|
| 1 | +# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. |
| 2 | +# SPDX-License-Identifier: Apache-2.0 |
| 3 | + |
| 4 | +import cuda.core.experimental |
| 5 | +import pytest |
| 6 | +from cuda.core.experimental import Device, DeviceMemoryResource |
| 7 | +from cuda.core.experimental._utils.cuda_utils import CUDAError |
| 8 | +from helpers.buffers import PatternGen, compare_buffer_to_constant, make_scratch_buffer |
| 9 | + |
| 10 | +NBYTES = 1024 |
| 11 | + |
| 12 | + |
| 13 | +def _mempool_device_impl(num): |
| 14 | + num_devices = len(cuda.core.experimental.system.devices) |
| 15 | + if num_devices < num: |
| 16 | + pytest.skip("Test requires at least {num} GPUs") |
| 17 | + |
| 18 | + devs = [Device(i) for i in range(num)] |
| 19 | + for i in reversed(range(num)): |
| 20 | + devs[i].set_current() |
| 21 | + |
| 22 | + if not all(devs[i].can_access_peer(j) for i in range(num) for j in range(num)): |
| 23 | + pytest.skip("Test requires GPUs with peer access") |
| 24 | + |
| 25 | + if not all(devs[i].properties.memory_pools_supported for i in range(num)): |
| 26 | + pytest.skip("Device does not support mempool operations") |
| 27 | + |
| 28 | + return devs |
| 29 | + |
| 30 | + |
| 31 | +@pytest.fixture |
| 32 | +def mempool_device_x2(): |
| 33 | + """Fixture that provides two devices if available, otherwise skips test.""" |
| 34 | + return _mempool_device_impl(2) |
| 35 | + |
| 36 | + |
| 37 | +@pytest.fixture |
| 38 | +def mempool_device_x3(): |
| 39 | + """Fixture that provides three devices if available, otherwise skips test.""" |
| 40 | + return _mempool_device_impl(3) |
| 41 | + |
| 42 | + |
| 43 | +def test_peer_access_basic(mempool_device_x2): |
| 44 | + """Basic tests for dmr.peer_accessible_by.""" |
| 45 | + dev0, dev1 = mempool_device_x2 |
| 46 | + zero_on_dev0 = make_scratch_buffer(dev0, 0, NBYTES) |
| 47 | + one_on_dev0 = make_scratch_buffer(dev0, 1, NBYTES) |
| 48 | + stream_on_dev0 = dev0.create_stream() |
| 49 | + dmr_on_dev1 = DeviceMemoryResource(dev1) |
| 50 | + buf_on_dev1 = dmr_on_dev1.allocate(NBYTES) |
| 51 | + |
| 52 | + # No access at first. |
| 53 | + assert 0 not in dmr_on_dev1.peer_accessible_by |
| 54 | + with pytest.raises(CUDAError, match="CUDA_ERROR_INVALID_VALUE"): |
| 55 | + one_on_dev0.copy_to(buf_on_dev1, stream=stream_on_dev0) |
| 56 | + |
| 57 | + with pytest.raises(CUDAError, match="CUDA_ERROR_INVALID_VALUE"): |
| 58 | + zero_on_dev0.copy_from(buf_on_dev1, stream=stream_on_dev0) |
| 59 | + |
| 60 | + # Allow access to device 1's allocations from device 0. |
| 61 | + dmr_on_dev1.peer_accessible_by = [dev0] |
| 62 | + assert 0 in dmr_on_dev1.peer_accessible_by |
| 63 | + compare_buffer_to_constant(zero_on_dev0, 0) |
| 64 | + one_on_dev0.copy_to(buf_on_dev1, stream=stream_on_dev0) |
| 65 | + zero_on_dev0.copy_from(buf_on_dev1, stream=stream_on_dev0) |
| 66 | + stream_on_dev0.sync() |
| 67 | + compare_buffer_to_constant(zero_on_dev0, 1) |
| 68 | + |
| 69 | + # Revoke access |
| 70 | + dmr_on_dev1.peer_accessible_by = [] |
| 71 | + assert 0 not in dmr_on_dev1.peer_accessible_by |
| 72 | + with pytest.raises(CUDAError, match="CUDA_ERROR_INVALID_VALUE"): |
| 73 | + one_on_dev0.copy_to(buf_on_dev1, stream=stream_on_dev0) |
| 74 | + |
| 75 | + with pytest.raises(CUDAError, match="CUDA_ERROR_INVALID_VALUE"): |
| 76 | + zero_on_dev0.copy_from(buf_on_dev1, stream=stream_on_dev0) |
| 77 | + |
| 78 | + |
| 79 | +def test_peer_access_property_x2(mempool_device_x2): |
| 80 | + """The the dmr.peer_accessible_by property (but not its functionality).""" |
| 81 | + # The peer access list is a sorted tuple and always excludes the self |
| 82 | + # device. |
| 83 | + dev0, dev1 = mempool_device_x2 |
| 84 | + dmr = DeviceMemoryResource(dev0) |
| 85 | + |
| 86 | + def check(expected): |
| 87 | + assert isinstance(dmr.peer_accessible_by, tuple) |
| 88 | + assert dmr.peer_accessible_by == expected |
