fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[Base/GPU] Cleanup float comparisons and NaN and -0 in clamping

Browse Source 
C++ relational operators are supposed to raise FE_INVALID if an argument is
NaN, use std::isless/greater[equal] instead where they were easy to locate
(though there are other places possibly, mostly min/max and clamp usage was
checked).

Also fixes a copy-paste error making the CPU shader interpreter execute
MINs as MAXs instead.
This commit is contained in:
Triang3l
2024-05-12 19:21:37 +03:00
parent f964290ea8
commit a3304d252f
9 changed files with 143 additions and 144 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
src/xenia/gpu/draw_util.cc
View File

@@ -399,16 +399,11 @@ void GetHostViewportInfo(const RegisterFile& regs,
float offset_axis = offset_base_xy[i] + offset_add_xy[i];
float scale_axis = scale_xy[i];
float scale_axis_abs = std::abs(scale_xy[i]);
float axis_0 = offset_axis - scale_axis_abs;
float axis_1 = offset_axis + scale_axis_abs;
float axis_max_unscaled_float = float(xy_max_unscaled[i]);
// max(0.0f, xy) drops NaN and < 0 - max picks the first argument in the
// !(a < b) case (always for NaN), min as float (axis_max_unscaled_float
// is well below 2^24) to safely drop very large values.
uint32_t axis_0_int =
uint32_t(std::min(axis_max_unscaled_float, std::max(0.0f, axis_0)));
uint32_t axis_1_int =
uint32_t(std::min(axis_max_unscaled_float, std::max(0.0f, axis_1)));
uint32_t axis_0_int = uint32_t(xe::clamp_float(
offset_axis - scale_axis_abs, 0.0f, axis_max_unscaled_float));
uint32_t axis_1_int = uint32_t(xe::clamp_float(
offset_axis + scale_axis_abs, 0.0f, axis_max_unscaled_float));
uint32_t axis_extent_int = axis_1_int - axis_0_int;
viewport_info_out.xy_offset[i] = axis_0_int * axis_resolution_scale;
viewport_info_out.xy_extent[i] = axis_extent_int * axis_resolution_scale;
@@ -511,8 +506,8 @@ void GetHostViewportInfo(const RegisterFile& regs,
// extension. But cases when this really matters are yet to be found -
// trying to fix this will result in more correct depth values, but
// incorrect clipping.
z_min = xe::saturate_unsigned(host_clip_offset_z);
z_max = xe::saturate_unsigned(host_clip_offset_z + host_clip_scale_z);
z_min = xe::saturate(host_clip_offset_z);
z_max = xe::saturate(host_clip_offset_z + host_clip_scale_z);
// Direct3D 12 doesn't allow reverse depth range - on some drivers it
// works, on some drivers it doesn't, actually, but it was never
// explicitly allowed by the specification.
@@ -877,10 +872,10 @@ bool GetResolveInfo(const RegisterFile& regs, const Memory& memory,
GetScissor(regs, scissor, false);
int32_t scissor_right = int32_t(scissor.offset[0] + scissor.extent[0]);
int32_t scissor_bottom = int32_t(scissor.offset[1] + scissor.extent[1]);
x0 = xe::clamp(x0, int32_t(scissor.offset[0]), scissor_right);
y0 = xe::clamp(y0, int32_t(scissor.offset[1]), scissor_bottom);
x1 = xe::clamp(x1, int32_t(scissor.offset[0]), scissor_right);
y1 = xe::clamp(y1, int32_t(scissor.offset[1]), scissor_bottom);
x0 = std::clamp(x0, int32_t(scissor.offset[0]), scissor_right);
y0 = std::clamp(y0, int32_t(scissor.offset[1]), scissor_bottom);
x1 = std::clamp(x1, int32_t(scissor.offset[0]), scissor_right);
y1 = std::clamp(y1, int32_t(scissor.offset[1]), scissor_bottom);
assert_true(x0 <= x1 && y0 <= y1);

