raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-21 03:16:54 +02:00
Code Issues Packages Projects Releases Wiki Activity

agc: merge resources from both shader stages

Browse Source
This commit is contained in:
nmzik
2026-09-05 07:39:30 +04:00
parent c7f7fd974b
commit 246212789f
2 changed files with 110 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/libs/agc.cpp
+45 -98
View File
@@ -714,78 +714,6 @@ int KYTY_SYSV_ABI AgcUnknownGetFusedShaderSize(SizeAlign* dst, const Shader* fro
return OK;
}
int KYTY_SYSV_ABI AgcUnknownFuseShaderHalves(Shader* fused_result, const Shader* front,
const Shader* back, void* scratch_mem) {
PRINT_NAME();
LOGF("\t fused_result = 0x%016" PRIx64 "\n"
"\t front = 0x%016" PRIx64 "\n"
"\t back = 0x%016" PRIx64 "\n"
"\t scratch_mem = 0x%016" PRIx64 "\n",
reinterpret_cast<uint64_t>(fused_result), reinterpret_cast<uint64_t>(front),
reinterpret_cast<uint64_t>(back), reinterpret_cast<uint64_t>(scratch_mem));
EXIT_NOT_IMPLEMENTED(fused_result == nullptr);
EXIT_NOT_IMPLEMENTED(front == nullptr);
EXIT_NOT_IMPLEMENTED(back == nullptr);
const auto front_type = static_cast<Prospero::ShaderBinaryType>(front->type);
const auto back_type = static_cast<Prospero::ShaderBinaryType>(back->type);
if (!((front_type == Prospero::ShaderBinaryType::kGsFront &&
back_type == Prospero::ShaderBinaryType::kGsBack) ||
(front_type == Prospero::ShaderBinaryType::kHsFront &&
back_type == Prospero::ShaderBinaryType::kHsBack))) {
return GRAPHICS5_ERROR_INVALID_SHADER_HALVES;
}
*fused_result = *back;
fused_result->type = static_cast<uint8_t>(front_type == Prospero::ShaderBinaryType::kGsFront
? Prospero::ShaderBinaryType::kGs
: Prospero::ShaderBinaryType::kHs);
if (front->specials != nullptr && back->specials != nullptr) {
const auto front_stages = front->specials->vgt_shader_stages_en.value;
const auto back_stages = back->specials->vgt_shader_stages_en.value;
const auto mismatch_bit =
(front_type == Prospero::ShaderBinaryType::kGsFront ? (1u << 22u) : (1u << 21u));
if (((front_stages ^ back_stages) & mismatch_bit) != 0) {
return GRAPHICS5_ERROR_INVALID_SHADER_HALVES;
}
}
if (scratch_mem != nullptr && back->sh_registers != nullptr && back->num_sh_registers != 0) {
auto* sh_registers = static_cast<ShaderRegister*>(scratch_mem);
memcpy(sh_registers, back->sh_registers,
static_cast<size_t>(back->num_sh_registers) * sizeof(ShaderRegister));
fused_result->sh_registers = sh_registers;
}
auto* fused_regs = fused_result->sh_registers;
const auto fused_reg_count = static_cast<uint32_t>(fused_result->num_sh_registers);
const auto front_reg_count = static_cast<uint32_t>(front->num_sh_registers);
if (front_type == Prospero::ShaderBinaryType::kGsFront) {
for (uint32_t occurrence = 0; occurrence < 2; occurrence++) {
auto* dst = find_shader_register(fused_regs, fused_reg_count,
Pm4::SPI_SHADER_PGM_CHKSUM_GS, occurrence);
const auto* src = find_shader_register(front->sh_registers, front_reg_count,
Pm4::SPI_SHADER_PGM_CHKSUM_GS, occurrence);
if (dst != nullptr && src != nullptr) {
dst->value = src->value;
}
}
patch_shader_register_address(fused_regs, fused_reg_count, Pm4::SPI_SHADER_PGM_LO_ES,
reinterpret_cast<uint64_t>(front->code));
} else {
patch_shader_register_address(fused_regs, fused_reg_count, Pm4::SPI_SHADER_PGM_LO_LS,
reinterpret_cast<uint64_t>(front->code));
}
