For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo function
wmma_cdna
def wmma_cdna[*, a_reuse: Bool = False, b_reuse: Bool = False](mut d: SIMD, a: SIMD, b: SIMD, c: SIMD)
Computes one CDNA5 WMMA, d = a * b + c, with the given reuse hints.
The instruction is picked from the dtypes and lane widths of the operands,
the same way the generic mma picks it. Every combination below is one
v_wmma instruction whose A and B operands live in registers, so the
fragments are the per-lane slices of the 16x16 output tile.
Supported combinations, as (A, B, C, D dtypes, lane widths of A/B/C/D):
- (float8_e4m3fn or float8_e5m2 for A and B, float16 or float32 for C and
D, 32/32/8/8),
16x16x64. A and B pair freely, giving thefp8.fp8,fp8.bf8,bf8.fp8andbf8.bf8instruction forms. - The same four pairings and accumulators at 64/64/8/8,
16x16x128. - (float16, float16, float16 or float32, 16/16/8/8) and (bfloat16,
bfloat16, bfloat16 or float32, 16/16/8/8),
16x16x32. - (bfloat16, bfloat16, float32, bfloat16, 16/16/8/8),
16x16x32: the mixed form whose f32 accumulator becomes a bf16 result. - (float32, float32, float32, 2/2/8/8),
16x16x4.
Constraints:
- Requires CDNA5 (gfx1250) or newer; the call fails to compile elsewhere.
- The lane widths of A, B, C and D must match the instruction shape for their dtypes, as listed above.
- A set reuse hint asserts that this WMMA reads the same A (or B) registers as the WMMA immediately before it in the final instruction order. The compiler may reorder WMMAs, so a caller must only set a hint where that order is guaranteed.
Parameters:
- a_reuse (
Bool): Hint that this WMMA reads the same A registers as the WMMA immediately before it in the final instruction order, which lets the matrix core skip reloading them. - b_reuse (
Bool): The same hint for the B registers.
Args: