raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 19:07:01 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: V_MUL_LO_U16

Browse Source
This commit is contained in:
nmzik
2026-09-18 11:21:57 +04:00
parent c2b7b0b58c
commit 394e638905
4 changed files with 56 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+51
View File
@@ -18159,6 +18159,51 @@ TestCase Vop3CvtPkI16I32Captured() {
return test;
}
TestCase Vop3MulLoU16CapturedAndSelectors() {
using O = ShaderOpcode;
TestCase test;
test.name = "Vop3MulLoU16CapturedAndSelectors";
// Distinct runtime halves expose source selection and prevent constant folding.
test.initial = {0x00138001u, 0x0005ffffu, 0xffffffffu};
test.expected = {0x00138001u, 0x0005ffffu, 0xffffffffu,
0xa5a57fffu, 0xa5a5ffedu, 0xa5a58005u, 0xa5a5005fu,
0x7fff1234u, 0xffed1234u, 0x80051234u, 0x005f1234u,
0xffff0001u, 0x00231234u};
auto &code = test.code;
AppendBufferLoadDword(&code, 12, 30);
AppendVMovU32(&code, 30, 4);
AppendBufferLoadDword(&code, 24, 30);
u32 output = 3;
for (u32 selectors : {0u, 1u, 2u, 3u, 8u, 9u, 10u, 11u}) {
AppendVMovLiteral(&code, 2, 0xa5a51234u);
if (selectors == 8u) {
code.push_back(0xd7054002u);
code.push_back(0x0002310cu); // Captured v_mul_lo_u16 v2.hi, v12.lo, v24.lo.
} else {
AppendVop3(&code, 0x305, 2, Vgpr(12), Vgpr(24), 0, 0, selectors);
}
AppendStoreVgpr(&code, 2, output++);
}
// 0xffff * 0xffff wraps to 1; aliasing preserves the original high half.
AppendVMovU32(&code, 30, 8);
AppendBufferLoadDword(&code, 12, 30);
AppendVop3(&code, 0x305, 12, Vgpr(12), Vgpr(12), 0, 0, 1);
AppendStoreVgpr(&code, 12, output++);
AppendVMovLiteral(&code, 2, 0xa5a51234u);
AppendVop3(&code, 0x305, 2, 255, Vgpr(24), 0, 0, 11);
code.push_back(0x0007deadu); // Select literal high half: 7 * 5 = 35.
AppendStoreVgpr(&code, 2, output);
AppendEnd(&code);
test.opcodes = {O::V_MOV_B32, O::BUFFER_LOAD_DWORD, O::V_MUL_LO_U16,
O::BUFFER_STORE_DWORD, O::S_ENDPGM};
test.decoded_counts = {{"V_MUL_LO_U16", 10}};
test.required_spirv = {"OpIMul"};
return test;
}
TestCase Vop3Med3I16Captured() {
using O = ShaderOpcode;
@@ -26909,6 +26954,7 @@ std::vector<TestCase> MakeCases() {
AddCase(Vop2SdwaLshrrevCapturedByte1Source);
AddCase(Vop2SdwaSubNcPreservesByteAndWordDestinations);
AddCase(Vop3CvtPkI16I32Captured);
AddCase(Vop3MulLoU16CapturedAndSelectors);
AddCase(Vop3Med3I16Captured);
AddCase(Vop2SdwaMinU32PreservesWordDestination);
AddCase(VectorShiftCountsMaskLowBits);
@@ -31913,6 +31959,11 @@ int main(int argc, char **argv) {
RunCase(&vulkan, Vop3Med3I16Captured());
return 0;
}
if (argc == 2 && std::strcmp(argv[1], "--mul-lo-u16-only") == 0) {
VulkanHarness vulkan;
RunCase(&vulkan, Vop3MulLoU16CapturedAndSelectors());
return 0;
}
if (argc == 2 && std::strcmp(argv[1], "--waitcnt-depctr-only") == 0) {
VulkanHarness vulkan;
RunCase(&vulkan, ScalarWaitcntDepctrCapturedVmVsrc());