fused_result->user_data = nullptr;
return OK;
}
static void merge_shader_register_max_field(ShaderRegister* dst, const ShaderRegister* src,
uint32_t shift, uint32_t mask) {
const auto dst_field = (dst->value >> shift) & mask;
@@ -796,10 +724,9 @@ static void merge_shader_register_max_field(ShaderRegister* dst, const ShaderReg
dst->value |= field << shift;
}
int KYTY_SYSV_ABI AgcUnknownNApJjpKNBl4(Shader* fused_result, const Shader* front,
const Shader* back, void* scratch_mem) {
PRINT_NAME();
static int fuse_shader_halves(Shader* fused_result, const Shader* front,
const Shader* back, void* scratch_mem,
bool recompute_shared_vgprs) {
LOGF("\t fused_result = 0x%016" PRIx64 "\n"
"\t front = 0x%016" PRIx64 "\n"
"\t back = 0x%016" PRIx64 "\n"
@@ -807,6 +734,10 @@ int KYTY_SYSV_ABI AgcUnknownNApJjpKNBl4(Shader* fused_result, const Shader* fron
reinterpret_cast<uint64_t>(fused_result), reinterpret_cast<uint64_t>(front),
reinterpret_cast<uint64_t>(back), reinterpret_cast<uint64_t>(scratch_mem));
EXIT_NOT_IMPLEMENTED(fused_result == nullptr);
EXIT_NOT_IMPLEMENTED(front == nullptr);
EXIT_NOT_IMPLEMENTED(back == nullptr);
const auto front_type = static_cast<Prospero::ShaderBinaryType>(front->type);
const auto is_gs = front_type == Prospero::ShaderBinaryType::kGsFront;
const auto is_hs = front_type == Prospero::ShaderBinaryType::kHsFront;
@@ -839,14 +770,12 @@ int KYTY_SYSV_ABI AgcUnknownNApJjpKNBl4(Shader* fused_result, const Shader* fron
const auto front_reg_count = static_cast<uint32_t>(front->num_sh_registers);
const auto checksum_offset =
is_gs ? Pm4::SPI_SHADER_PGM_CHKSUM_GS : Pm4::SPI_SHADER_PGM_CHKSUM_HS;
const auto* front_checksum0 =
find_shader_register(front->sh_registers, front_reg_count, checksum_offset, 0);
const auto* front_checksum1 =
find_shader_register(front->sh_registers, front_reg_count, checksum_offset, 1);
auto* fused_checksum0 = find_shader_register(fused_regs, fused_reg_count, checksum_offset, 0);
auto* fused_checksum1 = find_shader_register(fused_regs, fused_reg_count, checksum_offset, 1);
fused_checksum0->value = front_checksum0->value;
fused_checksum1->value = front_checksum1->value;
for (uint32_t occurrence = 0; occurrence < 2; occurrence++) {
const auto* src = find_shader_register(front->sh_registers, front_reg_count,
checksum_offset, occurrence);
auto* dst = find_shader_register(fused_regs, fused_reg_count, checksum_offset, occurrence);
dst->value = src->value;
}
const auto rsrc1_offset = is_gs ? Pm4::SPI_SHADER_PGM_RSRC1_GS : Pm4::SPI_SHADER_PGM_RSRC1_HS;
const auto rsrc2_offset = is_gs ? Pm4::SPI_SHADER_PGM_RSRC2_GS : Pm4::SPI_SHADER_PGM_RSRC2_HS;
@@ -857,33 +786,51 @@ int KYTY_SYSV_ABI AgcUnknownNApJjpKNBl4(Shader* fused_result, const Shader* fron
auto* fused_rsrc1 = find_shader_register(fused_regs, fused_reg_count, rsrc1_offset);
auto* fused_rsrc2 = find_shader_register(fused_regs, fused_reg_count, rsrc2_offset);
if (recompute_shared_vgprs) {
const auto front_vgprs = ((front_rsrc1->value & 0x3fu) + 1u) * 4u;
const auto back_vgprs = ((fused_rsrc1->value & 0x3fu) + 1u) * 4u;
const auto front_total = front_vgprs + (front_rsrc2->value >> 28u) * 8u;
const auto back_total = back_vgprs + (fused_rsrc2->value >> 28u) * 8u;
const auto max_total = std::max(front_total, back_total);
// sceAgcFuseShaderHalves reallocates shared VGPRs; the older export takes the maximum.
