b1262d13e0
The current implementation of the WMMA intrinsic ops as they are defined in the ROCDL tablegen is incorrect. They represent as operands what should be attributes such as `clamp`, `opsel`, `signA/signB`. This change performs a refactoring to bring it in line with what we expect. --------- Signed-off-by: Muzammiluddin Syed <muzasyed@amd.com>
41 lines
1.3 KiB
Python
41 lines
1.3 KiB
Python
# RUN: %PYTHON %s | FileCheck %s
|
|
# This is just a smoke test that the dialect is functional.
|
|
from array import array
|
|
|
|
from mlir.ir import *
|
|
from mlir.dialects import rocdl, arith
|
|
from mlir.extras import types as T
|
|
|
|
|
|
def constructAndPrintInModule(f):
|
|
print("\nTEST:", f.__name__)
|
|
with Context(), Location.unknown():
|
|
module = Module.create()
|
|
with InsertionPoint(module.body):
|
|
f()
|
|
print(module)
|
|
return f
|
|
|
|
|
|
# CHECK-LABEL: testSmoke
|
|
@constructAndPrintInModule
|
|
def testSmoke():
|
|
v_len = 16
|
|
f32 = F32Type.get()
|
|
# Note: this isn't actually the right type for the intrinsic (should be f16)
|
|
# but array doesn't support f16.
|
|
v16f32 = T.vector(v_len, f32)
|
|
f32_array = array("f", [0.0] * v_len)
|
|
a_frag = arith.constant(v16f32, f32_array)
|
|
b_frag = arith.constant(v16f32, f32_array)
|
|
c_frag = arith.constant(v16f32, f32_array)
|
|
|
|
c_frag = rocdl.wmma_f16_16x16x16_f16(v16f32, a_frag, b_frag, c_frag, opsel=False)
|
|
# CHECK: %{{.*}} = "rocdl.wmma.f16.16x16x16.f16"
|
|
print(c_frag)
|
|
assert isinstance(c_frag, OpView)
|
|
# CHECK: Value(%{{.*}} = "rocdl.wmma.f16.16x16x16.f16"
|
|
c_frag = rocdl.wmma_f16_16x16x16_f16_(v16f32, a_frag, b_frag, c_frag, opsel=False)
|
|
print(c_frag)
|
|
assert isinstance(c_frag, Value)
|