Dynamitos/Seele
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

The first frames work finally

Browse Source
This commit is contained in:
Dynamitos
2024-10-19 11:54:04 +02:00
parent 1193406dd8
commit 17746f8d20
11 changed files with 101 additions and 134 deletions
Download Patch File Download Diff File Expand all files Collapse all files
res/shaders/terrain/CBT.slang
-107
View File
@@ -230,7 +230,6 @@ uint bit_count(uint depth, uint element)
// decodes the position of the i-th one in the bitfield
uint decode_bit(uint handle)
{
#if defined(NAIVE_DECODE)
uint bitID = 1;
for (uint currentDepth = 0; currentDepth < WAVE_TREE_DEPTH; ++currentDepth)
{
@@ -242,65 +241,11 @@ uint decode_bit(uint handle)
}
return (bitID ^ OCBT_NUM_ELEMENTS);
#else
uint currentDepth = 0;
uint heapElementID = 1u;
for (currentDepth = 0; currentDepth < FIRST_VIRTUAL_LEVEL; ++currentDepth)
{
// Read the left element
uint heapValue = get_heap_element(2u * heapElementID);
// Does it fall in the right or left subtree?
uint b = handle < heapValue ? 0u : 1u;
// Pick a subtree
heapElementID = 2u * heapElementID + b;
// Move the iterator to exclude the right subtree if required
handle -= heapValue * b;
}
// Align with the internal depth
currentDepth++;
// Ok we have our subtree, now we need to pick the right bit
uint64_t heapValue = pParams.bitFieldBuffer[heapElementID - OCBT_LAST_LEVEL_SIZE * 2];
uint64_t mask = 0xffffffff;
uint32_t bitCount = 32;
for (; currentDepth < (WAVE_TREE_DEPTH + 1); ++currentDepth)
{
// Figure out the location of the first bit of this element
uint real_heap_id = 2 * heapElementID - 1;
uint level_first_element = (1u << currentDepth) - 1;
uint id_in_level = real_heap_id - level_first_element;
uint first_bit = bitCount * id_in_level;
uint local_id = first_bit % 64;
uint64_t target_bits = (heapValue >> local_id) & mask;
uint32_t high = uint(target_bits >> 32);
uint32_t low = uint(target_bits);
uint heapValue = countbits(high) + countbits(low);
// Does it fall in the right or left subtree?
uint b = handle < heapValue ? 0u : 1u;
// Pick a subtree
heapElementID = 2u * heapElementID + b;
// Move the iterator to exclude the right subtree if required
handle -= heapValue * b;
// Adjust the mask and bitcount
bitCount /= 2;
mask = mask >> bitCount;
}
return (heapElementID ^ OCBT_NUM_ELEMENTS);
#endif
}
// decodes the position of the i-th zero in the bitfield
uint decode_bit_complement(uint handle)
{
#if defined(NAIVE_DECODE)
uint bitID = 1u;
uint c = OCBT_NUM_ELEMENTS / 2u;
@@ -314,54 +259,6 @@ uint decode_bit_complement(uint handle)
}
return (bitID ^ OCBT_NUM_ELEMENTS);
#else
uint heapElementID = 1u;
uint c = OCBT_NUM_ELEMENTS / 2u;
uint currentDepth = 0;
for (currentDepth = 0; currentDepth < FIRST_VIRTUAL_LEVEL; ++currentDepth)
{
uint heapValue = c - get_heap_element(2u * heapElementID);
uint b = handle < heapValue ? 0u : 1u;
heapElementID = 2u * heapElementID + b;
handle -= heapValue * b;
c /= 2u;
}
// Align with the internal depth
currentDepth++;
// Ok we have our subtree, now we need to pick the right bit
uint64_t heapValue = pParams.bitFieldBuffer[heapElementID - OCBT_LAST_LEVEL_SIZE * 2];
uint64_t mask = 0xffffffff;
uint32_t bitCount = 32;
for (; currentDepth < (WAVE_TREE_DEPTH + 1); ++currentDepth)
{
// Figure out the location of the first bit of this element
uint real_heap_id = 2 * heapElementID - 1;
uint level_first_element = (1u << currentDepth) - 1;
uint id_in_level = real_heap_id - level_first_element;
uint first_bit = bitCount * id_in_level;
uint local_id = first_bit % 64;
uint64_t target_bits = (heapValue >> local_id) & mask;
uint32_t high = uint(target_bits >> 32);
uint32_t low = uint(target_bits);
uint heapValue = c - (countbits(high) + countbits(low));
uint b = handle < heapValue ? 0u : 1u;
heapElementID = 2u * heapElementID + b;
handle -= heapValue * b;
c /= 2u;
// Adjust the mask and bitcount
bitCount /= 2;
mask = mask >> bitCount;
}
return (heapElementID ^ OCBT_NUM_ELEMENTS);
#endif
}
void reduce(uint groupIndex)
@@ -578,11 +475,7 @@ void load_shared_memory_to_buffer(uint groupIndex)
// Load the bitfield to the LDS
for (uint e = 0; e < BUFFER_ELEMENT_PER_LANE; ++e)
{
#ifdef AMD
uint target_element = BUFFER_ELEMENT_PER_LANE * groupIndex + e;
#else
uint target_element = groupIndex + WORKGROUP_SIZE * e;
#endif
if (target_element < OCBT_TREE_NUM_SLOTS)
pParams.cbtBuffer[target_element] = gs_cbtTree[target_element];
}