Adding support for limiting ommArraySize (#16)

integration/omm-sdk-nvrhi/omm-sdk-nvrhi.cpp

@@ -649,8 +649,8 @@ omm::Gpu::DispatchConfigDesc GpuBakeNvrhiImpl::GetConfig(const GpuBakeNvrhi::Inp
 	config.indexCount							= (uint32_t)params.numIndices;
 	config.globalFormat							= params.format == nvrhi::rt::OpacityMicromapFormat::OC1_2_State ? Format::OC1_2_State : Format::OC1_4_State;
 	config.maxScratchMemorySize					= params.minimalMemoryMode ? Gpu::ScratchMemoryBudget::MB_4 : Gpu::ScratchMemoryBudget::MB_256;
+	config.maxOutOmmArraySize				    = params.maxOutOmmArraySize;
 	config.maxSubdivisionLevel					= params.maxSubdivisionLevel;
-	config.globalSubdivisionLevel				= params.maxSubdivisionLevel;
 	config.dynamicSubdivisionScale				= params.dynamicSubdivisionScale;
 	return config;
 }

integration/omm-sdk-nvrhi/omm-sdk-nvrhi.h

@@ -56,6 +56,7 @@ namespace omm
 			uint32_t							numIndices = 0;
 
 			uint32_t							maxSubdivisionLevel = 0;
+			uint32_t							maxOutOmmArraySize = 0xFFFFFFFF;
 			nvrhi::rt::OpacityMicromapFormat	format = nvrhi::rt::OpacityMicromapFormat::OC1_4_State;
 			float								dynamicSubdivisionScale = 0.5f;
 			bool								minimalMemoryMode = false;

omm-sdk/include/omm.h

@@ -16,7 +16,7 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 #define OMM_VERSION_MAJOR 0
 #define OMM_VERSION_MINOR 9
-#define OMM_VERSION_BUILD 1
+#define OMM_VERSION_BUILD 2
 
 #if defined(_MSC_VER)
     #define OMM_CALL __fastcall
@@ -788,7 +788,7 @@ typedef struct ommGpuPreDispatchInfo
    uint32_t       outOmmIndexCount;
    // Min required size of OUT_OMM_ARRAY_DATA. GetBakeInfo returns most conservative estimation while less conservative number
    // can be obtained via BakePrepass
-   size_t         outOmmArraySizeInBytes;
+   uint32_t       outOmmArraySizeInBytes;
    // Min required size of OUT_OMM_DESC_ARRAY. GetBakeInfo returns most conservative estimation while less conservative number
    // can be obtained via BakePrepass
    uint32_t       outOmmDescSizeInBytes;
@@ -848,12 +848,15 @@ typedef struct ommGpuDispatchConfigDesc
    float                     dynamicSubdivisionScale;
    // The global Format. May be overriden by the per-triangle config.
    ommFormat                 globalFormat;
-   // Micro triangle count is 4^N, where N is the subdivision level. Subdivision level must be in range [0,
-   // MaxSubdivisionLevel]. The global subdivisionLevel. May be overriden by the per-triangle subdivision level setting. The
-   // subdivision level to allow in dynamic mode and value is used to allocate appropriate scratch memory.
-   uint8_t                   globalSubdivisionLevel;
    uint8_t                   maxSubdivisionLevel;
-   uint8_t                   enableSubdivisionLevelBuffer;
+   ommBool                   enableSubdivisionLevelBuffer;
+   // The SDK will try to limit the omm array size of PreDispatchInfo::outOmmArraySizeInBytes and
+   // PostBakeInfo::outOmmArraySizeInBytes.
+   // Currently a greedy algorithm is implemented with a first come-first serve order.
+   // The SDK may (or may not) apply more sophisticated heuristics in the future.
+   // If no memory is available to allocate an OMM Array Block the state will default to Unknown Opaque (ignoring any bake
+   // flags do disable special indices).
+   uint32_t                  maxOutOmmArraySize;
    // Target scratch memory budget, The SDK will try adjust the sum of the transient pool buffers to match this value. Higher
    // budget more efficiently executes the baking operation. May return INSUFFICIENT_SCRATCH_MEMORY if set too low.
    ommGpuScratchMemoryBudget maxScratchMemorySize;
@@ -877,9 +880,9 @@ inline ommGpuDispatchConfigDesc ommGpuDispatchConfigDescDefault()
    v.alphaCutoff                   = 0.5f;
    v.dynamicSubdivisionScale       = 2;
    v.globalFormat                  = ommFormat_OC1_4_State;
-   v.globalSubdivisionLevel        = 4;
    v.maxSubdivisionLevel           = 8;
    v.enableSubdivisionLevelBuffer  = 0;
+   v.maxOutOmmArraySize            = 0xFFFFFFFF;
    v.maxScratchMemorySize          = ommGpuScratchMemoryBudget_Default;
    return v;
 }

