Scratch memory requests that are larger than one dword are using a
different memory layout than global instructions. Rather than being
placed contiguously, each dword is interleaved 64 lanes * 4 bytes away
as described in Section 9.1.5.2. "Swizzled Buffer Addressing" in the
MI300 specification. This was verified by comparing MI300 output (which
uses scratch_ instructions) with MI200 (which uses buffer instructions).
MI300 FashionMNIST bs=1 now matches CPU reference.
This requires several changes to the instruction implementations:
- For stores, data in the GPUDynInst can be swizzled before the data is
written to memory. This is easy to do using a helper method. This is
done in the template<int N> variant of initMemWrite. To use this x2
stores are changed to use template<int N> rather than loading a U64. The
swizzle function is renamed to swizzleAddr to avoid confusion with
swizzleData.
- For loads, data is unswizzled in completeAcc when writing register
values. This is not as easy to implement as a helper and is thus
implemented for the three load instructions that load more than one
dword.
- Accessing swizzled data requires at least one packet per dword. A new
GPU memory helper is added to create these packets for scratch requests
specifically. This is called in the template<int N> variant of
initMemRead / initMemWrite. Loads and stores of x2 are changed to use
this variant instead of accessing a U64.
The GPUDynInst status vector restrictions are increased to allow for
swizzled x4 accesses. For simplicity this does not currently support
misaligned swizzled accesses and will panic upon seeing such a case.
Change-Id: Ic686c51e28e0af029a043d5a5b3d4069f2cb94f9
7d46c50663