spectrometerHBH opened a new pull request, #14516:
URL: https://github.com/apache/tvm/pull/14516
This can enable us to compile a program with symbolic thread extent, which
is important to dyn kernels.
```py
@I.ir_module
class Module:
@T.prim_func
def matmul(var_rxplaceholder: T.handle, var_rxplaceholder_1: T.handle,
var_matmul: T.handle):
T.func_attr({"tir.noalias": T.bool(True)})
n = T.int64()
rxplaceholder = T.match_buffer(var_rxplaceholder, (n, T.int64(4)))
rxplaceholder_1 = T.match_buffer(var_rxplaceholder_1, (T.int64(4),
n))
matmul_1 = T.match_buffer(var_matmul, (n, n))
# with T.block("root"):
for i0 in T.thread_binding(n, thread="blockIdx.x"):
for i1 in T.thread_binding(n, thread="threadIdx.x"):
for k in range(T.int64(4)):
with T.block("matmul"):
v_i0, v_i1, v_k = T.axis.remap("SSR", [i0, i1, k])
T.reads(rxplaceholder[v_i0, v_k],
rxplaceholder_1[v_k, v_i1])
T.writes(matmul_1[v_i0, v_i1])
with T.init():
matmul_1[v_i0, v_i1] = T.float32(0)
matmul_1[v_i0, v_i1] = matmul_1[v_i0, v_i1] +
rxplaceholder[v_i0, v_k] * rxplaceholder_1[v_k, v_i1]
@R.function
def main(x: R.Tensor(("n", 4), dtype="float32"), y: R.Tensor((4, "n"),
dtype="float32")) -> R.Tensor(("n", "n"), dtype="float32"):
n = T.int64()
cls = Module
gv = R.call_tir(cls.matmul, (x, y), out_sinfo=R.Tensor((n, n),
dtype="float32"))
return gv
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
