fallback codegen: matrix transpose

src/backends/fallback/fallback_codegen.cpp

@@ -1815,6 +1815,47 @@ class FallbackCodegen {
         return b.CreateLoad(transform_type, transform_alloca, "");
     }
 
+
+	[[nodiscard]] llvm::Value *_translate_matrix_transpose(
+			CurrentFunction &current,
+			IRBuilder &b,
+			const xir::IntrinsicInst *inst) noexcept
+	{
+		// Lookup the operand (the matrix to be transposed)
+		auto matrix = inst->operand(0u);
+		auto dimension = inst->type()->dimension();
+
+		// Lookup the LLVM value
+		auto llvm_matrix = _lookup_value(current, b, matrix);
+
+		// Type validation
+		LUISA_ASSERT(matrix->type()->is_matrix(), "Matrix transpose type mismatch");
+		LUISA_ASSERT(matrix->type()->dimension()>2, "2x2 Matrix is yet unsupported");
+
+		// The resulting matrix will have the same dimension
+		llvm::Type *float_type = llvm::Type::getFloatTy(b.getContext());
+		llvm::ArrayType *vec4_type = llvm::ArrayType::get(float_type, 4); // Simulate vec4
+
+		// Initialize the resulting matrix with 'undef' values
+		llvm::ArrayType *result_type = llvm::ArrayType::get(vec4_type, dimension);
+		llvm::Value *result = llvm::UndefValue::get(result_type);
+
+		// Transpose the matrix by swapping rows and columns
+		for (unsigned i = 0; i < dimension; ++i) {
+			for (unsigned j = 0; j < dimension; ++j) {
+				// Extract value at position (i, j) in the original matrix
+				llvm::Value *value = b.CreateExtractValue(llvm_matrix, {i, j});
+
+				// Insert the value at position (j, i) in the transposed matrix
+				result = b.CreateInsertValue(result, value, {j, i});
+			}
+		}
+
+		return result;
+	}
+
+
+
     [[nodiscard]] llvm::Value *_translate_matrix_linalg_multiply(
         CurrentFunction &current,
         IRBuilder &b,
@@ -2273,7 +2314,7 @@ class FallbackCodegen {
             case xir::IntrinsicOp::MATRIX_COMP_DIV: break;
             case xir::IntrinsicOp::MATRIX_LINALG_MUL: return _translate_matrix_linalg_multiply(current, b, inst);
             case xir::IntrinsicOp::MATRIX_DETERMINANT: break;
-            case xir::IntrinsicOp::MATRIX_TRANSPOSE: break;
+            case xir::IntrinsicOp::MATRIX_TRANSPOSE: return _translate_matrix_transpose(current, b, inst);
             case xir::IntrinsicOp::MATRIX_INVERSE: break;
             case xir::IntrinsicOp::ATOMIC_EXCHANGE: break;
             case xir::IntrinsicOp::ATOMIC_COMPARE_EXCHANGE: break;