del 32 padding size with cann version check-pytorch-AtomGit

已合并

del 32 padding size with cann version check #33482

关龙锋创建于 4月10日从glf2016/pytorch_9325： v2.8.0 合入到 Ascend/pytorch： v2.8.0

del 32 padding size with cann version check #33482

已合并

关龙锋创建于 4月10日

从glf2016/pytorch_9325 v2.8.0合入到Ascend/pytorch v2.8.0

共 2 个文件变更+49-12

Mtest/npu/test_allocator_envs.py+28-7

	@@ -1,29 +1,50 @@
1	import os	1	import os
		2	+import math
2	import shutil	3	import shutil
3	import unittest	4	import unittest
4	import torch	5	import torch
5	import torch_npu	6	import torch_npu
6		7
		8	+import torch_npu.npu.utils as utils
		9	+
7	from torch_npu.testing.testcase import TestCase, run_tests	10	from torch_npu.testing.testcase import TestCase, run_tests
8	from torch_npu.testing.common_utils import SupportedDevices	11	from torch_npu.testing.common_utils import SupportedDevices
9		12
10		13
11	class TestAllocator(TestCase):	14	class TestAllocator(TestCase):
12	- @SupportedDevices(['Ascend910B'])
13	def test_huge_memory_alloc_20M(self):	15	def test_huge_memory_alloc_20M(self):
14	prev = torch_npu.npu.memory_allocated()	16	prev = torch_npu.npu.memory_allocated()
15	a = torch.rand(1024 * 1024 * 40, dtype=torch.float32).npu()	17	a = torch.rand(1024 * 1024 * 40, dtype=torch.float32).npu()
16	- torch.npu.synchronize()
17	# 实际申请1G内存	18	# 实际申请1G内存
18	- self.assertEqual(torch_npu.npu.memory_allocated(), prev + ((4 * 40 * 1024 * 1024 + 32) // 512 + 1) * 512)	19	+ version = utils.get_cann_version(module="CANN")
		20	+ if (utils.get_soc_version() >= 260 and version >= "9.1.0"):
		21	+ self.assertEqual(torch_npu.npu.memory_allocated(), prev + math.ceil((4 * 40 * 1024 * 1024) / 512) * 512)
		22	+ else:
		23	+ self.assertEqual(torch_npu.npu.memory_allocated(),
		24	+ prev + math.ceil((4 * 40 * 1024 * 1024 + 32) / 512) * 512)
19		25
20	- @SupportedDevices(['Ascend910B'])
21	def test_huge_memory_alloc_512B(self):	26	def test_huge_memory_alloc_512B(self):
		27	+ os.environ["PYTORCH_NPU_ALLOC_CONF"] = "expandable_segments:False"
22	prev = torch_npu.npu.memory_allocated()	28	prev = torch_npu.npu.memory_allocated()
23	- a = torch.rand(8 * 8 * 16, dtype=torch.float32).npu() # 512B	29	+ a = torch.rand(8 * 8 * 16, dtype=torch.float32).npu() # 512B
24	- torch.npu.synchronize()
25	# 实际申请1M内存	30	# 实际申请1M内存
26	- self.assertEqual(torch_npu.npu.memory_allocated(), prev + ((8 * 8 * 16 * 4 + 32) // 512 + 1) * 512)	31	+ version = utils.get_cann_version(module="CANN")
		32	+ if (utils.get_soc_version() >= 260 and version >= "9.1.0"):
		33	+ self.assertEqual(torch_npu.npu.memory_allocated(), prev + math.ceil((8 * 8 * 16 * 4) / 512) * 512)
		34	+ else:
		35	+ self.assertEqual(torch_npu.npu.memory_allocated(), prev + math.ceil((8 * 8 * 16 * 4 + 32) / 512) * 512)
		36	+
		37	+ def test_huge_memory_alloc_512B_by_vm(self):
		38	+ os.environ["PYTORCH_NPU_ALLOC_CONF"] = "expandable_segments:True"
		39	+ prev = torch_npu.npu.memory_allocated()
		40	+ a = torch.rand(8 * 8 * 16, dtype=torch.float32).npu() # 512B
		41	+ # 实际申请1M内存
		42	+ version = utils.get_cann_version(module="CANN")
		43	+ if (utils.get_soc_version() >= 260 and version >= "9.1.0"):
		44	+ self.assertEqual(torch_npu.npu.memory_allocated(), prev + math.ceil((8 * 8 * 16 * 4) / 512) * 512)
		45	+ else:
		46	+ self.assertEqual(torch_npu.npu.memory_allocated(), prev + math.ceil((8 * 8 * 16 * 4 + 32) / 512) * 512)
		47	+ del os.environ["PYTORCH_NPU_ALLOC_CONF"]
27		48
28	if __name__ == '__main__':	49	if __name__ == '__main__':
29	run_tests()	50	run_tests()

