mirror of
https://github.com/KytyPS5/KytyPS5.git
synced 2026-10-27 16:02:37 +00:00
332 lines
14 KiB
C++
332 lines
14 KiB
C++
#include "graphics/shader/recompiler/ImageOps.h"
|
|
|
|
#include <algorithm>
|
|
#include <fmt/format.h>
|
|
#include <iterator>
|
|
|
|
namespace Libs::Graphics::ShaderRecompiler::Decoder {
|
|
namespace {
|
|
|
|
struct MimgSampleInfo {
|
|
uint32_t opcode = 0;
|
|
const char* name = nullptr;
|
|
uint32_t flags = 0;
|
|
uint32_t address_components = 3;
|
|
};
|
|
|
|
struct MimgGatherInfo {
|
|
uint32_t opcode = 0;
|
|
const char* name = nullptr;
|
|
Opcode decoded = Opcode::Unsupported;
|
|
uint32_t flags = 0;
|
|
uint32_t address_components = 3;
|
|
};
|
|
|
|
struct MimgAtomicInfo {
|
|
uint32_t opcode = 0;
|
|
const char* name = nullptr;
|
|
Opcode decoded = Opcode::Unsupported;
|
|
};
|
|
|
|
constexpr ImageDimension DecodeImageDimension(uint32_t dim) {
|
|
switch (dim) {
|
|
case 1u: return ImageDimension::Dim2D;
|
|
case 2u: return ImageDimension::Dim3D;
|
|
case 3u: return ImageDimension::Dim2DArray;
|
|
case 5u:
|
|
case 7u: return ImageDimension::Dim2DArray;
|
|
case 6u: return ImageDimension::Dim2D;
|
|
default: return ImageDimension::Unknown;
|
|
}
|
|
}
|
|
|
|
constexpr uint32_t ImageCoordComponents(ImageDimension dimension) {
|
|
switch (dimension) {
|
|
case ImageDimension::Dim3D:
|
|
case ImageDimension::Dim2DArray: return 3u;
|
|
default: return 2u;
|
|
}
|
|
}
|
|
|
|
constexpr uint32_t ImageSampleAddressComponents(uint32_t flags, ImageDimension dimension) {
|
|
const auto coord_components = ImageCoordComponents(dimension);
|
|
uint32_t components = coord_components;
|
|
if ((flags & ImageSampleFlagOffset) != 0) {
|
|
components++;
|
|
}
|
|
if ((flags & ImageSampleFlagCompare) != 0) {
|
|
components++;
|
|
}
|
|
if ((flags & ImageSampleFlagBias) != 0) {
|
|
components++;
|
|
}
|
|
if ((flags & ImageSampleFlagLod) != 0) {
|
|
components++;
|
|
}
|
|
if ((flags & ImageSampleFlagDerivative) != 0) {
|
|
components += 4u;
|
|
}
|
|
return components;
|
|
}
|
|
|
|
constexpr MimgSampleInfo SampleInfo(uint32_t opcode, const char* name, uint32_t flags) {
|
|
return {opcode, name, flags, ImageSampleAddressComponents(flags, ImageDimension::Dim2D)};
|
|
}
|
|
|
|
constexpr MimgSampleInfo MIMG_SAMPLE_OPS[] = {
|
|
SampleInfo(0x20u, "image_sample", 0),
|
|
SampleInfo(0x21u, "image_sample_cl", ImageSampleFlagLodClamp),
|
|
SampleInfo(0x22u, "image_sample_d", ImageSampleFlagDerivative),
|
|
SampleInfo(0x23u, "image_sample_d_cl", ImageSampleFlagDerivative | ImageSampleFlagLodClamp),
|
|
SampleInfo(0x24u, "image_sample_l", ImageSampleFlagLod),
|
|
SampleInfo(0x25u, "image_sample_b", ImageSampleFlagBias),
|
|
SampleInfo(0x26u, "image_sample_b_cl", ImageSampleFlagBias | ImageSampleFlagLodClamp),
|
|
SampleInfo(0x27u, "image_sample_lz", ImageSampleFlagLevelZero),
