Skip to contentSkip to conversation

drm/amdkfd: Let VRAM allocations go to GTT domain on small APUs

PATCH Superseded Lang Yu 2 participants 7 messages Apr 26, 2024 – Apr 29, 2024

Subsystems (3)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Lang Yu

author
#0 lore raw

PATCH 1/2 drm/amdkfd: Let VRAM allocations go to GTT domain on small APUs

Small APUs(i.e., consumer, embedded products) usually have a small carveout device memory which can't satisfy most compute workloads memory allocation requirements.

We can't even run a Basic MNIST Example with a default 512MB carveout.
https://github.com/pytorch/examples/tree/main/mnist.

Though we can change BIOS settings to enlarge carveout size, which is inflexible and may bring complaint. On the other hand, the memory resource can't be effectively used between host and device.

The solution is MI300A approach, i.e., let VRAM allocations go to GTT.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c    |  6 +++++-
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 21 +++++++++++--------
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c      |  2 +-
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c          |  6 ++++--
 drivers/gpu/drm/amd/amdkfd/kfd_svm.h          |  3 ++-
 5 files changed, 24 insertions(+), 14 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +5 -1
index 7ba05f030dd1..3295838e9a1d 100644
@@ -456,7 +456,9 @@ void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev,
 			mem_info->local_mem_size_private =
 					KFD_XCP_MEMORY_SIZE(adev, xcp->id);
 	} else {
-		mem_info->local_mem_size_public = adev->gmc.visible_vram_size;
+		mem_info->local_mem_size_public = adev->flags & AMD_IS_APU ?
+						  (ttm_tt_pages_limit() << PAGE_SHIFT) :
+						  adev->gmc.visible_vram_size;
 		mem_info->local_mem_size_private = adev->gmc.real_vram_size -
 						adev->gmc.visible_vram_size;
 	}
@@ -824,6 +826,8 @@ u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id)
 		}
 		do_div(tmp, adev->xcp_mgr->num_xcp_per_mem_partition);
 		return ALIGN_DOWN(tmp, PAGE_SIZE);
+	} else if (adev->flags & AMD_IS_APU) {
+		return (ttm_tt_pages_limit() << PAGE_SHIFT);
 	} else {
 		return adev->gmc.real_vram_size;
 	}
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +12 -9
index c4f9960dafbb..7eb5afcc4895 100644
@@ -196,7 +196,7 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 			return -EINVAL;
 
 		vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
-		if (adev->gmc.is_app_apu) {
+		if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
 			system_mem_needed = size;
 			ttm_mem_needed = size;
 		}
@@ -232,7 +232,8 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 		  "adev reference can't be null when vram is used");
 	if (adev && xcp_id >= 0) {
 		adev->kfd.vram_used[xcp_id] += vram_needed;
-		adev->kfd.vram_used_aligned[xcp_id] += adev->gmc.is_app_apu ?
+		adev->kfd.vram_used_aligned[xcp_id] +=
+				(adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) ?
 				vram_needed :
 				ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
 	}
@@ -260,7 +261,7 @@ void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
 
 		if (adev) {
 			adev->kfd.vram_used[xcp_id] -= size;
-			if (adev->gmc.is_app_apu) {
+			if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
 				adev->kfd.vram_used_aligned[xcp_id] -= size;
 				kfd_mem_limit.system_mem_used -= size;
 				kfd_mem_limit.ttm_mem_used -= size;
@@ -889,7 +890,7 @@ static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
 	 * if peer device has large BAR. In contrast, access over xGMI is
 	 * allowed for both small and large BAR configurations of peer device
 	 */
-	if ((adev != bo_adev && !adev->gmc.is_app_apu) &&
+	if ((adev != bo_adev && !(adev->gmc.is_app_apu || adev->flags & AMD_IS_APU)) &&
 	    ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
@@ -1657,7 +1658,7 @@ size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
 		- atomic64_read(&adev->vram_pin_size)
 		- reserved_for_pt;
 
-	if (adev->gmc.is_app_apu) {
+	if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
 		system_mem_available = no_system_mem_limit ?
 					kfd_mem_limit.max_system_mem_limit :
 					kfd_mem_limit.max_system_mem_limit -
@@ -1669,6 +1670,7 @@ size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
 		available = min3(system_mem_available, ttm_mem_available,
 				 vram_available);
 		available = ALIGN_DOWN(available, PAGE_SIZE);
+
 	} else {
 		available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN);
 	}
@@ -1705,7 +1707,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
 
-		if (adev->gmc.is_app_apu) {
+		if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
 			domain = AMDGPU_GEM_DOMAIN_GTT;
 			alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
 			alloc_flags = 0;
@@ -1952,7 +1954,7 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
 	if (size) {
 		if (!is_imported &&
 		   (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM ||
-		   (adev->gmc.is_app_apu &&
+		   ((adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) &&
 		    mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT)))
 			*size = bo_size;
 		else
@@ -2374,8 +2376,9 @@ static int import_obj_create(struct amdgpu_device *adev,
 	(*mem)->dmabuf = dma_buf;
 	(*mem)->bo = bo;
 	(*mem)->va = va;
-	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ?
-		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
+	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) &&
+			 !(adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) ?
+			 AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
 
 	(*mem)->mapped_to_gpu_memory = 0;
 	(*mem)->process_info = avm->process_info;
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +1 -1
index 4bcfbeac48fb..4816fcb9803a 100644
@@ -1023,7 +1023,7 @@ int kgd2kfd_init_zone_device(struct amdgpu_device *adev)
 	if (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 1))
 		return -EINVAL;
 
