Assignment 08: XDNA GEMM Kernel
This week you will write an XDNA kernel to perform a matrix multiplication on the NPU. The XDNA kernel is a tensor kernel as it operates on tensor layouts.
Prerequisites: Run the following to enter the development environment and install torch.
nix develop nix-amd-npu#iron-full
iron-fhs
pip install torch # only once
Data Layout and Data Movement
In the main memory the matrices are stored in row-major order (in0: MK, in1: KN, and out: MN with M=16, N=16, and K=64).
During the data movement from L3 (main memory) to L1 (scratchpad), the matrices are tiled.
The dimensions are split as follows:
M->pmwithp=2,m=8,N->qnwithq=2,n=8, andK->rkwithr=8,k=8.
This yields the views in0: pmrk, in1: rkqn, and out: pmqn.
During the data movement to the L1 scratchpad memory, the layout is changed to in0: prmk, in1: rqkn, and out: pqmn.
The NPU scratchpad memory is zero-initialized during NPU setup, so within the tensor kernel you may assume that the output tensor memory is already zero.
When writing the output tensor, its layout is changed back to a matrix layout (out: MN).
Task 1 — Verify Function
Implement the verify() function for the matrix multiplication in src/driver.py.
Add torch.manual_seed(42) before the tensor initialization.
Set the maximal absolute error to 0.5 and the maximum relative error to 0.2 when comparing the tensors.
Task 2 — Instructions and Latencies
Fill in the table below with the instructions you will need for your tensor kernel.
Instruction |
Slot |
Latency |
|---|---|---|
Task 3 — Register Blocking
Choose a register blocking for your tensor kernel.
Assign the input and output tensors to the registers and explain your decision.
Keep in mind that the input tensors must be converted from BF16 to BFP16 (bfp16ebs8).
Note: The same register may be reused for multiple tensors when their lifetimes do not overlap.
Tensor |
Registers |
|---|---|
|
|
|
|
|
Task 4 — Data Layouts and Pointer Updates
Sketch the data layout and the required pointer updates corresponding to your register blocking.
Task 5 — Implementation
Implement the tensor kernel in
src/matmul.s. Do not use any control-flow instruction other than the finalret lr.Verify your kernel with:
make run_matmul
Task 6 — Performance
Count your instructions. What performance should your kernel achieve?
Argue whether your instruction count is minimal, or describe which optimizations could further reduce it.