|
|
SampleInfo(0x28u, "image_sample_c", ImageSampleFlagCompare),
|
|
SampleInfo(0x29u, "image_sample_c_cl", ImageSampleFlagCompare | ImageSampleFlagLodClamp),
|
|
SampleInfo(0x2au, "image_sample_c_d", ImageSampleFlagCompare | ImageSampleFlagDerivative),
|
|
SampleInfo(0x2bu, "image_sample_c_d_cl",
|
|
ImageSampleFlagCompare | ImageSampleFlagDerivative | ImageSampleFlagLodClamp),
|
|
SampleInfo(0x2cu, "image_sample_c_l", ImageSampleFlagCompare | ImageSampleFlagLod),
|
|
SampleInfo(0x2du, "image_sample_c_b", ImageSampleFlagCompare | ImageSampleFlagBias),
|
|
SampleInfo(0x2eu, "image_sample_c_b_cl",
|
|
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp),
|
|
SampleInfo(0x2fu, "image_sample_c_lz", ImageSampleFlagCompare | ImageSampleFlagLevelZero),
|
|
SampleInfo(0x30u, "image_sample_o", ImageSampleFlagOffset),
|
|
SampleInfo(0x31u, "image_sample_cl_o", ImageSampleFlagLodClamp | ImageSampleFlagOffset),
|
|
SampleInfo(0x32u, "image_sample_d_o", ImageSampleFlagDerivative | ImageSampleFlagOffset),
|
|
SampleInfo(0x33u, "image_sample_d_cl_o",
|
|
ImageSampleFlagDerivative | ImageSampleFlagLodClamp | ImageSampleFlagOffset),
|
|
SampleInfo(0x34u, "image_sample_l_o", ImageSampleFlagLod | ImageSampleFlagOffset),
|
|
SampleInfo(0x35u, "image_sample_b_o", ImageSampleFlagBias | ImageSampleFlagOffset),
|
|
SampleInfo(0x36u, "image_sample_b_cl_o",
|
|
ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagOffset),
|
|
SampleInfo(0x37u, "image_sample_lz_o", ImageSampleFlagLevelZero | ImageSampleFlagOffset),
|
|
SampleInfo(0x38u, "image_sample_c_o", ImageSampleFlagCompare | ImageSampleFlagOffset),
|
|
SampleInfo(0x39u, "image_sample_c_cl_o",
|
|
ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagOffset),
|
|
SampleInfo(0x3au, "image_sample_c_d_o",
|
|
ImageSampleFlagCompare | ImageSampleFlagDerivative | ImageSampleFlagOffset),
|
|
SampleInfo(0x3bu, "image_sample_c_d_cl_o",
|
|
ImageSampleFlagCompare | ImageSampleFlagDerivative | ImageSampleFlagLodClamp |
|
|
ImageSampleFlagOffset),
|
|
SampleInfo(0x3cu, "image_sample_c_l_o",
|
|
ImageSampleFlagCompare | ImageSampleFlagLod | ImageSampleFlagOffset),
|
|
SampleInfo(0x3du, "image_sample_c_b_o",
|
|
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagOffset),
|
|
SampleInfo(0x3eu, "image_sample_c_b_cl_o",
|
|
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
|
|
ImageSampleFlagOffset),
|
|
SampleInfo(0x3fu, "image_sample_c_lz_o",
|
|
ImageSampleFlagCompare | ImageSampleFlagLevelZero | ImageSampleFlagOffset),
|
|
SampleInfo(0x68u, "image_sample_cd", ImageSampleFlagDerivative | ImageSampleFlagCd),
|
|
SampleInfo(0x69u, "image_sample_cd_cl",
|
|
ImageSampleFlagDerivative | ImageSampleFlagCd | ImageSampleFlagLodClamp),
|
|
SampleInfo(0x6au, "image_sample_c_cd",
|
|