const auto shared = std::max(front_vgprs, back_vgprs) >= max_total
? 0u
: (max_total - std::min(front_total, back_total) + 7u) / 64u;
fused_rsrc2->value = (fused_rsrc2->value & 0x0fffffffu) | ((shared & 0xfu) << 28u);
} else {
merge_shader_register_max_field(fused_rsrc2, front_rsrc2, 28, 0x0fu);
}
merge_shader_register_max_field(fused_rsrc1, front_rsrc1, 0, 0x3fu);
merge_shader_register_max_field(fused_rsrc2, front_rsrc2, 28, 0x0fu);
if (is_gs) {
merge_shader_register_max_field(fused_rsrc1, front_rsrc1, 29, 0x03u);
merge_shader_register_max_field(fused_rsrc2, front_rsrc2, 16, 0x03u);
fused_rsrc2->value =
(fused_rsrc2->value & 0xf7ffffc1u) | (front_rsrc2->value & 0x0800003eu);
fused_rsrc2->value =
(fused_rsrc2->value & 0xfffbffffu) | (front_rsrc2->value & 0x00040000u);
} else {
merge_shader_register_max_field(fused_rsrc1, front_rsrc1, 28, 0x03u);
fused_rsrc2->value =
(fused_rsrc2->value & 0xf7ffffc1u) | (front_rsrc2->value & 0x0800003eu);
}
fused_rsrc2->value =
(fused_rsrc2->value & 0xf7ffffc1u) | (front_rsrc2->value & 0x0800003eu);
const auto program_lo_offset = is_gs ? Pm4::SPI_SHADER_PGM_LO_ES : Pm4::SPI_SHADER_PGM_LO_LS;
auto* program_lo = find_shader_register(fused_regs, fused_reg_count, program_lo_offset);
const auto address = reinterpret_cast<uint64_t>(front->code);
program_lo->value = static_cast<uint32_t>(address >> 8u);
(program_lo + 1)->value &= 0xffffff00u;
(program_lo + 1)->value |= static_cast<uint32_t>((address >> 40u) & 0xffu);
fused_result->user_data = front->user_data;
patch_shader_register_address(fused_regs, fused_reg_count,
is_gs ? Pm4::SPI_SHADER_PGM_LO_ES : Pm4::SPI_SHADER_PGM_LO_LS,
reinterpret_cast<uint64_t>(front->code));
fused_result->user_data = recompute_shared_vgprs ? nullptr : front->user_data;
return OK;
}
int KYTY_SYSV_ABI AgcUnknownFuseShaderHalves(Shader* fused_result, const Shader* front,
const Shader* back, void* scratch_mem) {
PRINT_NAME();
return fuse_shader_halves(fused_result, front, back, scratch_mem, true);
}
int KYTY_SYSV_ABI AgcUnknownNApJjpKNBl4(Shader* fused_result, const Shader* front,
const Shader* back, void* scratch_mem) {
PRINT_NAME();
return fuse_shader_halves(fused_result, front, back, scratch_mem, false);
}
static constexpr int GRAPHICS5_ERROR_INVALID_PACKET = static_cast<int>(0x8a6c000cu);
enum class RegIndirectPacket : uint32_t {