SPIRV: Simplify matrix->matrix constructor
When constructing a matrix from another matrix with smaller dimensions, there's no need to extract the scalars out of columns and rebuild the resulting matrix from scalars - instead, we can just construct shorter vectors with OpShuffle and combine them to the final result. This keeps the common casts such as mat3(mat4) in vector registers, which may improve performance for some GPUs, and cleans up output of translation tools like SPIRV-Cross. Fixes #1412.
This commit is contained in:
Arseny Kapoulkine
@@ -2031,83 +2031,113 @@ Id Builder::createMatrixConstructor(Decoration precision, const std::vector<Id>&
|
||||
Instruction* instr = module.getInstruction(componentTypeId);
|
||||
Id bitCount = instr->getIdOperand(0);
|
||||
|
||||
// Will use a two step process
|
||||
// 1. make a compile-time 2D array of values
|
||||
// 2. construct a matrix from that array
|
||||
|
||||
// Step 1.
|
||||
|
||||
// initialize the array to the identity matrix
|
||||
Id ids[maxMatrixSize][maxMatrixSize];
|
||||
Id one = (bitCount == 64 ? makeDoubleConstant(1.0) : makeFloatConstant(1.0));
|
||||
Id zero = (bitCount == 64 ? makeDoubleConstant(0.0) : makeFloatConstant(0.0));
|
||||
for (int col = 0; col < 4; ++col) {
|
||||
for (int row = 0; row < 4; ++row) {
|
||||
if (col == row)
|
||||
ids[col][row] = one;
|
||||
else
|
||||
ids[col][row] = zero;
|
||||
}
|
||||
}
|
||||
|
||||
// modify components as dictated by the arguments
|
||||
if (sources.size() == 1 && isScalar(sources[0])) {
|
||||
// a single scalar; resets the diagonals
|
||||
for (int col = 0; col < 4; ++col)
|
||||
ids[col][col] = sources[0];
|
||||
} else if (isMatrix(sources[0])) {
|
||||
// constructing from another matrix; copy over the parts that exist in both the argument and constructee
|
||||
if (isMatrix(sources[0]) && getNumColumns(sources[0]) >= numCols && getNumRows(sources[0]) >= numRows) {
|
||||
// To truncate the matrix to a smaller number of rows/columns, we need to:
|
||||
// 1. For each column, extract the column and truncate it to the required size using shuffle
|
||||
// 2. Assemble the resulting matrix from all columns
|
||||
Id matrix = sources[0];
|
||||
int minCols = std::min(numCols, getNumColumns(matrix));
|
||||
int minRows = std::min(numRows, getNumRows(matrix));
|
||||
for (int col = 0; col < minCols; ++col) {
|
||||
Id columnTypeId = getContainedTypeId(resultTypeId);
|
||||
Id sourceColumnTypeId = getContainedTypeId(getTypeId(matrix));
|
||||
|
||||
std::vector<unsigned> channels;
|
||||
for (int row = 0; row < numRows; ++row) {
|
||||
channels.push_back(row);
|
||||
}
|
||||
|
||||
std::vector<Id> matrixColumns;
|
||||
for (int col = 0; col < numCols; ++col) {
|
||||
std::vector<unsigned> indexes;
|
||||
indexes.push_back(col);
|
||||
for (int row = 0; row < minRows; ++row) {
|
||||
indexes.push_back(row);
|
||||
ids[col][row] = createCompositeExtract(matrix, componentTypeId, indexes);
|
||||
indexes.pop_back();
|
||||
setPrecision(ids[col][row], precision);
|
||||
Id colv = createCompositeExtract(matrix, sourceColumnTypeId, indexes);
|
||||
setPrecision(colv, precision);
|
||||
|
||||
if (numRows != getNumRows(matrix)) {
|
||||
matrixColumns.push_back(createRvalueSwizzle(precision, columnTypeId, colv, channels));
|
||||
} else {
|
||||
matrixColumns.push_back(colv);
|
||||
}
|
||||
}
|
||||
|
||||
return setPrecision(createCompositeConstruct(resultTypeId, matrixColumns), precision);
|
||||
} else {
|
||||
// fill in the matrix in column-major order with whatever argument components are available
|
||||
int row = 0;
|
||||
int col = 0;
|
||||
// Will use a two step process
|
||||
// 1. make a compile-time 2D array of values
|
||||
// 2. construct a matrix from that array
|
||||
|
||||
for (int arg = 0; arg < (int)sources.size(); ++arg) {
|
||||
Id argComp = sources[arg];
|
||||
for (int comp = 0; comp < getNumComponents(sources[arg]); ++comp) {
|
||||
if (getNumComponents(sources[arg]) > 1) {
|
||||
argComp = createCompositeExtract(sources[arg], componentTypeId, comp);
|
||||
setPrecision(argComp, precision);
|
||||
// Step 1.