omm-sdk/include/omm.hpp

@@ -683,7 +683,7 @@ namespace omm
          uint32_t    outOmmIndexCount                   = 0xFFFFFFFF;
          // Min required size of OUT_OMM_ARRAY_DATA. GetBakeInfo returns most conservative estimation while less conservative number
          // can be obtained via BakePrepass
-         size_t      outOmmArraySizeInBytes             = 0xFFFFFFFF;
+         uint32_t    outOmmArraySizeInBytes             = 0xFFFFFFFF;
          // Min required size of OUT_OMM_DESC_ARRAY. GetBakeInfo returns most conservative estimation while less conservative number
          // can be obtained via BakePrepass
          uint32_t    outOmmDescSizeInBytes              = 0xFFFFFFFF;
@@ -728,12 +728,15 @@ namespace omm
          float               dynamicSubdivisionScale       = 2;
          // The global Format. May be overriden by the per-triangle config.
          Format              globalFormat                  = Format::OC1_4_State;
-         // Micro triangle count is 4^N, where N is the subdivision level. Subdivision level must be in range [0,
-         // MaxSubdivisionLevel]. The global subdivisionLevel. May be overriden by the per-triangle subdivision level setting. The
-         // subdivision level to allow in dynamic mode and value is used to allocate appropriate scratch memory.
-         uint8_t             globalSubdivisionLevel        = 4;
          uint8_t             maxSubdivisionLevel           = 8;
-         uint8_t             enableSubdivisionLevelBuffer  = 0;
+         bool                enableSubdivisionLevelBuffer  = false;
+         // The SDK will try to limit the omm array size of PreDispatchInfo::outOmmArraySizeInBytes and
+         // PostBakeInfo::outOmmArraySizeInBytes.
+         // Currently a greedy algorithm is implemented with a first come-first serve order.
+         // The SDK may (or may not) apply more sophisticated heuristics in the future.
+         // If no memory is available to allocate an OMM Array Block the state will default to Unknown Opaque (ignoring any bake
+         // flags do disable special indices).
+         uint32_t            maxOutOmmArraySize            = 0xFFFFFFFF;
          // Target scratch memory budget, The SDK will try adjust the sum of the transient pool buffers to match this value. Higher
          // budget more efficiently executes the baking operation. May return INSUFFICIENT_SCRATCH_MEMORY if set too low.
          ScratchMemoryBudget maxScratchMemorySize          = ScratchMemoryBudget::Default;

