Gallivm: more instructions.

3 files changed, 73 insertions, 8 deletions

diff --git a/src/gallium/auxiliary/gallivm/instructions.cpp b/src/gallium/auxiliary/gallivm/instructions.cpp
index 5fdfe09d182..3eaf9aacf60 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<llvm::Value*> 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<const Type*> 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 8df30f62c8a..c3b28e97469 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 398fbd67bd1..fdfbb76c16a 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: