shader: separate exact descriptor row bounds from pointer offsets

This commit is contained in:
nmzik
2026-10-14 16:55:47 +02:00
parent 487bbf95fd
commit 04cb4c6553
4 changed files with 15 additions and 19 deletions
+2 -7
View File
@@ -525,12 +525,6 @@ struct ShaderInfo {
struct DescriptorSource { struct DescriptorSource {
struct IndirectDescriptor { struct IndirectDescriptor {
struct PointerTable {
uint32_t maximum = 0;
uint32_t offset = 0;
bool operator==(const PointerTable&) const = default;
};
struct SelectorRead { struct SelectorRead {
struct CountedRows { struct CountedRows {
uint32_t source = UINT32_MAX; uint32_t source = UINT32_MAX;
@@ -553,7 +547,8 @@ struct DescriptorSource {
bool operator==(const SelectorRead& other) const = default; bool operator==(const SelectorRead& other) const = default;
}; };
std::optional<SelectorRead> selector; std::optional<SelectorRead> selector;
std::optional<PointerTable> pointer_table; std::optional<uint32_t> table_maximum;
std::optional<uint32_t> pointer_offset;
uint32_t table_source = 0; uint32_t table_source = 0;
uint32_t table_offset = 0; uint32_t table_offset = 0;
uint32_t table_immediate = 0; uint32_t table_immediate = 0;
@@ -367,13 +367,12 @@ bool MaterializeIndirectDescriptor(const ResourcePlan& p
(table.Stride() & 3u) != 0u || table.Stride() < indirect.table_record_bytes))) (table.Stride() & 3u) != 0u || table.Stride() < indirect.table_record_bytes)))
return false; return false;
const bool affine = indirect.table_stride != 0u; const bool affine = indirect.table_stride != 0u;
const auto* pointer = indirect.pointer_table ? &*indirect.pointer_table : nullptr; const uint64_t step = indirect.table_maximum ? indirect.table_stride
const uint64_t step = pointer != nullptr ? indirect.table_stride
: affine ? uint64_t {1} << std::countr_zero(indirect.table_stride) : 4u; : affine ? uint64_t {1} << std::countr_zero(indirect.table_stride) : 4u;
const uint64_t first = ((pointer != nullptr ? indirect.table_offset const uint64_t first = ((indirect.table_maximum ? indirect.table_offset
: indirect.table_offset % step) & ~uint64_t {3}) + : indirect.table_offset % step) & ~uint64_t {3}) +
(indirect.table_immediate & ~3u); (indirect.table_immediate & ~3u);
const uint64_t domain = pointer != nullptr ? uint64_t {pointer->maximum} + 1u const uint64_t domain = indirect.table_maximum ? uint64_t {*indirect.table_maximum} + 1u
: (uint64_t {1} << 32u) / step; : (uint64_t {1} << 32u) / step;
const auto count = table_size >= first + 4u const auto count = table_size >= first + 4u
? std::min((table_size - first - 4u) / step + 1u, domain) ? std::min((table_size - first - 4u) / step + 1u, domain)
@@ -390,11 +389,11 @@ bool MaterializeIndirectDescriptor(const ResourcePlan& p
snapshot.flattened_srt.resize(mapping_offset + 2u, 0u); snapshot.flattened_srt.resize(mapping_offset + 2u, 0u);
for (uint32_t word = 0; word < count; ++word) { for (uint32_t word = 0; word < count; ++word) {
DescriptorValue candidate {.dword_count = dword_count}; DescriptorValue candidate {.dword_count = dword_count};
if (pointer != nullptr) { if (indirect.pointer_offset) {
std::array<uint32_t, 2> address {}; std::array<uint32_t, 2> address {};
if (!ReadScalarTable(table_base, table_size, first + word * step, runtime, address) || if (!ReadScalarTable(table_base, table_size, first + word * step, runtime, address) ||
!ReadScalarTable((uint64_t {address[1]} << 32u) | address[0], UINT64_MAX, !ReadScalarTable((uint64_t {address[1]} << 32u) | address[0], UINT64_MAX,
pointer->offset, runtime, *indirect.pointer_offset, runtime,
std::span(candidate.dwords).first(dword_count))) return false; std::span(candidate.dwords).first(dword_count))) return false;
} else if (affine) { } else if (affine) {
if (!ReadScalarTable(table_base, table_size, first + word * step, runtime, if (!ReadScalarTable(table_base, table_size, first + word * step, runtime,
@@ -400,7 +400,7 @@ public:
std::vector<std::pair<const IndirectDescriptorPlan*, Value>> buffer_selections; std::vector<std::pair<const IndirectDescriptorPlan*, Value>> buffer_selections;
for (const auto& plan: m_indirect_descriptors) { for (const auto& plan: m_indirect_descriptors) {
auto key = plan.key; auto key = plan.key;
if (m_sources[plan.source].indirect_descriptor->pointer_table) { if (m_sources[plan.source].indirect_descriptor->table_maximum) {
auto* block = plan.handle->Parent(); auto* block = plan.handle->Parent();
const auto where = std::ranges::find_if(block->Instructions(), const auto where = std::ranges::find_if(block->Instructions(),
[&](const Inst& inst) { return &inst == plan.handle; }); [&](const Inst& inst) { return &inst == plan.handle; });
@@ -974,7 +974,8 @@ private:
if (current.indirect_descriptor.has_value()) { if (current.indirect_descriptor.has_value()) {
const auto& a = *current.indirect_descriptor; const auto& a = *current.indirect_descriptor;
const auto& b = *descriptor.indirect_descriptor; const auto& b = *descriptor.indirect_descriptor;
if (a.selector != b.selector || a.pointer_table != b.pointer_table || if (a.selector != b.selector || a.table_maximum != b.table_maximum ||
a.pointer_offset != b.pointer_offset ||
a.table_source != b.table_source || a.table_source != b.table_source ||
a.table_offset != b.table_offset || a.table_immediate != b.table_immediate || a.table_offset != b.table_offset || a.table_immediate != b.table_immediate ||
a.table_stride != b.table_stride || a.indexed != b.indexed || a.table_stride != b.table_stride || a.indexed != b.indexed ||
@@ -2779,7 +2780,8 @@ private:
indirect.table_source = InternSource(table_source); indirect.table_source = InternSource(table_source);
indirect.table_immediate = low_memory->offset & ~3u; indirect.table_immediate = low_memory->offset & ~3u;
indirect.table_scalar = true; indirect.table_scalar = true;
indirect.pointer_table = {{maximum, descriptor_offset}}; indirect.table_maximum = maximum;
indirect.pointer_offset = descriptor_offset;
DescriptorSource source; DescriptorSource source;
source.dword_count = handle.NumArgs(); source.dword_count = handle.NumArgs();
source.dwords.fill(Value(0u)); source.dwords.fill(Value(0u));
+3 -3
View File
@@ -656,9 +656,9 @@ void TestNestedPointerImageTable() {
for (const auto source : {plan.info.images[0].source, plan.info.samplers[0].source}) { for (const auto source : {plan.info.images[0].source, plan.info.samplers[0].source}) {
const auto& descriptor = plan.descriptor_sources[source]; const auto& descriptor = plan.descriptor_sources[source];
const auto& indirect = descriptor.indirect_descriptor; const auto& indirect = descriptor.indirect_descriptor;
Check(indirect && !indirect->selector && indirect->pointer_table && Check(indirect && !indirect->selector && indirect->table_maximum == 65535u &&
indirect->pointer_table->maximum == 65535u && indirect->table_stride == 592u && indirect->table_stride == 592u &&
indirect->pointer_table->offset == (descriptor.dword_count == 8u ? 0x3d0u : 0x3f8u), indirect->pointer_offset == (descriptor.dword_count == 8u ? 0x3d0u : 0x3f8u),
"nested pointer descriptors lost their row geometry or captured the GPU selector payload"); "nested pointer descriptors lost their row geometry or captured the GPU selector payload");
} }