-	if (adev->gmc.is_app_apu)
+	if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU)
 		return 0;
 
 	pgmap = &kfddev->pgmap;
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +4 -2
index bfab16b43fec..238ac11bb97d 100644
@@ -2619,7 +2619,8 @@ svm_range_best_restore_location(struct svm_range *prange,
 		return -1;
 	}
 
-	if (node->adev->gmc.is_app_apu)
+	if (node->adev->gmc.is_app_apu ||
+	    node->adev->flags & AMD_IS_APU)
 		return 0;
 
 	if (prange->preferred_loc == gpuid ||
@@ -3337,7 +3338,8 @@ svm_range_best_prefetch_location(struct svm_range *prange)
 		goto out;
 	}
 
-	if (bo_node->adev->gmc.is_app_apu) {
+	if (bo_node->adev->gmc.is_app_apu ||
+	    bo_node->adev->flags & AMD_IS_APU) {
 		best_loc = 0;
 		goto out;
 	}
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +2 -1
index 026863a0abcd..9c37bd0567ef 100644
@@ -201,7 +201,8 @@ void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_s
  * is initialized to not 0 when page migration register device memory.
  */
 #define KFD_IS_SVM_API_SUPPORTED(adev) ((adev)->kfd.pgmap.type != 0 ||\
-					(adev)->gmc.is_app_apu)
+					(adev)->gmc.is_app_apu ||\
+					((adev)->flags & AMD_IS_APU))
 
 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo);
 
Signature · 1 line
2.25.1
author
#1 lore raw

PATCH 2/2 drm/amdkfd: Allow memory oversubscription on small APUs

The default ttm_tt_pages_limit is 1/2 of system memory. It is prone to out of memory with such a configuration.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c       |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       |  4 ++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 12 +++++++++---
 3 files changed, 12 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +1 -1
index 3295838e9a1d..c01c6f3ab562 100644
@@ -167,7 +167,7 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 	int i;
 	int last_valid_bit;
 
-	amdgpu_amdkfd_gpuvm_init_mem_limits();
+	amdgpu_amdkfd_gpuvm_init_mem_limits(adev);
 
 	if (adev->kfd.dev) {
 		struct kgd2kfd_shared_resources gpu_resources = {
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +2 -2
index 1de021ebdd46..13284dbd8c58 100644
@@ -363,7 +363,7 @@ u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id);
 
 
 #if IS_ENABLED(CONFIG_HSA_AMD)
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device *adev);
 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
 				struct amdgpu_vm *vm);
 
@@ -376,7 +376,7 @@ void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo);
 void amdgpu_amdkfd_reserve_system_mem(uint64_t size);
 #else
 static inline
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device *adev)
 {
 }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +9 -3
index 7eb5afcc4895..a3e623a320b3 100644
@@ -60,6 +60,7 @@ static struct {
 	int64_t system_mem_used;
 	int64_t ttm_mem_used;
 	spinlock_t mem_limit_lock;
+	bool alow_oversubscribe;
 } kfd_mem_limit;
 
 static const char * const domain_bit_to_string[] = {
@@ -110,7 +111,7 @@ static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_ad
  *  System (TTM + userptr) memory - 15/16th System RAM
  *  TTM memory - 3/8th System RAM
  */
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device *adev)
 {
 	struct sysinfo si;
 	uint64_t mem;
@@ -130,6 +131,7 @@ void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
 		kfd_mem_limit.max_system_mem_limit -= AMDGPU_RESERVE_MEM_LIMIT;
 
 	kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT;
+	kfd_mem_limit.alow_oversubscribe = !!(adev->flags & AMD_IS_APU);
 	pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
 		(kfd_mem_limit.max_system_mem_limit >> 20),
 		(kfd_mem_limit.max_ttm_mem_limit >> 20));
@@ -221,8 +223,12 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 	     kfd_mem_limit.max_ttm_mem_limit) ||
 	    (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed >
 	     vram_size - reserved_for_pt - atomic64_read(&adev->vram_pin_size))) {
-		ret = -ENOMEM;
-		goto release;
+		if (kfd_mem_limit.alow_oversubscribe) {
+			pr_warn_ratelimited("Memory is getting oversubscried.\n");
+		} else {
+			ret = -ENOMEM;
+			goto release;
+		}
 	}
 
 	/* Update memory accounting by decreasing available system
Signature · 1 line
2.25.1
maintainer
#2 lore raw

Re: [PATCH 2/2] drm/amdkfd: Allow memory oversubscription on small APUs

Quoted 2 lines · Lang Yu

The default ttm_tt_pages_limit is 1/2 of system memory. It is prone to out of memory with such a configuration.

Indiscriminately allowing the violation of all memory limits is not a good solution. It will lead to poor performance once you actually reach ttm_pages_limit and TTM starts swapping out BOs.

Regards,
  Felix

Quoted 87 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c       |  2 +-
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       |  4 ++--
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 12 +++++++++---
  3 files changed, 12 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +1 -1
index 3295838e9a1d..c01c6f3ab562 100644
@@ -167,7 +167,7 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
  	int i;
  	int last_valid_bit;
 
-	amdgpu_amdkfd_gpuvm_init_mem_limits();
+	amdgpu_amdkfd_gpuvm_init_mem_limits(adev);
 
  	if (adev->kfd.dev) {
  		struct kgd2kfd_shared_resources gpu_resources = {
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +2 -2
index 1de021ebdd46..13284dbd8c58 100644
@@ -363,7 +363,7 @@ u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id);
 
 
  #if IS_ENABLED(CONFIG_HSA_AMD)
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device *adev);
  void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
  				struct amdgpu_vm *vm);
 
