ShuriZma
/
xenia-canary
mirror of https://github.com/xenia-canary/xenia-canary.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

[DXBC] OpenGL snorm vfetch because why not

This commit is contained in:
Triang3l 2020-05-24 19:25:35 +03:00
parent 04ca28c902
commit aa73d652ba
5 changed files with 83 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

45
src/xenia/gpu/dxbc_shader_translator_fetch.cc
View File

@ -303,6 +303,8 @@ void DxbcShaderTranslator::ProcessVertexFetchInstruction(
DxbcOpIToF(result_unpacked_dest, result_src);
if (!instr.attributes.is_integer) {
float packed_scales[4] = {};
switch (instr.attributes.signed_rf_mode) {
case xenos::SignedRepeatingFractionMode::kZeroClampMinusOne: {
uint32_t packed_scales_mask = 0b0000;
for (uint32_t i = 0; i < 4; ++i) {
if (!(used_format_components & (1 << i))) {
@ -318,9 +320,26 @@ void DxbcShaderTranslator::ProcessVertexFetchInstruction(
DxbcOpMul(DxbcDest::R(system_temp_result_, packed_scales_mask),
result_src, DxbcSrc::LP(packed_scales));
}
// Treat both -(2^(n-1)) and -(2^(n-1)-1) as -1, according to Direct3D
// snorm to float conversion rules.
// Treat both -(2^(n-1)) and -(2^(n-1)-1) as -1.
DxbcOpMax(result_unpacked_dest, result_src, DxbcSrc::LF(-1.0f));
} break;
case xenos::SignedRepeatingFractionMode::kNoZero: {
float packed_zeros[4] = {};
for (uint32_t i = 0; i < 4; ++i) {
if (!(used_format_components & (1 << i))) {
continue;
}
assert_not_zero(packed_widths[i]);
packed_zeros[i] =
1.0f / float((uint32_t(1) << packed_widths[i]) - 1);
packed_scales[i] = 2.0f * packed_zeros[i];
}
DxbcOpMAd(result_unpacked_dest, result_src,
DxbcSrc::LP(packed_scales), DxbcSrc::LP(packed_zeros));
} break;
default:
assert_unhandled_case(instr.attributes.signed_rf_mode);
}
}
} else {
DxbcOpUBFE(result_unpacked_dest, DxbcSrc::LP(packed_widths),
@ -367,12 +386,26 @@ void DxbcShaderTranslator::ProcessVertexFetchInstruction(
DxbcOpUToF(result_unpacked_dest, result_src);
}
if (!instr.attributes.is_integer) {
DxbcOpMul(
result_unpacked_dest, result_src,
DxbcSrc::LF(instr.attributes.is_signed ? (1.0f / 2147483647.0f)
: (1.0f / 4294967295.0f)));
if (instr.attributes.is_signed) {
switch (instr.attributes.signed_rf_mode) {
case xenos::SignedRepeatingFractionMode::kZeroClampMinusOne:
DxbcOpMul(result_unpacked_dest, result_src,
DxbcSrc::LF(1.0f / 2147483647.0f));
// No need to clamp to -1 if signed - 1/(2^31-1) is rounded to
// 1/(2^31) as float32.
break;
case xenos::SignedRepeatingFractionMode::kNoZero:
DxbcOpMAd(result_unpacked_dest, result_src,
DxbcSrc::LF(1.0f / 2147483647.5f),
DxbcSrc::LF(0.5f / 2147483647.5f));
break;
default:
assert_unhandled_case(instr.attributes.signed_rf_mode);
}
} else {
DxbcOpMul(result_unpacked_dest, result_src,
DxbcSrc::LF(1.0f / 4294967295.0f));
}
}
break;
case VertexFormat::k_32_FLOAT:

2
src/xenia/gpu/shader.h
View File

@ -428,6 +428,8 @@ struct ParsedVertexFetchInstruction {
int32_t exp_adjust = 0;
// Prefetch count minus 1.
uint32_t prefetch_count = 0;
xenos::SignedRepeatingFractionMode signed_rf_mode =
xenos::SignedRepeatingFractionMode::kZeroClampMinusOne;
bool is_index_rounded = false;
bool is_signed = false;
bool is_integer = false;

1
src/xenia/gpu/shader_translator.cc
View File

@ -993,6 +993,7 @@ void ShaderTranslator::ParseVertexFetchInstruction(
i.attributes.is_index_rounded = op.is_index_rounded();
i.attributes.is_signed = op.is_signed();
i.attributes.is_integer = !op.is_normalized();
i.attributes.signed_rf_mode = op.signed_rf_mode();
// Store for later use by mini fetches.
if (!op.is_mini_fetch()) {

5
src/xenia/gpu/ucode.h
View File

@ -505,6 +505,9 @@ struct VertexFetchInstruction {
int exp_adjust() const { return data_.exp_adjust; }
bool is_signed() const { return data_.fomat_comp_all == 1; }
bool is_normalized() const { return data_.num_format_all == 0; }
xenos::SignedRepeatingFractionMode signed_rf_mode() const {
return data_.signed_rf_mode_all;
}
bool is_index_rounded() const { return data_.is_index_rounded == 1; }
// Dword stride, [0, 255].
uint32_t stride() const { return data_.stride; }
@ -536,7 +539,7 @@ struct VertexFetchInstruction {
uint32_t dst_swiz : 12;
uint32_t fomat_comp_all : 1;
uint32_t num_format_all : 1;
uint32_t signed_rf_mode_all : 1;
xenos::SignedRepeatingFractionMode signed_rf_mode_all : 1;
uint32_t is_index_rounded : 1;
VertexFormat format : 6;
uint32_t reserved2 : 2;

18
src/xenia/gpu/xenos.h
View File

@ -183,11 +183,11 @@ enum class IndexFormat : uint32_t {
kInt32,
};
// GPUSURFACENUMBER from a game .pdb. "Repeat" means repeating fraction, it's
// what ATI calls normalized.
// SurfaceNumberX from yamato_enum.h.
enum class SurfaceNumFormat : uint32_t {
kUnsignedRepeat = 0,
kSignedRepeat = 1,
kUnsignedRepeatingFraction = 0,
// Microsoft-style, scale factor (2^(n-1))-1.
kSignedRepeatingFraction = 1,
kUnsignedInteger = 2,
kSignedInteger = 3,
kFloat = 7,
@ -516,6 +516,14 @@ inline bool IsMajorModeExplicit(MajorMode major_mode,
primitive_type >= PrimitiveType::kExplicitMajorModeForceStart;
}
enum class SignedRepeatingFractionMode : uint32_t {
// Microsoft-style representation with two -1 representations (one is slightly
// past -1 but clamped).
kZeroClampMinusOne,
// OpenGL "alternate mapping" format lacking representation for zero.
kNoZero,
};
// instr_arbitrary_filter_t
enum class ArbitraryFilter : uint32_t {
k2x4Sym = 0,
@ -706,7 +714,7 @@ XEPACKEDUNION(xe_gpu_texture_fetch_t, {
ClampMode clamp_x : 3; // +10
ClampMode clamp_y : 3; // +13
ClampMode clamp_z : 3; // +16
uint32_t signed_rf_mode_all : 1; // +19
xenos::SignedRepeatingFractionMode signed_rf_mode_all : 1; // +19
// TODO(Triang3l): 1 or 2 dim_tbd bits?
uint32_t unk_0 : 2; // +20
uint32_t pitch : 9; // +22 byte_pitch >> 5