Non square matrices related fixes
- Implemented proper VariableRegisterCount (we were using row
count instead of column count to get the number of registers)
and VariableRegisterSize.
- Matrix to matrix copies now clear the correct rows of the
destination matrix when needed
- Added registerSize helper function to type to help clarify
this for matrices.
- Added missing member initializations in TType constructor
Change-Id: Ic880815515c7d12ad12e44f1392aa6892caa953f
Reviewed-on: https://swiftshader-review.googlesource.com/3718
Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
diff --git a/src/OpenGL/compiler/OutputASM.cpp b/src/OpenGL/compiler/OutputASM.cpp
index 4fb4126..428be87 100644
--- a/src/OpenGL/compiler/OutputASM.cpp
+++ b/src/OpenGL/compiler/OutputASM.cpp
@@ -964,11 +964,12 @@
if(visit == PostVisit)
{
TIntermTyped *arg0 = arg[0]->getAsTyped();
- const int dim = result->getNominalSize();
+ const int outCols = result->getNominalSize();
+ const int outRows = result->getSecondarySize();
if(arg0->isScalar() && arg.size() == 1) // Construct scale matrix
{
- for(int i = 0; i < dim; i++)
+ for(int i = 0; i < outCols; i++)
{
Instruction *init = emit(sw::Shader::OPCODE_MOV, result, &zero);
init->dst.index += i;
@@ -980,9 +981,12 @@
}
else if(arg0->isMatrix())
{
- for(int i = 0; i < dim; i++)
+ const int inCols = arg0->getNominalSize();
+ const int inRows = arg0->getSecondarySize();
+
+ for(int i = 0; i < outCols; i++)
{
- if(dim > dim2(arg0))
+ if(i >= inCols || outRows > inRows)
{
// Initialize to identity matrix
Constant col((i == 0 ? 1.0f : 0.0f), (i == 1 ? 1.0f : 0.0f), (i == 2 ? 1.0f : 0.0f), (i == 3 ? 1.0f : 0.0f));
@@ -990,11 +994,11 @@
mov->dst.index += i;
}
- if(i < dim2(arg0))
+ if(i < inCols)
{
Instruction *mov = emitCast(result, arg0);
mov->dst.index += i;
- mov->dst.mask = 0xF >> (4 - dim2(arg0));
+ mov->dst.mask = 0xF >> (4 - inRows);
argument(mov->src[0], arg0, i);
}
}
@@ -1018,9 +1022,9 @@
mov->src[0].swizzle = (readSwizzle(argi, size) << (row * 2)) + 0x55 * element;
int end = row + size - element;
- column = end >= dim ? column + 1 : column;
- element = element + dim - row;
- row = end >= dim ? 0 : end;
+ column = end >= outRows ? column + 1 : column;
+ element = element + outRows - row;
+ row = end >= outRows ? 0 : end;
}
}
}
@@ -1461,7 +1465,7 @@
}
else if(type.isMatrix())
{
- return registers * type.getSecondarySize();
+ return registers * type.registerSize();
}
UNREACHABLE(0);
@@ -1477,7 +1481,7 @@
return registerSize(*((*(type.getStruct()->fields().begin()))->type()), 0);
}
- return type.isMatrix() ? type.getSecondarySize() : type.getNominalSize();
+ return type.registerSize();
}
if(type.isArray() && registers >= type.elementRegisterCount())
@@ -2077,7 +2081,7 @@
if(var == -1)
{
var = allocate(varyings, varying);
- int componentCount = varying->getNominalSize();
+ int componentCount = varying->registerSize();
int registerCount = varying->totalRegisterCount();
if(pixelShader)
diff --git a/src/OpenGL/compiler/SymbolTable.cpp b/src/OpenGL/compiler/SymbolTable.cpp
index ad3488e..7932bd5 100644
--- a/src/OpenGL/compiler/SymbolTable.cpp
+++ b/src/OpenGL/compiler/SymbolTable.cpp
@@ -26,8 +26,9 @@
int TSymbolTableLevel::uniqueId = 0;
TType::TType(const TPublicType &p) :
- type(p.type), precision(p.precision), primarySize(p.primarySize), secondarySize(p.secondarySize), qualifier(p.qualifier), array(p.array), arraySize(p.arraySize),
- maxArraySize(0), arrayInformationType(0), structure(0), deepestStructNesting(0), mangled(0)
+ type(p.type), precision(p.precision), qualifier(p.qualifier), invariant(false), layoutQualifier(TLayoutQualifier::create()),
+ primarySize(p.primarySize), secondarySize(p.secondarySize), array(p.array), arraySize(p.arraySize), maxArraySize(0),
+ arrayInformationType(0), interfaceBlock(0), structure(0), deepestStructNesting(0), mangled(0)
{
if (p.userDef)
{
diff --git a/src/OpenGL/compiler/Types.h b/src/OpenGL/compiler/Types.h