omm-sdk/scripts/omm.json

@@ -1824,21 +1824,21 @@
                 },
                 "comment": "The global Format. May be overriden by the per-triangle config."
             },
-            {
-                "type": "uint8_t",
-                "name": "globalSubdivisionLevel",
-                "value": "4",
-                "comment": "Micro triangle count is 4^N, where N is the subdivision level. Subdivision level must be in range [0, MaxSubdivisionLevel]. The global subdivisionLevel. May be overriden by the per-triangle subdivision level setting. The subdivision level to allow in dynamic mode and value is used to allocate appropriate scratch memory."
-            },
             {
                 "type": "uint8_t",
                 "name": "maxSubdivisionLevel",
                 "value": "8"
             },
             {
-                "type": "uint8_t",
+                "type": "bool",
                 "name": "enableSubdivisionLevelBuffer",
-                "value": "0"
+                "value": "false"
+            },
+            {
+                "type": "uint32_t",
+                "name": "maxOutOmmArraySize",
+                "value": "0xFFFFFFFF",
+                "comment": "The SDK will try to limit the omm array size of PreDispatchInfo::outOmmArraySizeInBytes and PostBakeInfo::outOmmArraySizeInBytes.\nCurrently a greedy algorithm is implemented with a first come-first serve order.\nThe SDK may (or may not) apply more sophisticated heuristics in the future.\nIf no memory is available to allocate an OMM Array Block the state will default to Unknown Opaque (ignoring any bake flags do disable special indices)."
             },
             {
                 "type": "ScratchMemoryBudget",
@@ -1915,7 +1915,7 @@
                 "value": "0xFFFFFFFF"
             },
             {
-                "type": "size_t",
+                "type": "uint32_t",
                 "name": "outOmmArraySizeInBytes",
                 "value": "0xFFFFFFFF",
                 "comment": "Min required size of OUT_OMM_ARRAY_DATA. GetBakeInfo returns most conservative estimation while less conservative number can be obtained via BakePrepass"

omm-sdk/scripts/omm_header_c.txt

@@ -16,7 +16,7 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 #define OMM_VERSION_MAJOR 0
 #define OMM_VERSION_MINOR 9
-#define OMM_VERSION_BUILD 1
+#define OMM_VERSION_BUILD 2
 
 #if defined(_MSC_VER)
     #define OMM_CALL __fastcall

omm-sdk/shaders/omm_desc_patch.cs.hlsl

@@ -51,7 +51,7 @@ uint GetSourcePrimitiveIndex(uint primitiveIndex)
 
 	if (primitiveIndexOrHashTableEntryIndex < -4)
 	{
-		const uint hashTableEntryIndex = -(primitiveIndexOrHashTableEntryIndex + 4);
+		const uint hashTableEntryIndex = -(primitiveIndexOrHashTableEntryIndex + 5);
 		const uint primitiveIndexRef =  OMM_SUBRESOURCE_LOAD(HashTableBuffer, 8 * hashTableEntryIndex + 4); // [hash|primitiveIndex]
 		return primitiveIndexRef;
 	}

omm-sdk/shaders/omm_global_cb.hlsli

@@ -7,14 +7,12 @@ and any modifications thereto. Any use, reproduction, disclosure or
 distribution of this software and related documentation without an express
 license agreement from NVIDIA CORPORATION is strictly prohibited.
 */
-
-
 #define OMM_DECLARE_GLOBAL_CONSTANT_BUFFER						\
 OMM_CONSTANTS_START(GlobalConstants)							\
 	OMM_CONSTANT(uint, IndexCount)								\
 	OMM_CONSTANT(uint, PrimitiveCount)							\
 	OMM_CONSTANT(uint, MaxBatchCount)							\
-	OMM_CONSTANT(uint, GlobalSubdivisionLevel)					\
+	OMM_CONSTANT(uint, MaxOutOmmArraySize)						\
 																\
 	OMM_CONSTANT(uint, IsOmmIndexFormat16bit)					\
 	OMM_CONSTANT(uint, EnableSpecialIndices)					\
@@ -45,7 +43,6 @@ OMM_CONSTANTS_START(GlobalConstants)							\
 																\
 	OMM_CONSTANT(float2, TexSize)								\
 	OMM_CONSTANT(float2, InvTexSize)							\
-																\
 /*	---- Buffer offsets go here								  */\
 																\
 	OMM_CONSTANT(uint, IEBakeBufferOffset)						\

omm-sdk/shaders/omm_post_build_info.cs.hlsl

@@ -31,6 +31,6 @@ void main(uint3 tid : SV_DispatchThreadID)
 	const uint ommDescCount			= OMM_SUBRESOURCE_LOAD(OmmDescAllocatorCounterBuffer, 0);
 	const uint ommDescByteSize		= ommDescCount * 8;
 