@@ -376,7 +376,7 @@ void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo);
  void amdgpu_amdkfd_reserve_system_mem(uint64_t size);
  #else
  static inline
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device *adev)
  {
  }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +9 -3
index 7eb5afcc4895..a3e623a320b3 100644
@@ -60,6 +60,7 @@ static struct {
  	int64_t system_mem_used;
  	int64_t ttm_mem_used;
  	spinlock_t mem_limit_lock;
+	bool alow_oversubscribe;
  } kfd_mem_limit;
 
  static const char * const domain_bit_to_string[] = {
@@ -110,7 +111,7 @@ static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_ad
   *  System (TTM + userptr) memory - 15/16th System RAM
   *  TTM memory - 3/8th System RAM
   */
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device *adev)
  {
  	struct sysinfo si;
  	uint64_t mem;
@@ -130,6 +131,7 @@ void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
  		kfd_mem_limit.max_system_mem_limit -= AMDGPU_RESERVE_MEM_LIMIT;
 
  	kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT;
+	kfd_mem_limit.alow_oversubscribe = !!(adev->flags & AMD_IS_APU);
  	pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
  		(kfd_mem_limit.max_system_mem_limit >> 20),
  		(kfd_mem_limit.max_ttm_mem_limit >> 20));
@@ -221,8 +223,12 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
  	     kfd_mem_limit.max_ttm_mem_limit) ||
  	    (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed >
  	     vram_size - reserved_for_pt - atomic64_read(&adev->vram_pin_size))) {
-		ret = -ENOMEM;
-		goto release;
+		if (kfd_mem_limit.alow_oversubscribe) {
+			pr_warn_ratelimited("Memory is getting oversubscried.\n");
+		} else {
+			ret = -ENOMEM;
+			goto release;
+		}
  	}
 
  	/* Update memory accounting by decreasing available system
author
#3 lore raw

RE: [PATCH 2/2] drm/amdkfd: Allow memory oversubscription on small APUs

[Public]

Quoted 10 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Saturday, April 27, 2024 6:45 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Yang, Philip <[email protected]>; Koenig, Christian <[email protected]>; Zhang, Yifan <[email protected]>; Liu, Aaron <[email protected]>
Subject: Re: [PATCH 2/2] drm/amdkfd: Allow memory oversubscription on small APUs

Quoted 6 lines
Quoted 2 lines · Lang Yu

The default ttm_tt_pages_limit is 1/2 of system memory. It is prone to out of memory with such a configuration.

Indiscriminately allowing the violation of all memory limits is not a good solution. It will lead to poor performance once you actually reach ttm_pages_limit and TTM starts swapping out BOs.

Hi Felix,

I just feel it's like a bug that 1/2 of system memory is fee, the driver tells users out of memory. On the other hand, if memory is available, why not use it.

By the way, can we use USERPTR for VRAM allocations? Then we don't have ttm_tt_pages_limit limitations. Thanks.

I actually did some tests on Strix (12 CU@2100 MHz, 29412M 128bits LPDDR5@937MHz) with
https://github.com/ROCm/pytorch-micro-benchmarking.

Command: python micro_benchmarking_pytorch.py --network resnet50 --batch-size=64 --iterations=20

1, Run 1 resnet50 (FP32, batch size 64)
Memory usage:

        System mem used 6748M out of 29412M
        TTM mem used 6658M out of 15719M

Memory oversubscription percentage: 0
Throughput [img/sec] : 49.04

2, Run 2 resnet50 simultaneously (FP32, batch size 64) Memory usage:

        System mem used 13496M out of 29412M
        TTM mem used 13316M out of 15719M

Memory oversubscription percentage: 0 Throughput [img/sec] (respectively) : 25.27 / 26.70

3, Run 3 resnet50 simultaneously (FP32, batch size 64) Memory usage:

        System mem used 20245M out of 29412M
        TTM mem used 19974M out of 15719M

Memory oversubscription percentage: ~27%

Throughput [img/sec](respectively) : 10.62 / 7.47 / 6.90 (In theory: 16 / 16 / 16)

From my observations,

1, GPU is underutilized a lot, sometimes its loading is less than 50% and even 0, when running 3 resnet50 simultaneously with ~27% memory oversubscription. The driver is busying evicting and restoring process. It takes ~2-5 seconds to restore all the BOs for one process (swap in and out BOs, actually allocate and copy pages), even though the process doesn't need all the allocated BOs to be resident.

2, Sometimes, the fairness can't be guaranteed between process when memory is oversubscribed. They can't share the GPU equally when created with default priority.

3, The less GPU underutilization time during evicting and restoring, the less performance degradation under memory oversubscription.

Regards,
Lang

Quoted 109 lines

Regards,
Felix

Quoted 6 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c       |  2 +-
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       |  4 ++--
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 12

+++++++++---

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c void amdgpu_amdkfd_device_init(struct

Quoted 8 lines
  3 files changed, 12 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +0 -0
index 3295838e9a1d..c01c6f3ab562 100644
@@ -167,7 +167,7 @@ void amdgpu_amdkfd_device_init(struct

amdgpu_device *adev)

Quoted 14 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +1 -1
     int i;
     int last_valid_bit;
 
-    amdgpu_amdkfd_gpuvm_init_mem_limits();
+    amdgpu_amdkfd_gpuvm_init_mem_limits(adev);
 
     if (adev->kfd.dev) {
             struct kgd2kfd_shared_resources gpu_resources = { diff --git

a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 1de021ebdd46..13284dbd8c58 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h u64 amdgpu_amdkfd_xcp_memory_size(struct · line 367

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
@@ -363,7 +363,7 @@ u64 amdgpu_amdkfd_xcp_memory_size(struct
 amdgpu_device *adev, int xcp_id);
 
 
  #if IS_ENABLED(CONFIG_HSA_AMD)
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device

*adev);

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h void amdgpu_amdkfd_release_notify(struct

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0 -0
  void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
                             struct amdgpu_vm *vm);
@@ -376,7 +376,7 @@ void amdgpu_amdkfd_release_notify(struct

amdgpu_bo *bo);

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo);

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
  void amdgpu_amdkfd_reserve_system_mem(uint64_t size);
  #else
  static inline
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device

*adev)

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0 -0
  {
  }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
index 7eb5afcc4895..a3e623a320b3 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c static struct {

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +2 -0
@@ -60,6 +60,7 @@ static struct {
     int64_t system_mem_used;
     int64_t ttm_mem_used;
     spinlock_t mem_limit_lock;
+    bool alow_oversubscribe;
  } kfd_mem_limit;
 
  static const char * const domain_bit_to_string[] = { @@ -110,7
+111,7 @@ static bool reuse_dmamap(struct amdgpu_device *adev, struct

amdgpu_device *bo_ad

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c

Quoted 5 lines

* System (TTM + userptr) memory - 15/16th System RAM
* TTM memory - 3/8th System RAM
*/

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -1
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device

*adev)

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c void

Quoted 4 lines
  {
     struct sysinfo si;
     uint64_t mem;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
@@ -130,6 +131,7 @@ void

amdgpu_amdkfd_gpuvm_init_mem_limits(void)

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c void · line 131

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
             kfd_mem_limit.max_system_mem_limit -=

AMDGPU_RESERVE_MEM_LIMIT;

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c void amdgpu_amdkfd_gpuvm_init_mem_limits(void)

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
     kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() <<
 PAGE_SHIFT;
+    kfd_mem_limit.alow_oversubscribe = !!(adev->flags & AMD_IS_APU);
     pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
             (kfd_mem_limit.max_system_mem_limit >> 20),
             (kfd_mem_limit.max_ttm_mem_limit >> 20)); @@ -221,8

+223,12 @@ int

Quoted 3 lines

amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,

          kfd_mem_limit.max_ttm_mem_limit) ||
         (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] +

vram_needed >

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,

Quoted 6 lines

vram_size - reserved_for_pt - atomic64_read(&adev-
vram_pin_size))) {

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +2 -2
-            ret = -ENOMEM;
-            goto release;
+            if (kfd_mem_limit.alow_oversubscribe) {
+                    pr_warn_ratelimited("Memory is getting

oversubscried.\n");

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +4 -0
+            } else {
+                    ret = -ENOMEM;
+                    goto release;
+            }
     }
 
     /* Update memory accounting by decreasing available system
maintainer
#4 lore raw

Re: [PATCH 2/2] drm/amdkfd: Allow memory oversubscription on small APUs

Quoted 15 lines · Lang

[Public]

Quoted 13 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Saturday, April 27, 2024 6:45 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Yang, Philip <[email protected]>; Koenig, Christian <[email protected]>; Zhang, Yifan <[email protected]>; Liu, Aaron <[email protected]>
Subject: Re: [PATCH 2/2] drm/amdkfd: Allow memory oversubscription on small APUs

Quoted 2 lines · Lang Yu

The default ttm_tt_pages_limit is 1/2 of system memory. It is prone to out of memory with such a configuration.

Quoted 7 lines
Quoted 3 lines

Indiscriminately allowing the violation of all memory limits is not a good solution. It will lead to poor performance once you actually reach ttm_pages_limit and TTM starts swapping out BOs.

Hi Felix,

I just feel it's like a bug that 1/2 of system memory is fee, the driver tells users out of memory. On the other hand, if memory is available, why not use it.

TTM does not allow us to use more than 1/2 system memory. I believe that's because TTM needs additional memory to swap out BOs. Any GTT allocation through the render node APIs is subject to the same limitations.

Render node APIs can handle memory overcommitment more gracefully because the kernel mode driver is in the loop for command submissions and fences. That doesn't work for KFD with user mode queues. The memory limits in KFD are there to prevent overcommitting memory because we need all of our memory (per process) to be resident at the same time. If we let KFD exceed the TTM limits, we get into situations where we're thrashing (processes evicting each other constantly) or even worse, where we're just not able to make all memory resident. So we end up with suspended user mode queues and extremely poor performance or soft hangs.

Quoted 3 lines

By the way, can we use USERPTR for VRAM allocations? Then we don't have ttm_tt_pages_limit limitations. Thanks.

No. There is an expectation that VRAM BOs can be shared between processes through DMABufs (for HIP IPC APIs). You can't export userptrs as DMABufs.

You can try to raise the TTM pages limit using a TTM module parameter. But this is taking a risk for system stability when TTM gets into a situation where it needs to swap out a large BO.

Regards,
  Felix

Quoted 149 lines

I actually did some tests on Strix (12 CU@2100 MHz, 29412M 128bits LPDDR5@937MHz) with
https://github.com/ROCm/pytorch-micro-benchmarking.

Command: python micro_benchmarking_pytorch.py --network resnet50 --batch-size=64 --iterations=20

1, Run 1 resnet50 (FP32, batch size 64)
Memory usage:

         System mem used 6748M out of 29412M
         TTM mem used 6658M out of 15719M

Memory oversubscription percentage: 0
Throughput [img/sec] : 49.04

2, Run 2 resnet50 simultaneously (FP32, batch size 64) Memory usage:

         System mem used 13496M out of 29412M
         TTM mem used 13316M out of 15719M

Memory oversubscription percentage: 0 Throughput [img/sec] (respectively) : 25.27 / 26.70

3, Run 3 resnet50 simultaneously (FP32, batch size 64) Memory usage:

         System mem used 20245M out of 29412M
         TTM mem used 19974M out of 15719M

Memory oversubscription percentage: ~27%

Throughput [img/sec](respectively) : 10.62 / 7.47 / 6.90 (In theory: 16 / 16 / 16)

From my observations,

1, GPU is underutilized a lot, sometimes its loading is less than 50% and even 0, when running 3 resnet50 simultaneously with ~27% memory oversubscription. The driver is busying evicting and restoring process. It takes ~2-5 seconds to restore all the BOs for one process (swap in and out BOs, actually allocate and copy pages), even though the process doesn't need all the allocated BOs to be resident.

