raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 02:47:02 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: support packed D16 XYZW formatted buffer stores

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 19:35:08 +02:00
parent c914435f20
commit 4843d8d942
8 changed files with 107 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+70 -22
View File
@@ -31393,39 +31393,87 @@ std::vector<TestCase> BufferLoadFormatD16XyzwCases() {
return cases;
}
std::vector<TestCase> BufferStoreFormatD16XCases() {
std::vector<TestCase> BufferStoreFormatD16Cases() {
using F = Prospero::BufferFormat;
using O = ShaderOpcode;
struct StoreCase {
const char *name;
F format;
u32 input;
u32 expected;
std::array<u32, 2> input;
std::array<u32, 4> expected;
u32 words = 1;
bool wide = false;
};
constexpr StoreCase inputs[] = {
{"BufferStoreFormatD16XFloat16", F::k16Float, 0x7e63c000u, 0xffffc000u},
{"BufferStoreFormatD16XFloat32", F::k32Float, 0x7e633800u, 0x3f000000u},
{"BufferStoreFormatD16XUint32", F::k32UInt, 0x7e63fffeu, 0x0000fffeu},
{"BufferStoreFormatD16XSint32", F::k32SInt, 0x7e63fffeu, 0xfffffffeu},
{"BufferStoreFormatD16XUnorm8", F::k8UNorm, 0x7e633800u, 0xffffff80u},
{"BufferStoreFormatD16XSnorm16", F::k16SNorm, 0x7e63b800u, 0xffffc000u},
{"BufferStoreFormatD16XZeroFill", F::k16_16Float, 0x7e633800u, 0x00003800u},
{"BufferStoreFormatD16XFloat16", F::k16Float, {0x7e63c000u}, {0xffffc000u}},
{"BufferStoreFormatD16XFloat32", F::k32Float, {0x7e633800u}, {0x3f000000u}},
{"BufferStoreFormatD16XUint32", F::k32UInt, {0x7e63fffeu}, {0x0000fffeu}},
{"BufferStoreFormatD16XSint32", F::k32SInt, {0x7e63fffeu}, {0xfffffffeu}},
{"BufferStoreFormatD16XUnorm8", F::k8UNorm, {0x7e633800u}, {0xffffff80u}},
{"BufferStoreFormatD16XSnorm16", F::k16SNorm, {0x7e63b800u}, {0xffffc000u}},
{"BufferStoreFormatD16XZeroFill", F::k16_16Float, {0x7e633800u}, {0x00003800u}},
{"BufferStoreFormatD16XyzwFloat16", F::k16_16_16_16Float,
{0xc0003c00u, 0x34003800u}, {0xc0003c00u, 0x34003800u}, 2, true},
{"BufferStoreFormatD16XyzwFloat32", F::k32_32_32_32Float,
{0xc0003c00u, 0x34003800u}, {0x3f800000u, 0xc0000000u, 0x3f000000u, 0x3e800000u}, 4, true},
{"BufferStoreFormatD16XyzwUint32", F::k32_32_32_32UInt,
{0xffff0001u, 0x80007fffu}, {1u, 0xffffu, 0x7fffu, 0x8000u}, 4, true},
{"BufferStoreFormatD16XyzwSint32", F::k32_32_32_32SInt,
{0x8000ffffu, 0x7fff0001u}, {0xffffffffu, 0xffff8000u, 1u, 0x7fffu}, 4, true},
{"BufferStoreFormatD16XyzwSint8", F::k8_8_8_8SInt,
{0xff80ff7fu, 0x0080007fu}, {0x7f7f8080u}, 1, true},
{"BufferStoreFormatD16XyzwUnorm8", F::k8_8_8_8UNorm,
{0x3800c000u, 0x40003c00u}, {0xffff8000u}, 1, true},
};
std::vector<TestCase> cases;
for (const auto &input : inputs) {
for (const auto &input : inputs)
for (const u32 wave : {32u, 64u}) {
if (!input.wide && wave == 64u) continue;
TestCase test;
test.name = input.name;
AppendVMovLiteral(&test.code, 66, input.input);
AppendVMovU32(&test.code, 65, 1);
// Match the captured CS824586f35db5c80d instruction at pc 0x484.
test.code.push_back(EncodeMubuf0(0x84u, 0, true, false));
test.code.push_back(EncodeMubuf1(66, 0, 65));
AppendVMovLiteral(&test.code, 66, input.input[0]);
if (input.wide) {
AppendVMovLiteral(&test.code, 67, input.input[1]);
test.code.push_back(EncodeVop2(0x25, 65, InlineU32(1), 0));
AppendSMovLiteral(&test.code, 126, 0x55555555u);
AppendSMovLiteral(&test.code, 127, wave == 64u ? 0x55555555u : 0u);
} else AppendVMovU32(&test.code, 65, 1);
const auto store = [&] {
// XYZW matches captured e21c6000/8002000e with v66:67/v65 and s0:3.
test.code.push_back(EncodeMubuf0(input.wide ? 0x87u : 0x84u, 0, true, false, input.wide));
test.code.push_back(EncodeMubuf1(66, 0, 65));
};
store();
if (input.wide) {
AppendVMovU32(&test.code, 65, wave + 1u);
store(); // Active lanes must not write beyond the descriptor range.
