raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-19 18:36:53 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: share formatted loads and support packed D16 XYZW buffers

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 04:09:57 +02:00
parent 1905b28b3c
commit 6abfe3bbe4
8 changed files with 170 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+93 -2
View File
@@ -30951,6 +30951,7 @@ std::vector<TestCase> BufferLoadFormatD16XCases() {
u32 index = 0;
bool active = true;
bool overlap = false;
u32 stride = 4;
};
constexpr LoadCase inputs[] = {
{"BufferLoadFormatD16XFloat16", F::k16Float, 0x7e63c000u, 0xabcdc000u},
@@ -30971,6 +30972,8 @@ std::vector<TestCase> BufferLoadFormatD16XCases() {
{"BufferLoadFormatD16XOobFloatOne", F::k32Float, 0u, 0xabcd3c00u, 1, 4},
{"BufferLoadFormatD16XOobUintOne", F::k32UInt, 0u, 0xabcd0001u, 1, 4},
{"BufferLoadFormatD16XOobZero", F::k32Float, 0u, 0xabcd0000u, 4, 4},
{"BufferLoadFormatD16XZeroStrideFloatOne", F::k32Float, 0u, 0xabcd3c00u, 1, 0, true, false, 0},
{"BufferLoadFormatD16XZeroStrideUintOne", F::k32UInt, 0u, 0xabcd0001u, 1, 0, true, false, 0},
{"BufferLoadFormatD16XInactive", F::k32Float, 0u, 0xabcd1234u, 4, 0, false},
{"BufferLoadFormatD16XOverlappingAddress", F::k16UInt, 7u, 7u, 4, 0, true, true},
};
@@ -30993,8 +30996,8 @@ std::vector<TestCase> BufferLoadFormatD16XCases() {
AppendEnd(&test.code);
test.initial = {input.input, 0xdeadbeefu, 0u, 0xcafef00du};
test.expected = {input.input, 0xdeadbeefu, input.expected, 0xcafef00du};
test.user_data = MakeStructuredStorageBufferData(4, 4, false, BufferFormat(input.format));
test.user_data[3] = (test.user_data[3] & ~7u) | input.selector;
test.user_data = MakeStructuredStorageBufferData(input.stride, 4, false, BufferFormat(input.format));
test.user_data[3] = (test.user_data[3] & ~0x30000007u) | input.selector;
test.has_user_data = true;
test.opcodes = {O::V_MOV_B32, O::BUFFER_LOAD_FORMAT_D16_X,
O::BUFFER_STORE_DWORD, O::S_ENDPGM};
@@ -31003,6 +31006,88 @@ std::vector<TestCase> BufferLoadFormatD16XCases() {
return cases;
}
std::vector<TestCase> BufferLoadFormatD16XyzwCases() {
using F = Prospero::BufferFormat;
using O = ShaderOpcode;
struct LoadCase {
const char *name;
F format;
std::array<u32, 4> input;
std::array<u32, 2> expected;
u32 swizzle = DstSel(4, 5, 6, 7);
u32 stride = 16;
u32 index = 0;
bool active = true;
bool overlap = false;
};
constexpr LoadCase inputs[] = {
{"BufferLoadFormatD16XyzwFloat32", F::k32_32_32_32Float,
{0x3f800000u, 0x3f000000u, 0u, 0xc0000000u}, {0x38003c00u, 0xc0000000u}},
{"BufferLoadFormatD16XyzwFloat16", F::k16_16_16_16Float,
{0x38003c00u, 0xc0000000u}, {0x38003c00u, 0xc0000000u}},
{"BufferLoadFormatD16XyzwUintSaturates", F::k32_32_32_32UInt,
{0u, 0x10000u, 0xffffu, 7u}, {0xffff0000u, 0x0007ffffu}},
{"BufferLoadFormatD16XyzwSintSaturates", F::k32_32_32_32SInt,
{0xffff7fffu, 7u, 0x8000u, 0xffff8000u}, {0x00078000u, 0x80007fffu}},
{"BufferLoadFormatD16XyzwUnorm", F::k8_8_8_8UNorm,
{0xffff0000u}, {0u, 0x3c003c00u}},
{"BufferLoadFormatD16XyzwSwizzle", F::k32_32_32_32Float,
{0u, 0x3f800000u, 0xc0000000u, 0x3f000000u},
{0xc0003800u, 0x3c003c00u}, DstSel(7, 6, 5, 1)},
