aboutsummaryrefslogtreecommitdiff
path: root/gnu/packages/patches/python-pytorch-without-kineto.patch
blob: 896e11dcc26c50e7b8e36551945d29a8bcd6559e (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
Even when building without Kineto, the <ActivityType.h> header is still
imported and the ActivityType type is used. This patch was copied from
https://github.com/pytorch/pytorch/pull/111048 and adapted.

diff --git a/torch/csrc/profiler/kineto_shim.h b/torch/csrc/profiler/kineto_shim.h
index 7a3b788..7f49d18 100644
--- a/torch/csrc/profiler/kineto_shim.h
+++ b/torch/csrc/profiler/kineto_shim.h
@@ -12,7 +12,55 @@
 #undef USE_KINETO
 #endif
 
+#ifdef USE_KINETO
 #include <ActivityType.h>
+#else
+namespace libkineto {
+// copied from header
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+// Note : All activity types are not enabled by default. Please add them
+// at correct position in the enum
+enum class ActivityType {
+    // Activity types enabled by default
+    CPU_OP = 0, // cpu side ops
+    USER_ANNOTATION,
+    GPU_USER_ANNOTATION,
+    GPU_MEMCPY,
+    GPU_MEMSET,
+    CONCURRENT_KERNEL, // on-device kernels
+    EXTERNAL_CORRELATION,
+    CUDA_RUNTIME, // host side cuda runtime events
+    CUDA_DRIVER, // host side cuda driver events
+    CPU_INSTANT_EVENT, // host side point-like events
+    PYTHON_FUNCTION,
+    OVERHEAD, // CUPTI induced overhead events sampled from its overhead API.
+
+    // Optional Activity types
+    CUDA_SYNC, // synchronization events between runtime and kernels
+    GLOW_RUNTIME, // host side glow runtime events
+    MTIA_RUNTIME, // host side MTIA runtime events
+    CUDA_PROFILER_RANGE, // CUPTI Profiler range for performance metrics
+    MTIA_CCP_EVENTS, // MTIA ondevice CCP events
+    HPU_OP, // HPU host side runtime event
+    XPU_RUNTIME, // host side xpu runtime events
+    MTIA_WORKLOADD,
+
+    PRIVATEUSE1_RUNTIME,
+    PRIVATEUSE1_DRIVER,
+
+    ENUM_COUNT, // This is to add buffer and not used for any profiling logic. Add your new type before it.
+    OPTIONAL_ACTIVITY_TYPE_START = CUDA_SYNC,
+};
+}
+
+#endif
 
 #include <torch/csrc/Export.h>
 #include <torch/csrc/profiler/api.h>