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. Error Log:

"torch.cuda.OutOfMemoryError: HIP out of memory. Tried to allocate
84.00 MiB. GPU 0 has a total capacity of 512.00 MiB of which 0 bytes
is free. Of the allocated memory 103.83 MiB is allocated by PyTorch,
and 22.17 MiB is reserved by PyTorch but unallocated"

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.
Then device and host can flexibly and effectively share memory resource.

v2: Report local_mem_size_private as 0. (Felix)

Signed-off-by: Lang Yu <Lang.Yu@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

1 files changed, 11 insertions, 9 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 0535b07987d9..8975cf41a91a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -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 -
@@ -1705,7 +1706,7 @@ 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 +1953,7 @@ 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 +2375,9 @@ 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;