I have chatted with @minminsun and his team these days. Just as then mentioned
https://github.com/dmlc/tvm/issues/4105#issuecomment-542032766. We can have
different frontends but only one backend. In my previous implement, users can
only use fragments with 16x16x16 shape and row-major layout. To solve this
problem, Minmin uses `new_expr`. Here I proposed a new design here. We use
attributes to transmit metadata. Here is an example:
```
// attr [A.shared.wmma.matrix_a] storage_scope = "wmma.matrix_a"
// attr [A.shared.wmma.matrix_a] fragment_layout = "row_major"
// attr [A.shared.wmma.matrix_a] fragment_shape = "16, 16, 16"
allocate A.shared.wmma.matrix_a[float16 * 1024]
```
This sulotion has been accepted by Minmin and his team. Thanks for supporting
my propsal.
Users can set these configuration in tensor intrinsics
``` Python
def intrin_wmma_load_matrix(scope):
n = 16
A = tvm.placeholder((n, n), name='A', dtype='float16')
BA = tvm.decl_buffer(A.shape, A.dtype, scope='shared', data_alignment=32,
offset_factor=256)
C = tvm.compute((n, n), lambda i, j: A[i, j], name='C')
BC = tvm.decl_buffer(C.shape, C.dtype, scope=scope, data_alignment=32,
offset_factor=256)
def intrin_func(ins, outs):
ib = tvm.ir_builder.create()
BA = ins[0]
BC = outs[0]
# shape (n, n, n) and 'row_major' layout
ib.emit(tvm.call_intrin('handle', 'tvm_load_matrix_sync',
BC.data, n, n, n, BC.elem_offset // 256,
BA.access_ptr('r'), n, 'row_major'))
return ib.get()
return tvm.decl_tensor_intrin(C.op, intrin_func, binds={A: BA, C: BC})
```
There is a new IR_pass will transmit these information from intrinsic to
attributes.
I am really happy to cooperate with Minmin's team. Thank you again for
contributing to TVM.
cc @tqchen
--
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/dmlc/tvm/issues/4052#issuecomment-542488197
