Optimize GPU particles Bitonic sort to use separate buffers for indices and keys to avoid additional buffer copy

This commit is contained in:
Wojtek Figat
2025-08-08 18:24:44 +02:00
parent 519a9c0a14
commit 854f3acd4c
9 changed files with 115 additions and 169 deletions

View File

@@ -10,12 +10,12 @@
struct Item
{
float Key;
uint Value;
uint Index;
};
META_CB_BEGIN(0, Data)
float NullItemKey;
uint NullItemValue;
uint NullItemIndex;
uint CounterOffset;
uint MaxIterations;
uint LoopK;
@@ -40,12 +40,12 @@ uint InsertOneBit(uint value, uint oneBitMask)
// (effectively a negation) or leave the value alone. When the KeySign is
// 1, we are sorting descending, so when A < B, they should swap. For an
// ascending sort, -A < -B should swap.
bool ShouldSwap(Item a, Item b)
bool ShouldSwap(float a, float b)
{
//return (a ^ NullItem) < (b ^ NullItem);
//return (a.Key) < (b.Key);
return (a.Key * KeySign) < (b.Key * KeySign);
//return (a) < (b);
return (a * KeySign) < (b * KeySign);
//return asfloat(a) < asfloat(b);
//return (asfloat(a) * KeySign) < (asfloat(b) * KeySign);
}
@@ -93,7 +93,8 @@ void CS_IndirectArgs(uint groupIndex : SV_GroupIndex)
#if defined(_CS_PreSort) || defined(_CS_InnerSort)
RWStructuredBuffer<Item> SortBuffer : register(u0);
RWBuffer<uint> SortedIndices : register(u0);
RWBuffer<float> SortingKeys : register(u1);
groupshared Item SortData[THREAD_GROUP_SIZE * 2];
@@ -103,12 +104,13 @@ void LoadItem(uint element, uint count)
Item item;
if (element < count)
{
item = SortBuffer[element];
item.Key = SortingKeys[element];
item.Index = SortedIndices[element];
}
else
{
item.Key = NullItemKey;
item.Value = NullItemValue;
item.Index = NullItemIndex;
}
SortData[element & (THREAD_GROUP_SIZE * 2 - 1)] = item;
}
@@ -117,7 +119,9 @@ void StoreItem(uint element, uint count)
{
if (element < count)
{
SortBuffer[element] = SortData[element & 2047];
Item item = SortData[element & ((THREAD_GROUP_SIZE * 2 - 1))];
SortingKeys[element] = item.Key;
SortedIndices[element] = item.Index;
}
}
@@ -153,7 +157,7 @@ void CS_PreSort(uint3 groupID : SV_GroupID, uint groupIndex : SV_GroupIndex)
Item a = SortData[index1];
Item b = SortData[index2];
if (ShouldSwap(a, b))
if (ShouldSwap(a.Key, b.Key))
{
// Swap the items
SortData[index1] = b;
@@ -197,7 +201,7 @@ void CS_InnerSort(uint3 groupID : SV_GroupID, uint groupIndex : SV_GroupIndex)
Item a = SortData[index1];
Item b = SortData[index2];
if (ShouldSwap(a, b))
if (ShouldSwap(a.Key, b.Key))
{
// Swap the items
SortData[index1] = b;
@@ -215,7 +219,8 @@ void CS_InnerSort(uint3 groupID : SV_GroupID, uint groupIndex : SV_GroupIndex)
#ifdef _CS_OuterSort
RWStructuredBuffer<Item> SortBuffer : register(u0);
RWBuffer<uint> SortedIndices : register(u0);
RWBuffer<float> SortingKeys : register(u1);
META_CS(true, FEATURE_LEVEL_SM5)
[numthreads(1024, 1, 1)]
@@ -230,35 +235,19 @@ void CS_OuterSort(uint3 dispatchThreadId : SV_DispatchThreadID)
if (index2 >= count)
return;
Item a = SortBuffer[index1];
Item b = SortBuffer[index2];
float aKey = SortingKeys[index1];
float bKey = SortingKeys[index2];
if (ShouldSwap(a, b))
if (ShouldSwap(aKey, bKey))
{
// Swap the items
SortBuffer[index1] = b;
SortBuffer[index2] = a;
SortingKeys[index1] = bKey;
SortingKeys[index2] = aKey;
uint aIndex = SortedIndices[index1];
uint bIndex = SortedIndices[index2];
SortedIndices[index1] = bIndex;
SortedIndices[index2] = aIndex;
}
}
#endif
#ifdef _CS_CopyIndices
StructuredBuffer<Item> SortBuffer : register(t1);
RWBuffer<uint> SortedIndices : register(u0);
META_CS(true, FEATURE_LEVEL_SM5)
[numthreads(1024, 1, 1)]
void CS_CopyIndices(uint3 dispatchThreadId : SV_DispatchThreadID)
{
const uint count = CounterBuffer.Load(CounterOffset);
uint index = dispatchThreadId.x;
if (index >= count)
return;
Item element = SortBuffer[index];
SortedIndices[index] = element.Value;
}
#endif

View File

@@ -20,13 +20,9 @@ META_CB_END
// Particles data buffer
ByteAddressBuffer ParticlesData : register(t0);
// Output sorting keys buffer (index + key)
struct Item
{
float Key;
uint Value;
};
RWStructuredBuffer<Item> SortingKeys : register(u0);
// Sorting data (per-particle)
RWBuffer<uint> SortedIndices : register(u0);
RWBuffer<float> SortingKeys : register(u1);
float GetParticleFloat(uint particleIndex, int offset)
{
@@ -78,8 +74,6 @@ void CS_Sort(uint3 dispatchThreadId : SV_DispatchThreadID)
#endif
// Write sorting index-key pair
Item item;
item.Key = sortKey;
item.Value = index;
SortingKeys[index] = item;
SortedIndices[index] = index;
SortingKeys[index] = sortKey;
}