Merge remote-tracking branch 'origin/master' into elementwise-hyper-and-trig-funcs

Author: oleksandr-pavlyk

.github/workflows/conda-package.yml

@@ -18,7 +18,7 @@ jobs:
 
     strategy:
       matrix:
-        python: ['3.8', '3.9', '3.10']
+        python: ['3.8', '3.9', '3.10', '3.11']
     steps:
       - uses: actions/checkout@v3
         with:
@@ -63,7 +63,7 @@ jobs:
 
     strategy:
       matrix:
-        python: ['3.8', '3.9', '3.10']
+        python: ['3.8', '3.9', '3.10', '3.11']
     env:
       conda-bld: C:\Miniconda\conda-bld\win-64\
     steps:
@@ -102,7 +102,7 @@ jobs:
 
     strategy:
       matrix:
-        python: ['3.8', '3.9', '3.10']
+        python: ['3.8', '3.9', '3.10', '3.11']
         experimental: [false]
         runner: [ubuntu-20.04]
     continue-on-error: ${{ matrix.experimental }}
@@ -185,7 +185,7 @@ jobs:
         shell: cmd /C CALL {0}
     strategy:
       matrix:
-        python: ['3.8', '3.9', '3.10']
+        python: ['3.8', '3.9', '3.10', '3.11']
         experimental: [false]
         runner: [windows-latest]
     continue-on-error: ${{ matrix.experimental }}
@@ -300,7 +300,7 @@ jobs:
     runs-on: ubuntu-20.04
     strategy:
       matrix:
-        python: ['3.8', '3.9', '3.10']
+        python: ['3.8', '3.9', '3.10', '3.11']
     steps:
       - name: Download artifact
         uses: actions/download-artifact@v3
@@ -324,7 +324,7 @@ jobs:
     runs-on: windows-latest
     strategy:
       matrix:
-        python: ['3.8', '3.9', '3.10']
+        python: ['3.8', '3.9', '3.10', '3.11']
     steps:
       - name: Download artifact
         uses: actions/download-artifact@v3

conda-recipe/build.sh

@@ -1,10 +1,7 @@
 #!/bin/bash
 
-# Workaround to Klocwork overwriting LD_LIBRARY_PATH that was modified
-# by DPC++ compiler conda packages. Will need to be added to DPC++ compiler
-# activation scripts.
-export LDFLAGS="$LDFLAGS -Wl,-rpath,$PREFIX/lib"
-export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$PREFIX/compiler/lib/intel64_lin:$PREFIX/compiler/lib:$PREFIX/lib"
+# This is necessary to help DPC++ find Intel libraries such as SVML, IRNG, etc in build prefix
+export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:${BUILD_PREFIX}/lib"
 
 # Intel LLVM must cooperate with compiler and sysroot from conda
 echo "--gcc-toolchain=${BUILD_PREFIX} --sysroot=${BUILD_PREFIX}/${HOST}/sysroot -target ${HOST}" > icpx_for_conda.cfg

dpctl/tensor/_reduction.py

@@ -123,6 +123,10 @@ def sum(arr, axis=None, dtype=None, keepdims=False):
 
     res_usm_type = arr.usm_type
     if arr.size == 0:
+        if keepdims:
+            res_shape = res_shape + (1,) * red_nd
+            inv_perm = sorted(range(nd), key=lambda d: perm[d])
+            res_shape = tuple(res_shape[i] for i in inv_perm)
         return dpt.zeros(
             res_shape, dtype=res_dt, usm_type=res_usm_type, sycl_queue=q
         )

dpctl/tests/elementwise/test_floor_ceil_trunc.py

@@ -57,7 +57,7 @@ def test_floor_ceil_trunc_usm_type(np_call, dpt_call, usm_type):
     q = get_queue_or_skip()
 
     arg_dt = np.dtype("f4")
-    input_shape = (2, 2, 2, 10)
+    input_shape = (10, 10, 10, 10)
     X = dpt.empty(input_shape, dtype=arg_dt, usm_type=usm_type, sycl_queue=q)
     X[..., 0::2] = -0.4
     X[..., 1::2] = 0.7
@@ -67,8 +67,7 @@ def test_floor_ceil_trunc_usm_type(np_call, dpt_call, usm_type):
     assert Y.sycl_queue == X.sycl_queue
     assert Y.flags.c_contiguous
 
