OpenPunk/citra
1
0
Fork 0
mirror of https://github.com/citra-emu/citra.git synced 2024-11-25 03:50:14 +00:00
Code Issues Projects Releases Wiki Activity

Move the shader epilogue into the jitted shader as well.

Browse Source
This commit is contained in:
Henrik Rydgard 2016-04-30 18:58:24 +02:00
parent 4f802f8d26
commit 8b48a9a0ae
2 changed files with 83 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

64
src/video_core/shader/shader.cpp
View File

@ -67,9 +67,6 @@ MICROPROFILE_DEFINE(GPU_VertexShader, "GPU", "Vertex Shader", MP_RGB(50, 50, 240
OutputVertex Run(UnitState<false>& state, const InputVertex& input, int num_attributes) { OutputVertex Run(UnitState<false>& state, const InputVertex& input, int num_attributes) {
MICROPROFILE_SCOPE(GPU_VertexShader); MICROPROFILE_SCOPE(GPU_VertexShader);
state.debug.max_offset = 0;
state.debug.max_opdesc_id = 0;
// Setup output data // Setup output data
OutputVertex ret; OutputVertex ret;
@ -79,6 +76,9 @@ OutputVertex Run(UnitState<false>& state, const InputVertex& input, int num_attr
} else } else
#endif // ARCHITECTURE_x86_64 #endif // ARCHITECTURE_x86_64
{ {
state.debug.max_offset = 0;
state.debug.max_opdesc_id = 0;
auto& config = g_state.regs.vs; auto& config = g_state.regs.vs;
// Setup input register table // Setup input register table
state.program_counter = config.main_offset; state.program_counter = config.main_offset;
@ -105,43 +105,43 @@ OutputVertex Run(UnitState<false>& state, const InputVertex& input, int num_attr
if (num_attributes > 15) state.registers.input[attribute_register_map.attribute15_register] = input.attr[15]; if (num_attributes > 15) state.registers.input[attribute_register_map.attribute15_register] = input.attr[15];
RunInterpreter(state); RunInterpreter(state);
}
// TODO(neobrain): Under some circumstances, up to 16 attributes may be output. We need to // TODO(neobrain): Under some circumstances, up to 16 attributes may be output. We need to
// figure out what those circumstances are and enable the remaining outputs then. // figure out what those circumstances are and enable the remaining outputs then.
unsigned index = 0; unsigned index = 0;
for (unsigned i = 0; i < 7; ++i) { for (unsigned i = 0; i < 7; ++i) {
if (index >= g_state.regs.vs_output_total) if (index >= g_state.regs.vs_output_total)
break; break;
if ((g_state.regs.vs.output_mask & (1 << i)) == 0) if ((g_state.regs.vs.output_mask & (1 << i)) == 0)
continue; continue;
const auto& output_register_map = g_state.regs.vs_output_attributes[index]; // TODO: Don't hardcode VS here const auto& output_register_map = g_state.regs.vs_output_attributes[index]; // TODO: Don't hardcode VS here
u32 semantics[4] = { u32 semantics[4] = {
output_register_map.map_x, output_register_map.map_y, output_register_map.map_x, output_register_map.map_y,
output_register_map.map_z, output_register_map.map_w output_register_map.map_z, output_register_map.map_w
}; };
for (unsigned comp = 0; comp < 4; ++comp) { for (unsigned comp = 0; comp < 4; ++comp) {
float24* out = ((float24*)&ret) + semantics[comp]; float24* out = ((float24*)&ret) + semantics[comp];
if (semantics[comp] != Regs::VSOutputAttributes::INVALID) { if (semantics[comp] != Regs::VSOutputAttributes::INVALID) {
*out = state.registers.output[i][comp]; *out = state.registers.output[i][comp];
} else { } else {
// Zero output so that attributes which aren't output won't have denormals in them, // Zero output so that attributes which aren't output won't have denormals in them,
// which would slow us down later. // which would slow us down later.
memset(out, 0, sizeof(*out)); memset(out, 0, sizeof(*out));
}
} }
index++;
} }
index++; // The hardware takes the absolute and saturates vertex colors like this, *before* doing interpolation
} for (unsigned i = 0; i < 4; ++i) {
ret.color[i] = float24::FromFloat32(
// The hardware takes the absolute and saturates vertex colors like this, *before* doing interpolation std::fmin(std::fabs(ret.color[i].ToFloat32()), 1.0f));
for (unsigned i = 0; i < 4; ++i) { }
ret.color[i] = float24::FromFloat32(
std::fmin(std::fabs(ret.color[i].ToFloat32()), 1.0f));
} }
LOG_TRACE(HW_GPU, "Output vertex: pos(%.2f, %.2f, %.2f, %.2f), quat(%.2f, %.2f, %.2f, %.2f), " LOG_TRACE(HW_GPU, "Output vertex: pos(%.2f, %.2f, %.2f, %.2f), quat(%.2f, %.2f, %.2f, %.2f), "

