Add memory module

Signed-off-by: Andrew Lewycky <Andrew.Lewycky@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@amd.com>


[ROCm/ROCR-Runtime commit: 6a8a380ba3]

projects/rocr-runtime/src/memory.c

@@ -28,6 +28,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include <assert.h>
+#include "fmm.h"
 
 HSAKMT_STATUS
 HSAKMTAPI
@@ -39,9 +40,49 @@ hsaKmtSetMemoryPolicy(
 	HSAuint64 MemorySizeInBytes
 	)
 {
+	HSAKMT_STATUS result;
+	uint32_t gpu_id;
+
 	CHECK_KFD_OPEN();
 
-	return HSAKMT_STATUS_NOT_SUPPORTED;
+	result = validate_nodeid(Node, &gpu_id);
+	if (result != HSAKMT_STATUS_SUCCESS)
+		return result;
+
+	// We accept any legal policy and alternate address location. You get CC everywhere anyway.
+	if ((DefaultPolicy != HSA_CACHING_CACHED && DefaultPolicy != HSA_CACHING_NONCACHED)
+	    || (AlternatePolicy != HSA_CACHING_CACHED && AlternatePolicy != HSA_CACHING_NONCACHED))
+	{
+		return HSAKMT_STATUS_INVALID_PARAMETER;
+	}
+
+	CHECK_PAGE_MULTIPLE(MemoryAddressAlternate);
+	CHECK_PAGE_MULTIPLE(MemorySizeInBytes);
+
+	struct kfd_ioctl_set_memory_policy_args args;
+	memset(&args, 0, sizeof(args));
+
+	args.gpu_id = gpu_id;
+	args.default_policy = (DefaultPolicy == HSA_CACHING_CACHED) ? KFD_IOC_CACHE_POLICY_COHERENT : KFD_IOC_CACHE_POLICY_NONCOHERENT;
+	args.alternate_policy = (AlternatePolicy == HSA_CACHING_CACHED) ? KFD_IOC_CACHE_POLICY_COHERENT : KFD_IOC_CACHE_POLICY_NONCOHERENT;
+	args.alternate_aperture_base = (uintptr_t)MemoryAddressAlternate;
+	args.alternate_aperture_size = MemorySizeInBytes;
+
+	int err = kfd_ioctl(KFD_IOC_SET_MEMORY_POLICY, &args);
+
+	return (err == -1) ? HSAKMT_STATUS_ERROR : HSAKMT_STATUS_SUCCESS;
+}
+
+static HSAuint32 PageSizeFromFlags(unsigned int pageSizeFlags)
+{
+	switch (pageSizeFlags)
+	{
+	case HSA_PAGE_SIZE_4KB: return 4*1024;
+	case HSA_PAGE_SIZE_64KB: return 64*1024;
+	case HSA_PAGE_SIZE_2MB: return 2*1024*1024;
+	case HSA_PAGE_SIZE_1GB: return 1024*1024*1024;
+	default: assert(false); return 4*1024;
+	}
 }
 
 HSAKMT_STATUS
@@ -54,8 +95,29 @@ hsaKmtAllocMemory(
     )
 {
 	CHECK_KFD_OPEN();
+	HSAKMT_STATUS result;
+	uint32_t gpu_id;
+
+	result = validate_nodeid(PreferredNode, &gpu_id);
+	if (result != HSAKMT_STATUS_SUCCESS)
+		return result;
+
+	// The required size should be page aligned (GDS?)
+	HSAuint64 page_size = PageSizeFromFlags(MemFlags.ui32.PageSize);
+	if ((SizeInBytes & (page_size-1)) && !MemFlags.ui32.GDSMemory){
+		return HSAKMT_STATUS_INVALID_PARAMETER;
+	}
+
+	if (MemFlags.ui32.HostAccess && !MemFlags.ui32.NonPaged){
+	    int err = posix_memalign(MemoryAddress, page_size, SizeInBytes);
+		if (err == 0)
+			return HSAKMT_STATUS_SUCCESS;
+		else
+			return HSAKMT_STATUS_NO_MEMORY;
+	}
+	else
+		return HSAKMT_STATUS_INVALID_PARAMETER;
 
-	return HSAKMT_STATUS_NOT_SUPPORTED;
 }
 
 HSAKMT_STATUS
@@ -65,9 +127,17 @@ hsaKmtFreeMemory(
     HSAuint64   SizeInBytes         //IN
     )
 {
+	HSAKMT_STATUS hsa_status = HSAKMT_STATUS_SUCCESS;
 	CHECK_KFD_OPEN();
 
-	return HSAKMT_STATUS_NOT_SUPPORTED;
+	if (fmm_is_inside_some_aperture(MemoryAddress)){
+		if (fmm_release( MemoryAddress, SizeInBytes))
+			hsa_status = HSAKMT_STATUS_INVALID_PARAMETER;
+	}
+	else
+		free(MemoryAddress);
+
+	return hsa_status;
 }
 
 HSAKMT_STATUS
@@ -79,7 +149,7 @@ hsaKmtRegisterMemory(
 {
 	CHECK_KFD_OPEN();
 
-	return HSAKMT_STATUS_NOT_SUPPORTED;
+	return HSAKMT_STATUS_SUCCESS;
 }
 
 HSAKMT_STATUS
@@ -90,7 +160,7 @@ hsaKmtDeregisterMemory(
 {
 	CHECK_KFD_OPEN();
 
-	return HSAKMT_STATUS_NOT_SUPPORTED;
+	return HSAKMT_STATUS_SUCCESS;
 }
 
 HSAKMT_STATUS
@@ -103,7 +173,13 @@ hsaKmtMapMemoryToGPU(
 {
 	CHECK_KFD_OPEN();
 
-	return HSAKMT_STATUS_NOT_SUPPORTED;
+	// We don't support GPUVM in the stub, there should never be a request for a GPUVA.
+	if (AlternateVAGPU)
+	{
+		*AlternateVAGPU = 0;
+	}
+
+	return HSAKMT_STATUS_SUCCESS;
 }
 
 HSAKMT_STATUS
@@ -114,5 +190,5 @@ hsaKmtUnmapMemoryToGPU(
 {
 	CHECK_KFD_OPEN();
 
-	return HSAKMT_STATUS_NOT_SUPPORTED;
+	return HSAKMT_STATUS_SUCCESS;
 }