2, Sometimes, the fairness can't be guaranteed between process when memory is oversubscribed. They can't share the GPU equally when created with default priority.

3, The less GPU underutilization time during evicting and restoring, the less performance degradation under memory oversubscription.

Regards,
Lang

Quoted 107 lines

Regards,
Felix

Quoted 5 lines
---
   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c       |  2 +-
   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       |  4 ++--
   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 12

+++++++++---

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c void amdgpu_amdkfd_device_init(struct

Quoted 8 lines
   3 files changed, 12 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +0 -0
index 3295838e9a1d..c01c6f3ab562 100644
@@ -167,7 +167,7 @@ void amdgpu_amdkfd_device_init(struct

amdgpu_device *adev)

Quoted 14 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +1 -1
      int i;
      int last_valid_bit;
 
-    amdgpu_amdkfd_gpuvm_init_mem_limits();
+    amdgpu_amdkfd_gpuvm_init_mem_limits(adev);
 
      if (adev->kfd.dev) {
              struct kgd2kfd_shared_resources gpu_resources = { diff --git

a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 1de021ebdd46..13284dbd8c58 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h u64 amdgpu_amdkfd_xcp_memory_size(struct · line 367

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
@@ -363,7 +363,7 @@ u64 amdgpu_amdkfd_xcp_memory_size(struct
 amdgpu_device *adev, int xcp_id);
 
 
   #if IS_ENABLED(CONFIG_HSA_AMD)
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device

*adev);

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h void amdgpu_amdkfd_release_notify(struct

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0 -0
   void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
                              struct amdgpu_vm *vm);
@@ -376,7 +376,7 @@ void amdgpu_amdkfd_release_notify(struct

amdgpu_bo *bo);

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo);

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
   void amdgpu_amdkfd_reserve_system_mem(uint64_t size);
   #else
   static inline
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device

*adev)

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0 -0
   {
   }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
index 7eb5afcc4895..a3e623a320b3 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c static struct {

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +2 -0
@@ -60,6 +60,7 @@ static struct {
      int64_t system_mem_used;
      int64_t ttm_mem_used;
      spinlock_t mem_limit_lock;
+    bool alow_oversubscribe;
   } kfd_mem_limit;
 
   static const char * const domain_bit_to_string[] = { @@ -110,7
+111,7 @@ static bool reuse_dmamap(struct amdgpu_device *adev, struct

amdgpu_device *bo_ad

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c

Quoted 5 lines

* System (TTM + userptr) memory - 15/16th System RAM
* TTM memory - 3/8th System RAM
*/

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -1
-void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
+void amdgpu_amdkfd_gpuvm_init_mem_limits(struct amdgpu_device

*adev)

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c void

Quoted 4 lines
   {
      struct sysinfo si;
      uint64_t mem;
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
@@ -130,6 +131,7 @@ void

amdgpu_amdkfd_gpuvm_init_mem_limits(void)

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c void · line 131

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
              kfd_mem_limit.max_system_mem_limit -=

AMDGPU_RESERVE_MEM_LIMIT;

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c void amdgpu_amdkfd_gpuvm_init_mem_limits(void)

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
      kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() <<
 PAGE_SHIFT;
+    kfd_mem_limit.alow_oversubscribe = !!(adev->flags & AMD_IS_APU);
      pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
              (kfd_mem_limit.max_system_mem_limit >> 20),
              (kfd_mem_limit.max_ttm_mem_limit >> 20)); @@ -221,8

+223,12 @@ int

Quoted 3 lines

amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,

           kfd_mem_limit.max_ttm_mem_limit) ||
          (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] +

vram_needed >

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,

Quoted 6 lines

vram_size - reserved_for_pt - atomic64_read(&adev-
vram_pin_size))) {

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +2 -2
-            ret = -ENOMEM;
-            goto release;
+            if (kfd_mem_limit.alow_oversubscribe) {
+                    pr_warn_ratelimited("Memory is getting

oversubscried.\n");

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +4 -0
+            } else {
+                    ret = -ENOMEM;
+                    goto release;
+            }
      }
 
      /* Update memory accounting by decreasing available system
maintainer
#5 lore raw
In reply to Lang Yu
Quoted 7 lines · Lang Yu

Small APUs(i.e., consumer, embedded products) usually have a small carveout device memory which can't satisfy most compute workloads memory allocation requirements.

We can't even run a Basic MNIST Example with a default 512MB carveout.
https://github.com/pytorch/examples/tree/main/mnist.

Quoted 7 lines

Though we can change BIOS settings to enlarge carveout size, which is inflexible and may bring complaint. On the other hand, the memory resource can't be effectively used between host and device.

The solution is MI300A approach, i.e., let VRAM allocations go to GTT.

Two nit-picks inline. Other than that, this patch looks reasonable to me.