170
src/xenia/gpu/shader_interpreter.cc
View File

@@ -346,16 +346,18 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
} break;
case ucode::AluVectorOpcode::kMax: {
for (uint32_t i = 0; i < 4; ++i) {
vector_result[i] = vector_operands[0][i] >= vector_operands[1][i]
? vector_operands[0][i]
: vector_operands[1][i];
vector_result[i] =
std::isgreaterequal(vector_operands[0][i], vector_operands[1][i])
? vector_operands[0][i]
: vector_operands[1][i];
}
} break;
case ucode::AluVectorOpcode::kMin: {
for (uint32_t i = 0; i < 4; ++i) {
vector_result[i] = vector_operands[0][i] < vector_operands[1][i]
? vector_operands[0][i]
: vector_operands[1][i];
vector_result[i] =
std::isless(vector_operands[0][i], vector_operands[1][i])
? vector_operands[0][i]
: vector_operands[1][i];
}
} break;
case ucode::AluVectorOpcode::kSeq: {
@@ -366,14 +368,14 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
} break;
case ucode::AluVectorOpcode::kSgt: {
for (uint32_t i = 0; i < 4; ++i) {
vector_result[i] =
float(vector_operands[0][i] > vector_operands[1][i]);
vector_result[i] = float(
std::isgreater(vector_operands[0][i], vector_operands[1][i]));
}
} break;
case ucode::AluVectorOpcode::kSge: {
for (uint32_t i = 0; i < 4; ++i) {
vector_result[i] =
float(vector_operands[0][i] >= vector_operands[1][i]);
vector_result[i] = float(std::isgreaterequal(vector_operands[0][i],
vector_operands[1][i]));
}
} break;
case ucode::AluVectorOpcode::kSne: {
@@ -419,14 +421,14 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
} break;
case ucode::AluVectorOpcode::kCndGe: {
for (uint32_t i = 0; i < 4; ++i) {
vector_result[i] = vector_operands[0][i] >= 0.0f
vector_result[i] = std::isgreaterequal(vector_operands[0][i], 0.0f)
? vector_operands[1][i]
: vector_operands[2][i];
}
} break;
case ucode::AluVectorOpcode::kCndGt: {
for (uint32_t i = 0; i < 4; ++i) {
vector_result[i] = vector_operands[0][i] > 0.0f
vector_result[i] = std::isgreater(vector_operands[0][i], 0.0f)
? vector_operands[1][i]
: vector_operands[2][i];
}
@@ -478,32 +480,38 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
float x_abs = std::abs(x), y_abs = std::abs(y), z_abs = std::abs(z);
// Result is T coordinate, S coordinate, 2 * major axis, face ID.
if (z_abs >= x_abs && z_abs >= y_abs) {
bool z_negative = std::isless(z, 0.0f);
vector_result[0] = -y;
vector_result[1] = z < 0.0f ? -x : x;
vector_result[1] = z_negative ? -x : x;
vector_result[2] = z;
vector_result[3] = z < 0.0f ? 5.0f : 4.0f;
vector_result[3] = z_negative ? 5.0f : 4.0f;
} else if (y_abs >= x_abs) {
vector_result[0] = y < 0.0f ? -z : z;
bool y_negative = std::isless(y, 0.0f);
vector_result[0] = y_negative ? -z : z;
vector_result[1] = x;
vector_result[2] = y;
vector_result[3] = y < 0.0f ? 3.0f : 2.0f;
vector_result[3] = y_negative ? 3.0f : 2.0f;
} else {
bool x_negative = std::isless(x, 0.0f);
vector_result[0] = -y;
vector_result[1] = x < 0.0f ? z : -z;
vector_result[1] = x_negative ? z : -z;
vector_result[2] = x;
vector_result[3] = x < 0.0f ? 1.0f : 0.0f;
vector_result[3] = x_negative ? 1.0f : 0.0f;
}
vector_result[2] *= 2.0f;
} break;
case ucode::AluVectorOpcode::kMax4: {
if (vector_operands[0][0] >= vector_operands[0][1] &&
vector_operands[0][0] >= vector_operands[0][2] &&
vector_operands[0][0] >= vector_operands[0][3]) {
if (std::isgreaterequal(vector_operands[0][0], vector_operands[0][1]) &&
std::isgreaterequal(vector_operands[0][0], vector_operands[0][2]) &&