{"BufferLoadFormatD16XyzwPartialRecord", F::k32_32_32_32Float,
{0x3f800000u, 0x3f800000u, 0x3f800000u, 0x3f800000u},
{0u, 0u}, DstSel(4, 5, 6, 7), 12},
{"BufferLoadFormatD16XyzwOobOne", F::k32_32_32_32Float,
{}, {0x3c00u, 0x3c00u}, DstSel(1, 4, 1, 4), 16, 1},
{"BufferLoadFormatD16XyzwZeroStrideFloatOne", F::k32_32_32_32Float,
{}, {0x3c003c00u, 0x3c000000u}, DstSel(1, 1, 4, 1), 0},
{"BufferLoadFormatD16XyzwZeroStrideUintOne", F::k32_32_32_32UInt,
{}, {0x10000u, 1u}, DstSel(4, 1, 1, 4), 0},
{"BufferLoadFormatD16XyzwInactive", F::k32_32_32_32Float,
{}, {0xabcd1234u, 0xbeef5678u}, DstSel(4, 5, 6, 7), 16, 0, false},
{"BufferLoadFormatD16XyzwOverlappingAddress", F::k32_32_32_32Float,
{0x3f800000u, 0x3f000000u, 0u, 0xc0000000u},
{0x38003c00u, 0xc0000000u}, DstSel(4, 5, 6, 7), 16, 0, true, true},
};
std::vector<TestCase> cases;
for (const auto &input : inputs) {
TestCase test;
test.name = input.name;
AppendVMovLiteral(&test.code, 16, input.overlap ? input.index : 0xabcd1234u);
AppendVMovLiteral(&test.code, 17, 0xbeef5678u);
AppendVMovLiteral(&test.code, 18, 0xcafef00du);
AppendVMovU32(&test.code, 12, input.index);
if (!input.active) test.code.push_back(EncodeSop1(0x04, 126, InlineU32(0)));
// Captured b21e373811f28e31 pc 0x1914: e20c2000 8001100c.
test.code.push_back(EncodeMubuf0(0x83u, 0, true, false));
test.code.push_back(EncodeMubuf1(16, 1, input.overlap ? 16 : 12));
if (!input.active) {
test.code.push_back(EncodeSMovB32(126, InlineU32(1)));
test.code.push_back(EncodeSMovB32(127, InlineU32(0)));
}
for (u32 word = 0; word < 3; word++) AppendStoreVgpr(&test.code, 16 + word, 4 + word);
AppendEnd(&test.code);
test.initial.assign(input.input.begin(), input.input.end());
test.initial.resize(8, 0xdeadbeefu);
test.expected = test.initial;
test.expected[4] = input.expected[0];
test.expected[5] = input.expected[1];
test.expected[6] = 0xcafef00du;
const auto descriptor = MakeStructuredStorageBufferData(
input.stride, 1, false, BufferFormat(input.format));
std::copy_n(descriptor.begin(), 4, test.user_data.begin() + 4);
test.user_data[7] = (test.user_data[7] & ~0x30000fffu) | input.swizzle;
test.has_user_data = true;
if (input.stride != 0 && input.active) {
test.storage_buffer_ranges = {{0, input.stride}, {0, 32}};
}
test.opcodes = {O::V_MOV_B32, O::BUFFER_LOAD_FORMAT_D16_XYZW,
O::BUFFER_STORE_DWORD, O::S_ENDPGM};
cases.push_back(std::move(test));
}
return cases;
}
std::vector<TestCase> BufferStoreFormatD16XCases() {
using F = Prospero::BufferFormat;
using O = ShaderOpcode;
@@ -39315,6 +39400,9 @@ std::vector<TestCase> MakeCases() {
for (auto &test : BufferLoadFormatD16XCases()) {
cases.push_back(std::move(test));
}
for (auto &test : BufferLoadFormatD16XyzwCases()) {
cases.push_back(std::move(test));
}
for (auto &test : BufferStoreFormatD16XCases()) {
cases.push_back(std::move(test));
}
@@ -44982,6 +45070,9 @@ int main(int argc, char **argv) {
for (const auto &test : BufferLoadFormatD16XCases()) {
RunCase(&vulkan, test);
}
for (const auto &test : BufferLoadFormatD16XyzwCases()) {
RunCase(&vulkan, test);
}
for (const auto &test : BufferStoreFormatD16XCases()) {
RunCase(&vulkan, test);
}