Quoted 13 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c    |  6 +++++-
  .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 21 +++++++++++--------
  drivers/gpu/drm/amd/amdkfd/kfd_migrate.c      |  2 +-
  drivers/gpu/drm/amd/amdkfd/kfd_svm.c          |  6 ++++--
  drivers/gpu/drm/amd/amdkfd/kfd_svm.h          |  3 ++-
  5 files changed, 24 insertions(+), 14 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +0 -0
index 7ba05f030dd1..3295838e9a1d 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev, · line 463

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +3 -1
@@ -456,7 +456,9 @@ void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev,
  			mem_info->local_mem_size_private =
  					KFD_XCP_MEMORY_SIZE(adev, xcp->id);
  	} else {
-		mem_info->local_mem_size_public = adev->gmc.visible_vram_size;
+		mem_info->local_mem_size_public = adev->flags & AMD_IS_APU ?
+						  (ttm_tt_pages_limit() << PAGE_SHIFT) :
+						  adev->gmc.visible_vram_size;
  		mem_info->local_mem_size_private = adev->gmc.real_vram_size -
  						adev->gmc.visible_vram_size;

On an APU the private size should be reported as 0.

Quoted 61 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +2 -0
  	}
@@ -824,6 +826,8 @@ u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id)
  		}
  		do_div(tmp, adev->xcp_mgr->num_xcp_per_mem_partition);
  		return ALIGN_DOWN(tmp, PAGE_SIZE);
+	} else if (adev->flags & AMD_IS_APU) {
+		return (ttm_tt_pages_limit() << PAGE_SHIFT);
  	} else {
  		return adev->gmc.real_vram_size;
  	}
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +6 -5
index c4f9960dafbb..7eb5afcc4895 100644
@@ -196,7 +196,7 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
  			return -EINVAL;
 
  		vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
-		if (adev->gmc.is_app_apu) {
+		if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
  			system_mem_needed = size;
  			ttm_mem_needed = size;
  		}
@@ -232,7 +232,8 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
  		  "adev reference can't be null when vram is used");
  	if (adev && xcp_id >= 0) {
  		adev->kfd.vram_used[xcp_id] += vram_needed;
-		adev->kfd.vram_used_aligned[xcp_id] += adev->gmc.is_app_apu ?
+		adev->kfd.vram_used_aligned[xcp_id] +=
+				(adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) ?
  				vram_needed :
  				ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
  	}
@@ -260,7 +261,7 @@ void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
 
  		if (adev) {
  			adev->kfd.vram_used[xcp_id] -= size;
-			if (adev->gmc.is_app_apu) {
+			if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
  				adev->kfd.vram_used_aligned[xcp_id] -= size;
  				kfd_mem_limit.system_mem_used -= size;
  				kfd_mem_limit.ttm_mem_used -= size;
@@ -889,7 +890,7 @@ static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
  	 * if peer device has large BAR. In contrast, access over xGMI is
  	 * allowed for both small and large BAR configurations of peer device
  	 */
-	if ((adev != bo_adev && !adev->gmc.is_app_apu) &&
+	if ((adev != bo_adev && !(adev->gmc.is_app_apu || adev->flags & AMD_IS_APU)) &&
  	    ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
  	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
  	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
@@ -1657,7 +1658,7 @@ size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
  		- atomic64_read(&adev->vram_pin_size)
  		- reserved_for_pt;
 
-	if (adev->gmc.is_app_apu) {
+	if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
  		system_mem_available = no_system_mem_limit ?
  					kfd_mem_limit.max_system_mem_limit :
  					kfd_mem_limit.max_system_mem_limit -

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev, · line 1673

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
@@ -1669,6 +1670,7 @@ size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
  		available = min3(system_mem_available, ttm_mem_available,
  				 vram_available);
  		available = ALIGN_DOWN(available, PAGE_SIZE);
+

Unnecessary whitespace change.

Regards,
  Felix

Quoted 84 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +5 -4
  	} else {
  		available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN);
  	}
@@ -1705,7 +1707,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
  	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
  		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
 
-		if (adev->gmc.is_app_apu) {
+		if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
  			domain = AMDGPU_GEM_DOMAIN_GTT;
  			alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
  			alloc_flags = 0;
@@ -1952,7 +1954,7 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
  	if (size) {
  		if (!is_imported &&
  		   (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM ||
-		   (adev->gmc.is_app_apu &&
+		   ((adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) &&
  		    mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT)))
  			*size = bo_size;
  		else
@@ -2374,8 +2376,9 @@ static int import_obj_create(struct amdgpu_device *adev,
  	(*mem)->dmabuf = dma_buf;
  	(*mem)->bo = bo;
  	(*mem)->va = va;
-	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ?
-		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
+	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) &&
+			 !(adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) ?
+			 AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
 
  	(*mem)->mapped_to_gpu_memory = 0;
  	(*mem)->process_info = avm->process_info;
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +1 -1
index 4bcfbeac48fb..4816fcb9803a 100644
@@ -1023,7 +1023,7 @@ int kgd2kfd_init_zone_device(struct amdgpu_device *adev)
  	if (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 1))
  		return -EINVAL;
 
-	if (adev->gmc.is_app_apu)
+	if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU)
  		return 0;
 
  	pgmap = &kfddev->pgmap;
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +4 -2
index bfab16b43fec..238ac11bb97d 100644
@@ -2619,7 +2619,8 @@ svm_range_best_restore_location(struct svm_range *prange,
  		return -1;
  	}
 
-	if (node->adev->gmc.is_app_apu)
+	if (node->adev->gmc.is_app_apu ||
+	    node->adev->flags & AMD_IS_APU)
  		return 0;
 
  	if (prange->preferred_loc == gpuid ||
@@ -3337,7 +3338,8 @@ svm_range_best_prefetch_location(struct svm_range *prange)
  		goto out;
  	}
 