54
src/video_core/shader/shader_jit_x64.cpp
View File

@ -94,7 +94,7 @@ const JitFunction instr_table[64] = {
// purposes, as documented below: // purposes, as documented below:
/// Pointer to the uniform memory /// Pointer to the uniform memory
static const X64Reg UNIFORMS = R9; static const X64Reg UNIFORMS = R15;
/// The two 32-bit VS address offset registers set by the MOVA instruction /// The two 32-bit VS address offset registers set by the MOVA instruction
static const X64Reg ADDROFFS_REG_0 = R10; static const X64Reg ADDROFFS_REG_0 = R10;
static const X64Reg ADDROFFS_REG_1 = R11; static const X64Reg ADDROFFS_REG_1 = R11;
@ -109,7 +109,12 @@ static const X64Reg COND0 = R13;
/// Result of the previous CMP instruction for the Y-component comparison /// Result of the previous CMP instruction for the Y-component comparison
static const X64Reg COND1 = R14; static const X64Reg COND1 = R14;
/// Pointer to the UnitState instance for the current VS unit /// Pointer to the UnitState instance for the current VS unit
static const X64Reg REGISTERS = R15; static const X64Reg REGISTERS = ABI_PARAM1;
// Clashes with others, but only used in epilogue.
static const X64Reg ISCRATCH1 = RSI;
static const X64Reg ISCRATCH2 = RDI;
/// Pointer to the input data. Aliased over LOOPCOUNT as this is before any loops execute. /// Pointer to the input data. Aliased over LOOPCOUNT as this is before any loops execute.
static const X64Reg INPUT = RSI; static const X64Reg INPUT = RSI;
/// SIMD scratch register /// SIMD scratch register
@ -825,7 +830,7 @@ void JitShader::Compile() {
ABI_PushRegistersAndAdjustStack(ABI_ALL_CALLEE_SAVED, 8); ABI_PushRegistersAndAdjustStack(ABI_ALL_CALLEE_SAVED, 8);
// Prologue: Scatter inputs into registers according to map // Prologue: Scatter inputs into registers according to map
MOV(PTRBITS, R(REGISTERS), R(ABI_PARAM1)); // MOV(PTRBITS, R(REGISTERS), R(ABI_PARAM1));
MOV(PTRBITS, R(INPUT), R(ABI_PARAM2)); MOV(PTRBITS, R(INPUT), R(ABI_PARAM2));
for (int i = 0; i < num_attributes; i++) { for (int i = 0; i < num_attributes; i++) {
MOVAPS(SCRATCH, MDisp(INPUT, i * 16)); MOVAPS(SCRATCH, MDisp(INPUT, i * 16));
@ -850,7 +855,50 @@ void JitShader::Compile() {
MOVAPS(NEGBIT, MatR(RAX)); MOVAPS(NEGBIT, MatR(RAX));
// Call the start of the shader program. // Call the start of the shader program.
CALLptr(R(ABI_PARAM3)); CALLptr(R(ABI_PARAM3));
// Alright, back from the program. Now we can do the epilogue. // Alright, back from the program. Now we can do the epilogue.
X64Reg OUTPUT = ABI_PARAM4;
// TODO(neobrain): Under some circumstances, up to 16 attributes may be output. We need to
// figure out what those circumstances are and enable the remaining outputs then.
unsigned index = 0;
for (unsigned i = 0; i < 7; ++i) {
if (index >= g_state.regs.vs_output_total)
break;
if ((g_state.regs.vs.output_mask & (1 << i)) == 0)
continue;
const auto& output_register_map = g_state.regs.vs_output_attributes[index];
u32 semantics[4] = {
output_register_map.map_x, output_register_map.map_y,
output_register_map.map_z, output_register_map.map_w
};
if (semantics[1] == semantics[0] + 1 && semantics[2] == semantics[1] + 1 && semantics[3] == semantics[2] + 1) {
MOVAPS(SCRATCH, MDisp(REGISTERS, UnitState<false>::OutputOffset(i)));
MOVAPS(MDisp(OUTPUT, semantics[0] * 4), SCRATCH);
} else {
for (unsigned comp = 0; comp < 4; ++comp) {
int outOffset = semantics[comp] * 4;
if (semantics[comp] != Regs::VSOutputAttributes::INVALID) {
MOV(32, R(ISCRATCH1), MDisp(REGISTERS, UnitState<false>::OutputOffset(i) + comp * 4));
MOV(32, MDisp(OUTPUT, outOffset), R(ISCRATCH1));
} else {
// Zero output so that attributes which aren't output won't have denormals in them,
// which would slow us down later.
MOV(32, MDisp(OUTPUT, outOffset), Imm32(0));
}
}
}
index++;
}
// Saturate/Clamp color, specifically.
// These effectively do an ABS.
MOVAPS(SCRATCH, R(NEGBIT));
ANDNPS(SCRATCH, MDisp(OUTPUT, Regs::VSOutputAttributes::Semantic::COLOR_R * 4));
// Clamp to 1.0.
MINPS(SCRATCH, R(ONE));
ABI_PopRegistersAndAdjustStack(ABI_ALL_CALLEE_SAVED, 8); ABI_PopRegistersAndAdjustStack(ABI_ALL_CALLEE_SAVED, 8);
RET(); RET();