mirror of
https://github.com/KytyPS5/KytyPS5.git
synced 2026-10-27 13:26:52 +01:00
shader: read bit-scan records only at ranks their mask buffer words set
This commit is contained in:
@@ -331,6 +331,18 @@ bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::In
|
||||
keys.resize(count);
|
||||
return true;
|
||||
};
|
||||
// Reads the key of each record whose rank is set.
|
||||
const auto read_ranks = [&](uint64_t base, uint64_t size, uint64_t first, uint64_t step,
|
||||
uint32_t ranks) {
|
||||
keys.reserve(std::popcount(ranks));
|
||||
for (; ranks != 0u; ranks &= ranks - 1u) {
|
||||
const auto offset = first + std::countr_zero(ranks) * step;
|
||||
uint32_t key = 0;
|
||||
if (offset > UINT32_MAX || !read_selector(base, size, offset, {&key, 1})) return false;
|
||||
keys.push_back(key);
|
||||
}
|
||||
return true;
|
||||
};
|
||||
if (selector->rows) {
|
||||
const auto& rows = *selector->rows;
|
||||
DescriptorValue count_value;
|
||||
@@ -450,8 +462,32 @@ bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::In
|
||||
// Out-of-bounds scalar reads return zero.
|
||||
const auto end = count != 0u ? last * step + selector->offset + 4u : 0u;
|
||||
if (end > uint64_t {UINT32_MAX} + 1u) return false;
|
||||
if (!read_keys(base, size, uint64_t {first} * step + selector->offset, step, count))
|
||||
const auto start = uint64_t {first} * step + selector->offset;
|
||||
if (selector->ranks) {
|
||||
// A scan selects the ranks some word of its mask's address class sets.
|
||||
const auto& mask = *selector->ranks;
|
||||
DescriptorValue mask_value;
|
||||
ShaderBufferResource masks;
|
||||
if (count > 32u || !WalkedDescriptor(program, mask.source, mask_value) ||
|
||||
!DecodeBufferDescriptor(mask_value, masks)) return false;
|
||||
// Each chunk holds whole strides of the power-of-two class, or one of its words.
|
||||
const auto words = mask.stride / 4u;
|
||||
const auto chunk =
|
||||
words <= MaxIndirectDescriptorProbes ? MaxIndirectDescriptorProbes : 1u;
|
||||
uint32_t ranks = 0;
|
||||
for (auto at = uint64_t {mask.offset}; at + 4u <= masks.GetSize();
|
||||
at += uint64_t {std::max(chunk, words)} * 4u) {
|
||||
keys.resize(std::min<uint64_t>((masks.GetSize() - at) / 4u, chunk));
|
||||
if (!ReadScalarTable(masks.Base48(), masks.GetSize(), at, runtime, keys))
|
||||
return false;
|
||||
for (size_t word = 0; word < keys.size(); word += words) ranks |= keys[word];
|
||||
}
|
||||
keys.clear();
|
||||
if (count < 32u) ranks &= (1u << count) - 1u;
|
||||
if (!read_ranks(base, size, start, step, ranks)) return false;
|
||||
} else if (!read_keys(base, size, start, step, count)) {
|
||||
return false;
|
||||
}
|
||||
} else if (!indirect.selector_mask.IsEmpty()) {
|
||||
uint32_t mask = 0;
|
||||
uint32_t count = 0;
|
||||
@@ -463,18 +499,8 @@ bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::In
|
||||
if (count < 32u) mask &= (1u << count) - 1u;
|
||||
const auto material_base =
|
||||
(static_cast<uint64_t>(material_value.dwords[1]) << 32u) | material_value.dwords[0];
|
||||
keys.reserve(std::popcount(mask));
|
||||
while (mask != 0u) {
|
||||
const auto index = std::countr_zero(mask);
|
||||
const auto offset = static_cast<uint64_t>(selector->offset) +
|
||||
static_cast<uint64_t>(index) * selector->stride;
|
||||
if (offset > UINT32_MAX) return false;
|
||||
uint32_t key = 0;
|
||||
if (!read_selector(material_base, UINT64_MAX, static_cast<uint32_t>(offset),
|
||||
{&key, 1})) return false;
|
||||
keys.push_back(key);
|
||||
mask &= mask - 1u;
|
||||
}
|
||||
if (!read_ranks(material_base, UINT64_MAX, selector->offset, selector->stride, mask))
|
||||
return false;
|
||||
} else if (selector->stride == 0u) {
|
||||
ShaderBufferResource material;
|
||||
if (!DecodeBufferDescriptor(material_value, material) ||
|
||||
|
||||
Reference in New Issue
Block a user