[mlir][nvgpu] Add tma last dim bytes check (#153451)
Add the check the number of bytes in the last dimension of Tma must be a multiple of 16.
This commit is contained in:
parent
87de48d11f
commit
6d08a39eeb
@ -422,6 +422,12 @@ std::optional<InFlightDiagnostic> verifyTmaDescriptorWithMemref(
|
||||
<< descMemref << " != " << dstMemref;
|
||||
}
|
||||
|
||||
int lastDimBytes =
|
||||
descMemref.getShape().back() * descMemref.getElementTypeBitWidth() / 8;
|
||||
if (lastDimBytes % 16 != 0) {
|
||||
return op->emitError() << "the bytes in the last dimension of the tensor "
|
||||
"map must be a multiple of 16";
|
||||
}
|
||||
return std::nullopt;
|
||||
}
|
||||
|
||||
|
@ -378,3 +378,14 @@ func.func @check_matrixC_dim(%arg0: vector<4x4xf16>, %arg1: vector<2x2xf16>, %ar
|
||||
%d = nvgpu.mma.sync (%arg0, %arg1, %arg2) {mmaShape = [16, 8, 16]} : (vector<4x4xf16>, vector<2x2xf16>, vector<4xf16>) -> vector<2x2xf16>
|
||||
return %d : vector<2x2xf16>
|
||||
}
|
||||
|
||||
// -----
|
||||
|
||||
!desc = !nvgpu.tensormap.descriptor<tensor = memref<32x8xi8,3>, swizzle=none, l2promo = none, oob = zero, interleave = none>
|
||||
!mbarrier = !nvgpu.mbarrier.group<memorySpace = #gpu.address_space<workgroup>>
|
||||
func.func @tma_last_dim_bytes(%desc: !desc, %buffer: memref<32x8xi8,3>, %mbarrier: !mbarrier) {
|
||||
%c0 = arith.constant 0 : index
|
||||
// expected-error @+1 {{the bytes in the last dimension of the tensor map must be a multiple of 16}}
|
||||
nvgpu.tma.async.load %desc[%c0, %c0], %mbarrier[%c0] to %buffer : !desc, !mbarrier -> memref<32x8xi8,3>
|
||||
return
|
||||
}
|
||||
|
Loading…
x
Reference in New Issue
Block a user