ImageSampleFlagCompare | ImageSampleFlagDerivative | ImageSampleFlagCd),
|
|
SampleInfo(0x6bu, "image_sample_c_cd_cl",
|
|
ImageSampleFlagCompare | ImageSampleFlagDerivative | ImageSampleFlagCd |
|
|
ImageSampleFlagLodClamp),
|
|
SampleInfo(0x6cu, "image_sample_cd_o",
|
|
ImageSampleFlagDerivative | ImageSampleFlagCd | ImageSampleFlagOffset),
|
|
SampleInfo(0x6du, "image_sample_cd_cl_o",
|
|
ImageSampleFlagDerivative | ImageSampleFlagCd | ImageSampleFlagLodClamp |
|
|
ImageSampleFlagOffset),
|
|
SampleInfo(0x6eu, "image_sample_c_cd_o",
|
|
ImageSampleFlagCompare | ImageSampleFlagDerivative | ImageSampleFlagCd |
|
|
ImageSampleFlagOffset),
|
|
SampleInfo(0x6fu, "image_sample_c_cd_cl_o",
|
|
ImageSampleFlagCompare | ImageSampleFlagDerivative | ImageSampleFlagCd |
|
|
ImageSampleFlagLodClamp | ImageSampleFlagOffset),
|
|
SampleInfo(0xa0u, "image_sample_a", 0),
|
|
SampleInfo(0xa1u, "image_sample_cl_a", ImageSampleFlagLodClamp),
|
|
SampleInfo(0xa5u, "image_sample_b_a", ImageSampleFlagBias),
|
|
SampleInfo(0xa6u, "image_sample_b_cl_a", ImageSampleFlagBias | ImageSampleFlagLodClamp),
|
|
SampleInfo(0xa8u, "image_sample_c_a", ImageSampleFlagCompare),
|
|
SampleInfo(0xa9u, "image_sample_c_cl_a", ImageSampleFlagCompare | ImageSampleFlagLodClamp),
|
|
SampleInfo(0xadu, "image_sample_c_b_a", ImageSampleFlagCompare | ImageSampleFlagBias),
|
|
SampleInfo(0xaeu, "image_sample_c_b_cl_a",
|
|
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp),
|
|
SampleInfo(0xb0u, "image_sample_o_a", ImageSampleFlagOffset),
|
|
SampleInfo(0xb1u, "image_sample_cl_o_a", ImageSampleFlagLodClamp | ImageSampleFlagOffset),
|
|
SampleInfo(0xb5u, "image_sample_b_o_a", ImageSampleFlagBias | ImageSampleFlagOffset),
|
|
SampleInfo(0xb6u, "image_sample_b_cl_o_a",
|
|
ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagOffset),
|
|
SampleInfo(0xb8u, "image_sample_c_o_a", ImageSampleFlagCompare | ImageSampleFlagOffset),
|
|
SampleInfo(0xb9u, "image_sample_c_cl_o_a",
|
|
ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagOffset),
|
|
SampleInfo(0xbdu, "image_sample_c_b_o_a",
|
|
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagOffset),
|
|
SampleInfo(0xbeu, "image_sample_c_b_cl_o_a",
|
|
ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
|
|
ImageSampleFlagOffset),
|
|
};
|
|
|
|
constexpr MimgGatherInfo MIMG_GATHER_OPS[] = {
|
|
{0x47u, "image_gather4_lz", Opcode::ImageGather4Lz, ImageSampleFlagLevelZero, 2u},
|
|
{0x48u, "image_gather4_c", Opcode::ImageGather4C, ImageSampleFlagCompare, 3u},
|
|
{0x4fu, "image_gather4_c_lz", Opcode::ImageGather4CLz,
|
|
ImageSampleFlagCompare | ImageSampleFlagLevelZero, 3u},
|
|
{0x57u, "image_gather4_lz_o", Opcode::ImageGather4LzO,
|
|
ImageSampleFlagLevelZero | ImageSampleFlagOffset, 3u},
|
|
{0x58u, "image_gather4_c_o", Opcode::ImageGather4CO,
|
|
ImageSampleFlagCompare | ImageSampleFlagOffset, 4u},