AppendVMovU32(&test.code, 65, 0u);
test.code.push_back(EncodeSop1(0x04, 126, InlineU32(0)));
store(); // EXEC-zero must preserve the leading record.
}
AppendEnd(&test.code);
test.initial = {0xdeadbeefu, 0xffffffffu, 0xcafef00du};
test.expected = {0xdeadbeefu, input.expected, 0xcafef00du};
test.user_data = MakeStructuredStorageBufferData(4, 3, false, BufferFormat(input.format));
if (input.wide) {
test.initial.assign((wave + 2u) * input.words, 0xdeadbeefu);
test.expected = test.initial;
for (u32 lane = 0; lane + 2u < wave; lane += 2u)
std::copy_n(input.expected.begin(), input.words,
test.expected.begin() + (lane + 1u) * input.words);
test.storage_buffer_range_bytes = wave * input.words * 4u - 1u;
test.has_compute_info = true;
test.compute_info.wave_size = wave;
test.compute_info.threads_num[0] = wave;
test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1u;
test.compute_info.thread_ids_num = 1u;
test.required_spirv = {"Volatile"};
} else {
test.initial = {0xdeadbeefu, 0xffffffffu, 0xcafef00du};
test.expected = {0xdeadbeefu, input.expected[0], 0xcafef00du};
}
test.user_data = MakeStructuredStorageBufferData(input.words * 4u, input.wide ? wave : 3u,
false, BufferFormat(input.format));
test.has_user_data = true;
test.opcodes = {O::V_MOV_B32, O::BUFFER_STORE_FORMAT_D16_X, O::S_ENDPGM};
test.opcodes = {O::V_MOV_B32, input.wide ? O::BUFFER_STORE_FORMAT_D16_XYZW :
O::BUFFER_STORE_FORMAT_D16_X, O::S_ENDPGM};
cases.push_back(std::move(test));
}
return cases;
@@ -40597,7 +40645,7 @@ std::vector<TestCase> MakeCases() {
for (auto &test : BufferLoadFormatD16XyzwCases()) {
cases.push_back(std::move(test));
}
for (auto &test : BufferStoreFormatD16XCases()) {
for (auto &test : BufferStoreFormatD16Cases()) {
cases.push_back(std::move(test));
}
for (auto &test : FormattedStoreConversionCases()) {
@@ -46271,7 +46319,7 @@ int main(int argc, char **argv) {
for (const auto &test : BufferLoadFormatD16XyzwCases()) {
RunCase(&vulkan, test);
}
for (const auto &test : BufferStoreFormatD16XCases()) {
for (const auto &test : BufferStoreFormatD16Cases()) {
RunCase(&vulkan, test);
}
for (const auto &test : FormattedStoreConversionCases()) {
tests/shaderCfgTests.cpp
+19
View File
@@ -15375,6 +15375,24 @@ void TestCapturedBufferLoadUbyteD16Decode() {
"captured BUFFER_LOAD_UBYTE_D16 lost its preserved half, address, or coherence flags");
}
void TestCapturedBufferStoreD16XyzwDecode() {
namespace Decoder = ShaderRecompiler::Decoder;
const uint32_t code[] = {0xe21c6000u, 0x8002000eu};
Decoder::Instruction inst;
Decoder::DecodeInstruction(code, 0, inst);
Check(inst.opcode == Decoder::Opcode::BUFFER_STORE_FORMAT_D16_XYZW &&
inst.word_count == 2 && inst.data_dwords == 2 && inst.data_bits == 16 &&
inst.data_components == 4 && inst.formatted && !inst.typed,
"captured D16 XYZW store lost its four components in two packed VGPRs");
Check(inst.dst.kind == Decoder::OperandKind::Vgpr && inst.dst.reg == 0 &&
inst.dst.sdwa_sel == 6 &&
inst.src0.kind == Decoder::OperandKind::Vgpr && inst.src0.reg == 14 &&
inst.src1.kind == Decoder::OperandKind::Sgpr && inst.src1.reg == 8 &&
inst.src2.kind == Decoder::OperandKind::IntegerInlineConstant && inst.src2.value == 0 &&
inst.idxen && !inst.offen && inst.offset == 0 && inst.glc && !inst.dlc && !inst.slc,
"captured D16 XYZW store lost its full source VGPRs, index, or coherence flags");
}
void TestNewShaderRecompilerUnsupportedMemoryDecode() {
const uint32_t mubuf_unknown[] = {EncodeMubuf0(0x7b), EncodeMubuf1(0, 0, 1),
0xbf810000u};
@@ -16059,6 +16077,7 @@ int main() {
TestPixelProgramCacheBindingIdentity();
TestGraphicsPushConstantPlacement();
TestCapturedBufferLoadUbyteD16Decode();
TestCapturedBufferStoreD16XyzwDecode();
TestNewShaderRecompilerUnsupportedMemoryDecode();
TestNewShaderRecompilerVop3LaneReadDestinationEncoding();