-	if (bo_node->adev->gmc.is_app_apu) {
+	if (bo_node->adev->gmc.is_app_apu ||
+	    bo_node->adev->flags & AMD_IS_APU) {
  		best_loc = 0;
  		goto out;
  	}
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +2 -1
index 026863a0abcd..9c37bd0567ef 100644
@@ -201,7 +201,8 @@ void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_s
   * is initialized to not 0 when page migration register device memory.
   */
  #define KFD_IS_SVM_API_SUPPORTED(adev) ((adev)->kfd.pgmap.type != 0 ||\
-					(adev)->gmc.is_app_apu)
+					(adev)->gmc.is_app_apu ||\
+					((adev)->flags & AMD_IS_APU))
 
  void svm_range_bo_unref_async(struct svm_range_bo *svm_bo);
 
author
#6 lore raw

[Public]

Quoted 23 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Saturday, April 27, 2024 6:52 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Yang, Philip <[email protected]>; Koenig, Christian <[email protected]>; Zhang, Yifan <[email protected]>; Liu, Aaron <[email protected]>
Subject: Re: [PATCH 1/2] drm/amdkfd: Let VRAM allocations go to GTT domain on small APUs

Quoted 11 lines · Lang Yu

Small APUs(i.e., consumer, embedded products) usually have a small carveout device memory which can't satisfy most compute workloads memory allocation requirements.

We can't even run a Basic MNIST Example with a default 512MB carveout.
https://github.com/pytorch/examples/tree/main/mnist.

Though we can change BIOS settings to enlarge carveout size, which is inflexible and may bring complaint. On the other hand, the memory resource can't be effectively used between host and device.

Quoted 5 lines
Quoted 3 lines

The solution is MI300A approach, i.e., let VRAM allocations go to GTT.

Two nit-picks inline. Other than that, this patch looks reasonable to me.

Thanks. Will update them accordingly.

Regards,
Lang

Quoted 227 lines
Quoted 4 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c    |  6 +++++-
  .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 21 +++++++++++-

-------

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c void amdgpu_amdkfd_get_local_mem_info(struct

Quoted 11 lines
  drivers/gpu/drm/amd/amdkfd/kfd_migrate.c      |  2 +-
  drivers/gpu/drm/amd/amdkfd/kfd_svm.c          |  6 ++++--
  drivers/gpu/drm/amd/amdkfd/kfd_svm.h          |  3 ++-
  5 files changed, 24 insertions(+), 14 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +0 -0
index 7ba05f030dd1..3295838e9a1d 100644
@@ -456,7 +456,9 @@ void amdgpu_amdkfd_get_local_mem_info(struct

amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c void amdgpu_amdkfd_get_local_mem_info(struct · line 461

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +1 -1
                     mem_info->local_mem_size_private =
                                     KFD_XCP_MEMORY_SIZE(adev, xcp-
 id);
     } else {
-            mem_info->local_mem_size_public = adev-
 gmc.visible_vram_size;
+            mem_info->local_mem_size_public = adev->flags &

AMD_IS_APU ?

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c void amdgpu_amdkfd_get_local_mem_info(struct · line 462

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +1 -0
+                                              (ttm_tt_pages_limit() <<

PAGE_SHIFT) :

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +1 -0
+                                              adev-
 gmc.visible_vram_size;
             mem_info->local_mem_size_private = adev-
gmc.real_vram_size -
                                             adev->gmc.visible_vram_size;

On an APU the private size should be reported as 0.

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c u64 amdgpu_amdkfd_xcp_memory_size(struct

Quoted 2 lines
     }
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +0 -0
@@ -824,6 +826,8 @@ u64 amdgpu_amdkfd_xcp_memory_size(struct

amdgpu_device *adev, int xcp_id)

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_reserve_mem_limit(struct

Quoted 14 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +2 -0
             }
             do_div(tmp, adev->xcp_mgr->num_xcp_per_mem_partition);
             return ALIGN_DOWN(tmp, PAGE_SIZE);
+    } else if (adev->flags & AMD_IS_APU) {
+            return (ttm_tt_pages_limit() << PAGE_SHIFT);
     } else {
             return adev->gmc.real_vram_size;
     }
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
index c4f9960dafbb..7eb5afcc4895 100644
@@ -196,7 +196,7 @@ int amdgpu_amdkfd_reserve_mem_limit(struct

amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_reserve_mem_limit(struct

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -1
                     return -EINVAL;
 
             vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
-            if (adev->gmc.is_app_apu) {
+            if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
                     system_mem_needed = size;
                     ttm_mem_needed = size;
             }
@@ -232,7 +232,8 @@ int amdgpu_amdkfd_reserve_mem_limit(struct

amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -1
               "adev reference can't be null when vram is used");
     if (adev && xcp_id >= 0) {
             adev->kfd.vram_used[xcp_id] += vram_needed;
-            adev->kfd.vram_used_aligned[xcp_id] += adev-

gmc.is_app_apu ?

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +2 -0
+            adev->kfd.vram_used_aligned[xcp_id] +=
+                            (adev->gmc.is_app_apu || adev->flags &

AMD_IS_APU) ?