std::isgreaterequal(vector_operands[0][0], vector_operands[0][3])) {
vector_result[0] = vector_operands[0][0];
} else if (vector_operands[0][1] >= vector_operands[0][2] &&
vector_operands[0][1] >= vector_operands[0][3]) {
} else if (std::isgreaterequal(vector_operands[0][1],
vector_operands[0][2]) &&
std::isgreaterequal(vector_operands[0][1],
vector_operands[0][3])) {
vector_result[0] = vector_operands[0][1];
} else if (vector_operands[0][2] >= vector_operands[0][3]) {
} else if (std::isgreaterequal(vector_operands[0][2],
vector_operands[0][3])) {
vector_result[0] = vector_operands[0][2];
} else {
vector_result[0] = vector_operands[0][3];
@@ -529,21 +537,21 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
replicate_vector_result_x = true;
} break;
case ucode::AluVectorOpcode::kSetpGtPush: {
state_.predicate =
vector_operands[0][3] == 0.0f && vector_operands[1][3] > 0.0f;
vector_result[0] =
(vector_operands[0][0] == 0.0f && vector_operands[1][0] > 0.0f)
? 0.0f
: vector_operands[0][0] + 1.0f;
state_.predicate = vector_operands[0][3] == 0.0f &&
std::isgreater(vector_operands[1][3], 0.0f);
vector_result[0] = (vector_operands[0][0] == 0.0f &&
std::isgreater(vector_operands[1][0], 0.0f))
? 0.0f
: vector_operands[0][0] + 1.0f;
replicate_vector_result_x = true;
} break;
case ucode::AluVectorOpcode::kSetpGePush: {
state_.predicate =
vector_operands[0][3] == 0.0f && vector_operands[1][3] >= 0.0f;
vector_result[0] =
(vector_operands[0][0] == 0.0f && vector_operands[1][0] >= 0.0f)
? 0.0f
: vector_operands[0][0] + 1.0f;
state_.predicate = vector_operands[0][3] == 0.0f &&
std::isgreaterequal(vector_operands[1][3], 0.0f);
vector_result[0] = (vector_operands[0][0] == 0.0f &&
std::isgreaterequal(vector_operands[1][0], 0.0f))
? 0.0f
: vector_operands[0][0] + 1.0f;
replicate_vector_result_x = true;
} break;
// Not implementing pixel kill currently, the interpreter is currently
@@ -557,19 +565,19 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
replicate_vector_result_x = true;
} break;
case ucode::AluVectorOpcode::kKillGt: {
vector_result[0] =
float(vector_operands[0][0] > vector_operands[1][0] ||
vector_operands[0][1] > vector_operands[1][1] ||
vector_operands[0][2] > vector_operands[1][2] ||
vector_operands[0][3] > vector_operands[1][3]);
vector_result[0] = float(
std::isgreater(vector_operands[0][0], vector_operands[1][0]) ||
std::isgreater(vector_operands[0][1], vector_operands[1][1]) ||
std::isgreater(vector_operands[0][2], vector_operands[1][2]) ||
std::isgreater(vector_operands[0][3], vector_operands[1][3]));
replicate_vector_result_x = true;
} break;
case ucode::AluVectorOpcode::kKillGe: {
vector_result[0] =
float(vector_operands[0][0] >= vector_operands[1][0] ||
vector_operands[0][1] >= vector_operands[1][1] ||
vector_operands[0][2] >= vector_operands[1][2] ||
vector_operands[0][3] >= vector_operands[1][3]);
vector_result[0] = float(
std::isgreaterequal(vector_operands[0][0], vector_operands[1][0]) ||
std::isgreaterequal(vector_operands[0][1], vector_operands[1][1]) ||
std::isgreaterequal(vector_operands[0][2], vector_operands[1][2]) ||
std::isgreaterequal(vector_operands[0][3], vector_operands[1][3]));
replicate_vector_result_x = true;
} break;
case ucode::AluVectorOpcode::kKillNe: {
@@ -590,14 +598,13 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
vector_result[3] = vector_operands[1][3];
} break;
case ucode::AluVectorOpcode::kMaxA: {
