Forks/citra
1
0
Fork 0
mirror of https://github.com/PabloMK7/citra.git synced 2025-10-31 13:50:03 +00:00
Code Issues Projects Releases Packages Wiki Activity

Pica/VertexShader: Cleanup flow control logic and implement CMP/IFU instructions.

Browse source
This commit is contained in:
Tony Wasserka 2014-12-13 21:30:13 +01:00
parent 22afb9d830
commit 6bd41de276
1 changed files with 57 additions and 51 deletions
Download patch file Download diff file Expand all files Collapse all files

108
src/video_core/vertex_shader.cpp
View file

@ -2,6 +2,8 @@
// Licensed under GPLv2
// Refer to the license.txt file included.
#include <stack>
#include <boost/range/algorithm.hpp>
#include <common/file_util.h>
@ -65,9 +67,6 @@ const std::array<u32, 1024>& GetSwizzlePatterns()
return swizzle_data;
}
// TODO: Is there actually a limit on hardware?
const int if_stack_size = 8;
struct VertexShaderState {
u32* program_counter;
@ -84,14 +83,14 @@ struct VertexShaderState {
enum {
INVALID_ADDRESS = 0xFFFFFFFF
};
u32 call_stack[8]; // TODO: What is the maximal call stack depth?
u32* call_stack_pointer;
struct IfStackElement {
u32 else_addr;
u32 else_instructions;
} if_stack[if_stack_size];
IfStackElement* if_stack_pointer;
struct CallStackElement {
u32 final_address;
u32 return_address;
};
// TODO: Is there a maximal size for this?
std::stack<CallStackElement> call_stack;
struct {
u32 max_offset; // maximum program counter ever reached
@ -101,12 +100,27 @@ struct VertexShaderState {
static void ProcessShaderCode(VertexShaderState& state) {
while (true) {
bool increment_pc = true;
if (!state.call_stack.empty()) {
if (state.program_counter - shader_memory.data() == state.call_stack.top().final_address) {
state.program_counter = &shader_memory[state.call_stack.top().return_address];
state.call_stack.pop();
// TODO: Is "trying again" accurate to hardware?
continue;
}
}
bool exit_loop = false;
const Instruction& instr = *(const Instruction*)state.program_counter;
const SwizzlePattern& swizzle = *(SwizzlePattern*)&swizzle_data[instr.common.operand_desc_id];
state.debug.max_offset = std::max<u32>(state.debug.max_offset, 1 + (state.program_counter - shader_memory.data()));
auto call = [&](std::stack<VertexShaderState::CallStackElement>& stack, u32 offset, u32 num_instructions, u32 return_offset) {
state.program_counter = &shader_memory[offset] - 1; // -1 to make sure when incrementing the PC we end up at the correct offset
stack.push({ offset + num_instructions, return_offset });
};
u32 binary_offset = state.program_counter - shader_memory.data();
state.debug.max_offset = std::max<u32>(state.debug.max_offset, 1 + binary_offset);
auto LookupSourceRegister = [&](const SourceRegister& source_reg) -> const float24* {
switch (source_reg.GetRegisterType()) {
@ -328,30 +342,33 @@ static void ProcessShaderCode(VertexShaderState& state) {
default:
// Handle each instruction on its own
switch (instr.opcode) {
// NOP is currently used as a heuristic for leaving from a function.
// TODO: This is completely incorrect.
case Instruction::OpCode::NOP:
if (*state.call_stack_pointer == VertexShaderState::INVALID_ADDRESS) {
exit_loop = true;
} else {
// Jump back to call stack position, invalidate call stack entry, move up call stack pointer
state.program_counter = &shader_memory[*state.call_stack_pointer];
*state.call_stack_pointer-- = VertexShaderState::INVALID_ADDRESS;
}
case Instruction::OpCode::END:
exit_loop = true;
break;
case Instruction::OpCode::CALL:
increment_pc = false;
_dbg_assert_(HW_GPU, state.call_stack_pointer - state.call_stack < sizeof(state.call_stack));
*++state.call_stack_pointer = state.program_counter - shader_memory.data();
state.program_counter = &shader_memory[instr.flow_control.dest_offset];
call(state.call_stack,
instr.flow_control.dest_offset,
instr.flow_control.num_instructions,
binary_offset + 1);
break;
case Instruction::OpCode::END:
// TODO
case Instruction::OpCode::NOP:
break;
case Instruction::OpCode::IFU:
if (shader_uniforms.b[instr.flow_control.bool_uniform_id]) {
call(state.call_stack,
binary_offset + 1,
instr.flow_control.dest_offset - binary_offset - 1,
instr.flow_control.dest_offset + instr.flow_control.num_instructions);
} else {
call(state.call_stack,
instr.flow_control.dest_offset,
instr.flow_control.num_instructions,
instr.flow_control.dest_offset + instr.flow_control.num_instructions);
}
break;
case Instruction::OpCode::IFC:
@ -381,12 +398,15 @@ static void ProcessShaderCode(VertexShaderState& state) {
}
if (results[2]) {
++state.if_stack_pointer;
state.if_stack_pointer->else_addr = instr.flow_control.dest_offset;
state.if_stack_pointer->else_instructions = instr.flow_control.num_instructions;
call(state.call_stack,
binary_offset + 1,
instr.flow_control.dest_offset - binary_offset - 1,
instr.flow_control.dest_offset + instr.flow_control.num_instructions);
} else {
state.program_counter = &shader_memory[instr.flow_control.dest_offset] - 1;
call(state.call_stack,
instr.flow_control.dest_offset,
instr.flow_control.num_instructions,
instr.flow_control.dest_offset + instr.flow_control.num_instructions);
}
break;
@ -401,15 +421,7 @@ static void ProcessShaderCode(VertexShaderState& state) {
break;
}
if (increment_pc)
++state.program_counter;
if (state.if_stack_pointer >= &state.if_stack[0]) {
if (state.program_counter - shader_memory.data() == state.if_stack_pointer->else_addr) {
state.program_counter += state.if_stack_pointer->else_instructions;
state.if_stack_pointer--;
}
}
++state.program_counter;
if (exit_loop)
break;
@ -462,12 +474,6 @@ OutputVertex RunShader(const InputVertex& input, int num_attributes)
state.conditional_code[0] = false;
state.conditional_code[1] = false;
boost::fill(state.call_stack, VertexShaderState::INVALID_ADDRESS);
state.call_stack_pointer = &state.call_stack[0];
std::fill(state.if_stack, state.if_stack + sizeof(state.if_stack) / sizeof(state.if_stack[0]),
VertexShaderState::IfStackElement{VertexShaderState::INVALID_ADDRESS, VertexShaderState::INVALID_ADDRESS});
state.if_stack_pointer = state.if_stack - 1; // Meh. TODO: Make this less ugly
ProcessShaderCode(state);
DebugUtils::DumpShader(shader_memory.data(), state.debug.max_offset, swizzle_data.data(),