Quoted 2 lines

vram_needed :
ALIGN(vram_needed,

VRAM_AVAILABLITY_ALIGN);

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c void

Quoted 2 lines
     }
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
@@ -260,7 +261,7 @@ void

amdgpu_amdkfd_unreserve_mem_limit(struct

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -1
 amdgpu_device *adev,
 
             if (adev) {
                     adev->kfd.vram_used[xcp_id] -= size;
-                    if (adev->gmc.is_app_apu) {
+                    if (adev->gmc.is_app_apu || adev->flags &

AMD_IS_APU) {

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
                             adev->kfd.vram_used_aligned[xcp_id] -= size;
                             kfd_mem_limit.system_mem_used -= size;
                             kfd_mem_limit.ttm_mem_used -= size; @@ -

889,7 +890,7 @@ static

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,

Quoted 10 lines

int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
* if peer device has large BAR. In contrast, access over xGMI is
* allowed for both small and large BAR configurations of peer device */

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +2 -1
-    if ((adev != bo_adev && !adev->gmc.is_app_apu) &&
+    if ((adev != bo_adev && !(adev->gmc.is_app_apu || adev->flags &
+AMD_IS_APU)) &&
         ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
          (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
          (mem->alloc_flags &

KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) { @@

Quoted 1 line

-1657,7 +1658,7 @@ size_t amdgpu_amdkfd_get_available_memory(struct

amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev, · line 1662

Quoted 7 lines

- atomic64_read(&adev->vram_pin_size)
- reserved_for_pt;

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -1
-    if (adev->gmc.is_app_apu) {
+    if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
             system_mem_available = no_system_mem_limit ?

kfd_mem_limit.max_system_mem_limit :

Quoted 1 line
      kfd_mem_limit.max_system_mem_limit - @@ -1669,6 +1670,7 @@
Quoted 1 line

size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device

*adev,

Quoted 4 lines
             available = min3(system_mem_available, ttm_mem_available,
                              vram_available);
             available = ALIGN_DOWN(available, PAGE_SIZE);

+

Unnecessary whitespace change.

Regards,
Felix

Quoted 2 lines

} else {
available = ALIGN_DOWN(vram_available,

VRAM_AVAILABLITY_ALIGN);

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int

Quoted 2 lines
     }
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
@@ -1705,7 +1707,7 @@ int

amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -1
     if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
             domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
 
-            if (adev->gmc.is_app_apu) {
+            if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
                     domain = AMDGPU_GEM_DOMAIN_GTT;
                     alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
                     alloc_flags = 0;
@@ -1952,7 +1954,7 @@ int

amdgpu_amdkfd_gpuvm_free_memory_of_gpu(

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
     if (size) {
             if (!is_imported &&
                (mem->bo->preferred_domains ==

AMDGPU_GEM_DOMAIN_VRAM ||

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -1
-               (adev->gmc.is_app_apu &&
+               ((adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) &&
                 mem->bo->preferred_domains ==

AMDGPU_GEM_DOMAIN_GTT)))

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c static int import_obj_create(struct

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
                     *size = bo_size;
             else
@@ -2374,8 +2376,9 @@ static int import_obj_create(struct

amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c static int import_obj_create(struct amdgpu_device *adev,

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -1
     (*mem)->dmabuf = dma_buf;
     (*mem)->bo = bo;
     (*mem)->va = va;
-    (*mem)->domain = (bo->preferred_domains &

AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ?

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c static int import_obj_create(struct · line 2378

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -1
-            AMDGPU_GEM_DOMAIN_VRAM :

AMDGPU_GEM_DOMAIN_GTT;

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c static int import_obj_create(struct · line 2379

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
+    (*mem)->domain = (bo->preferred_domains &

AMDGPU_GEM_DOMAIN_VRAM) &&

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c static int import_obj_create(struct · line 2380

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
+                     !(adev->gmc.is_app_apu || adev->flags &

AMD_IS_APU) ?

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c static int import_obj_create(struct · line 2381

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -0
+                     AMDGPU_GEM_DOMAIN_VRAM :

AMDGPU_GEM_DOMAIN_GTT;

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c int kgd2kfd_init_zone_device(struct

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
     (*mem)->mapped_to_gpu_memory = 0;
     (*mem)->process_info = avm->process_info; diff --git

a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
index 4bcfbeac48fb..4816fcb9803a 100644

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +0 -0
@@ -1023,7 +1023,7 @@ int kgd2kfd_init_zone_device(struct

amdgpu_device *adev)

drivers/gpu/drm/amd/amdkfd/kfd_svm.c svm_range_best_restore_location(struct

Quoted 14 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +1 -1
     if (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 1))
             return -EINVAL;
 
-    if (adev->gmc.is_app_apu)
+    if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU)
             return 0;
 
     pgmap = &kfddev->pgmap;
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +0 -0
index bfab16b43fec..238ac11bb97d 100644
@@ -2619,7 +2619,8 @@ svm_range_best_restore_location(struct

svm_range *prange,

drivers/gpu/drm/amd/amdkfd/kfd_svm.h void svm_range_list_lock_and_flush_work(struct

Quoted 25 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +2 -1
             return -1;
     }
 
-    if (node->adev->gmc.is_app_apu)
+    if (node->adev->gmc.is_app_apu ||
+        node->adev->flags & AMD_IS_APU)
             return 0;
 
     if (prange->preferred_loc == gpuid || @@ -3337,7 +3338,8 @@

svm_range_best_prefetch_location(struct svm_range *prange)

             goto out;
     }
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +2 -1
-    if (bo_node->adev->gmc.is_app_apu) {
+    if (bo_node->adev->gmc.is_app_apu ||
+        bo_node->adev->flags & AMD_IS_APU) {
             best_loc = 0;
             goto out;
     }
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +0 -0
index 026863a0abcd..9c37bd0567ef 100644
@@ -201,7 +201,8 @@ void svm_range_list_lock_and_flush_work(struct

svm_range_list *svms, struct mm_s

drivers/gpu/drm/amd/amdkfd/kfd_svm.h void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_s

Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +0 -0
   * is initialized to not 0 when page migration register device memory.
   */
  #define KFD_IS_SVM_API_SUPPORTED(adev) ((adev)->kfd.pgmap.type != 0
||\
> -                                    (adev)->gmc.is_app_apu)
> +                                    (adev)->gmc.is_app_apu ||\
> +                                    ((adev)->flags & AMD_IS_APU))
Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +0 -0
  void svm_range_bo_unref_async(struct svm_range_bo *svm_bo);