|
|
{0x5fu, "image_gather4_c_lz_o", Opcode::ImageGather4CLzO,
|
|
ImageSampleFlagCompare | ImageSampleFlagLevelZero | ImageSampleFlagOffset, 4u},
|
|
{0x61u, "image_gather4h", Opcode::ImageGather4H, ImageSampleFlagGatherHorizontal, 2u},
|
|
};
|
|
|
|
constexpr MimgAtomicInfo MIMG_ATOMIC_OPS[] = {
|
|
{0x11u, "image_atomic_add", Opcode::ImageAtomicAdd},
|
|
{0x15u, "image_atomic_umin", Opcode::ImageAtomicUMin},
|
|
{0x17u, "image_atomic_umax", Opcode::ImageAtomicUMax},
|
|
{0x18u, "image_atomic_and", Opcode::ImageAtomicAnd},
|
|
{0x19u, "image_atomic_or", Opcode::ImageAtomicOr},
|
|
{0x1au, "image_atomic_xor", Opcode::ImageAtomicXor},
|
|
};
|
|
|
|
const MimgSampleInfo* LookupSample(uint32_t opcode) {
|
|
for (const auto& info: MIMG_SAMPLE_OPS) {
|
|
if (info.opcode == opcode) {
|
|
return &info;
|
|
}
|
|
}
|
|
return nullptr;
|
|
}
|
|
|
|
const MimgGatherInfo* LookupGather(uint32_t opcode) {
|
|
for (const auto& info: MIMG_GATHER_OPS) {
|
|
if (info.opcode == opcode) {
|
|
return &info;
|
|
}
|
|
}
|
|
return nullptr;
|
|
}
|
|
|
|
const MimgAtomicInfo* LookupAtomic(uint32_t opcode) {
|
|
for (const auto& info: MIMG_ATOMIC_OPS) {
|
|
if (info.opcode == opcode) {
|
|
return &info;
|
|
}
|
|
}
|
|
return nullptr;
|
|
}
|
|
|
|
uint32_t CountDmaskComponents(uint32_t dmask) {
|
|
uint32_t count = 0;
|
|
for (uint32_t i = 0; i < 4u; i++) {
|
|
count += (dmask >> i) & 1u;
|
|
}
|
|
return count != 0 ? count : 1u;
|
|
}
|
|
|
|
bool IsSingleDmaskBit(uint32_t dmask) {
|
|
return dmask != 0u && (dmask & (dmask - 1u)) == 0u;
|
|
}
|
|
|
|
} // namespace
|
|
|
|
bool DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index, Instruction* inst,
|
|
std::string* error) {
|
|
if (word_index + 1u >= code.size()) {
|
|
if (error != nullptr) {
|
|
*error = fmt::format("truncated MIMG instruction at pc 0x{:08x}", pc);
|
|
}
|
|
return false;
|
|
}
|
|
|
|
const uint32_t word0 = code[word_index];
|
|
const uint32_t word1 = code[word_index + 1u];
|
|
const uint32_t opcode = ((word0 >> 18u) & 0x7fu) | ((word0 & 1u) << 7u);
|
|
const uint32_t nsa_dwords = (word0 >> 1u) & 0x3u;
|
|
const auto dimension = DecodeImageDimension((word0 >> 3u) & 0x7u);
|
|
const uint32_t word_count = 2u + nsa_dwords;
|
|
if (word_index + word_count > code.size()) {
|
|
if (error != nullptr) {
|
|
*error = fmt::format("truncated MIMG NSA instruction at pc 0x{:08x}", pc);
|
|
}
|
|
return false;
|
|
}
|
|
|
|
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 a16 = ((word1 >> 30u) & 0x1u) != 0u;
|
|
const auto* sample = LookupSample(opcode);
|
|
const auto* gather = LookupGather(opcode);
|
|
const auto* atomic = LookupAtomic(opcode);
|
|
|
|
inst->pc = pc;
|
|
inst->word = word0;
|
|
inst->word_count = word_count;
|
|
inst->family = Family::MIMG;
|
|
inst->opcode_id = opcode;
|
|
