Params=*/fsqrtArgs, @@ -1058,7 +1058,7 @@ llvm::Value *Instructions::callFSqrt(llvm::Value *val) /*Linkage=*/GlobalValue::ExternalLinkage, /*Name=*/"llvm.sqrt.f32", m_mod); m_llvmFSqrt->setCallingConv(CallingConv::C); - m_llvmFSqrt->setParamAttrs(fsqrtPal); + m_llvmFSqrt->setAttributes(fsqrtPal); } CallInst *call = m_builder.CreateCall(m_llvmFSqrt, val, name("sqrt")); @@ -1074,7 +1074,7 @@ llvm::Value * Instructions::callPow(llvm::Value *val1, llvm::Value *val2) std::vector powArgs; powArgs.push_back(Type::FloatTy); powArgs.push_back(Type::FloatTy); - PAListPtr powPal; + AttrListPtr powPal; FunctionType* powType = FunctionType::get( /*Result=*/Type::FloatTy, /*Params=*/powArgs, @@ -1084,7 +1084,7 @@ llvm::Value * Instructions::callPow(llvm::Value *val1, llvm::Value *val2) /*Linkage=*/GlobalValue::ExternalLinkage, /*Name=*/"llvm.pow.f32", m_mod); m_llvmPow->setCallingConv(CallingConv::C); - m_llvmPow->setParamAttrs(powPal); + m_llvmPow->setAttributes(powPal); } std::vector params; params.push_back(val1); @@ -1126,7 +1126,7 @@ llvm::Value * Instructions::constVector(float x, float y, float z, float w) llvm::Function * Instructions::declarePrintf() { std::vector args; - PAListPtr params; + AttrListPtr params; FunctionType* funcTy = FunctionType::get( /*Result=*/IntegerType::get(32), /*Params=*/args, @@ -1136,7 +1136,7 @@ llvm::Function * Instructions::declarePrintf() /*Linkage=*/GlobalValue::ExternalLinkage, /*Name=*/"printf", m_mod); func_printf->setCallingConv(CallingConv::C); - func_printf->setParamAttrs(params); + func_printf->setAttributes(params); return func_printf; } @@ -1148,7 +1148,7 @@ llvm::Function * Instructions::declareFunc(int label) args.push_back(vecPtr); args.push_back(vecPtr); args.push_back(vecPtr); - PAListPtr params; + AttrListPtr params; FunctionType *funcType = FunctionType::get( /*Result=*/Type::VoidTy, /*Params=*/args, @@ -1159,7 +1159,7 @@ llvm::Function * Instructions::declareFunc(int label) /*Linkage=*/GlobalValue::ExternalLinkage, /*Name=*/name.c_str(), m_mod); func->setCallingConv(CallingConv::C); - func->setParamAttrs(params); + func->setAttributes(params); return func; } diff --git a/src/gallium/auxiliary/gallivm/instructions.h b/src/gallium/auxiliary/gallivm/instructions.h index c3b28e9746..e18571251e 100644 --- a/src/gallium/auxiliary/gallivm/instructions.h +++ b/src/gallium/auxiliary/gallivm/instructions.h @@ -146,7 +146,7 @@ private: llvm::Module *m_mod; llvm::Function *m_func; char m_name[32]; - llvm::IRBuilder m_builder; + llvm::IRBuilder<> m_builder; int m_idx; llvm::VectorType *m_floatVecType; diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.cpp b/src/gallium/auxiliary/gallivm/instructionssoa.cpp index 9a3ed9f538..5863f37095 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.cpp +++ b/src/gallium/auxiliary/gallivm/instructionssoa.cpp @@ -259,7 +259,7 @@ void InstructionsSoa::createBuiltins() { MemoryBuffer *buffer = MemoryBuffer::getMemBuffer( (const char*)&soabuiltins_data[0], - (const char*)&soabuiltins_data[Elements(soabuiltins_data)-1]); + (const char*)&soabuiltins_data[Elements(soabuiltins_data)]); m_builtins = ParseBitcodeFile(buffer); std::cout<<"Builtins created at "<getFunctionType(), GlobalValue::ExternalLinkage, originalFunc->getName(), currentModule()); func->setCallingConv(CallingConv::C); - const PAListPtr pal; - func->setParamAttrs(pal); + const AttrListPtr pal; + func->setAttributes(pal); currentModule()->dump(); } else { DenseMap val; diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.h b/src/gallium/auxiliary/gallivm/instructionssoa.h index 3e20b652dd..20cab3b3bb 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.h +++ b/src/gallium/auxiliary/gallivm/instructionssoa.h @@ -96,7 +96,7 @@ private: const std::vector in3); void injectFunction(llvm::Function *originalFunc, int op = TGSI_OPCODE_LAST); private: - llvm::IRBuilder m_builder; + llvm::IRBuilder<> m_builder; StorageSoa *m_storage; std::map m_functionsMap; -- cgit v1.2.3 From a6ff215777da2181d7099284f2da28eff78273a9 Mon Sep 17 00:00:00 2001 From: Stephane Marchesin Date: Wed, 1 Oct 2008 00:00:58 +0200 Subject: Gallivm: add slt. glxgears should be running, except it isn't. --- src/gallium/auxiliary/gallivm/instructionssoa.cpp | 9 ++ src/gallium/auxiliary/gallivm/instructionssoa.h | 2 + src/gallium/auxiliary/gallivm/soabuiltins.c | 155 +++++++++++++--------- src/gallium/auxiliary/gallivm/tgsitollvm.cpp | 1 + 4 files changed, 101 insertions(+), 66 deletions(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.cpp b/src/gallium/auxiliary/gallivm/instructionssoa.cpp index 5863f37095..a658072551 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.cpp +++ b/src/gallium/auxiliary/gallivm/instructionssoa.cpp @@ -181,6 +181,7 @@ void InstructionsSoa::createFunctionMap() m_functionsMap[TGSI_OPCODE_POWER] = "pow"; m_functionsMap[TGSI_OPCODE_LIT] = "lit"; m_functionsMap[TGSI_OPCODE_RSQ] = "rsq"; + m_functionsMap[TGSI_OPCODE_SLT] = "slt"; } void InstructionsSoa::createDependencies() @@ -280,6 +281,14 @@ std::vector InstructionsSoa::dp3(const std::vector i return callBuiltin(func, in1, in2); } + +std::vector InstructionsSoa::slt(const std::vector in1, + const std::vector in2) +{ + llvm::Function *func = function(TGSI_OPCODE_SLT); + return callBuiltin(func, in1, in2); +} + llvm::Value * InstructionsSoa::allocaTemp() { VectorType *vector = VectorType::get(Type::FloatTy, 4); diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.h b/src/gallium/auxiliary/gallivm/instructionssoa.h index 20cab3b3bb..3817fdc904 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.h +++ b/src/gallium/auxiliary/gallivm/instructionssoa.h @@ -69,6 +69,8 @@ public: std::vector pow(const std::vector in1, const std::vector in2); std::vector rsq(const std::vector in1); + std::vector slt(const std::vector in1, + const std::vector in2); std::vector sub(const std::vector in1, const std::vector in2); void end(); diff --git a/src/gallium/auxiliary/gallivm/soabuiltins.c b/src/gallium/auxiliary/gallivm/soabuiltins.c index 78f84510e2..cb85e1734e 100644 --- a/src/gallium/auxiliary/gallivm/soabuiltins.c +++ b/src/gallium/auxiliary/gallivm/soabuiltins.c @@ -36,6 +36,8 @@ typedef __attribute__(( ext_vector_type(4) )) float float4; extern float fabsf(float val); +/* helpers */ + float4 absvec(float4 vec) { float4 res; @@ -47,6 +49,58 @@ float4 absvec(float4 vec) return res; } +float4 maxvec(float4 a, float4 b) +{ + return (float4){(a.x > b.x) ? a.x : b.x, + (a.y > b.y) ? a.y : b.y, + (a.z > b.z) ? a.z : b.z, + (a.w > b.w) ? a.w : b.w}; +} + +float4 minvec(float4 a, float4 b) +{ + return (float4){(a.x < b.x) ? a.x : b.x, + (a.y < b.y) ? a.y : b.y, + (a.z < b.z) ? a.z : b.z, + (a.w < b.w) ? a.w : b.w}; +} + +extern float powf(float num, float p); +extern float sqrtf(float x); + +float4 powvec(float4 vec, float4 q) +{ + float4 p; + p.x = powf(vec.x, q.x); + p.y = powf(vec.y, q.y); + p.z = powf(vec.z, q.z); + p.w = powf(vec.w, q.w); + return p; +} + +float4 sqrtvec(float4 vec) +{ + float4 p; + p.x = sqrtf(vec.x); + p.y = sqrtf(vec.y); + p.z = sqrtf(vec.z); + p.w = sqrtf(vec.w); + return p; +} + +float4 sltvec(float4 v1, float4 v2) +{ + float4 p; + p.x = (v1.x < v2.x) ? 