Filling in a MTLAccelerationStructureInstanceDescriptor

Question

Created Jul ’20

Replies 3

Boosts 0

Participants 3

While translating the Accelerating Ray Tracing Using Metal sample project into Swift, I have run into an issue within the createAccelerationStructures function.

I am trying to fill in the first three rows of the instance transformation matrix. In the Objective C example, it looks like:

Code Block for (int column = 0; column < 4; column)
	for (int row = 0; row < 3; row)
		instanceDescriptors[instanceIndex].transformationMatrix.columns[column][row] = instance.transform.columns[column][row];

The matrix targeted is a MTLPackedFloat4x3. It has defied me pulling out the needed elements from my instance transform and placing them in the descriptor. I have tried the obvious:

Code Block for column in 0..<( 4 )
	{
	for row in 0..<( 3 )
		{
		anInstanceDescriptorPointer.pointee.transformationMatrix.columns[column][row] = aNodeInstance.transform.columns[column][row]
		}
	}

I have tried more obscure ideas related to simd matrices and Swift. The compiler errors generally warn me that MTLPackedFloat4x3 has no subscripts.

Question - Is there an accepted way to assign three rows from the instance transform to the descriptor transformationMatrix in Swift?

Thank you for your earlier help with all of this. I should mention that the new intersector feature in Metal has worked really well for me when feeding it a primitiveaccelerationstructure. Very elegant! But using instances seems to offer more flexibility going forward.

Thank you very much!

Boost

Answer 1

OOPer OP

Jul ’20

The compiler errors generally warn me that MTLPackedFloat4x3 has no subscripts.

Columns in MTLPackedFloat4x3 are represented as a fixed-sized C-array, and imported into Swift as a tuple.
So, you need to access each column by tuple member access like .0, .1, .2 or .3.
Or else access whole columns as a tuple.

I would do it in this way.

First, prepare an initializer taking SIMD matrix.

Code Block extension MTLPackedFloat4x3 {
		init(_ mat4x4: matrix_float4x4) {
				self.init(columns: (
						MTLPackedFloat3(mat4x4[0]),
						MTLPackedFloat3(mat4x4[1]),
						MTLPackedFloat3(mat4x4[2]),
						MTLPackedFloat3(mat4x4[3])
				))
		}
}
extension MTLPackedFloat3 {
		init(_ vec4: simd_float4) {
				self.init()
				self.elements = (vec4[0], vec4[1], vec4[2])
		}
}

And set whole transformationMatrix at once.

Code Block             anInstanceDescriptorPointer.pointee.transformationMatrix =  MTLPackedFloat4x3(aNodeInstance.transform)

0

Answer 2

John Lobe OP

Jul ’20

That is pretty interesting in a few different ways - And it compiles!
Thanks so much for your response & solution, particularly in the WWDC aftermath ~ JL

0

Answer 3

otabuzzman OP

Dec ’22

@OOPer, thx for sharing your code.

@JL, in a one-to-one port of Apple‘s ray tracer sample (Accelerating ray tracing using Metal) one would use the transpose of the matrix because the sample uses row-first order and MTLAccelerationStructureInstanceDescriptor.transformationMatrix expects column-first.

anInstanceDescriptorPointer.pointee.transformationMatrix = MTLPackedFloat4x3(aNodeInstance.transform.transpose)

Without transpose the render result seems to be a single Cornell Box but in fact are nine instances stacked at the same position.

A bit late, but in case anyone else will come across…

0