inst->opcode = (sample != nullptr ? Opcode::ImageSample
|
|
: gather != nullptr ? gather->decoded
|
|
: atomic != nullptr ? atomic->decoded
|
|
: opcode == 0x00u ? Opcode::ImageLoad
|
|
: opcode == 0x01u ? Opcode::ImageLoadMip
|
|
: opcode == 0x08u ? Opcode::ImageStore
|
|
: opcode == 0x09u ? Opcode::ImageStoreMip
|
|
: opcode == 0x0eu ? Opcode::ImageGetResinfo
|
|
: opcode == 0x60u ? Opcode::ImageGetLod
|
|
: Opcode::Unsupported);
|
|
inst->dmask = (word0 >> 8u) & 0xfu;
|
|
inst->data_dwords = gather != nullptr ? 4u : CountDmaskComponents(inst->dmask);
|
|
inst->glc = ((word0 >> 13u) & 1u) != 0;
|
|
inst->slc = ((word0 >> 25u) & 1u) != 0;
|
|
inst->image_sample_flags = (sample != nullptr ? sample->flags
|
|
: gather != nullptr ? gather->flags
|
|
: 0u) |
|
|
(a16 ? ImageSampleFlagA16 : 0u);
|
|
inst->image_dimension = dimension;
|
|
inst->image_nsa_dwords = nsa_dwords;
|
|
for (uint32_t i = 0; i < nsa_dwords * 4u && i < MaxImageNsaAddressComponents; i++) {
|
|
inst->image_nsa_addr[i] = (code[word_index + 2u + i / 4u] >> ((i % 4u) * 8u)) & 0xffu;
|
|
}
|
|
inst->image_address_components =
|
|
sample != nullptr ? ImageSampleAddressComponents(sample->flags, dimension)
|
|
: gather != nullptr ? ImageSampleAddressComponents(gather->flags, dimension)
|
|
: atomic != nullptr ? 3u
|
|
: opcode == 0x60u ? ImageCoordComponents(dimension)
|
|
: opcode == 0x01u ? ImageCoordComponents(dimension) + 1u
|
|
: opcode == 0x00u ? ImageCoordComponents(dimension)
|
|
: (opcode == 0x09u ? ImageCoordComponents(dimension) + 1u
|
|
: opcode == 0x08u ? ImageCoordComponents(dimension)
|
|
: 0u);
|
|
SetRawWords(inst, code, word_index, word_count);
|
|
|
|
if (inst->opcode == Opcode::Unsupported) {
|
|
SetUnsupported(inst, Family::MIMG, opcode, "MIMG opcode is not implemented");
|
|
}
|
|
if (gather != nullptr && !IsSingleDmaskBit(inst->dmask)) {
|
|
SetUnsupported(inst, Family::MIMG, opcode,
|
|
"MIMG image gather requires exactly one dmask bit");
|
|
}
|
|
|
|
DecodeVectorGpr(vdata, &inst->dst, nullptr);
|
|
DecodeVectorGpr(vaddr, &inst->src0, nullptr);
|
|
DecodeScalarSource(srsrc * 4u, pc, &inst->src1, nullptr);
|
|
DecodeScalarSource(ssamp * 4u, pc, &inst->src2, nullptr);
|
|
inst->src_count = 3;
|
|
return true;
|
|
}
|
|
|
|
const char* MimgSampleOpcodeName(uint32_t opcode) {
|
|
const auto* sample = LookupSample(opcode);
|
|
return sample != nullptr ? sample->name : nullptr;
|
|
}
|
|
|
|
uint32_t MimgSampleOpcodeFlags(uint32_t opcode) {
|
|
const auto* sample = LookupSample(opcode);
|
|
return sample != nullptr ? sample->flags : 0u;
|
|
}
|
|
|
|
uint32_t MimgSampleAddressComponents(uint32_t opcode) {
|
|
const auto* sample = LookupSample(opcode);
|
|
return sample != nullptr ? sample->address_components : 0u;
|
|
}
|
|
|
|
} // namespace Libs::Graphics::ShaderRecompiler::Decoder
|