| 89 | + |
| 90 | + # No access to begin with. |
| 91 | + check(expected=()) |
| 92 | + # fmt: off |
| 93 | + dmr.peer_accessible_by = (0,) ; check(expected=()) # noqa: E702 |
| 94 | + dmr.peer_accessible_by = (1,) ; check(expected=(1,)) # noqa: E702 |
| 95 | + dmr.peer_accessible_by = (0, 1) ; check(expected=(1,)) # noqa: E702 |
| 96 | + dmr.peer_accessible_by = () ; check(expected=()) # noqa: E702 |
| 97 | + dmr.peer_accessible_by = [0, 1] ; check(expected=(1,)) # noqa: E702 |
| 98 | + dmr.peer_accessible_by = set() ; check(expected=()) # noqa: E702 |
| 99 | + dmr.peer_accessible_by = [1, 1, 1, 1, 1] ; check(expected=(1,)) # noqa: E702 |
| 100 | + # fmt: on |
| 101 | + |
| 102 | + with pytest.raises(ValueError, match=r"device_id must be \>\= 0"): |
| 103 | + dmr.peer_accessible_by = [-1] # device ID out of bounds |
| 104 | + |
| 105 | + num_devices = len(cuda.core.experimental.system.devices) |
| 106 | + |
| 107 | + with pytest.raises(ValueError, match=r"device_id must be within \[0, \d+\)"): |
| 108 | + dmr.peer_accessible_by = [num_devices] # device ID out of bounds |
| 109 | + |
| 110 | + |
| 111 | +def test_peer_access_transitions(mempool_device_x3): |
| 112 | + """Advanced tests for dmr.peer_accessible_by.""" |
| 113 | + |
| 114 | + # Check all transitions between peer access states. The implementation |
| 115 | + # performs transactions that add or remove access as needed. This test |
| 116 | + # ensures that that is working as expected. |
| 117 | + |
| 118 | + # Doing everything from the point-of-view of device 0, there are four |
| 119 | + # access states: |
| 120 | + # |
| 121 | + # [(), (1,), (2,), (1, 2)] |
| 122 | + # |
| 123 | + # and 4^2-4 = 12 non-identity transitions. |
| 124 | + |
| 125 | + devs = mempool_device_x3 # Three devices |
| 126 | + |
| 127 | + # Allocate per-device resources. |
| 128 | + streams = [dev.create_stream() for dev in devs] |
| 129 | + pgens = [PatternGen(devs[i], NBYTES, streams[i]) for i in range(3)] |
| 130 | + dmrs = [DeviceMemoryResource(dev) for dev in devs] |
| 131 | + bufs = [dmr.allocate(NBYTES) for dmr in dmrs] |
| 132 | + |
| 133 | + def verify_state(state, pattern_seed): |
| 134 | + """ |
| 135 | + Verify an access state from the POV of device 0. E.g., (1,) means |
| 136 | + device 1 has access but device 2 does not. |
| 137 | + """ |
| 138 | + # Populate device 0's buffer with a new pattern. |
| 139 | + devs[0].set_current() |
| 140 | + pgens[0].fill_buffer(bufs[0], seed=pattern_seed) |
| 141 | + streams[0].sync() |
| 142 | + |
| 143 | + for peer in [1, 2]: |
| 144 | + devs[peer].set_current() |
| 145 | + if peer in state: |
| 146 | + # Peer device has access to 0's allocation |
| 147 | + bufs[peer].copy_from(bufs[0], stream=streams[peer]) |
| 148 | + # Check the result on the peer device. |
| 149 | + pgens[peer].verify_buffer(bufs[peer], seed=pattern_seed) |
| 150 | + else: |
| 151 | + # Peer device has no access to 0's allocation |
| 152 | + with pytest.raises(CUDAError, match="CUDA_ERROR_INVALID_VALUE"): |
| 153 | + bufs[peer].copy_from(bufs[0], stream=streams[peer]) |
| 154 | + |
| 155 | + # For each transition, set the access state before and after, checking for |
| 156 | + # the expected peer access capabilities at each stop. |
| 157 | + pattern_seed = 0 |
| 158 | + states = [(), (1,), (2,), (1, 2)] |
| 159 | + transitions = [(s0, s1) for s0 in states for s1 in states if s0 != s1] |
| 160 | + for init_state, final_state in transitions: |
| 161 | + dmrs[0].peer_accessible_by = init_state |
| 162 | + assert dmrs[0].peer_accessible_by == init_state |
| 163 | + verify_state(init_state, pattern_seed) |
| 164 | + pattern_seed += 1 |
| 165 | + |
| 166 | + dmrs[0].peer_accessible_by = final_state |
| 167 | + assert dmrs[0].peer_accessible_by == final_state |
| 168 | + verify_state(final_state, pattern_seed) |
| 169 | + pattern_seed += 1 |
0 commit comments