Mirrors/shadPS4
1
0
Fork 0
mirror of https://github.com/shadps4-emu/shadPS4.git synced 2025-05-18 17:34:52 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

control_flow_graph: Initial divergence handling (#434)

Browse source 
* control_flow_graph: Initial divergence handling

* cfg: Handle additional case

* spirv: Handle tgid enable bits

* clang format

* spirv: Use proper format

* translator: Add more instructions
This commit is contained in:
TheTurtle 2024-08-16 20:05:37 +03:00 • committed by GitHub
parent ff33b00c3a
commit 1d1c88ad31
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
14 changed files with 154 additions and 36 deletions
Download patch file Download diff file Expand all files Collapse all files

120
src/shader_recompiler/frontend/control_flow_graph.cpp
View file

@ -35,15 +35,22 @@ static IR::Condition MakeCondition(Opcode opcode) {
return IR::Condition::Execz;
case Opcode::S_CBRANCH_EXECNZ:
return IR::Condition::Execnz;
case Opcode::S_AND_SAVEEXEC_B64:
case Opcode::S_ANDN2_B64:
return IR::Condition::Execnz;
default:
return IR::Condition::True;
}
}
static constexpr size_t LabelReserveSize = 32;
CFG::CFG(Common::ObjectPool<Block>& block_pool_, std::span<const GcnInst> inst_list_)
: block_pool{block_pool_}, inst_list{inst_list_} {
index_to_pc.resize(inst_list.size() + 1);
labels.reserve(LabelReserveSize);
EmitLabels();
EmitDivergenceLabels();
EmitBlocks();
LinkBlocks();
}
@ -51,14 +58,7 @@ CFG::CFG(Common::ObjectPool<Block>& block_pool_, std::span<const GcnInst> inst_l
void CFG::EmitLabels() {
// Always set a label at entry point.
u32 pc = 0;
labels.push_back(pc);
const auto add_label = [this](u32 address) {
const auto it = std::ranges::find(labels, address);
if (it == labels.end()) {
labels.push_back(address);
}
};
AddLabel(pc);
// Iterate instruction list and add labels to branch targets.
for (u32 i = 0; i < inst_list.size(); i++) {
@ -66,15 +66,15 @@ void CFG::EmitLabels() {
const GcnInst inst = inst_list[i];
if (inst.IsUnconditionalBranch()) {
const u32 target = inst.BranchTarget(pc);
add_label(target);
AddLabel(target);
} else if (inst.IsConditionalBranch()) {
const u32 true_label = inst.BranchTarget(pc);
const u32 false_label = pc + inst.length;
add_label(true_label);
add_label(false_label);
AddLabel(true_label);
AddLabel(false_label);
} else if (inst.opcode == Opcode::S_ENDPGM) {
const u32 next_label = pc + inst.length;
add_label(next_label);
AddLabel(next_label);
}
pc += inst.length;
}
@ -84,16 +84,70 @@ void CFG::EmitLabels() {
std::ranges::sort(labels);
}
void CFG::EmitBlocks() {
const auto get_index = [this](Label label) -> size_t {
if (label == 0) {
return 0ULL;
}
const auto it_index = std::ranges::lower_bound(index_to_pc, label);
ASSERT(it_index != index_to_pc.end() || label > index_to_pc.back());
return std::distance(index_to_pc.begin(), it_index);
void CFG::EmitDivergenceLabels() {
const auto is_open_scope = [](const GcnInst& inst) {
// An open scope instruction is an instruction that modifies EXEC
// but also saves the previous value to restore later. This indicates
// we are entering a scope.
return inst.opcode == Opcode::S_AND_SAVEEXEC_B64 ||
// While this instruction does not save EXEC it is often used paired
// with SAVEEXEC to mask the threads that didn't pass the condition
// of initial branch.
inst.opcode == Opcode::S_ANDN2_B64;
};
const auto is_close_scope = [](const GcnInst& inst) {
// Closing an EXEC scope can be either a branch instruction
// (typical case when S_AND_SAVEEXEC_B64 is right before a branch)
// or by a move instruction to EXEC that restores the backup.
return (inst.opcode == Opcode::S_MOV_B64 && inst.dst[0].field == OperandField::ExecLo) ||
// Sometimes compiler might insert instructions between the SAVEEXEC and the branch.
// Those instructions need to be wrapped in the condition as well so allow branch
// as end scope instruction.
inst.opcode == Opcode::S_CBRANCH_EXECZ || inst.opcode == Opcode::S_ANDN2_B64;
};
// Since we will be adding new labels, avoid iterating those as well.
const size_t end_size = labels.size();
for (u32 l = 0; l < end_size; l++) {
const Label start = labels[l];
// Stop if we reached end of existing labels.
if (l == end_size - 1) {
break;
}
const Label end = labels[l + 1];
const size_t end_index = GetIndex(end);
s32 curr_begin = -1;
for (size_t index = GetIndex(start); index < end_index; index++) {
const auto& inst = inst_list[index];
if (is_close_scope(inst) && curr_begin != -1) {
// If there are no instructions inside scope don't do anything.
if (index - curr_begin == 1) {
curr_begin = -1;
continue;
}
// Add a label to the instruction right after the open scope call.
// It is the start of a new basic block.
const auto& save_inst = inst_list[curr_begin];
const Label label = index_to_pc[curr_begin] + save_inst.length;
AddLabel(label);
// Add a label to the close scope instruction as well.
AddLabel(index_to_pc[index]);
// Reset scope begin.
curr_begin = -1;
}
// Mark a potential start of an exec scope.
if (is_open_scope(inst)) {
curr_begin = index;
}
}
}
// Sort labels to make sure block insertion is correct.
std::ranges::sort(labels);
}
void CFG::EmitBlocks() {
for (auto it = labels.begin(); it != labels.end(); it++) {
const Label start = *it;
const auto next_it = std::next(it);
@ -102,8 +156,10 @@ void CFG::EmitBlocks() {
// Last label is special.
return;
}
// The end label is the start instruction of next block.
// The end instruction of this block is the previous one.
const Label end = *next_it;
const size_t end_index = get_index(end) - 1;
const size_t end_index = GetIndex(end) - 1;
const auto& end_inst = inst_list[end_index];
// Insert block between the labels using the last instruction
@ -111,7 +167,7 @@ void CFG::EmitBlocks() {
Block* block = block_pool.Create();
block->begin = start;
block->end = end;
block->begin_index = get_index(start);
block->begin_index = GetIndex(start);
block->end_index = end_index;
block->end_inst = end_inst;
block->cond = MakeCondition(end_inst.opcode);
@ -126,8 +182,26 @@ void CFG::LinkBlocks() {
return &*it;
};
for (auto& block : blocks) {
for (auto it = blocks.begin(); it != blocks.end(); it++) {
auto& block = *it;
const auto end_inst{block.end_inst};
// Handle divergence block inserted here.
if (end_inst.opcode == Opcode::S_AND_SAVEEXEC_B64 ||
end_inst.opcode == Opcode::S_ANDN2_B64) {
// Blocks are stored ordered by address in the set
auto next_it = std::next(it);
auto* target_block = &(*next_it);
++target_block->num_predecessors;
block.branch_true = target_block;
auto merge_it = std::next(next_it);
auto* merge_block = &(*merge_it);
++merge_block->num_predecessors;
block.branch_false = merge_block;
block.end_class = EndClass::Branch;
continue;
}
// If the block doesn't end with a branch we simply
// need to link with the next block.
if (!end_inst.IsTerminateInstruction()) {

