raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 10:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: support RDNA2 BVH64 ray intersections with shared traversal

Browse Source
This commit is contained in:
nmzik
2026-10-04 11:59:42 +04:00
parent 33e74cab2c
commit dbf66b7f4f
6 changed files with 77 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRayTracingGpuTests.inc
+26 -9
View File
@@ -83,13 +83,15 @@ TestCase FlatStackApertures(u32 wave_size) {
return test;
}
TestCase BvhIntersections(bool barycentrics, bool sorted, u32 grow = 0) {
TestCase BvhIntersections(bool barycentrics, bool sorted, u32 grow = 0, bool wide = false) {
TestCase test;
test.name = barycentrics ? (sorted ? "BvhBarycentricSorted" : "BvhBarycentricUnsorted")
: (sorted ? "BvhTriangleIdSorted" : "BvhTriangleIdUnsorted");
if (grow != 0) test.name = "BvhBoxExitGrowth";
if (wide) test.name = "Bvh64HighNodePointers";
constexpr u32 NodeStart = 1024, OutputStart = 512, LaneCount = 32;
constexpr uint64_t GuestBase = 0x10000;
const u32 node_base = wide ? 1u << 29 : 0;
const auto bits = [](float value) { return std::bit_cast<u32>(value); };
constexpr u32 Infinity = 0x7f800000, Invalid = UINT32_MAX, Flag = 0x04040404;
test.initial.resize(NodeStart + 258 * 16);
@@ -183,6 +185,19 @@ TestCase BvhIntersections(bool barycentrics, bool sorted, u32 grow = 0) {
} else if (lane == 29) {
data[input] = (255 << 3) | 5;
}
if (wide) {
// Keep the same small backing while placing nodes beyond the 32-bit pointer range.
std::copy_backward(data.begin() + input + 1, data.begin() + input + 11,
data.begin() + input + 12);
data[input + 1] = 1;
if (lane == 14 || lane == 15) {
data[input] = lane == 14 ? 0 : 5;
// Neither an oversized index nor a shifted-address wrap may read node zero.
data[input + 1] = lane == 14 ? 0x00001000 : 0x20000000;
expected[lane] = lane == 14 ? std::array<u32, 4>{Infinity, bits(1), Invalid, 0}
: std::array<u32, 4>{Invalid, Invalid, Invalid, Invalid};
}
}
}
test.expected = test.initial;
for (u32 lane = 0; lane < LaneCount; ++lane) {
@@ -198,29 +213,31 @@ TestCase BvhIntersections(bool barycentrics, bool sorted, u32 grow = 0) {
// The resource is a GPU value, not a materialized host image descriptor.
AppendSMovLiteral(&code, 8, GuestBase >> 8);
AppendSMovLiteral(&code, 9, (sorted ? 0x80000000 : 0) | (grow << 23));
AppendSMovLiteral(&code, 10, 256);
AppendSMovLiteral(&code, 10, node_base + 256);
AppendSMovLiteral(&code, 11, barycentrics ? 0x81000000 : 0x80000000);
code.push_back(EncodeVop2(0x1a, 1, InlineU32(6), 0));
for (u32 offset : {0u, 4u, 8u}) {
code.insert(code.end(), {EncodeMubuf0(offset == 8 ? 0x0f : 0x0e, offset * 4),
code.insert(code.end(), {EncodeMubuf0(offset == 8 && !wide ? 0x0f : 0x0e, offset * 4),
EncodeMubuf1(4 + offset, 0, 1)});
}
code.insert(code.end(), {EncodeMimg0(0xe6, 15, 0, false, 0, true),
code.insert(code.end(), {EncodeMimg0(wide ? 0xe7 : 0xe6, 15, 0, false, 0, true),
EncodeMimg1(20, 4, 2)});
code.push_back(EncodeVop2(0x1a, 2, InlineU32(2), 0));
for (u32 i = 0; i < 4; ++i) AppendStoreVgprAtLaneDwordOffset(&code, 20 + i, 2, OutputStart + i);
// NSA addressing, the FP32 second-block bound, and inactive destination lanes.
AppendSMovLiteral(&code, 10, 2);
AppendSMovLiteral(&code, 10, node_base + 2);
for (u32 i = 0; i < 4; ++i) AppendVMovLiteral(&code, 24 + i, 0x12345678 + i);
AppendSMovLiteral(&code, 126, 0x000fffff);
code.insert(code.end(), {EncodeMimg0(0xe6, 15, 3, false, 0, true),
EncodeMimg1(24, 4, 2), 0x08070605, 0x0c0b0a09, 0x00000e0d});
code.insert(code.end(), {EncodeMimg0(wide ? 0xe7 : 0xe6, 15, 3, false, 0, true),
EncodeMimg1(24, 4, 2), 0x08070605, 0x0c0b0a09,
wide ? 0x000f0e0du : 0x00000e0du});
code.push_back(EncodeSop1(0x04, 126, 193));
for (u32 i = 0; i < 4; ++i)
AppendStoreVgprAtLaneDwordOffset(&code, 24 + i, 2, OutputStart + LaneCount * 4 + i);
AppendEnd(&code);
test.bda_mappings = {{GuestBase, NodeStart * sizeof(u32)},
{GuestBase + 0x4000, (NodeStart + 257 * 16) * sizeof(u32)}};
const auto nodes_address = GuestBase + (uint64_t{node_base} << 6);
test.bda_mappings = {{nodes_address, NodeStart * sizeof(u32)},
{nodes_address + 0x4000, (NodeStart + 257 * 16) * sizeof(u32)}};
test.compute_info.wave_size = LaneCount;
test.compute_info.threads_num[0] = LaneCount;
test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1;