At-operator is matmul not dot (#1140)

* At-operator is matmul not dot

* Reduce dimensionality of inputs, to work under default max-dim=4

cmake/thirdparty/get_tblis.cmake

@@ -171,11 +171,11 @@ function(find_or_configure_tblis)
 endfunction()
 
 if(NOT DEFINED cunumeric_TBLIS_BRANCH)
-  set(cunumeric_TBLIS_BRANCH master)
+  set(cunumeric_TBLIS_BRANCH arm-build)
 endif()
 
 if(NOT DEFINED cunumeric_TBLIS_REPOSITORY)
-  set(cunumeric_TBLIS_REPOSITORY https://github.com/devinamatthews/tblis.git)
+  set(cunumeric_TBLIS_REPOSITORY https://github.com/nv-legate/tblis.git)
 endif()
 
 find_or_configure_tblis(VERSION          1.2.0

cunumeric/array.py

@@ -1122,6 +1122,20 @@ def __ilshift__(self, rhs: Any) -> ndarray:
 
         return left_shift(self, rhs, out=self)
 
+    def __imatmul__(self, rhs: Any) -> ndarray:
+        """a.__imatmul__(value, /)
+
+        Return ``self@=value``.
+
+        Availability
+        --------
+        Multiple GPUs, Multiple CPUs
+
+        """
+        from .module import matmul
+
+        return matmul(self, rhs, out=self)
+
     def __imod__(self, rhs: Any) -> ndarray:
         """a.__imod__(value, /)
 
@@ -1329,7 +1343,9 @@ def __matmul__(self, value: Any) -> ndarray:
         Multiple GPUs, Multiple CPUs
 
         """
-        return self.dot(value)
+        from .module import matmul
+
+        return matmul(self, value)
 
     def __mod__(self, rhs: Any) -> ndarray:
         """a.__mod__(value, /)
@@ -1551,6 +1567,20 @@ def __rfloordiv__(self, lhs: Any) -> ndarray:
 
         return floor_divide(lhs, self)
 
+    def __rmatmul__(self, lhs: Any) -> ndarray:
+        """a.__rmatmul__(value, /)
+
+        Return ``value@self``.
+
+        Availability
+        --------
+        Multiple GPUs, Multiple CPUs
+
+        """
+        from .module import matmul
+
+        return matmul(lhs, self)
+
     def __rmod__(self, lhs: Any) -> ndarray:
         """a.__rmod__(value, /)
 

tests/integration/test_matmul.py

@@ -16,6 +16,7 @@
 import numpy as np
 import pytest
 from legate.core import LEGATE_MAX_DIM
+from utils.comparisons import allclose
 from utils.contractions import (
     check_default,
     check_permutations,
@@ -29,7 +30,7 @@
 
 @pytest.mark.parametrize("a_ndim", range(1, LEGATE_MAX_DIM + 1))
 @pytest.mark.parametrize("b_ndim", range(1, LEGATE_MAX_DIM + 1))
-def test(a_ndim, b_ndim):
+def test_function(a_ndim, b_ndim):
     name = f"matmul({a_ndim} x {b_ndim})"
     modes = matmul_modes(a_ndim, b_ndim)
 
@@ -43,6 +44,57 @@ def operation(lib, *args, **kwargs):
         check_types(name, modes, operation)
 
 
+@pytest.mark.parametrize(
+    "a_shape",
+    (
+        (3, 4, 5),
+        (4, 5),
+        (5,),
+    ),
+)
+@pytest.mark.parametrize(
+    "b_shape",
+    (
+        (3, 5, 6),
+        (5, 6),
+        (5,),
+    ),
+)
+def test_operator(a_shape, b_shape):
+    np_a = np.random.random(a_shape)
+    np_b = np.random.random(b_shape)
+    num_a = num.array(np_a)
+    num_b = num.array(np_b)
+    assert allclose(np_a @ np_b, num_a @ num_b)
+
+
+@pytest.mark.parametrize(
+    "a_shape",
+    (
+        (3, 4, 5),
+        (4, 5),
+        (5,),
+    ),
+)
+@pytest.mark.parametrize(
+    "b_shape",
+    (
+        (3, 5, 5),
+        (5, 5),
+    ),
+)
+def test_inplace_operator(a_shape, b_shape):
+    if len(a_shape) < len(b_shape):
+        return
+    np_a = np.random.random(a_shape)
+    np_b = np.random.random(b_shape)
+    num_a = num.array(np_a)
+    num_b = num.array(np_b)
+    np_a @= np_b
+    num_a @= num_b
+    assert allclose(np_a, num_a)
+
+
 class TestMatmulErrors:
     @pytest.mark.parametrize(
         "shapesAB",