1.0 : 0.0; + p.y = (v1.y < v2.y) ? 1.0 : 0.0; + p.z = (v1.z < v2.z) ? 1.0 : 0.0; + p.w = (v1.w < v2.w) ? 1.0 : 0.0; + return p; +} + + +/* instructions */ + void abs(float4 *res, float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w) { @@ -69,7 +123,6 @@ void dp3(float4 *res, res[3] = dot; } - void dp4(float4 *res, float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) @@ -83,35 +136,25 @@ void dp4(float4 *res, res[3] = dot; } -extern float powf(float num, float p); -extern float sqrtf(float x); - -float4 powvec(float4 vec, float4 q) -{ - float4 p; - p.x = powf(vec.x, q.x); - p.y = powf(vec.y, q.y); - p.z = powf(vec.z, q.z); - p.w = powf(vec.w, q.w); - return p; -} - -void pow(float4 *res, - float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, - float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) +void lit(float4 *res, + float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w) { - res[0] = powvec(tmp0x, tmp1x); - res[1] = res[0]; - res[2] = res[0]; - res[3] = res[0]; -} + const float4 zerovec = (float4) {0.0, 0.0, 0.0, 0.0}; + const float4 min128 = (float4) {-128.f, -128.f, -128.f, -128.f}; + const float4 plus128 = (float4) {128.f, 128.f, 128.f, 128.f}; -float4 minvec(float4 a, float4 b) -{ - return (float4){(a.x < b.x) ? a.x : b.x, - (a.y < b.y) ? a.y : b.y, - (a.z < b.z) ? a.z : b.z, - (a.w < b.w) ? a.w : b.w}; + res[0] = (float4){1.0, 1.0, 1.0, 1.0}; + if (tmp0x.x > 0) { + float4 tmpy = maxvec(tmp0y, zerovec); + float4 tmpw = minvec(tmp0w, plus128); + tmpw = maxvec(tmpw, min128); + res[1] = tmp0x; + res[2] = powvec(tmpy, tmpw); + } else { + res[1] = zerovec; + res[2] = zerovec; + } + res[3] = (float4){1.0, 1.0, 1.0, 1.0}; } void min(float4 *res, @@ -125,14 +168,6 @@ void min(float4 *res, } -float4 maxvec(float4 a, float4 b) -{ - return (float4){(a.x > b.x) ? a.x : b.x, - (a.y > b.y) ? a.y : b.y, - (a.z > b.z) ? a.z : b.z, - (a.w > b.w) ? a.w : b.w}; -} - void max(float4 *res, float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) @@ -143,37 +178,14 @@ void max(float4 *res, res[3] = maxvec(tmp0w, tmp1w); } - -void lit(float4 *res, - float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w) -{ - const float4 zerovec = (float4) {0.0, 0.0, 0.0, 0.0}; - const float4 min128 = (float4) {-128.f, -128.f, -128.f, -128.f}; - const float4 plus128 = (float4) {128.f, 128.f, 128.f, 128.f}; - - res[0] = (float4){1.0, 1.0, 1.0, 1.0}; - if (tmp0x.x > 0) { - float4 tmpy = maxvec(tmp0y, zerovec); - float4 tmpw = minvec(tmp0w, plus128); - tmpw = maxvec(tmpw, min128); - res[1] = tmp0x; - res[2] = powvec(tmpy, tmpw); - } else { - res[1] = zerovec; - res[2] = zerovec; - } - res[3] = (float4){1.0, 1.0, 1.0, 1.0}; -} - - -float4 sqrtvec(float4 vec) +void pow(float4 *res, + float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, + float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) { - float4 p; - p.x = sqrtf(vec.x); - p.y = sqrtf(vec.y); - p.z = sqrtf(vec.z); - p.w = sqrtf(vec.w); - return p; + res[0] = powvec(tmp0x, tmp1x); + res[1] = res[0]; + res[2] = res[0]; + res[3] = res[0]; } void rsq(float4 *res, @@ -185,3 +197,14 @@ void rsq(float4 *res, res[2] = onevec/sqrtvec(absvec(tmp0z)); res[3] = onevec/sqrtvec(absvec(tmp0w)); } + +void slt(float4 *res, + float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, + float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) +{ + res[0] = sltvec(tmp0x, tmp1x); + res[1] = sltvec(tmp0y, tmp1y); + res[2] = sltvec(tmp0z, tmp1z); + res[3] = sltvec(tmp0w, tmp1w); +} + diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp index fdfbb76c16..7292c0e366 100644 --- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp +++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp @@ -767,6 +767,7 @@ translate_instructionir(llvm::Module *module, } break; case TGSI_OPCODE_SLT: { + out = instr->slt(inputs[0], inputs[1]); } break; case TGSI_OPCODE_SGE: { -- cgit v1.2.3 From dd7e5a498066e4ebdb7ad40773de48e5bc993164 Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Wed, 1 Oct 2008 13:34:38 +0100 Subject: draw: add streamlined paths for fetching linear verts --- src/gallium/auxiliary/draw/draw_vs_aos.c | 44 +++++---- src/gallium/auxiliary/draw/draw_vs_aos.h | 19 ++-- src/gallium/auxiliary/draw/draw_vs_aos_io.c | 137 +++++++++++++++++++++------- 3 files changed, 134 insertions(+), 66 deletions(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c index a556477a76..4c794e0e23 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos.c @@ -92,9 +92,9 @@ struct x86_reg aos_get_x86( struct aos_compilation *cp, assert(which_reg == 1); offset = Offset(struct aos_machine, constants); break; - case X86_ATTRIBS: + case X86_BUFFERS: assert(which_reg == 0); - offset = Offset(struct aos_machine, attrib); + offset = Offset(struct aos_machine, buffer); break; default: assert(0); @@ -1939,6 +1939,8 @@ static boolean build_vertex_program( struct draw_vs_varient_aos_sse *varient, save_fpu_state( &cp ); set_fpu_round_nearest( &cp ); + aos_init_inputs( &cp, linear ); + /* Note address for loop jump */ label = x86_get_label(cp.func); @@ -2018,13 +2020,7 @@ static boolean build_vertex_program( struct draw_vs_varient_aos_sse *varient, /* Incr index */ - if (linear) { - x86_inc(cp.func, cp.idx_EBX); - } - else { - x86_lea(cp.func, cp.idx_EBX, x86_make_disp(cp.idx_EBX, 4)); - } - + aos_incr_inputs( &cp, linear ); } /* decr count, loop if not zero */ @@ -2065,14 +2061,10 @@ static void vaos_set_buffer( struct draw_vs_varient *varient, unsigned stride ) { struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient; - unsigned i; - for (i = 0; i < vaos->base.key.nr_inputs; i++) { - if (vaos->base.key.element[i].in.buffer == buf) { - vaos->attrib[i].input_ptr = ((char *)ptr + - vaos->base.key.element[i].in.offset); - vaos->attrib[i].input_stride = stride; - } + if (buf < vaos->nr_vb) { + vaos->buffer[buf].base_ptr = (char *)ptr; + vaos->buffer[buf].stride = stride; } } @@ -2089,7 +2081,7 @@ static void PIPE_CDECL vaos_run_elts( struct draw_vs_varient *varient, machine->internal[IMM_PSIZE][0] = vaos->draw->rasterizer->point_size; machine->constants = vaos->draw->vs.aligned_constants; machine->immediates = vaos->base.vs->immediates; - machine->attrib = vaos->attrib; + machine->buffer = vaos->buffer; vaos->gen_run_elts( machine, elts, @@ -2108,7 +2100,7 @@ static void PIPE_CDECL vaos_run_linear( struct draw_vs_varient *varient, machine->internal[IMM_PSIZE][0] = vaos->draw->rasterizer->point_size; machine->constants = vaos->draw->vs.aligned_constants; machine->immediates = vaos->base.vs->immediates; - machine->attrib = vaos->attrib; + machine->buffer = vaos->buffer; vaos->gen_run_linear( machine, start, @@ -2127,7 +2119,7 @@ static void vaos_destroy( struct draw_vs_varient *varient ) { struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient; - FREE( vaos->attrib ); + FREE( vaos->buffer ); x86_release_func( &vaos->func[0] ); x86_release_func( &vaos->func[1] ); @@ -2140,6 +2132,7 @@ static void vaos_destroy( struct draw_vs_varient *varient ) static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs, const struct draw_vs_varient_key *key ) { + unsigned i; struct draw_vs_varient_aos_sse *vaos = CALLOC_STRUCT(draw_vs_varient_aos_sse); if (!vaos) @@ -2154,10 +2147,15 @@ static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs, vaos->draw = vs->draw; - vaos->attrib = MALLOC( key->nr_inputs * sizeof(vaos->attrib[0]) ); - if (!vaos->attrib) + for (i = 0; i < key->nr_inputs; i++) + vaos->nr_vb = MAX2( vaos->nr_vb, key->element[i].in.buffer + 1 ); + + vaos->buffer = MALLOC( vaos->nr_vb * sizeof(vaos->buffer[0]) ); + if (!vaos->buffer) goto fail; + debug_printf("nr_vb: %d\n", vaos->nr_vb); + #if 0 tgsi_dump(vs->state.tokens, 0); #endif @@ -2179,8 +2177,8 @@ static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs, return &vaos->base; fail: - if (vaos && vaos->attrib) - FREE(vaos->attrib); + if (vaos && vaos->buffer) + FREE(vaos->buffer); if (vaos) x86_release_func( &vaos->func[0] ); diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.h b/src/gallium/auxiliary/draw/draw_vs_aos.h index 7fe6f79db0..306392e5d6 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos.h +++ b/src/gallium/auxiliary/draw/draw_vs_aos.h @@ -87,9 +87,10 @@ struct lit_info { #define MAX_SHINE_TAB 4 #define MAX_LIT_INFO 16 -struct aos_attrib { - const void *input_ptr; - unsigned input_stride; +struct aos_buffer { + const void *base_ptr; + unsigned stride; + void *ptr; /* updated per vertex */ }; @@ -123,7 +124,7 @@ struct aos_machine { const float (*immediates)[4]; /* points to shader data */ const float (*constants)[4]; /* points to draw data */ - const struct aos_attrib *attrib; /* points to ? */ + const struct aos_buffer *buffer; /* points to ? */ }; @@ -179,8 +180,9 @@ struct x86_reg aos_get_shader_reg( struct aos_compilation *cp, unsigned file, unsigned idx ); -boolean aos_fetch_inputs( struct aos_compilation *cp, - boolean linear ); +boolean aos_init_inputs( struct aos_compilation *cp, boolean linear ); +boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear ); +boolean aos_incr_inputs( struct aos_compilation *cp, boolean linear ); boolean aos_emit_outputs( struct aos_compilation *cp ); @@ -210,7 +212,7 @@ do { \ #define X86_NULL 0 #define X86_IMMEDIATES 1 #define X86_CONSTANTS 2 -#define X86_ATTRIBS 3 +#define X86_BUFFERS 3 struct x86_reg aos_get_x86( struct aos_compilation *cp, unsigned which_reg, @@ -232,7 +234,8 @@ struct draw_vs_varient_aos_sse { struct draw_vs_varient base; struct draw_context *draw; - struct aos_attrib *attrib; + struct aos_buffer *buffer; + unsigned nr_vb; vaos_run_linear_func gen_run_linear; vaos_run_elts_func gen_run_elts; diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_io.c b/src/gallium/auxiliary/draw/draw_vs_aos_io.c index 26297c74f8..8e08b9285f 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos_io.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos_io.c @@ -95,28 +95,6 @@ static void emit_load_R8G8B8A8_UNORM( struct aos_compilation *cp, -static void get_src_ptr( struct aos_compilation *cp, - struct x86_reg src, - struct x86_reg elt, - unsigned a ) -{ - struct x86_reg attrib = x86_make_disp(aos_get_x86( cp, 0, X86_ATTRIBS ), - a * sizeof(struct aos_attrib)); - - struct x86_reg input_ptr = x86_make_disp(attrib, - Offset(struct aos_attrib, input_ptr)); - - struct x86_reg input_stride = x86_make_disp(attrib, - Offset(struct aos_attrib, input_stride)); - - /* Calculate pointer to current attrib: - */ - x86_mov(cp->func, src, input_stride); - x86_imul(cp->func, src, elt); - x86_add(cp->func, src, input_ptr); -} - - /* Extended swizzles? Maybe later. */ static void emit_swizzle( struct aos_compilation *cp, @@ -128,22 +106,44 @@ static void emit_swizzle( struct aos_compilation *cp, } + +static boolean get_buffer_ptr( struct aos_compilation *cp, + unsigned buf_idx, + struct x86_reg elt, + struct x86_reg ptr) +{ + struct x86_reg buf = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ), + buf_idx * sizeof(struct aos_buffer)); + + struct x86_reg buf_base_ptr = x86_make_disp(buf, + Offset(struct aos_buffer, base_ptr)); + + struct x86_reg buf_stride = x86_make_disp(buf, + Offset(struct aos_buffer, stride)); + + /* Calculate pointer to current attrib: + */ + x86_mov(cp->func, ptr, buf_stride); + x86_imul(cp->func, ptr, elt); + x86_add(cp->func, ptr, buf_base_ptr); + + return TRUE; +} + + + + static boolean load_input( struct aos_compilation *cp, unsigned idx, - boolean linear ) + struct x86_reg bufptr ) { unsigned format = cp->vaos->base.key.element[idx].in.format; - struct x86_reg src = cp->tmp_EAX; + unsigned offset = cp->vaos->base.key.element[idx].in.offset; struct x86_reg dataXMM = aos_get_xmm_reg(cp); /* Figure out source pointer address: */ - get_src_ptr(cp, - src, - linear ? cp->idx_EBX : x86_deref(cp->idx_EBX), - idx); - - src = x86_deref(src); + struct x86_reg src = x86_make_disp(bufptr, offset); aos_adopt_xmm_reg( cp, dataXMM, @@ -179,20 +179,87 @@ static boolean load_input( struct aos_compilation *cp, return TRUE; } - -boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear ) +static boolean load_inputs( struct aos_compilation *cp, + unsigned buffer, + struct x86_reg ptr ) { unsigned i; - + for (i = 0; i < cp->vaos->base.key.nr_inputs; i++) { - if (!load_input( cp, i, linear )) + if (cp->vaos->base.key.element[i].in.buffer == buffer) { + + if (!load_input( cp, i, ptr )) + return FALSE; + + cp->insn_counter++; + } + } + + return TRUE; +} + +boolean aos_init_inputs( struct aos_compilation *cp, boolean linear ) +{ + if (linear && cp->vaos->nr_vb == 1) { + + struct x86_reg elt = cp->idx_EBX; + struct x86_reg ptr = cp->tmp_EAX; + + if (!get_buffer_ptr( cp, 0, elt, ptr )) return FALSE; - cp->insn_counter++; + + /* In the linear, single buffer case, keep the buffer pointer + * instead of the index number. + */ + x86_mov( cp->func, elt, ptr ); + } + + return TRUE; +} + +boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear ) +{ + if (linear && cp->vaos->nr_vb == 1) { + + load_inputs( cp, 0, cp->idx_EBX ); + + } + else { + struct x86_reg elt = linear ? cp->idx_EBX : x86_deref(cp->idx_EBX); + unsigned j; + + for (j = 0; j < cp->vaos->nr_vb; j++) { + struct x86_reg ptr = cp->tmp_EAX; + + if (!get_buffer_ptr( cp, j, elt, ptr )) + return FALSE; + + cp->insn_counter++; + + if (!load_inputs( cp, j, ptr )) + return FALSE; + } } return TRUE; } +boolean aos_incr_inputs( struct aos_compilation *cp, boolean linear ) +{ + if (linear && cp->vaos->nr_vb == 1) { + struct x86_reg stride = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ), + (0 * sizeof(struct aos_buffer) + + Offset(struct aos_buffer, stride))); + + x86_add(cp->func, cp->idx_EBX, stride); + } + else if (linear) { + x86_inc(cp->func, cp->idx_EBX); + } + else { + x86_lea(cp->func, cp->idx_EBX, x86_make_disp(cp->idx_EBX, 4)); + } +} -- cgit v1.2.3 From 102daee1b8971cf39235e220b9524bec1e4a7089 Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Thu, 2 Oct 2008 12:46:01 +0100 Subject: rtasm: add prefetch instructions --- src/gallium/auxiliary/rtasm/rtasm_x86sse.c | 26 ++++++++++++++++++++++++++ src/gallium/auxiliary/rtasm/rtasm_x86sse.h | 5 +++++ 2 files changed, 31 insertions(+) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c index 6d4c081e04..9085f4cc0e 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c +++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c @@ -629,6 +629,32 @@ void x86_and( struct x86_function *p, * SSE instructions */ +void sse_prefetchnta( struct x86_function *p, struct x86_reg ptr) +{ + DUMP_R( ptr ); + assert(ptr.mod != mod_REG); + emit_2ub(p, 0x0f, 0x18); + emit_modrm_noreg(p, 0, ptr); +} + +void sse_prefetch0( struct x86_function *p, struct x86_reg ptr) +{ + DUMP_R( ptr ); + assert(ptr.mod != mod_REG); + emit_2ub(p, 0x0f, 0x18); + emit_modrm_noreg(p, 1, ptr); +} + +void sse_prefetch1( struct x86_function *p, struct x86_reg ptr) +{ + DUMP_R( ptr ); + assert(ptr.mod != mod_REG); + emit_2ub(p, 0x0f, 0x18); + emit_modrm_noreg(p, 2, ptr); +} + + + void sse_movss( struct x86_function *p, struct x86_reg dst, diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.h b/src/gallium/auxiliary/rtasm/rtasm_x86sse.h index af94577aab..2d7715f965 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.h +++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.h @@ -184,6 +184,11 @@ void sse2_pshufd( struct x86_function *p, struct x86_reg dest, struct x86_reg ar void sse2_rcpps( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); void sse2_rcpss( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); + +void sse_prefetchnta( struct x86_function *p, struct x86_reg ptr); +void sse_prefetch0( struct x86_function *p, struct x86_reg ptr); +void sse_prefetch1( struct x86_function *p, struct x86_reg ptr); + void sse_addps( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); void sse_addss( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); void sse_cvtps2pi( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); -- cgit v1.2.3 From af9cfea9cc80411351f9879d8eeb525bf7b4ca50 Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Wed, 1 Oct 2008 18:40:01 +0100 Subject: draw: don't keep refetching constant inputs --- .../auxiliary/draw/draw_pt_fetch_shade_emit.c | 37 +++--- src/gallium/auxiliary/draw/draw_vs.h | 4 +- src/gallium/auxiliary/draw/draw_vs_aos.c | 26 ++++- src/gallium/auxiliary/draw/draw_vs_aos.h | 2 + src/gallium/auxiliary/draw/draw_vs_aos_io.c | 127 +++++++++++++++------ src/gallium/auxiliary/draw/draw_vs_varient.c | 10 +- 6 files changed, 144 insertions(+), 62 deletions(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c index 73fc70c1bc..a0e08dd10a 100644 --- a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c +++ b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c @@ -79,6 +79,7 @@ static void fse_prepare( struct draw_pt_middle_end *middle, unsigned num_vs_inputs = draw->vs.vertex_shader->info.num_inputs; const struct vertex_info *vinfo; unsigned i; + unsigned nr_vbs = 0; if (!draw->render->set_primitive( draw->render, @@ -102,7 +103,7 @@ static void fse_prepare( struct draw_pt_middle_end *middle, fse->key.viewport = !draw->identity_viewport; fse->key.clip = !draw->bypass_clipping; - fse->key.pad = 0; + fse->key.const_vbuffers = 0; memset(fse->key.element, 0, fse->key.nr_elements * sizeof(fse->key.element[0])); @@ -116,9 +117,16 @@ static void fse_prepare( struct draw_pt_middle_end *middle, */ fse->key.element[i].in.buffer = src->vertex_buffer_index; fse->key.element[i].in.offset = src->src_offset; + nr_vbs = MAX2(nr_vbs, src->vertex_buffer_index + 1); } + for (i = 0; i < 5 && i < nr_vbs; i++) { + if (draw->pt.vertex_buffer[i].pitch == 0) + fse->key.const_vbuffers |= (1<key.const_vbuffers); + { unsigned dst_offset = 0; @@ -162,13 +170,7 @@ static void fse_prepare( struct draw_pt_middle_end *middle, } } - - /* Would normally look up a vertex shader and peruse its list of - * varients somehow. We omitted that step and put all the - * hardcoded "shaders" into an array. We're just making the - * assumption that this happens to be a matching shader... ie - * you're running isosurf, aren't you? - */ + fse->active = draw_vs_lookup_varient( draw->vs.vertex_shader, &fse->key ); @@ -177,18 +179,17 @@ static void fse_prepare( struct draw_pt_middle_end *middle, return ; } + if (0) debug_printf("%s: found const_vbuffers: %x\n", __FUNCTION__, + fse->active->key.const_vbuffers); + /* Now set buffer pointers: */ - for (i = 0; i < num_vs_inputs; i++) { - unsigned buf = draw->pt.vertex_element[i].vertex_buffer_index; - - fse->active->set_input( fse->active, - i, - - ((const ubyte *) draw->pt.user.vbuffer[buf] + - draw->pt.vertex_buffer[buf].buffer_offset), - - draw->pt.vertex_buffer[buf].pitch ); + for (i = 0; i < draw->pt.nr_vertex_buffers; i++) { + fse->active->set_buffer( fse->active, + i, + ((const ubyte *) draw->pt.user.vbuffer[i] + + draw->pt.vertex_buffer[i].buffer_offset), + draw->pt.vertex_buffer[i].pitch ); } *max_vertices = (draw->render->max_vertex_buffer_bytes / diff --git a/src/gallium/auxiliary/draw/draw_vs.h b/src/gallium/auxiliary/draw/draw_vs.h index 45992d1986..68c24abad3 100644 --- a/src/gallium/auxiliary/draw/draw_vs.h +++ b/src/gallium/auxiliary/draw/draw_vs.h @@ -64,7 +64,7 @@ struct draw_vs_varient_key { unsigned nr_outputs:8; unsigned viewport:1; unsigned clip:1; - unsigned pad:5; + unsigned const_vbuffers:5; struct draw_varient_element element[PIPE_MAX_ATTRIBS]; }; @@ -76,7 +76,7 @@ struct draw_vs_varient { struct draw_vertex_shader *vs; - void (*set_input)( struct draw_vs_varient *, + void (*set_buffer)( struct draw_vs_varient *, unsigned i, const void *ptr, unsigned stride ); diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c index 4c794e0e23..87232865e2 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos.c @@ -196,6 +196,18 @@ static void spill( struct aos_compilation *cp, unsigned idx ) } +void aos_spill_all( struct aos_compilation *cp ) +{ + unsigned i; + + for (i = 0; i < 8; i++) { + if (cp->xmm[i].dirty) + spill(cp, i); + aos_release_xmm_reg(cp, i); + } +} + + static struct x86_reg get_xmm_writable( struct aos_compilation *cp, struct x86_reg reg ) { @@ -1941,6 +1953,9 @@ static boolean build_vertex_program( struct draw_vs_varient_aos_sse *varient, aos_init_inputs( &cp, linear ); + cp.x86_reg[0] = 0; + cp.x86_reg[1] = 0; + /* Note address for loop jump */ label = x86_get_label(cp.func); @@ -2066,6 +2081,8 @@ static void vaos_set_buffer( struct draw_vs_varient *varient, vaos->buffer[buf].base_ptr = (char *)ptr; vaos->buffer[buf].stride = stride; } + + if (0) debug_printf("%s %d/%d: %p %d\n", __FUNCTION__, buf, vaos->nr_vb, ptr, stride); } @@ -2078,6 +2095,8 @@ static void PIPE_CDECL vaos_run_elts( struct draw_vs_varient *varient, struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient; struct aos_machine *machine = vaos->draw->vs.aos_machine; + if (0) debug_printf("%s %d\n", __FUNCTION__, count); + machine->internal[IMM_PSIZE][0] = vaos->draw->rasterizer->point_size; machine->constants = vaos->draw->vs.aligned_constants; machine->immediates = vaos->base.vs->immediates; @@ -2097,6 +2116,9 @@ static void PIPE_CDECL vaos_run_linear( struct draw_vs_varient *varient, struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient; struct aos_machine *machine = vaos->draw->vs.aos_machine; + if (0) debug_printf("%s %d %d const: %x\n", __FUNCTION__, start, count, + vaos->base.key.const_vbuffers); + machine->internal[IMM_PSIZE][0] = vaos->draw->rasterizer->point_size; machine->constants = vaos->draw->vs.aligned_constants; machine->immediates = vaos->base.vs->immediates; @@ -2140,7 +2162,7 @@ static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs, vaos->base.key = *key; vaos->base.vs = vs; - vaos->base.set_input = vaos_set_buffer; + vaos->base.set_buffer = vaos_set_buffer; vaos->base.destroy = vaos_destroy; vaos->base.run_linear = vaos_run_linear; vaos->base.run_elts = vaos_run_elts; @@ -2154,7 +2176,7 @@ static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs, if (!vaos->buffer) goto fail; - debug_printf("nr_vb: %d\n", vaos->nr_vb); + debug_printf("nr_vb: %d const: %x\n", vaos->nr_vb, vaos->base.key.const_vbuffers); #if 0 tgsi_dump(vs->state.tokens, 0); diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.h b/src/gallium/auxiliary/draw/draw_vs_aos.h index 306392e5d6..264387517b 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos.h +++ b/src/gallium/auxiliary/draw/draw_vs_aos.h @@ -176,6 +176,8 @@ void aos_adopt_xmm_reg( struct aos_compilation *cp, unsigned idx, unsigned dirty ); +void aos_spill_all( struct aos_compilation *cp ); + struct x86_reg aos_get_shader_reg( struct aos_compilation *cp, unsigned file, unsigned idx ); diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_io.c b/src/gallium/auxiliary/draw/draw_vs_aos_io.c index 8e08b9285f..b0c51d7fa1 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos_io.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos_io.c @@ -108,29 +108,45 @@ static void emit_swizzle( struct aos_compilation *cp, static boolean get_buffer_ptr( struct aos_compilation *cp, - unsigned buf_idx, - struct x86_reg elt, - struct x86_reg ptr) + boolean linear, + unsigned buf_idx, + struct x86_reg elt, + struct x86_reg ptr) { struct x86_reg buf = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ), buf_idx * sizeof(struct aos_buffer)); - struct x86_reg buf_base_ptr = x86_make_disp(buf, - Offset(struct aos_buffer, base_ptr)); - struct x86_reg buf_stride = x86_make_disp(buf, Offset(struct aos_buffer, stride)); + if (linear) { + struct x86_reg buf_ptr = x86_make_disp(buf, + Offset(struct aos_buffer, ptr)); - /* Calculate pointer to current attrib: - */ - x86_mov(cp->func, ptr, buf_stride); - x86_imul(cp->func, ptr, elt); - x86_add(cp->func, ptr, buf_base_ptr); - return TRUE; -} + /* Calculate pointer to current attrib: + */ + x86_mov(cp->func, ptr, buf_ptr); + x86_mov(cp->func, elt, buf_stride); + x86_add(cp->func, elt, ptr); + sse_prefetchnta(cp->func, x86_deref(elt)); + x86_mov(cp->func, buf_ptr, elt); + } + else { + struct x86_reg buf_base_ptr = x86_make_disp(buf, + Offset(struct aos_buffer, base_ptr)); + + + /* Calculate pointer to current attrib: + */ + x86_mov(cp->func, ptr, buf_stride); + x86_imul(cp->func, ptr, elt); + x86_add(cp->func, ptr, buf_base_ptr); + } + cp->insn_counter++; + return TRUE; +} static boolean load_input( struct aos_compilation *cp, @@ -200,18 +216,57 @@ static boolean load_inputs( struct aos_compilation *cp, boolean aos_init_inputs( struct aos_compilation *cp, boolean linear ) { - if (linear && cp->vaos->nr_vb == 1) { + unsigned i; + for (i = 0; i < cp->vaos->nr_vb; i++) { + struct x86_reg buf = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ), + i * sizeof(struct aos_buffer)); - struct x86_reg elt = cp->idx_EBX; - struct x86_reg ptr = cp->tmp_EAX; + struct x86_reg buf_base_ptr = x86_make_disp(buf, + Offset(struct aos_buffer, base_ptr)); - if (!get_buffer_ptr( cp, 0, elt, ptr )) - return FALSE; + if (cp->vaos->base.key.const_vbuffers & (1<tmp_EAX; - /* In the linear, single buffer case, keep the buffer pointer - * instead of the index number. - */ - x86_mov( cp->func, elt, ptr ); + x86_mov(cp->func, ptr, buf_base_ptr); + + /* Load all inputs for this constant vertex buffer + */ + load_inputs( cp, i, x86_deref(ptr) ); + + /* Then just force them out to aos_machine.input[] + */ + aos_spill_all( cp ); + + } + else if (linear) { + + struct x86_reg elt = cp->idx_EBX; + struct x86_reg ptr = cp->tmp_EAX; + + struct x86_reg buf_stride = x86_make_disp(buf, + Offset(struct aos_buffer, stride)); + + struct x86_reg buf_ptr = x86_make_disp(buf, + Offset(struct aos_buffer, ptr)); + + + /* Calculate pointer to current attrib: + */ + x86_mov(cp->func, ptr, buf_stride); + x86_imul(cp->func, ptr, elt); + x86_add(cp->func, ptr, buf_base_ptr); + + + /* In the linear case, keep the buffer pointer instead of the + * index number. + */ + if (cp->vaos->nr_vb == 1) + x86_mov( cp->func, elt, ptr ); + else + x86_mov( cp->func, buf_ptr, ptr ); + + cp->insn_counter++; + } } return TRUE; @@ -219,23 +274,22 @@ boolean aos_init_inputs( struct aos_compilation *cp, boolean linear ) boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear ) { - if (linear && cp->vaos->nr_vb == 1) { - - load_inputs( cp, 0, cp->idx_EBX ); + unsigned j; - } - else { - struct x86_reg elt = linear ? cp->idx_EBX : x86_deref(cp->idx_EBX); - unsigned j; - - for (j = 0; j < cp->vaos->nr_vb; j++) { + for (j = 0; j < cp->vaos->nr_vb; j++) { + if (cp->vaos->base.key.const_vbuffers & (1<vaos->nr_vb == 1) { + load_inputs( cp, 0, cp->idx_EBX ); + } + else { + struct x86_reg elt = linear ? cp->idx_EBX : x86_deref(cp->idx_EBX); struct x86_reg ptr = cp->tmp_EAX; - if (!get_buffer_ptr( cp, j, elt, ptr )) + if (!get_buffer_ptr( cp, linear, j, elt, ptr )) return FALSE; - cp->insn_counter++; - if (!load_inputs( cp, j, ptr )) return FALSE; } @@ -252,13 +306,16 @@ boolean aos_incr_inputs( struct aos_compilation *cp, boolean linear ) Offset(struct aos_buffer, stride))); x86_add(cp->func, cp->idx_EBX, stride); + sse_prefetchnta(cp->func, x86_deref(cp->idx_EBX)); } else if (linear) { - x86_inc(cp->func, cp->idx_EBX); + /* Nothing to do */ } else { x86_lea(cp->func, cp->idx_EBX, x86_make_disp(cp->idx_EBX, 4)); } + + return TRUE; } diff --git a/src/gallium/auxiliary/draw/draw_vs_varient.c b/src/gallium/auxiliary/draw/draw_vs_varient.c index 4daf05dae7..7ee567d478 100644 --- a/src/gallium/auxiliary/draw/draw_vs_varient.c +++ b/src/gallium/auxiliary/draw/draw_vs_varient.c @@ -64,10 +64,10 @@ struct draw_vs_varient_generic { -static void vsvg_set_input( struct draw_vs_varient *varient, - unsigned buffer, - const void *ptr, - unsigned stride ) +static void vsvg_set_buffer( struct draw_vs_varient *varient, + unsigned buffer, + const void *ptr, + unsigned stride ) { struct draw_vs_varient_generic *vsvg = (struct draw_vs_varient_generic *)varient; @@ -265,7 +265,7 @@ struct draw_vs_varient *draw_vs_varient_generic( struct draw_vertex_shader *vs, vsvg->base.key = *key; vsvg->base.vs = vs; - vsvg->base.set_input = vsvg_set_input; + vsvg->base.set_buffer = vsvg_set_buffer; vsvg->base.run_elts = vsvg_run_elts; vsvg->base.run_linear = vsvg_run_linear; vsvg->base.destroy = vsvg_destroy; -- cgit v1.2.3 From 918a444913435bdee33214e25811875100f873b0 Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Thu, 2 Oct 2008 12:53:11 +0100 Subject: draw: modify prefetching slightly --- src/gallium/auxiliary/draw/draw_vs_aos_io.c | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_io.c b/src/gallium/auxiliary/draw/draw_vs_aos_io.c index b0c51d7fa1..dd79bc799a 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos_io.