-    expected_Y = np.empty(input_shape, dtype=arg_dt)
-    expected_Y = np_call(np.float32(X))
+    expected_Y = np_call(dpt.asnumpy(X))
     tol = 8 * dpt.finfo(Y.dtype).resolution
     assert_allclose(dpt.asnumpy(Y), expected_Y, atol=tol, rtol=tol)
 
@@ -125,12 +124,6 @@ def test_floor_ceil_trunc_errors(dpt_call):
         y,
     )
 
-    x = dpt.zeros(2)
-    y = x
-    assert_raises_regex(
-        TypeError, "Input and output arrays have memory overlap", dpt_call, x, y
-    )
-
     x = dpt.zeros(2, dtype="float32")
     y = np.empty_like(x)
     assert_raises_regex(

dpctl/tests/test_sycl_kernel_submit.py

@@ -167,16 +167,21 @@ def test_async_submit():
     assert isinstance(kern2Kernel, dpctl_prog.SyclKernel)
 
     status_complete = dpctl.event_status_type.complete
-    n = 256 * 1024
-    X = dpt.empty((3, n), dtype="u4", usm_type="device", sycl_queue=q)
+
+    # choose input size based on capability of the device
+    f = q.sycl_device.max_work_group_size
+    n = f * 1024
+    n_alloc = 4 * n
+
+    X = dpt.empty((3, n_alloc), dtype="u4", usm_type="device", sycl_queue=q)
     first_row = dpctl_mem.as_usm_memory(X[0])
     second_row = dpctl_mem.as_usm_memory(X[1])
     third_row = dpctl_mem.as_usm_memory(X[2])
 
     p1, p2 = 17, 27
 
     async_detected = False
-    for _ in range(5):
+    for attempt in range(5):
         e1 = q.submit(
             kern1Kernel,
             [
@@ -209,19 +214,22 @@ def test_async_submit():
         e3_st = e3.execution_status
         e2_st = e2.execution_status
         e1_st = e1.execution_status
-        if not all(
-            [
-                e == status_complete
-                for e in (
-                    e1_st,
-                    e2_st,
-                    e3_st,
-                )
-            ]
-        ):
+        are_complete = [
+            e == status_complete
+            for e in (
+                e1_st,
+                e2_st,
+                e3_st,
+            )
+        ]
+        e3.wait()
+        if not all(are_complete):
             async_detected = True
-            e3.wait()
             break
+        else:
+            n = n * (1 if attempt % 2 == 0 else 2)
+            if n > n_alloc:
+                break
 
     assert async_detected, "No evidence of async submission detected, unlucky?"
     Xnp = dpt.asnumpy(X)
@@ -231,4 +239,4 @@ def test_async_submit():
         Xref[1, i] = (i * i * i) % p2
         Xref[2, i] = min(Xref[0, i], Xref[1, i])
 
-    assert np.array_equal(Xnp, Xref)
+    assert np.array_equal(Xnp[:, :n], Xref[:, :n])

dpctl/tests/test_tensor_sum.py

@@ -133,3 +133,26 @@ def test_sum_arg_out_dtype_scalar(arg_dtype, out_dtype):
     assert isinstance(r, dpt.usm_ndarray)
     assert r.dtype == dpt.dtype(out_dtype)
     assert dpt.asnumpy(r) == 1
+
+
+def test_sum_keepdims_zero_size():
+    """See gh-1293"""
+    get_queue_or_skip()
+    n = 10
+    a = dpt.ones((n, 0, n))
+
+    s1 = dpt.sum(a, keepdims=True)
+    assert s1.shape == (1, 1, 1)
+
+    s2 = dpt.sum(a, axis=(0, 1), keepdims=True)
+    assert s2.shape == (1, 1, n)
+
+    s3 = dpt.sum(a, axis=(1, 2), keepdims=True)
+    assert s3.shape == (n, 1, 1)
+
+    s4 = dpt.sum(a, axis=(0, 2), keepdims=True)
+    assert s4.shape == (1, 0, 1)
+
+    a0 = a[0]
+    s5 = dpt.sum(a0, keepdims=True)
+    assert s5.shape == (1, 1)