19
src/shader_recompiler/frontend/control_flow_graph.h
View file

@ -3,11 +3,13 @@
#pragma once
#include <algorithm>
#include <span>
#include <string>
#include <boost/container/small_vector.hpp>
#include <boost/intrusive/set.hpp>
#include "common/assert.h"
#include "common/object_pool.h"
#include "common/types.h"
#include "shader_recompiler/frontend/instruction.h"
@ -55,9 +57,26 @@ public:
private:
void EmitLabels();
void EmitDivergenceLabels();
void EmitBlocks();
void LinkBlocks();
void AddLabel(Label address) {
const auto it = std::ranges::find(labels, address);
if (it == labels.end()) {
labels.push_back(address);
}
};
size_t GetIndex(Label label) {
if (label == 0) {
return 0ULL;
}
const auto it_index = std::ranges::lower_bound(index_to_pc, label);
ASSERT(it_index != index_to_pc.end() || label > index_to_pc.back());
return std::distance(index_to_pc.begin(), it_index);
};
public:
Common::ObjectPool<Block>& block_pool;
std::span<const GcnInst> inst_list;

2
src/shader_recompiler/frontend/translate/scalar_alu.cpp
View file

@ -29,6 +29,8 @@ void Translator::EmitScalarAlu(const GcnInst& inst) {
return S_CMP(ConditionOp::LG, true, inst);
case Opcode::S_CMP_GT_I32:
return S_CMP(ConditionOp::GT, true, inst);
case Opcode::S_CMP_LE_I32:
return S_CMP(ConditionOp::LE, true, inst);
case Opcode::S_CMP_GE_I32:
return S_CMP(ConditionOp::GE, true, inst);
case Opcode::S_CMP_EQ_I32:

12
src/shader_recompiler/frontend/translate/translate.cpp
View file

@ -64,9 +64,15 @@ void Translator::EmitPrologue() {
ir.SetVectorReg(dst_vreg++, ir.GetAttributeU32(IR::Attribute::LocalInvocationId, 1));
ir.SetVectorReg(dst_vreg++, ir.GetAttributeU32(IR::Attribute::LocalInvocationId, 2));
ir.SetScalarReg(dst_sreg++, ir.GetAttributeU32(IR::Attribute::WorkgroupId, 0));
ir.SetScalarReg(dst_sreg++, ir.GetAttributeU32(IR::Attribute::WorkgroupId, 1));
ir.SetScalarReg(dst_sreg++, ir.GetAttributeU32(IR::Attribute::WorkgroupId, 2));
if (info.tgid_enable[0]) {
ir.SetScalarReg(dst_sreg++, ir.GetAttributeU32(IR::Attribute::WorkgroupId, 0));
}
if (info.tgid_enable[1]) {
ir.SetScalarReg(dst_sreg++, ir.GetAttributeU32(IR::Attribute::WorkgroupId, 1));
}
if (info.tgid_enable[2]) {
ir.SetScalarReg(dst_sreg++, ir.GetAttributeU32(IR::Attribute::WorkgroupId, 2));
}
break;
default:
throw NotImplementedException("Unknown shader stage");

5
src/shader_recompiler/frontend/translate/vector_memory.cpp
View file

@ -91,6 +91,11 @@ void Translator::EmitVectorMemory(const GcnInst& inst) {
case Opcode::BUFFER_STORE_FORMAT_XYZW:
return BUFFER_STORE_FORMAT(4, false, true, inst);
case Opcode::TBUFFER_STORE_FORMAT_X:
return BUFFER_STORE_FORMAT(1, true, true, inst);
case Opcode::TBUFFER_STORE_FORMAT_XYZ:
return BUFFER_STORE_FORMAT(3, true, true, inst);
case Opcode::BUFFER_STORE_DWORD:
return BUFFER_STORE_FORMAT(1, false, false, inst);
case Opcode::BUFFER_STORE_DWORDX2: