shader: decode every PS5 image_gather4 encoding

Co-authored-by: shadowbeat070 <70843316+shadowbeat070@users.noreply.github.com>
This commit is contained in:
nmzik
2026-10-27 01:13:46 +01:00
parent 1b70938052
commit ffa3992c1a
7 changed files with 249 additions and 130 deletions
@@ -14,12 +14,6 @@ struct MimgSampleInfo {
uint32_t flags = 0;
};
struct MimgGatherInfo {
uint32_t encoding = 0;
Opcode decoded = Opcode::UNSUPPORTED;
uint32_t flags = 0;
};
struct MimgAtomicInfo {
uint32_t encoding = 0;
Opcode decoded = Opcode::UNSUPPORTED;
@@ -181,23 +175,72 @@ constexpr MimgSampleInfo MIMG_SAMPLE_OPCODE_LIST[] = {
ImageSampleFlagOffset | ImageSampleFlagAdjust},
};
constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
{0x44u, Opcode::IMAGE_GATHER4_L, ImageSampleFlagLod},
{0x47u, Opcode::IMAGE_GATHER4_LZ, ImageSampleFlagLevelZero},
{0x48u, Opcode::IMAGE_GATHER4_C, ImageSampleFlagCompare},
{0x4cu, Opcode::IMAGE_GATHER4_C_L, ImageSampleFlagCompare | ImageSampleFlagLod},
{0x4fu, Opcode::IMAGE_GATHER4_C_LZ,
ImageSampleFlagCompare | ImageSampleFlagLevelZero},
{0x57u, Opcode::IMAGE_GATHER4_LZ_O,
ImageSampleFlagLevelZero | ImageSampleFlagOffset},
{0x58u, Opcode::IMAGE_GATHER4_C_O,
ImageSampleFlagCompare | ImageSampleFlagOffset},
{0x5fu, Opcode::IMAGE_GATHER4_C_LZ_O,
constexpr MimgSampleInfo MIMG_GATHER_OPCODE_LIST[] = {
{0x40u, "image_gather4", 0},
{0x41u, "image_gather4_cl", ImageSampleFlagLodClamp},
{0x44u, "image_gather4_l", ImageSampleFlagLod},
{0x45u, "image_gather4_b", ImageSampleFlagBias},
{0x46u, "image_gather4_b_cl", ImageSampleFlagBias | ImageSampleFlagLodClamp},
{0x47u, "image_gather4_lz", ImageSampleFlagLevelZero},
{0x48u, "image_gather4_c", ImageSampleFlagCompare},
{0x49u, "image_gather4_c_cl", ImageSampleFlagCompare | ImageSampleFlagLodClamp},
{0x4cu, "image_gather4_c_l", ImageSampleFlagCompare | ImageSampleFlagLod},
{0x4du, "image_gather4_c_b", ImageSampleFlagCompare | ImageSampleFlagBias},
{0x4eu, "image_gather4_c_b_cl",
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp},
{0x4fu, "image_gather4_c_lz", ImageSampleFlagCompare | ImageSampleFlagLevelZero},
{0x50u, "image_gather4_o", ImageSampleFlagOffset},
{0x51u, "image_gather4_cl_o", ImageSampleFlagLodClamp | ImageSampleFlagOffset},
{0x54u, "image_gather4_l_o", ImageSampleFlagLod | ImageSampleFlagOffset},
{0x55u, "image_gather4_b_o", ImageSampleFlagBias | ImageSampleFlagOffset},
{0x56u, "image_gather4_b_cl_o",
ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagOffset},
{0x57u, "image_gather4_lz_o", ImageSampleFlagLevelZero | ImageSampleFlagOffset},
{0x58u, "image_gather4_c_o", ImageSampleFlagCompare | ImageSampleFlagOffset},
{0x59u, "image_gather4_c_cl_o",
ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagOffset},
{0x5cu, "image_gather4_c_l_o",
ImageSampleFlagCompare | ImageSampleFlagLod | ImageSampleFlagOffset},
{0x5du, "image_gather4_c_b_o",
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagOffset},
{0x5eu, "image_gather4_c_b_cl_o",
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
ImageSampleFlagOffset},
{0x5fu, "image_gather4_c_lz_o",
ImageSampleFlagCompare | ImageSampleFlagLevelZero | ImageSampleFlagOffset},
{0x61u, Opcode::IMAGE_GATHER4H, ImageSampleFlagGatherHorizontal},
{0xc8u, Opcode::IMAGE_GATHER4_C, ImageSampleFlagCompare | ImageSampleFlagAdjust},
{0xd8u, Opcode::IMAGE_GATHER4_C_O,
{0x61u, "image_gather4h", ImageSampleFlagGatherHorizontal},
{0xc0u, "image_gather4_a", ImageSampleFlagAdjust},
{0xc1u, "image_gather4_cl_a", ImageSampleFlagLodClamp | ImageSampleFlagAdjust},
{0xc5u, "image_gather4_b_a", ImageSampleFlagBias | ImageSampleFlagAdjust},
{0xc6u, "image_gather4_b_cl_a",
ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagAdjust},
{0xc8u, "image_gather4_c_a", ImageSampleFlagCompare | ImageSampleFlagAdjust},
{0xc9u, "image_gather4_c_cl_a",
ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagAdjust},
{0xcdu, "image_gather4_c_b_a",
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagAdjust},
{0xceu, "image_gather4_c_b_cl_a",
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
ImageSampleFlagAdjust},
{0xd0u, "image_gather4_o_a", ImageSampleFlagOffset | ImageSampleFlagAdjust},
{0xd1u, "image_gather4_cl_o_a",
ImageSampleFlagLodClamp | ImageSampleFlagOffset | ImageSampleFlagAdjust},
{0xd5u, "image_gather4_b_o_a",
ImageSampleFlagBias | ImageSampleFlagOffset | ImageSampleFlagAdjust},
{0xd6u, "image_gather4_b_cl_o_a",
ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagOffset |
ImageSampleFlagAdjust},
{0xd8u, "image_gather4_c_o_a",
ImageSampleFlagCompare | ImageSampleFlagOffset | ImageSampleFlagAdjust},
{0xd9u, "image_gather4_c_cl_o_a",
ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagOffset |
ImageSampleFlagAdjust},
{0xddu, "image_gather4_c_b_o_a",
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagOffset |
ImageSampleFlagAdjust},
{0xdeu, "image_gather4_c_b_cl_o_a",
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
ImageSampleFlagOffset | ImageSampleFlagAdjust},
};
constexpr MimgAtomicInfo MIMG_ATOMIC_OPCODE_LIST[] = {
@@ -219,13 +262,13 @@ constexpr auto MIMG_SAMPLE_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_SAMPLE_OPCO
constexpr auto MIMG_GATHER_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_GATHER_OPCODE_LIST);
constexpr auto MIMG_ATOMIC_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_ATOMIC_OPCODE_LIST);
Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const MimgGatherInfo* gather,
Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const MimgSampleInfo* gather,
const MimgAtomicInfo* atomic) {
if (sample != nullptr) {
return Opcode::IMAGE_SAMPLE;
}
if (gather != nullptr) {
return gather->decoded;
return Opcode::IMAGE_GATHER4;
}
if (atomic != nullptr) {
return atomic->decoded;
@@ -244,24 +287,10 @@ Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const Mim
}
}
uint32_t DecodeMimgSampleFlags(const MimgSampleInfo* sample, const MimgGatherInfo* gather) {
if (sample != nullptr) {
return sample->flags;
}
if (gather != nullptr) {
return gather->flags;
}
return 0;
}
uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
const MimgSampleInfo* sample, const MimgGatherInfo* gather,
const MimgAtomicInfo* atomic) {
if (sample != nullptr) {
return ImageSampleAddressComponents(sample->flags, dimension);
}
if (gather != nullptr) {
return ImageSampleAddressComponents(gather->flags, dimension);
const MimgSampleInfo* sampled, const MimgAtomicInfo* atomic) {
if (sampled != nullptr) {
return ImageSampleAddressComponents(sampled->flags, dimension);
}
if (atomic != nullptr) {
return ImageCoordComponents(dimension);
@@ -329,16 +358,17 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
const auto dimension = DecodeImageDimension((word0 >> 3u) & 0x7u);
const uint32_t word_count = 2u + nsa_dwords;
const uint32_t vdata = (word1 >> 8u) & 0xffu;
const uint32_t vaddr = word1 & 0xffu;
const uint32_t srsrc = (word1 >> 16u) & 0x1fu;
const uint32_t ssamp = (word1 >> 21u) & 0x1fu;
const bool r128 = ((word0 >> 15u) & 0x1u) != 0u;
const bool a16 = ((word1 >> 30u) & 0x1u) != 0u;
const bool d16 = ((word1 >> 31u) & 0x1u) != 0u;
const auto* sample = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
const auto* gather = Detail::FindOpcode(MIMG_GATHER_OPS, opcode);
const auto* atomic = Detail::FindOpcode(MIMG_ATOMIC_OPS, opcode);
const uint32_t vdata = (word1 >> 8u) & 0xffu;
const uint32_t vaddr = word1 & 0xffu;
const uint32_t srsrc = (word1 >> 16u) & 0x1fu;
const uint32_t ssamp = (word1 >> 21u) & 0x1fu;
const bool r128 = ((word0 >> 15u) & 0x1u) != 0u;
const bool a16 = ((word1 >> 30u) & 0x1u) != 0u;
const bool d16 = ((word1 >> 31u) & 0x1u) != 0u;
const auto* sample = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
const auto* gather = Detail::FindOpcode(MIMG_GATHER_OPS, opcode);
const auto* atomic = Detail::FindOpcode(MIMG_ATOMIC_OPS, opcode);
const auto* sampled = sample != nullptr ? sample : gather;
inst.pc = pc;
inst.word_count = word_count;
@@ -351,11 +381,11 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
inst.data_dwords = d16 ? (inst.data_components + 1u) / 2u : inst.data_components;
inst.glc = ((word0 >> 13u) & 1u) != 0;
inst.slc = ((word0 >> 25u) & 1u) != 0;
inst.image_sample_flags = DecodeMimgSampleFlags(sample, gather);
inst.image_sample_flags = sampled != nullptr ? sampled->flags : 0u;
if (a16) {
inst.image_sample_flags |= ImageSampleFlagA16;
}
if ((sample != nullptr || gather != nullptr) && (word0 & (1u << 12u)) != 0u) {
if (sampled != nullptr && (word0 & (1u << 12u)) != 0u) {
inst.image_sample_flags |= ImageSampleFlagUnnormalized;
}
inst.image_dimension = dimension;
@@ -364,8 +394,7 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
for (uint32_t i = 0; i < nsa_dwords * 4u; i++) {
inst.image_nsa_addr[i] = (code[word_index + 2u + i / 4u] >> ((i % 4u) * 8u)) & 0xffu;
}
inst.image_address_components =
DecodeMimgAddressComponents(opcode, dimension, sample, gather, atomic);
inst.image_address_components = DecodeMimgAddressComponents(opcode, dimension, sampled, atomic);
SetRawWords(inst, code, word_index, word_count);
if (inst.opcode == Opcode::UNSUPPORTED) {
@@ -389,8 +418,8 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
}
}
}
const bool supports_d16 = sample != nullptr || gather != nullptr || opcode == 0x00u ||
opcode == 0x01u || opcode == 0x08u || opcode == 0x09u;
const bool supports_d16 = sampled != nullptr || opcode == 0x00u || opcode == 0x01u ||
opcode == 0x08u || opcode == 0x09u;
if (d16 && !supports_d16) {
SetUnsupported(inst, Family::MIMG, opcode, "MIMG opcode does not support D16 data");
}
@@ -407,9 +436,10 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
inst.src_count = 3;
}
const char* MimgSampleOpcodeName(uint32_t opcode) {
const auto* sample = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
return sample != nullptr ? sample->name : nullptr;
const char* MimgSampledOpcodeName(uint32_t opcode) {
const auto* sample = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
const auto* sampled = sample != nullptr ? sample : Detail::FindOpcode(MIMG_GATHER_OPS, opcode);
return sampled != nullptr ? sampled->name : nullptr;
}
} // namespace Libs::Graphics::ShaderRecompiler::Decoder
@@ -8,7 +8,7 @@ namespace Libs::Graphics::ShaderRecompiler::Decoder {
void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
Instruction& inst);
const char* MimgSampleOpcodeName(uint32_t opcode);
const char* MimgSampledOpcodeName(uint32_t opcode);
struct ImageAddressComponent {
uint32_t bit_offset;
@@ -123,10 +123,9 @@ std::string ImageSampleFlagsToString(uint32_t flags) {
}
std::string FormatMimg(const Instruction& inst) {
const char* sample_name =
inst.opcode == Opcode::IMAGE_SAMPLE ? MimgSampleOpcodeName(inst.opcode_id) : nullptr;
const char* sampled_name = MimgSampledOpcodeName(inst.opcode_id);
const std::string_view name =
sample_name != nullptr ? sample_name : magic_enum::enum_name(inst.opcode);
sampled_name != nullptr ? sampled_name : magic_enum::enum_name(inst.opcode);
std::string text =
fmt::format("0x{:08x}: {} {}, {}, {}, {} ; dmask=0x{:x} image_dim={}", inst.pc, name,
OperandToString(inst.dst).c_str(), OperandToString(inst.src0).c_str(),
@@ -138,22 +137,10 @@ std::string FormatMimg(const Instruction& inst) {
if (inst.image_r128) {
text += " r128=1";
}
switch (inst.opcode) {
case Opcode::IMAGE_SAMPLE:
case Opcode::IMAGE_GATHER4_L:
case Opcode::IMAGE_GATHER4_LZ:
case Opcode::IMAGE_GATHER4_C:
case Opcode::IMAGE_GATHER4_C_L:
case Opcode::IMAGE_GATHER4_C_LZ:
case Opcode::IMAGE_GATHER4_LZ_O:
case Opcode::IMAGE_GATHER4_C_O:
case Opcode::IMAGE_GATHER4_C_LZ_O:
case Opcode::IMAGE_GATHER4H:
text += fmt::format(" sample_flags={} addr_components={}",
ImageSampleFlagsToString(inst.image_sample_flags).c_str(),
inst.image_address_components);
break;
default: break;
if (sampled_name != nullptr) {
text += fmt::format(" sample_flags={} addr_components={}",
ImageSampleFlagsToString(inst.image_sample_flags).c_str(),
inst.image_address_components);
}
if (inst.image_nsa_dwords != 0) {
text += fmt::format(" nsa_dwords={} nsa_addr=", inst.image_nsa_dwords);
@@ -616,15 +603,7 @@ std::string InstructionToString(const Instruction& inst) {
case Opcode::IMAGE_LOAD_MIP:
case Opcode::IMAGE_GET_RESINFO:
case Opcode::IMAGE_GET_LOD:
case Opcode::IMAGE_GATHER4_L:
case Opcode::IMAGE_GATHER4_LZ:
case Opcode::IMAGE_GATHER4_C:
case Opcode::IMAGE_GATHER4_C_L:
case Opcode::IMAGE_GATHER4_C_LZ:
case Opcode::IMAGE_GATHER4_LZ_O:
case Opcode::IMAGE_GATHER4_C_O:
case Opcode::IMAGE_GATHER4_C_LZ_O:
case Opcode::IMAGE_GATHER4H: return WithUnsupportedReason(inst, FormatMimg(inst));
case Opcode::IMAGE_GATHER4: return WithUnsupportedReason(inst, FormatMimg(inst));
case Opcode::S_LOAD_DWORD:
case Opcode::S_LOAD_DWORDX2:
case Opcode::S_LOAD_DWORDX4:
@@ -708,15 +708,7 @@ enum class Opcode {
IMAGE_ATOMIC_FMIN,
IMAGE_ATOMIC_FMAX,
IMAGE_SAMPLE,
IMAGE_GATHER4_L,
IMAGE_GATHER4_LZ,
IMAGE_GATHER4_C,
IMAGE_GATHER4_C_L,
IMAGE_GATHER4_C_LZ,
IMAGE_GATHER4_LZ_O,
IMAGE_GATHER4_C_O,
IMAGE_GATHER4_C_LZ_O,
IMAGE_GATHER4H,
IMAGE_GATHER4,
V_INTERP_P1_F32,
V_INTERP_P2_F32,
V_INTERP_MOV_F32,
@@ -1166,15 +1166,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
case Decoder::Opcode::IMAGE_STORE:
case Decoder::Opcode::IMAGE_STORE_MIP: return IMAGE_STORE(inst);
case Decoder::Opcode::IMAGE_SAMPLE: return IMAGE_SAMPLE(inst);
case Decoder::Opcode::IMAGE_GATHER4_L:
case Decoder::Opcode::IMAGE_GATHER4_LZ:
case Decoder::Opcode::IMAGE_GATHER4_C:
case Decoder::Opcode::IMAGE_GATHER4_C_L:
case Decoder::Opcode::IMAGE_GATHER4_C_LZ:
case Decoder::Opcode::IMAGE_GATHER4_LZ_O:
case Decoder::Opcode::IMAGE_GATHER4_C_O:
case Decoder::Opcode::IMAGE_GATHER4_C_LZ_O:
case Decoder::Opcode::IMAGE_GATHER4H: return IMAGE_GATHER(inst);
case Decoder::Opcode::IMAGE_GATHER4: return IMAGE_GATHER(inst);
case Decoder::Opcode::DS_MIN_F32:
return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicFMin32, false);