mirror of
https://github.com/KytyPS5/KytyPS5.git
synced 2026-10-20 19:07:01 +02:00
graphics: implement RDNA2 64-bit first-set-bit
This commit is contained in:
@@ -11080,6 +11080,19 @@ TestCase ScalarBitfieldPack() {
|
||||
code.push_back(EncodeSop2(0x0a, 17, InlineU32(1), InlineU32(0)));
|
||||
code.push_back(EncodeSopc(0x0c, InlineU32(4), InlineU32(1)));
|
||||
code.push_back(EncodeSop2(0x0a, 18, InlineU32(1), InlineU32(0)));
|
||||
AppendSMovLiteral(&code, 25, 0u);
|
||||
AppendSMovLiteral(&code, 26, 0x00000100u);
|
||||
code.push_back(EncodeSopc(0x06, InlineU32(0), InlineU32(1)));
|
||||
code.push_back(EncodeSop1(0x14, 31, 25));
|
||||
code.push_back(EncodeSop2(0x0a, 34, InlineU32(1), InlineU32(0)));
|
||||
AppendSMovLiteral(&code, 27, 0x00000020u);
|
||||
AppendSMovLiteral(&code, 28, 0x00000100u);
|
||||
code.push_back(EncodeSopc(0x06, InlineU32(1), InlineU32(1)));
|
||||
code.push_back(EncodeSop1(0x14, 32, 27));
|
||||
code.push_back(EncodeSop2(0x0a, 35, InlineU32(1), InlineU32(0)));
|
||||
AppendSMovLiteral(&code, 29, 0u);
|
||||
AppendSMovLiteral(&code, 30, 0u);
|
||||
code.push_back(EncodeSop1(0x14, 33, 29));
|
||||
|
||||
AppendStoreSgpr(&code, 1, 0);
|
||||
AppendStoreSgpr(&code, 3, 1);
|
||||
@@ -11095,15 +11108,22 @@ TestCase ScalarBitfieldPack() {
|
||||
AppendStoreSgpr(&code, 15, 12);
|
||||
AppendStoreSgpr(&code, 17, 13);
|
||||
AppendStoreSgpr(&code, 18, 14);
|
||||
AppendStoreSgpr(&code, 31, 15);
|
||||
AppendStoreSgpr(&code, 32, 16);
|
||||
AppendStoreSgpr(&code, 33, 17);
|
||||
AppendStoreSgpr(&code, 34, 18);
|
||||
AppendStoreSgpr(&code, 35, 19);
|
||||
AppendEnd(&code);
|
||||
|
||||
return {"ScalarBitfieldPack",
|
||||
code,
|
||||
{},
|
||||
{0xf0000000u, 8, 9, 1, 0, 0, 0x0000000fu, 0, 0x00000f00u, 0x0000000fu,
|
||||
0xddddbbbbu, 0xccccbbbbu, 0xccccaaaau, 1, 1},
|
||||
0xddddbbbbu, 0xccccbbbbu, 0xccccaaaau, 1, 1, 40, 5, 0xffffffffu, 0,
|
||||
1},
|
||||
{O::SMovB32, O::SBrevB32, O::SBcnt1I32B32, O::SBcnt1I32B64,
|
||||
O::SBitreplicateB64B32, O::SBfmB32, O::SBfeU32, O::SPackLlB32B16,
|
||||
O::SFf1I32B64, O::SBitreplicateB64B32, O::SBfmB32, O::SBfeU32,
|
||||
O::SPackLlB32B16,
|
||||
O::SPackLhB32B16, O::SPackHhB32B16, O::SBitcmp0B32, O::SBitcmp1B32,
|
||||
O::SCselectB32, O::VMovB32, O::BufferStoreDword, O::SEndpgm}};
|
||||
}
|
||||
@@ -20269,6 +20289,11 @@ int main(int argc, char **argv) {
|
||||
RunCase(&vulkan, VectorAlignByteUsesFiveBitByteOffset());
|
||||
return 0;
|
||||
}
|
||||
if (argc == 2 && std::strcmp(argv[1], "--sff1-b64-only") == 0) {
|
||||
VulkanHarness vulkan;
|
||||
RunCase(&vulkan, ScalarBitfieldPack());
|
||||
return 0;
|
||||
}
|
||||
if (argc == 2 && std::strcmp(argv[1], "--context-state-only") == 0) {
|
||||
VulkanHarness vulkan;
|
||||
CheckPm4PolygonOffsetRegisters(vulkan.RuntimeRenderer());
|
||||
|
||||
@@ -2931,6 +2931,7 @@ void TestNewShaderRecompilerScalarB64Alu() {
|
||||
EncodeSop2(0x21, 26, 24, 0), // s_lshr_b64 s[26:27], s[24:25], s0
|
||||
EncodeSop2(0x25, 28, 132, 130), // s_bfm_b64 s[28:29], 4, 2
|
||||
EncodeSop1(0x10, 30, 26), // s_bcnt1_i32_b64 s30, s[26:27]
|
||||
EncodeSop1(0x14, 31, 26), // s_ff1_i32_b64 s31, s[26:27]
|
||||
EncodeSop1(0x16, 106, 26), // s_flbit_i32_b64 vcc_lo, s[26:27]
|
||||
EncodeSop1(0x3b, 32, 30), // s_bitreplicate_b64_b32 s[32:33], s30
|
||||
EncodeSop2(0x29, 34, 32, 255), // s_bfe_u64 s[34:35], s[32:33], 0x00040002
|
||||
@@ -2989,6 +2990,8 @@ void TestNewShaderRecompilerScalarB64Alu() {
|
||||
"new decoder did not decode old-backed S_BFM_B64");
|
||||
Check(Common::ContainsStr(result.decoded_dump, "s_bcnt1_i32_b64 s30, s26"),
|
||||
"new decoder did not decode old-backed S_BCNT1_I32_B64");
|
||||
Check(Common::ContainsStr(result.decoded_dump, "s_ff1_i32_b64 s31, s26"),
|
||||
"new decoder did not decode RDNA2 S_FF1_I32_B64");
|
||||
Check(Common::ContainsStr(result.decoded_dump, "s_flbit_i32_b64 vcc_lo, s26"),
|
||||
"new decoder did not decode RDNA2 S_FLBIT_I32_B64");
|
||||
Check(Common::ContainsStr(result.decoded_dump,
|
||||
@@ -3042,6 +3045,8 @@ void TestNewShaderRecompilerScalarB64Alu() {
|
||||
"S_BFM_B64 did not lower to paired-dword IR");
|
||||
Check(Common::ContainsStr(result.ir_dump, "BitCountU64 s30, s26"),
|
||||
"S_BCNT1_I32_B64 did not lower to paired-source bit count IR");
|
||||
Check(Common::ContainsStr(result.ir_dump, "FindLsbU64 s31, s26"),
|
||||
"S_FF1_I32_B64 did not lower to paired-source bit search IR");
|
||||
Check(Common::ContainsStr(result.ir_dump, "FindMsbFromHighU64 vcc_lo, s26"),
|
||||
"S_FLBIT_I32_B64 did not lower to paired-source leading-zero IR");
|
||||
Check(Common::ContainsStr(result.ir_dump, "BitReplicateB64B32 s32, s30"),
|
||||
|
||||
Reference in New Issue
Block a user