// std::max is `a < b ? b : a`, thus in case of NaN, the first argument
// (-256.0f) is always the result.
state_.address_register = int32_t(std::floor(
std::min(255.0f, std::max(-256.0f, vector_operands[0][3])) + 0.5f));
xe::clamp_float(vector_operands[0][3], -256.0f, 255.0f) + 0.5f));
for (uint32_t i = 0; i < 4; ++i) {
vector_result[i] = vector_operands[0][i] >= vector_operands[1][i]
? vector_operands[0][i]
: vector_operands[1][i];
vector_result[i] =
std::isgreaterequal(vector_operands[0][i], vector_operands[1][i])
? vector_operands[0][i]
: vector_operands[1][i];
}
} break;
default: {
@@ -702,7 +709,8 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
case ucode::AluScalarOpcode::kMulsPrev2: {
if (state_.previous_scalar == -FLT_MAX ||
!std::isfinite(state_.previous_scalar) ||
!std::isfinite(scalar_operands[1]) || scalar_operands[1] <= 0.0f) {
!std::isfinite(scalar_operands[1]) ||
std::islessequal(scalar_operands[1], 0.0f)) {
state_.previous_scalar = -FLT_MAX;
} else {
// Direct3D 9 behavior (0 or denormal * anything = +0).
@@ -713,23 +721,26 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
}
} break;
case ucode::AluScalarOpcode::kMaxs: {
state_.previous_scalar = scalar_operands[0] >= scalar_operands[1]
? scalar_operands[0]
: scalar_operands[1];
state_.previous_scalar =
std::isgreaterequal(scalar_operands[0], scalar_operands[1])
? scalar_operands[0]
: scalar_operands[1];
} break;
case ucode::AluScalarOpcode::kMins: {
state_.previous_scalar = scalar_operands[0] >= scalar_operands[1]
? scalar_operands[0]
: scalar_operands[1];
state_.previous_scalar =
std::isless(scalar_operands[0], scalar_operands[1])
? scalar_operands[0]
: scalar_operands[1];
} break;
case ucode::AluScalarOpcode::kSeqs: {
state_.previous_scalar = float(scalar_operands[0] == 0.0f);
} break;
case ucode::AluScalarOpcode::kSgts: {
state_.previous_scalar = float(scalar_operands[0] > 0.0f);
state_.previous_scalar = float(std::isgreater(scalar_operands[0], 0.0f));
} break;
case ucode::AluScalarOpcode::kSges: {
state_.previous_scalar = float(scalar_operands[0] >= 0.0f);
state_.previous_scalar =
float(std::isgreaterequal(scalar_operands[0], 0.0f));
} break;
case ucode::AluScalarOpcode::kSnes: {
state_.previous_scalar = float(scalar_operands[0] != 0.0f);
@@ -795,22 +806,20 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
state_.previous_scalar = 1.0f / std::sqrt(scalar_operands[0]);
} break;
case ucode::AluScalarOpcode::kMaxAs: {
// std::max is `a < b ? b : a`, thus in case of NaN, the first argument
// (-256.0f) is always the result.
state_.address_register = int32_t(std::floor(
std::min(255.0f, std::max(-256.0f, scalar_operands[0])) + 0.5f));
state_.previous_scalar = scalar_operands[0] >= scalar_operands[1]
? scalar_operands[0]
: scalar_operands[1];
xe::clamp_float(scalar_operands[0], -256.0f, 255.0f) + 0.5f));
state_.previous_scalar =
std::isgreaterequal(scalar_operands[0], scalar_operands[1])
? scalar_operands[0]
: scalar_operands[1];
} break;
case ucode::AluScalarOpcode::kMaxAsf: {
// std::max is `a < b ? b : a`, thus in case of NaN, the first argument
// (-256.0f) is always the result.
state_.address_register = int32_t(
std::floor(std::min(255.0f, std::max(-256.0f, scalar_operands[0]))));
state_.previous_scalar = scalar_operands[0] >= scalar_operands[1]
? scalar_operands[0]
: scalar_operands[1];
std::floor(xe::clamp_float(scalar_operands[0], -256.0f, 255.0f)));
state_.previous_scalar =
std::isgreaterequal(scalar_operands[0], scalar_operands[1])
? scalar_operands[0]
: scalar_operands[1];
} break;
case ucode::AluScalarOpcode::kSubs:
case ucode::AluScalarOpcode::kSubsc0:
@@ -829,11 +838,11 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
state_.previous_scalar = float(!state_.predicate);
} break;
case ucode::AluScalarOpcode::kSetpGt: {
state_.predicate = scalar_operands[0] > 0.0f;
state_.predicate = std::isgreater(scalar_operands[0], 0.0f);
state_.previous_scalar = float(!state_.predicate);
} break;
case ucode::AluScalarOpcode::kSetpGe: {
state_.predicate = scalar_operands[0] >= 0.0f;
state_.predicate = std::isgreaterequal(scalar_operands[0], 0.0f);
state_.previous_scalar = float(!state_.predicate);
} break;
case ucode::AluScalarOpcode::kSetpInv: {
@@ -845,7 +854,7 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
} break;
case ucode::AluScalarOpcode::kSetpPop: {
float new_counter = scalar_operands[0] - 1.0f;
state_.predicate = new_counter <= 0.0f;
state_.predicate = std::islessequal(new_counter, 0.0f);
state_.previous_scalar = state_.predicate ? 0.0f : new_counter;
} break;
case ucode::AluScalarOpcode::kSetpClr: {
@@ -862,10 +871,11 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
state_.previous_scalar = float(scalar_operands[0] == 0.0f);
} break;
case ucode::AluScalarOpcode::kKillsGt: {
state_.previous_scalar = float(scalar_operands[0] > 0.0f);
state_.previous_scalar = float(std::isgreater(scalar_operands[0], 0.0f));
} break;
case ucode::AluScalarOpcode::kKillsGe: {
state_.previous_scalar = float(scalar_operands[0] >= 0.0f);
state_.previous_scalar =
float(std::isgreaterequal(scalar_operands[0], 0.0f));
} break;
case ucode::AluScalarOpcode::kKillsNe: {
state_.previous_scalar = float(scalar_operands[0] != 0.0f);
@@ -891,11 +901,11 @@ void ShaderInterpreter::ExecuteAluInstruction(ucode::AluInstruction instr) {
if (instr.vector_clamp()) {
for (uint32_t i = 0; i < 4; ++i) {
vector_result[i] = xe::saturate_unsigned(vector_result[i]);
vector_result[i] = xe::saturate(vector_result[i]);
}
}
float scalar_result = instr.scalar_clamp()
? xe::saturate_unsigned(state_.previous_scalar)
? xe::saturate(state_.previous_scalar)
: state_.previous_scalar;
uint32_t scalar_result_write_mask = instr.GetScalarOpResultWriteMask();

2
src/xenia/gpu/trace_viewer.cc
View File

@@ -1066,7 +1066,7 @@ void ProgressBar(float frac, float width, float height = 0,
if (height == 0) {
height = ImGui::GetTextLineHeightWithSpacing();
}
frac = xe::saturate_unsigned(frac);
frac = xe::saturate(frac);
const auto fontAtlas = ImGui::GetIO().Fonts;

4
src/xenia/gpu/xenos.cc
View File

@@ -27,7 +27,7 @@ namespace xenos {
float PWLGammaToLinear(float gamma) {
// Not found in game executables, so just using the logic similar to that in
// the Source Engine.
gamma = xe::saturate_unsigned(gamma);
gamma = xe::saturate(gamma);
float scale, offset;
// While the compiled code for linear to gamma conversion uses `vcmpgtfp
// constant, value` comparison (constant > value, or value < constant), it's
@@ -68,7 +68,7 @@ float PWLGammaToLinear(float gamma) {
}
float LinearToPWLGamma(float linear) {
linear = xe::saturate_unsigned(linear);
linear = xe::saturate(linear);
float scale, offset;
// While the compiled code uses `vcmpgtfp constant, value` comparison
// (constant > value, or value < constant), it's preferable to use `value >=