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos_io.c @@ -54,6 +54,7 @@ static void emit_load_R32G32B32( struct aos_compilation *cp, struct x86_reg data, struct x86_reg src_ptr ) { +#if 1 sse_movss(cp->func, data, x86_make_disp(src_ptr, 8)); /* data = z ? ? ? */ sse_shufps(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ), SHUF(X,Y,Z,W) ); @@ -62,6 +63,16 @@ static void emit_load_R32G32B32( struct aos_compilation *cp, /* data = ? 0 z 1 */ sse_movlps(cp->func, data, src_ptr); /* data = x y z 1 */ +#else + sse_movups(cp->func, data, src_ptr); + /* data = x y z ? */ + sse2_pshufd(cp->func, data, data, SHUF(W,X,Y,Z) ); + /* data = ? x y z */ + sse_movss(cp->func, data, aos_get_internal_xmm( cp, IMM_ONES ) ); + /* data = 1 x y z */ + sse2_pshufd(cp->func, data, data, SHUF(Y,Z,W,X) ); + /* data = x y z 1 */ +#endif } static void emit_load_R32G32( struct aos_compilation *cp, @@ -128,7 +139,7 @@ static boolean get_buffer_ptr( struct aos_compilation *cp, x86_mov(cp->func, ptr, buf_ptr); x86_mov(cp->func, elt, buf_stride); x86_add(cp->func, elt, ptr); - sse_prefetchnta(cp->func, x86_deref(elt)); + if (buf_idx == 0) sse_prefetchnta(cp->func, x86_make_disp(elt, 192)); x86_mov(cp->func, buf_ptr, elt); } else { @@ -306,7 +317,7 @@ boolean aos_incr_inputs( struct aos_compilation *cp, boolean linear ) Offset(struct aos_buffer, stride))); x86_add(cp->func, cp->idx_EBX, stride); - sse_prefetchnta(cp->func, x86_deref(cp->idx_EBX)); + sse_prefetchnta(cp->func, x86_make_disp(cp->idx_EBX, 192)); } else if (linear) { /* Nothing to do */ @@ -327,7 +338,7 @@ static void emit_store_R32G32B32A32( struct aos_compilation *cp, struct x86_reg dst_ptr, struct x86_reg dataXMM ) { - sse_movups(cp->func, dst_ptr, dataXMM); + sse_movaps(cp->func, dst_ptr, dataXMM); } static void emit_store_R32G32B32( struct aos_compilation *cp, @@ -430,7 +441,7 @@ boolean aos_emit_outputs( struct aos_compilation *cp ) if (data.file != file_XMM) { struct x86_reg tmp = aos_get_xmm_reg( cp ); - sse_movups(cp->func, tmp, data); + sse_movaps(cp->func, tmp, data); data = tmp; } -- cgit v1.2.3 From 6965532e14717f71a6f4353fb683c5070c6b7d7a Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Fri, 3 Oct 2008 13:50:34 +0100 Subject: rtasm: add sse_movntps --- src/gallium/auxiliary/rtasm/rtasm_x86sse.c | 12 ++++++++++++ src/gallium/auxiliary/rtasm/rtasm_x86sse.h | 2 ++ 2 files changed, 14 insertions(+) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c index 9085f4cc0e..cc5871f873 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c +++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c @@ -653,6 +653,18 @@ void sse_prefetch1( struct x86_function *p, struct x86_reg ptr) emit_modrm_noreg(p, 2, ptr); } +void sse_movntps( struct x86_function *p, + struct x86_reg dst, + struct x86_reg src) +{ + DUMP_RR( dst, reg ); + + assert(dst.mod != mod_REG); + assert(src.mod == mod_REG); + emit_2ub(p, 0x0f, 0x2b); + emit_modrm(p, src, dst); +} + diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.h b/src/gallium/auxiliary/rtasm/rtasm_x86sse.h index 2d7715f965..af79f07dd3 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.h +++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.h @@ -189,6 +189,8 @@ void sse_prefetchnta( struct x86_function *p, struct x86_reg ptr); void sse_prefetch0( struct x86_function *p, struct x86_reg ptr); void sse_prefetch1( struct x86_function *p, struct x86_reg ptr); +void sse_movntps( struct x86_function *p, struct x86_reg dst, struct x86_reg src); + void sse_addps( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); void sse_addss( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); void sse_cvtps2pi( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); -- cgit v1.2.3 From 53d4706c6c0922160f310834daaec5718ff1c511 Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Wed, 10 Sep 2008 11:39:43 +0100 Subject: make draw's vertex_info struct smaller/quicker to compare with memcmp() --- src/gallium/auxiliary/draw/draw_pipe_vbuf.c | 4 +- src/gallium/auxiliary/draw/draw_pt_emit.c | 4 +- src/gallium/auxiliary/draw/draw_pt_fetch_emit.c | 4 +- .../auxiliary/draw/draw_pt_fetch_shade_emit.c | 6 +-- src/gallium/auxiliary/draw/draw_vertex.c | 6 +-- src/gallium/auxiliary/draw/draw_vertex.h | 44 ++++++++++++++++++---- src/gallium/drivers/i915simple/i915_prim_emit.c | 4 +- .../drivers/i915simple/i915_state_derived.c | 4 +- src/gallium/drivers/softpipe/sp_setup.c | 12 +++--- 9 files changed, 59 insertions(+), 29 deletions(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/draw/draw_pipe_vbuf.c b/src/gallium/auxiliary/draw/draw_pipe_vbuf.c index c0cf4269db..9825e116c3 100644 --- a/src/gallium/auxiliary/draw/draw_pipe_vbuf.c +++ b/src/gallium/auxiliary/draw/draw_pipe_vbuf.c @@ -231,9 +231,9 @@ vbuf_set_prim( struct vbuf_stage *vbuf, uint prim ) unsigned emit_sz = 0; unsigned src_buffer = 0; unsigned output_format; - unsigned src_offset = (vbuf->vinfo->src_index[i] * 4 * sizeof(float) ); + unsigned src_offset = (vbuf->vinfo->attrib[i].src_index * 4 * sizeof(float) ); - switch (vbuf->vinfo->emit[i]) { + switch (vbuf->vinfo->attrib[i].emit) { case EMIT_4F: output_format = PIPE_FORMAT_R32G32B32A32_FLOAT; emit_sz = 4 * sizeof(float); diff --git a/src/gallium/auxiliary/draw/draw_pt_emit.c b/src/gallium/auxiliary/draw/draw_pt_emit.c index d4eca80588..d520b05869 100644 --- a/src/gallium/auxiliary/draw/draw_pt_emit.c +++ b/src/gallium/auxiliary/draw/draw_pt_emit.c @@ -84,11 +84,11 @@ void draw_pt_emit_prepare( struct pt_emit *emit, unsigned emit_sz = 0; unsigned src_buffer = 0; unsigned output_format; - unsigned src_offset = (vinfo->src_index[i] * 4 * sizeof(float) ); + unsigned src_offset = (vinfo->attrib[i].src_index * 4 * sizeof(float) ); - switch (vinfo->emit[i]) { + switch (vinfo->attrib[i].emit) { case EMIT_4F: output_format = PIPE_FORMAT_R32G32B32A32_FLOAT; emit_sz = 4 * sizeof(float); diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c b/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c index 5a4db6cfe5..3966ad48ba 100644 --- a/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c +++ b/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c @@ -121,7 +121,7 @@ static void fetch_emit_prepare( struct draw_pt_middle_end *middle, memset(&key, 0, sizeof(key)); for (i = 0; i < vinfo->num_attribs; i++) { - const struct pipe_vertex_element *src = &draw->pt.vertex_element[vinfo->src_index[i]]; + const struct pipe_vertex_element *src = &draw->pt.vertex_element[vinfo->attrib[i].src_index]; unsigned emit_sz = 0; unsigned input_format = src->src_format; @@ -129,7 +129,7 @@ static void fetch_emit_prepare( struct draw_pt_middle_end *middle, unsigned input_offset = src->src_offset; unsigned output_format; - switch (vinfo->emit[i]) { + switch (vinfo->attrib[i].emit) { case EMIT_4F: output_format = PIPE_FORMAT_R32G32B32A32_FLOAT; emit_sz = 4 * sizeof(float); diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c index a0e08dd10a..f7e6a1a8ee 100644 --- a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c +++ b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c @@ -133,7 +133,7 @@ static void fse_prepare( struct draw_pt_middle_end *middle, for (i = 0; i < vinfo->num_attribs; i++) { unsigned emit_sz = 0; - switch (vinfo->emit[i]) { + switch (vinfo->attrib[i].emit) { case EMIT_4F: emit_sz = 4 * sizeof(float); break; @@ -161,8 +161,8 @@ static void fse_prepare( struct draw_pt_middle_end *middle, * numbers, not to positions in the hw vertex description -- * that's handled by the output_offset field. */ - fse->key.element[i].out.format = vinfo->emit[i]; - fse->key.element[i].out.vs_output = vinfo->src_index[i]; + fse->key.element[i].out.format = vinfo->attrib[i].emit; + fse->key.element[i].out.vs_output = vinfo->attrib[i].src_index; fse->key.element[i].out.offset = dst_offset; dst_offset += emit_sz; diff --git a/src/gallium/auxiliary/draw/draw_vertex.c b/src/gallium/auxiliary/draw/draw_vertex.c index 1446f785c5..3214213e44 100644 --- a/src/gallium/auxiliary/draw/draw_vertex.c +++ b/src/gallium/auxiliary/draw/draw_vertex.c @@ -49,7 +49,7 @@ draw_compute_vertex_size(struct vertex_info *vinfo) vinfo->size = 0; for (i = 0; i < vinfo->num_attribs; i++) { - switch (vinfo->emit[i]) { + switch (vinfo->attrib[i].emit) { case EMIT_OMIT: break; case EMIT_4UB: @@ -81,8 +81,8 @@ draw_dump_emitted_vertex(const struct vertex_info *vinfo, const uint8_t *data) unsigned i, j; for (i = 0; i < vinfo->num_attribs; i++) { - j = vinfo->src_index[i]; - switch (vinfo->emit[i]) { + j = vinfo->attrib[i].src_index; + switch (vinfo->attrib[i].emit) { case EMIT_OMIT: debug_printf("EMIT_OMIT:"); break; diff --git a/src/gallium/auxiliary/draw/draw_vertex.h b/src/gallium/auxiliary/draw/draw_vertex.h index 16c65c4317..dca6158128 100644 --- a/src/gallium/auxiliary/draw/draw_vertex.h +++ b/src/gallium/auxiliary/draw/draw_vertex.h @@ -75,12 +75,41 @@ struct vertex_info { uint num_attribs; uint hwfmt[4]; /**< hardware format info for this format */ - enum interp_mode interp_mode[PIPE_MAX_SHADER_INPUTS]; - enum attrib_emit emit[PIPE_MAX_SHADER_INPUTS]; /**< EMIT_x */ - uint src_index[PIPE_MAX_SHADER_INPUTS]; /**< map to post-xform attribs */ uint size; /**< total vertex size in dwords */ + + /* Keep this small and at the end of the struct to allow quick + * memcmp() comparisons. + */ + struct { + ubyte interp_mode:4; /**< INTERP_x */ + ubyte emit:4; /**< EMIT_x */ + ubyte src_index; /**< map to post-xform attribs */ + } attrib[PIPE_MAX_SHADER_INPUTS]; }; +static inline int +draw_vinfo_size( const struct vertex_info *a ) +{ + return ((const char *)&a->attrib[a->num_attribs] - + (const char *)a); +} + +static inline int +draw_vinfo_compare( const struct vertex_info *a, + const struct vertex_info *b ) +{ + unsigned sizea = draw_vinfo_size( a ); + return memcmp( a, b, sizea ); +} + +static inline void +draw_vinfo_copy( struct vertex_info *dst, + const struct vertex_info *src ) +{ + unsigned size = draw_vinfo_size( src ); + memcpy( dst, src, size ); +} + /** @@ -91,14 +120,15 @@ struct vertex_info */ static INLINE uint draw_emit_vertex_attr(struct vertex_info *vinfo, - enum attrib_emit emit, enum interp_mode interp, + enum attrib_emit emit, + enum interp_mode interp, /* only used by softpipe??? */ uint src_index) { const uint n = vinfo->num_attribs; assert(n < PIPE_MAX_SHADER_INPUTS); - vinfo->emit[n] = emit; - vinfo->interp_mode[n] = interp; - vinfo->src_index[n] = src_index; + vinfo->attrib[n].emit = emit; + vinfo->attrib[n].interp_mode = interp; + vinfo->attrib[n].src_index = src_index; vinfo->num_attribs++; return n; } diff --git a/src/gallium/drivers/i915simple/i915_prim_emit.c b/src/gallium/drivers/i915simple/i915_prim_emit.c index d194c2fb15..8f1f58b2dd 100644 --- a/src/gallium/drivers/i915simple/i915_prim_emit.c +++ b/src/gallium/drivers/i915simple/i915_prim_emit.c @@ -77,9 +77,9 @@ emit_hw_vertex( struct i915_context *i915, assert(!i915->dirty); for (i = 0; i < vinfo->num_attribs; i++) { - const uint j = vinfo->src_index[i]; + const uint j = vinfo->attrib[i].src_index; const float *attrib = vertex->data[j]; - switch (vinfo->emit[i]) { + switch (vinfo->attrib[i].emit) { case EMIT_1F: OUT_BATCH( fui(attrib[0]) ); count++; diff --git a/src/gallium/drivers/i915simple/i915_state_derived.c b/src/gallium/drivers/i915simple/i915_state_derived.c index 488615067c..178d4e8781 100644 --- a/src/gallium/drivers/i915simple/i915_state_derived.c +++ b/src/gallium/drivers/i915simple/i915_state_derived.c @@ -88,12 +88,12 @@ static void calculate_vertex_layout( struct i915_context *i915 ) if (needW) { draw_emit_vertex_attr(&vinfo, EMIT_4F, INTERP_LINEAR, src); vinfo.hwfmt[0] |= S4_VFMT_XYZW; - vinfo.emit[0] = EMIT_4F; + vinfo.attrib[0].emit = EMIT_4F; } else { draw_emit_vertex_attr(&vinfo, EMIT_3F, INTERP_LINEAR, src); vinfo.hwfmt[0] |= S4_VFMT_XYZ; - vinfo.emit[0] = EMIT_3F; + vinfo.attrib[0].emit = EMIT_3F; } /* hardware point size */ diff --git a/src/gallium/drivers/softpipe/sp_setup.c b/src/gallium/drivers/softpipe/sp_setup.c index bc8263c33e..13d8017393 100644 --- a/src/gallium/drivers/softpipe/sp_setup.c +++ b/src/gallium/drivers/softpipe/sp_setup.c @@ -773,10 +773,10 @@ static void setup_tri_coefficients( struct setup_context *setup ) /* setup interpolation for all the remaining attributes: */ for (fragSlot = 0; fragSlot < spfs->info.num_inputs; fragSlot++) { - const uint vertSlot = vinfo->src_index[fragSlot]; + const uint vertSlot = vinfo->attrib[fragSlot].src_index; uint j; - switch (vinfo->interp_mode[fragSlot]) { + switch (vinfo->attrib[fragSlot].interp_mode) { case INTERP_CONSTANT: for (j = 0; j < NUM_CHANNELS; j++) const_coeff(setup, &setup->coef[fragSlot], vertSlot, j); @@ -1084,10 +1084,10 @@ setup_line_coefficients(struct setup_context *setup, /* setup interpolation for all the remaining attributes: */ for (fragSlot = 0; fragSlot < spfs->info.num_inputs; fragSlot++) { - const uint vertSlot = vinfo->src_index[fragSlot]; + const uint vertSlot = vinfo->attrib[fragSlot].src_index; uint j; - switch (vinfo->interp_mode[fragSlot]) { + switch (vinfo->attrib[fragSlot].interp_mode) { case INTERP_CONSTANT: for (j = 0; j < NUM_CHANNELS; j++) const_coeff(setup, &setup->coef[fragSlot], vertSlot, j); @@ -1331,10 +1331,10 @@ setup_point( struct setup_context *setup, const_coeff(setup, &setup->posCoef, 0, 3); for (fragSlot = 0; fragSlot < spfs->info.num_inputs; fragSlot++) { - const uint vertSlot = vinfo->src_index[fragSlot]; + const uint vertSlot = vinfo->attrib[fragSlot].src_index; uint j; - switch (vinfo->interp_mode[fragSlot]) { + switch (vinfo->attrib[fragSlot].interp_mode) { case INTERP_CONSTANT: /* fall-through */ case INTERP_LINEAR: -- cgit v1.2.3 From 7053f8c902e904495dffbbf6ea55f414cec780e7 Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Mon, 6 Oct 2008 11:54:22 +0100 Subject: rtasm: fix debug build --- src/gallium/auxiliary/rtasm/rtasm_x86sse.