Mtorch_npu/csrc/core/npu/NPUCachingAllocator.cpp+21-5

	@@ -104,6 +104,7 @@ constexpr size_t kRoundUpPowerOfTwoEnd = 1ULL << 36; // 64 GB
104	constexpr size_t kRoundUpPowerOfTwoIntervals = 16;	104	constexpr size_t kRoundUpPowerOfTwoIntervals = 16;
105	const std::string kMinCannVersion = "8.1.RC1"; // minimum cann version which supports 1g mem 8.1.RC1	105	const std::string kMinCannVersion = "8.1.RC1"; // minimum cann version which supports 1g mem 8.1.RC1
106	const std::string kMinDriverVersion = "25.0.RC1"; // minimum driver version which supports 1g mem 25.0.RC1	106	const std::string kMinDriverVersion = "25.0.RC1"; // minimum driver version which supports 1g mem 25.0.RC1
		107	+const std::string kMinDel32PaddingSizeCannVersion = "9.1.0";// minimum cann version which supports del 32 paddingsize
107	const std::string kCannModule = "CANN"; // cann module name	108	const std::string kCannModule = "CANN"; // cann module name
108	constexpr int kPrecision = 4; // precision of the memory usage information	109	constexpr int kPrecision = 4; // precision of the memory usage information
109	constexpr size_t kLazyQuerySize = 512; // lazy query event size	110	constexpr size_t kLazyQuerySize = 512; // lazy query event size
	@@ -186,6 +187,23 @@ bool IsMallocPage1GMem(bool is_small_pool)
186	return !is_small_pool && is_support_page_size_1g;	187	return !is_small_pool && is_support_page_size_1g;
187	}	188	}
188		189
		190	+size_t AddPadSize()
		191	+{
		192	+ static size_t add_size = -1;
		193	+ if (add_size == -1) {
		194	+ // 新增cann版本兼容
		195	+ if (GetSocVersion() >= SocVersion::Ascend950 &&
		196	+ IsGteCANNVersion(kMinDel32PaddingSizeCannVersion, kCannModule)) {
		197	+ add_size = 0;
		198	+ } else {
		199	+ TORCH_NPU_WARN_ONCE(
		200	+ "The current CANN and Soc versions require processing for 32 padding size, with memory allocation.");
		201	+ add_size = 32;
		202	+ }
		203	+ }
		204	+ return add_size;
		205	+}
		206	+
189	struct Block;	207	struct Block;
190	struct PrivatePool;	208	struct PrivatePool;
191	using Comparison = bool ()(const Block , const Block *);	209	using Comparison = bool ()(const Block , const Block *);
	@@ -2073,9 +2091,7 @@ public:
2073		2091
2074	static size_t round_size(size_t size)	2092	static size_t round_size(size_t size)
2075	{	2093	{
2076	- constexpr size_t kPadSize = 32;	2094	+ size += AddPadSize();
2077	- size += kPadSize;
2078	-
2079	if (size < kMinBlockSize) {	2095	if (size < kMinBlockSize) {
2080	return kMinBlockSize;	2096	return kMinBlockSize;
2081	} else {	2097	} else {
	@@ -3540,7 +3556,7 @@ public:
3540	if (size != 0) {	3556	if (size != 0) {
3541	if (c10_npu::option::OptionsManager::CheckForceUncached()) {	3557	if (c10_npu::option::OptionsManager::CheckForceUncached()) {
3542	deleteFunc = &uncached_delete;	3558	deleteFunc = &uncached_delete;
3543	- size_t alloc_size = size + 32;	3559	+ size_t alloc_size = size + AddPadSize();
3544	NPU_CHECK_ERROR(c10_npu::acl::AclrtMallocAlign32(&devPtr, alloc_size,	3560	NPU_CHECK_ERROR(c10_npu::acl::AclrtMallocAlign32(&devPtr, alloc_size,
3545	aclrtMemMallocPolicy::ACL_MEM_MALLOC_HUGE_FIRST));	3561	aclrtMemMallocPolicy::ACL_MEM_MALLOC_HUGE_FIRST));
3546	TORCH_NPU_MEMORY_LOGD("Without NPUCachingAllocator, malloc by "	3562	TORCH_NPU_MEMORY_LOGD("Without NPUCachingAllocator, malloc by "
	@@ -3570,7 +3586,7 @@ public:
3570	if (size != 0) {	3586	if (size != 0) {
3571	if (c10_npu::option::OptionsManager::CheckForceUncached()) {	3587	if (c10_npu::option::OptionsManager::CheckForceUncached()) {
3572	deleteFunc = &uncached_delete;	3588	deleteFunc = &uncached_delete;
3573	- size_t alloc_size = size + 32 + aligned;	3589	+ size_t alloc_size = size + AddPadSize() + aligned;
3574	NPU_CHECK_ERROR(c10_npu::acl::AclrtMallocAlign32(&realPtr, alloc_size,	3590	NPU_CHECK_ERROR(c10_npu::acl::AclrtMallocAlign32(&realPtr, alloc_size,
3575	aclrtMemMallocPolicy::ACL_MEM_MALLOC_HUGE_FIRST));	3591	aclrtMemMallocPolicy::ACL_MEM_MALLOC_HUGE_FIRST));
3576	TORCH_NPU_MEMORY_LOGD("Without NPUCachingAllocator, malloc by AclrtMallocAlign32: size=%zu", alloc_size);	3592	TORCH_NPU_MEMORY_LOGD("Without NPUCachingAllocator, malloc by AclrtMallocAlign32: size=%zu", alloc_size);