umangyadav
commented
3 months ago For exmaple:
Following is the code snippet before fuse_mlir pass for the test_gru_bidirct_3args test where @22 is a dot operation and only used once where it is sliced.
@22 = gpu::mlir_op[op=dot](@0,r), [mlir_dot1] -> float_type, {2, 10}, {10, 1}
@23 = slice[axes={1},starts={0},ends={5}](@21) -> float_type, {2, 5}, {15, 1}
@24 = contiguous(@23) -> float_type, {2, 5}, {5, 1}
@25 = slice[axes={1},starts={10},ends={15}](@21) -> float_type, {2, 5}, {15, 1}
@26 = contiguous(@25) -> float_type, {2, 5}, {5, 1}
@27 = slice[axes={1},starts={0},ends={5}](@22) -> float_type, {2, 5}, {10, 1}
RNN has patterns where Output of GEMM is immediately Sliced.
MIGraphX can slice inputs of the GEMMs instead and only compute portion of the GEMM that is useful instead.