-	u_postBuildInfo.Store(0, ommArrayByteSize);
+	u_postBuildInfo.Store(0, min(ommArrayByteSize, g_GlobalConstants.MaxOutOmmArraySize));
 	u_postBuildInfo.Store(4, ommDescByteSize);
 }

omm-sdk/shaders/omm_work_setup_common.hlsli

@@ -203,19 +203,25 @@ uint GetSubdivisionLevel(TexCoords texCoords)
 	}
 	else
 	{
-		return g_GlobalConstants.GlobalSubdivisionLevel;
+		return g_GlobalConstants.MaxSubdivisionLevel;
 	}
 }
 
 int GetOmmDescOffset(ByteAddressBuffer ommIndexBuffer, uint primitiveIndex)
 {
-	// TODO: support 16-bit indices.
 	if (g_GlobalConstants.IsOmmIndexFormat16bit)
 	{
 		const uint dwOffset = primitiveIndex.x >> 1u;
-		const uint shift = (primitiveIndex.x & 1u) << 4u; // 0 or 16
-		const uint val = ommIndexBuffer.Load(4 * dwOffset);
-		return (val >> shift) & 0xFFFF;
+		const uint shift	= (primitiveIndex.x & 1u) << 4u; // 0 or 16
+		const uint raw		= ommIndexBuffer.Load(4 * dwOffset);
+		const uint raw16	= (raw >> shift) & 0xFFFFu;
+
+		if (raw16 > 0xFFFB) // e.g special index
+		{
+			return (raw16 - 0xFFFF) - 1; // -1, -2, -3 or -4
+		}
+
+		return raw16;
 	}
 	else
 	{

omm-sdk/shaders/omm_work_setup_cs.cs.hlsl

@@ -39,7 +39,8 @@ void main(uint3 tid : SV_DispatchThreadID)
 	uint hashTableEntryIndex;
 	hashTable::Result result		= FindOrInsertOMMEntry(texCoords, subdivisionLevel, hashTableEntryIndex);
 
-	uint vmDescOffset = 0;
+	int vmDescOffset = (int)SpecialIndex::FullyUnknownOpaque;
+
 	if (result == hashTable::Result::Null ||
 		result == hashTable::Result::Inserted || 
 		result == hashTable::Result::ReachedMaxAttemptCount)
@@ -66,9 +67,7 @@ void main(uint3 tid : SV_DispatchThreadID)
 			OMM_SUBRESOURCE_INTERLOCKEDADD(OmmArrayAllocatorCounterBuffer, 0, vmDataByteSize, vmArrayOffset);
 		}
 
-		const uint kMaxVmArrayBudget = 0xFFFFFFFF;
-
-		if ((vmArrayOffset + vmDataByteSize) < kMaxVmArrayBudget)
+		if ((vmArrayOffset + vmDataByteSize) <= g_GlobalConstants.MaxOutOmmArraySize)
 		{
 			// Allocate new VM-desc for the vmArrayOffset
 			{
@@ -147,7 +146,7 @@ void main(uint3 tid : SV_DispatchThreadID)
 	else // if (status == hashTable::Result::Found
 	{
 		// Store the hash-table offset and patch up the pointers later.
-		vmDescOffset = (uint)(-hashTableEntryIndex - 4);
+		vmDescOffset = (uint)(-hashTableEntryIndex - 5);
 	}
 
 	OMM_SUBRESOURCE_STORE(TempOmmIndexBuffer, 4 * primitiveIndex, vmDescOffset);