citra-emu
/
citra-nightly
Archived
1
0
Fork 0
Code

shader_interpreter: Fix control flow edge cases and implement break/breakc (#6844)

This commit is contained in:
GPUCode 2023-08-22 04:12:46 +03:00 committed by GitHub
parent 894c1c85a5
commit a94297922b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 128 additions and 72 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
externals/boost vendored

@ -1 +1 @@
Subproject commit 700ae2eff3134792f09cea2b051666688b1d5b97
Subproject commit 3c27c785ad0f8a742af02e620dc225673f3a12d8

198
src/video_core/shader/shader_interpreter.cpp
View File

@ -6,6 +6,7 @@
#include <array>
#include <cmath>
#include <numeric>
#include <boost/circular_buffer.hpp>
#include <boost/container/static_vector.hpp>
#include <nihstro/shader_bytecode.h>
#include "common/assert.h"
@ -26,32 +27,64 @@ using nihstro::SwizzlePattern;
namespace Pica::Shader {
struct IfStackElement {
u32 else_address;
u32 end_address;
};
struct CallStackElement {
u32 final_address; // Address upon which we jump to return_address
u32 return_address; // Where to jump when leaving scope
u8 repeat_counter; // How often to repeat until this call stack element is removed
u8 loop_increment; // Which value to add to the loop counter after an iteration
// TODO: Should this be a signed value? Does it even matter?
u32 loop_address; // The address where we'll return to after each loop iteration
u32 end_address;
u32 return_address;
};
struct LoopStackElement {
u32 entry_address;
u32 end_address;
u8 loop_downcounter;
u8 address_increment;
u8 previous_aL;
};
template <bool Debug>
static void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData<Debug>& debug_data,
unsigned offset) {
// TODO: Is there a maximal size for this?
boost::container::static_vector<CallStackElement, 16> call_stack;
u32 program_counter = offset;
unsigned entry_point) {
boost::circular_buffer<IfStackElement> if_stack(8);
boost::circular_buffer<CallStackElement> call_stack(4);
boost::circular_buffer<LoopStackElement> loop_stack(4);
u32 program_counter = entry_point;
state.conditional_code[0] = false;
state.conditional_code[1] = false;
auto call = [&program_counter, &call_stack](u32 offset, u32 num_instructions, u32 return_offset,
u8 repeat_count, u8 loop_increment) {
// -1 to make sure when incrementing the PC we end up at the correct offset
program_counter = offset - 1;
ASSERT(call_stack.size() < call_stack.capacity());
call_stack.push_back(
{offset + num_instructions, return_offset, repeat_count, loop_increment, offset});
const auto do_if = [&](Instruction instr, bool condition) {
if (condition) {
if_stack.push_back({
.else_address = instr.flow_control.dest_offset,
.end_address = instr.flow_control.dest_offset + instr.flow_control.num_instructions,
});
} else {
program_counter = instr.flow_control.dest_offset - 1;
}
};
const auto do_call = [&](Instruction instr) {
call_stack.push_back({
.end_address = instr.flow_control.dest_offset + instr.flow_control.num_instructions,
.return_address = program_counter + 1,
});
program_counter = instr.flow_control.dest_offset - 1;
};
const auto do_loop = [&](Instruction instr, const Common::Vec4<u8>& loop_param) {
const u8 previous_aL = static_cast<u8>(state.address_registers[2]);
loop_stack.push_back({
.entry_address = program_counter + 1,
.end_address = instr.flow_control.dest_offset + 1,
.loop_downcounter = loop_param.x,
.address_increment = loop_param.z,
.previous_aL = previous_aL,
});
state.address_registers[2] = loop_param.y;
};
auto evaluate_condition = [&state](Instruction::FlowControlType flow_control) {
@ -82,25 +115,11 @@ static void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData
// Placeholder for invalid inputs
static f24 dummy_vec4_float24[4];
unsigned iteration = 0;
bool exit_loop = false;
while (!exit_loop) {
if (!call_stack.empty()) {
auto& top = call_stack.back();
if (program_counter == top.final_address) {
state.address_registers[2] += top.loop_increment;
if (top.repeat_counter-- == 0) {
program_counter = top.return_address;
call_stack.pop_back();
} else {
program_counter = top.loop_address;
}
// TODO: Is "trying again" accurate to hardware?
continue;
}
}
u32 iteration = 0;
bool should_stop = false;
while (!should_stop) {
bool is_break = false;
const u32 old_program_counter = program_counter;
const Instruction instr = {program_code[program_counter]};
const SwizzlePattern swizzle = {swizzle_data[instr.common.operand_desc_id]};
@ -538,7 +557,7 @@ static void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData
// Handle each instruction on its own
switch (instr.opcode.Value()) {
case OpCode::Id::END:
exit_loop = true;
should_stop = true;
break;
case OpCode::Id::JMPC:
@ -559,72 +578,68 @@ static void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData
break;
case OpCode::Id::CALL:
call(instr.flow_control.dest_offset, instr.flow_control.num_instructions,
program_counter + 1, 0, 0);
do_call(instr);
break;
case OpCode::Id::CALLU:
Record<DebugDataRecord::COND_BOOL_IN>(
debug_data, iteration, uniforms.b[instr.flow_control.bool_uniform_id]);
if (uniforms.b[instr.flow_control.bool_uniform_id]) {
call(instr.flow_control.dest_offset, instr.flow_control.num_instructions,
program_counter + 1, 0, 0);
do_call(instr);
}
break;
case OpCode::Id::CALLC:
Record<DebugDataRecord::COND_CMP_IN>(debug_data, iteration, state.conditional_code);
if (evaluate_condition(instr.flow_control)) {
call(instr.flow_control.dest_offset, instr.flow_control.num_instructions,
program_counter + 1, 0, 0);
do_call(instr);
}
break;
case OpCode::Id::NOP:
break;
case OpCode::Id::IFU:
case OpCode::Id::IFU: {
Record<DebugDataRecord::COND_BOOL_IN>(
debug_data, iteration, uniforms.b[instr.flow_control.bool_uniform_id]);
if (uniforms.b[instr.flow_control.bool_uniform_id]) {
call(program_counter + 1, instr.flow_control.dest_offset - program_counter - 1,
instr.flow_control.dest_offset + instr.flow_control.num_instructions, 0,
0);
} else {
call(instr.flow_control.dest_offset, instr.flow_control.num_instructions,
instr.flow_control.dest_offset + instr.flow_control.num_instructions, 0,
0);
}
const bool cond = uniforms.b[instr.flow_control.bool_uniform_id];
do_if(instr, cond);
break;
}
case OpCode::Id::IFC: {
// TODO: Do we need to consider swizzlers here?
Record<DebugDataRecord::COND_CMP_IN>(debug_data, iteration, state.conditional_code);
if (evaluate_condition(instr.flow_control)) {
call(program_counter + 1, instr.flow_control.dest_offset - program_counter - 1,
instr.flow_control.dest_offset + instr.flow_control.num_instructions, 0,
0);
} else {
call(instr.flow_control.dest_offset, instr.flow_control.num_instructions,
instr.flow_control.dest_offset + instr.flow_control.num_instructions, 0,
0);
}
const bool cond = evaluate_condition(instr.flow_control);
do_if(instr, cond);
break;
}
case OpCode::Id::LOOP: {
Common::Vec4<u8> loop_param(uniforms.i[instr.flow_control.int_uniform_id].x,
uniforms.i[instr.flow_control.int_uniform_id].y,
uniforms.i[instr.flow_control.int_uniform_id].z,
uniforms.i[instr.flow_control.int_uniform_id].w);
const Common::Vec4<u8>& loop_param = uniforms.i[instr.flow_control.int_uniform_id];
state.address_registers[2] = loop_param.y;
Record<DebugDataRecord::LOOP_INT_IN>(debug_data, iteration, loop_param);
call(program_counter + 1, instr.flow_control.dest_offset - program_counter,
instr.flow_control.dest_offset + 1, loop_param.x, loop_param.z);
do_loop(instr, loop_param);
Record<DebugDataRecord::ADDR_REG_OUT>(debug_data, iteration,
state.address_registers);
break;
}
case OpCode::Id::BREAK: {
is_break = true;
Record<DebugDataRecord::ADDR_REG_OUT>(debug_data, iteration,
state.address_registers);
break;
}
case OpCode::Id::BREAKC: {
Record<DebugDataRecord::COND_CMP_IN>(debug_data, iteration, state.conditional_code);
if (evaluate_condition(instr.flow_control)) {
is_break = true;
}
Record<DebugDataRecord::ADDR_REG_OUT>(debug_data, iteration,
state.address_registers);
break;
}
@ -657,6 +672,47 @@ static void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData
++program_counter;
++iteration;
// Stacks are checked in the order CALL -> IF -> LOOP. The CALL stack
// can be popped multiple times per instruction. A JMP at the end of a
// scope is never taken, this is why we compare against
// old_program_counter + 1 here.
u32 next_program_counter = old_program_counter + 1;
for (u32 i = 0; i < 4; i++) {
if (call_stack.empty() || call_stack.back().end_address != next_program_counter)
break;
// Hardware bug: when popping four CALL scopes at once, the last
// one doesn't update the program counter
if (i < 3) {
program_counter = call_stack.back().return_address;
next_program_counter = program_counter;
}
call_stack.pop_back();
}
// The other two stacks can only pop one entry per instruction. They
// are checked against the original program counter before any CALL
// scopes were closed and they overwrite any previous program counter
// updates.
if (!if_stack.empty() && if_stack.back().else_address == old_program_counter + 1) {
program_counter = if_stack.back().end_address;
if_stack.pop_back();
}
if (!loop_stack.empty() &&
(loop_stack.back().end_address == old_program_counter + 1 || is_break)) {
auto& loop = loop_stack.back();
state.address_registers[2] += loop.address_increment;
if (!is_break && loop.loop_downcounter--) {
program_counter = loop.entry_address;
} else {
program_counter = loop.end_address;
// Only restore previous value if there is a surrounding LOOP scope.
if (loop_stack.size() > 1)
state.address_registers[2] = loop.previous_aL;
loop_stack.pop_back();
}
}
}
}