|
||||
|
||||
// initialize the array to the identity matrix
|
||||
Id ids[maxMatrixSize][maxMatrixSize];
|
||||
Id one = (bitCount == 64 ? makeDoubleConstant(1.0) : makeFloatConstant(1.0));
|
||||
Id zero = (bitCount == 64 ? makeDoubleConstant(0.0) : makeFloatConstant(0.0));
|
||||
for (int col = 0; col < 4; ++col) {
|
||||
for (int row = 0; row < 4; ++row) {
|
||||
if (col == row)
|
||||
ids[col][row] = one;
|
||||
else
|
||||
ids[col][row] = zero;
|
||||
}
|
||||
}
|
||||
|
||||
// modify components as dictated by the arguments
|
||||
if (sources.size() == 1 && isScalar(sources[0])) {
|
||||
// a single scalar; resets the diagonals
|
||||
for (int col = 0; col < 4; ++col)
|
||||
ids[col][col] = sources[0];
|
||||
} else if (isMatrix(sources[0])) {
|
||||
// constructing from another matrix; copy over the parts that exist in both the argument and constructee
|
||||
Id matrix = sources[0];
|
||||
int minCols = std::min(numCols, getNumColumns(matrix));
|
||||
int minRows = std::min(numRows, getNumRows(matrix));
|
||||
for (int col = 0; col < minCols; ++col) {
|
||||
std::vector<unsigned> indexes;
|
||||
indexes.push_back(col);
|
||||
for (int row = 0; row < minRows; ++row) {
|
||||
indexes.push_back(row);
|
||||
ids[col][row] = createCompositeExtract(matrix, componentTypeId, indexes);
|
||||
indexes.pop_back();
|
||||
setPrecision(ids[col][row], precision);
|
||||
}
|
||||
ids[col][row++] = argComp;
|
||||
if (row == numRows) {
|
||||
row = 0;
|
||||
col++;
|
||||
}
|
||||
} else {
|
||||
// fill in the matrix in column-major order with whatever argument components are available
|
||||
int row = 0;
|
||||
int col = 0;
|
||||
|
||||
for (int arg = 0; arg < (int)sources.size(); ++arg) {
|
||||
Id argComp = sources[arg];
|
||||
for (int comp = 0; comp < getNumComponents(sources[arg]); ++comp) {
|
||||
if (getNumComponents(sources[arg]) > 1) {
|
||||
argComp = createCompositeExtract(sources[arg], componentTypeId, comp);
|
||||
setPrecision(argComp, precision);
|
||||
}
|
||||
ids[col][row++] = argComp;
|
||||
if (row == numRows) {
|
||||
row = 0;
|
||||
col++;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Step 2: Construct a matrix from that array.
|
||||
// First make the column vectors, then make the matrix.
|
||||
|
||||
// make the column vectors
|
||||
Id columnTypeId = getContainedTypeId(resultTypeId);
|
||||
std::vector<Id> matrixColumns;
|
||||
for (int col = 0; col < numCols; ++col) {
|
||||
std::vector<Id> vectorComponents;
|
||||
for (int row = 0; row < numRows; ++row)
|
||||
vectorComponents.push_back(ids[col][row]);
|
||||
Id column = createCompositeConstruct(columnTypeId, vectorComponents);
|
||||
setPrecision(column, precision);
|
||||
matrixColumns.push_back(column);
|
||||
}
|
||||
|
||||
// make the matrix
|
||||
return setPrecision(createCompositeConstruct(resultTypeId, matrixColumns), precision);
|
||||
}
|
||||
|
||||
// Step 2: Construct a matrix from that array.
|
||||
// First make the column vectors, then make the matrix.
|
||||
|
||||
// make the column vectors
|
||||
Id columnTypeId = getContainedTypeId(resultTypeId);
|
||||
std::vector<Id> matrixColumns;
|
||||
for (int col = 0; col < numCols; ++col) {
|
||||
std::vector<Id> vectorComponents;
|
||||
for (int row = 0; row < numRows; ++row)
|
||||
vectorComponents.push_back(ids[col][row]);
|
||||
Id column = createCompositeConstruct(columnTypeId, vectorComponents);
|
||||
setPrecision(column, precision);
|
||||
matrixColumns.push_back(column);
|
||||
}
|
||||
|
||||
// make the matrix
|
||||
return setPrecision(createCompositeConstruct(resultTypeId, matrixColumns), precision);
|
||||
}
|
||||
|
||||
// Comments in header
|
||||
|
||||
Reference in New Issue
Block a user