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c index cc5871f873..dd26d4d9ed 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c +++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c @@ -657,7 +657,7 @@ void sse_movntps( struct x86_function *p, struct x86_reg dst, struct x86_reg src) { - DUMP_RR( dst, reg ); + DUMP_RR( dst, src ); assert(dst.mod != mod_REG); assert(src.mod == mod_REG); -- cgit v1.2.3 From 9b827018133868e84ddc0998a5b5387584c7478c Mon Sep 17 00:00:00 2001 From: Michal Krol Date: Mon, 6 Oct 2008 13:23:56 +0200 Subject: draw: Fix compiler errors on Windows. --- src/gallium/auxiliary/draw/draw_vertex.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/draw/draw_vertex.h b/src/gallium/auxiliary/draw/draw_vertex.h index dca6158128..a943607d7e 100644 --- a/src/gallium/auxiliary/draw/draw_vertex.h +++ b/src/gallium/auxiliary/draw/draw_vertex.h @@ -87,14 +87,14 @@ struct vertex_info } attrib[PIPE_MAX_SHADER_INPUTS]; }; -static inline int +static INLINE int draw_vinfo_size( const struct vertex_info *a ) { return ((const char *)&a->attrib[a->num_attribs] - (const char *)a); } -static inline int +static INLINE int draw_vinfo_compare( const struct vertex_info *a, const struct vertex_info *b ) { @@ -102,7 +102,7 @@ draw_vinfo_compare( const struct vertex_info *a, return memcmp( a, b, sizea ); } -static inline void +static INLINE void draw_vinfo_copy( struct vertex_info *dst, const struct vertex_info *src ) { -- cgit v1.2.3 From f7ee3c979261b4a2b77365b47c7147f69fbfd606 Mon Sep 17 00:00:00 2001 From: Brian Date: Mon, 6 Oct 2008 18:31:56 -0600 Subject: gallium: replace assertion with conditional/recovery code The assertion failed when we ran out of exec memory. Found with conform texcombine test. --- src/gallium/auxiliary/rtasm/rtasm_x86sse.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c index dd26d4d9ed..ad9d8f8ced 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c +++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c @@ -370,7 +370,11 @@ void x86_jcc( struct x86_function *p, DUMP_I(cc); if (offset < 0) { - assert(p->csr - p->store > -offset); + /*assert(p->csr - p->store > -offset);*/ + if (p->csr - p->store <= -offset) { + /* probably out of memory (using the error_overflow buffer) */ + return; + } } if (offset <= 127 && offset >= -128) { -- cgit v1.2.3 From c48da7d78b4e7bdbe056b3c9668756d49019be06 Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Mon, 6 Oct 2008 12:22:55 +0100 Subject: draw: add switch for drivers to force vertex data passthrough --- src/gallium/auxiliary/draw/draw_context.c | 8 +++++++ src/gallium/auxiliary/draw/draw_context.h | 3 +++ src/gallium/auxiliary/draw/draw_private.h | 3 +++ src/gallium/auxiliary/draw/draw_pt.c | 38 +++++++++++++++---------------- 4 files changed, 33 insertions(+), 19 deletions(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/draw/draw_context.c b/src/gallium/auxiliary/draw/draw_context.c index 36751c2621..41a4cba1dd 100644 --- a/src/gallium/auxiliary/draw/draw_context.c +++ b/src/gallium/auxiliary/draw/draw_context.c @@ -272,6 +272,14 @@ draw_enable_point_sprites(struct draw_context *draw, boolean enable) } +void +draw_set_force_passthrough( struct draw_context *draw, boolean enable ) +{ + draw_do_flush( draw, DRAW_FLUSH_STATE_CHANGE ); + draw->force_passthrough = enable; +} + + /** * Ask the draw module for the location/slot of the given vertex attribute in * a post-transformed vertex. diff --git a/src/gallium/auxiliary/draw/draw_context.h b/src/gallium/auxiliary/draw/draw_context.h index 0ab3681b64..3eeb453531 100644 --- a/src/gallium/auxiliary/draw/draw_context.h +++ b/src/gallium/auxiliary/draw/draw_context.h @@ -160,6 +160,9 @@ void draw_set_render( struct draw_context *draw, void draw_set_driver_clipping( struct draw_context *draw, boolean bypass_clipping ); +void draw_set_force_passthrough( struct draw_context *draw, + boolean enable ); + /******************************************************************************* * Draw pipeline */ diff --git a/src/gallium/auxiliary/draw/draw_private.h b/src/gallium/auxiliary/draw/draw_private.h index 626a2e3e30..37c4c87f87 100644 --- a/src/gallium/auxiliary/draw/draw_private.h +++ b/src/gallium/auxiliary/draw/draw_private.h @@ -163,12 +163,15 @@ struct draw_context struct { boolean bypass_clipping; + boolean bypass_vs; } driver; boolean flushing; /**< debugging/sanity */ boolean suspend_flushing; /**< internally set */ boolean bypass_clipping; /**< set if either api or driver bypass_clipping true */ + boolean force_passthrough; /**< never clip or shade */ + /* pipe state that we need: */ const struct pipe_rasterizer_state *rasterizer; struct pipe_viewport_state viewport; diff --git a/src/gallium/auxiliary/draw/draw_pt.c b/src/gallium/auxiliary/draw/draw_pt.c index 669c11c993..87ec6ae20c 100644 --- a/src/gallium/auxiliary/draw/draw_pt.c +++ b/src/gallium/auxiliary/draw/draw_pt.c @@ -69,26 +69,26 @@ draw_pt_arrays(struct draw_context *draw, return TRUE; } - - if (!draw->render) { - opt |= PT_PIPELINE; - } - - if (draw_need_pipeline(draw, - draw->rasterizer, - prim)) { - opt |= PT_PIPELINE; - } - - if (!draw->bypass_clipping && !draw->pt.test_fse) { - opt |= PT_CLIPTEST; + if (!draw->force_passthrough) { + if (!draw->render) { + opt |= PT_PIPELINE; + } + + if (draw_need_pipeline(draw, + draw->rasterizer, + prim)) { + opt |= PT_PIPELINE; + } + + if (!draw->bypass_clipping && !draw->pt.test_fse) { + opt |= PT_CLIPTEST; + } + + if (!draw->rasterizer->bypass_vs) { + opt |= PT_SHADE; + } } - - if (!draw->rasterizer->bypass_vs) { - opt |= PT_SHADE; - } - - + if (opt == 0) middle = draw->pt.middle.fetch_emit; else if (opt == PT_SHADE && !draw->pt.no_fse) -- cgit v1.2.3 From 23cc303994eb630c56b1224dfdac51dcea41ed03 Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Tue, 7 Oct 2008 16:44:24 +0100 Subject: draw: don't assume output buffer pointer is aligned --- src/gallium/auxiliary/draw/draw_vs_aos_io.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src/gallium/auxiliary') diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_io.c b/src/gallium/auxiliary/draw/draw_vs_aos_io.c index dd79bc799a..39f75b50b7 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos_io.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos_io.c @@ -338,7 +338,7 @@ static void emit_store_R32G32B32A32( struct aos_compilation *cp, struct x86_reg dst_ptr, struct x86_reg dataXMM ) { - sse_movaps(cp->func, dst_ptr, dataXMM); + sse_movups(cp->func, dst_ptr, dataXMM); } static void emit_store_R32G32B32( struct aos_compilation *cp, -- cgit v1.2.3