index cb2722c..e58c996 100644
--- a/src/OpenGL/compiler/Types.h
+++ b/src/OpenGL/compiler/Types.h
@@ -345,6 +345,11 @@
}
}
+ int registerSize() const
+ {
+ return isMatrix() ? secondarySize : primarySize;
+ }
+
bool isMatrix() const { return secondarySize > 1; }
void setSecondarySize(int s1) { secondarySize = s1; }
int getSecondarySize() const { return secondarySize; }
diff --git a/src/OpenGL/compiler/intermediate.h b/src/OpenGL/compiler/intermediate.h
index 99d126d..8b72d30 100644
--- a/src/OpenGL/compiler/intermediate.h
+++ b/src/OpenGL/compiler/intermediate.h
@@ -321,6 +321,7 @@
int totalRegisterCount() const { return type.totalRegisterCount(); }
int elementRegisterCount() const { return type.elementRegisterCount(); }
+ int registerSize() const { return type.registerSize(); }
int getArraySize() const { return type.getArraySize(); }
protected:
diff --git a/src/OpenGL/libGLESv2/Program.cpp b/src/OpenGL/libGLESv2/Program.cpp
index dd4c598..b8ba80b 100644
--- a/src/OpenGL/libGLESv2/Program.cpp
+++ b/src/OpenGL/libGLESv2/Program.cpp
@@ -65,7 +65,7 @@
int Uniform::registerCount() const
{
- return size() * VariableRowCount(type);
+ return size() * VariableRegisterCount(type);
}
UniformBlock::UniformBlock(const std::string &name, unsigned int elementIndex, unsigned int dataSize) :
@@ -1128,8 +1128,8 @@
{
int in = input->reg;
int out = output->reg;
- int components = VariableColumnCount(output->type);
- int registers = VariableRowCount(output->type) * output->size();
+ int components = VariableRegisterSize(output->type);
+ int registers = VariableRegisterCount(output->type) * output->size();
ASSERT(in >= 0);
@@ -1270,7 +1270,7 @@
linkedAttribute[location] = *attribute;
- int rows = VariableRowCount(attribute->type);
+ int rows = VariableRegisterCount(attribute->type);
if(rows + location > MAX_VERTEX_ATTRIBS)
{
@@ -1292,7 +1292,7 @@
if(location == -1) // Not set by glBindAttribLocation
{
- int rows = VariableRowCount(attribute->type);
+ int rows = VariableRegisterCount(attribute->type);
int availableIndex = AllocateFirstFreeBits(&usedLocations, rows, MAX_VERTEX_ATTRIBS);
if(availableIndex == -1 || availableIndex + rows > MAX_VERTEX_ATTRIBS)
@@ -1308,7 +1308,7 @@
for(int attributeIndex = 0; attributeIndex < MAX_VERTEX_ATTRIBS; )
{
int index = vertexShader->getSemanticIndex(linkedAttribute[attributeIndex].name);
- int rows = std::max(VariableRowCount(linkedAttribute[attributeIndex].type), 1);
+ int rows = std::max(VariableRegisterCount(linkedAttribute[attributeIndex].type), 1);
for(int r = 0; r < rows; r++)
{
diff --git a/src/OpenGL/libGLESv2/utilities.cpp b/src/OpenGL/libGLESv2/utilities.cpp
index ea19daf..bdb8e46 100644
--- a/src/OpenGL/libGLESv2/utilities.cpp
+++ b/src/OpenGL/libGLESv2/utilities.cpp
@@ -278,6 +278,19 @@
return 0;
}
+ int VariableRegisterCount(GLenum type)
+ {
+ // Number of registers used is the number of columns for matrices or 1 for scalars and vectors
+ return (VariableRowCount(type) > 1) ? VariableColumnCount(type) : 1;
+ }
+
+ int VariableRegisterSize(GLenum type)
+ {
+ // Number of components per register is the number of rows for matrices or columns for scalars and vectors
+ int nbRows = VariableRowCount(type);
+ return (nbRows > 1) ? nbRows : VariableColumnCount(type);
+ }
+
int AllocateFirstFreeBits(unsigned int *bits, unsigned int allocationSize, unsigned int bitsSize)
{
ASSERT(allocationSize <= bitsSize);
diff --git a/src/OpenGL/libGLESv2/utilities.h b/src/OpenGL/libGLESv2/utilities.h
index baa699f..82f5fa9 100644
--- a/src/OpenGL/libGLESv2/utilities.h
+++ b/src/OpenGL/libGLESv2/utilities.h
@@ -33,6 +33,8 @@
bool IsSamplerUniform(GLenum type);
int VariableRowCount(GLenum type);
int VariableColumnCount(GLenum type);
+ int VariableRegisterCount(GLenum type);
+ int VariableRegisterSize(GLenum type);
int AllocateFirstFreeBits(unsigned int *bits, unsigned int allocationSize, unsigned int bitsSize);
