*
*/
+#include <algorithm>
+
#include "aco_ir.h"
+#include <stack>
namespace aco {
namespace {
-struct NOP_ctx {
+struct NOP_ctx_gfx8_9 {
enum chip_class chip_class;
unsigned vcc_physical;
- /* pre-GFX10 */
/* just initialize these with something less than max NOPs */
int VALU_wrexec = -10;
int VALU_wrvcc = -10;
int VALU_wrsgpr = -10;
- /* GFX10 */
- int last_VMEM_since_scalar_write = -1;
+ NOP_ctx_gfx8_9(Program* program) : chip_class(program->chip_class) {
+ vcc_physical = program->config->num_sgprs - 2;
+ }
+};
+
+struct NOP_ctx_gfx10 {
bool has_VOPC = false;
+ bool has_nonVALU_exec_read = false;
+ bool has_VMEM = false;
+ bool has_branch_after_VMEM = false;
+ bool has_DS = false;
+ bool has_branch_after_DS = false;
+ std::bitset<128> sgprs_read_by_VMEM;
+ std::bitset<128> sgprs_read_by_SMEM;
+
+ void join(const NOP_ctx_gfx10 &other) {
+ has_VOPC |= other.has_VOPC;
+ has_nonVALU_exec_read |= other.has_nonVALU_exec_read;
+ has_VMEM |= other.has_VMEM;
+ has_branch_after_VMEM |= other.has_branch_after_VMEM;
+ has_DS |= other.has_DS;
+ has_branch_after_DS |= other.has_branch_after_DS;
+ sgprs_read_by_VMEM |= other.sgprs_read_by_VMEM;
+ sgprs_read_by_SMEM |= other.sgprs_read_by_SMEM;
+ }
- NOP_ctx(Program* program) : chip_class(program->chip_class) {
- vcc_physical = program->config->num_sgprs - 2;
+ bool operator==(const NOP_ctx_gfx10 &other)
+ {
+ return
+ has_VOPC == other.has_VOPC &&
+ has_nonVALU_exec_read == other.has_nonVALU_exec_read &&
+ has_VMEM == other.has_VMEM &&
+ has_branch_after_VMEM == other.has_branch_after_VMEM &&
+ has_DS == other.has_DS &&
+ has_branch_after_DS == other.has_branch_after_DS &&
+ sgprs_read_by_VMEM == other.sgprs_read_by_VMEM &&
+ sgprs_read_by_SMEM == other.sgprs_read_by_SMEM;
}
};
+template <std::size_t N>
+bool check_written_regs(const aco_ptr<Instruction> &instr, const std::bitset<N> &check_regs)
+{
+ return std::any_of(instr->definitions.begin(), instr->definitions.end(), [&check_regs](const Definition &def) -> bool {
+ bool writes_any = false;
+ for (unsigned i = 0; i < def.size(); i++) {
+ unsigned def_reg = def.physReg() + i;
+ writes_any |= def_reg < check_regs.size() && check_regs[def_reg];
+ }
+ return writes_any;
+ });
+}
+
+template <std::size_t N>
+void mark_read_regs(const aco_ptr<Instruction> &instr, std::bitset<N> ®_reads)
+{
+ for (const Operand &op : instr->operands) {
+ for (unsigned i = 0; i < op.size(); i++) {
+ unsigned reg = op.physReg() + i;
+ if (reg < reg_reads.size())
+ reg_reads.set(reg);
+ }
+ }
+}
+
bool VALU_writes_sgpr(aco_ptr<Instruction>& instr)
{
if ((uint32_t) instr->format & (uint32_t) Format::VOPC)
return false;
}
+bool instr_reads_exec(const aco_ptr<Instruction>& instr)
+{
+ return std::any_of(instr->operands.begin(), instr->operands.end(), [](const Operand &op) -> bool {
+ return op.physReg() == exec_lo || op.physReg() == exec_hi;
+ });
+}
+
+bool instr_writes_exec(const aco_ptr<Instruction>& instr)
+{
+ return std::any_of(instr->definitions.begin(), instr->definitions.end(), [](const Definition &def) -> bool {
+ return def.physReg() == exec_lo || def.physReg() == exec_hi;
+ });
+}
+
+bool instr_writes_sgpr(const aco_ptr<Instruction>& instr)
+{
+ return std::any_of(instr->definitions.begin(), instr->definitions.end(), [](const Definition &def) -> bool {
+ return def.getTemp().type() == RegType::sgpr;
+ });
+}
+
+inline bool instr_is_branch(const aco_ptr<Instruction>& instr)
+{
+ return instr->opcode == aco_opcode::s_branch ||
+ instr->opcode == aco_opcode::s_cbranch_scc0 ||
+ instr->opcode == aco_opcode::s_cbranch_scc1 ||
+ instr->opcode == aco_opcode::s_cbranch_vccz ||
+ instr->opcode == aco_opcode::s_cbranch_vccnz ||
+ instr->opcode == aco_opcode::s_cbranch_execz ||
+ instr->opcode == aco_opcode::s_cbranch_execnz ||
+ instr->opcode == aco_opcode::s_cbranch_cdbgsys ||
+ instr->opcode == aco_opcode::s_cbranch_cdbguser ||
+ instr->opcode == aco_opcode::s_cbranch_cdbgsys_or_user ||
+ instr->opcode == aco_opcode::s_cbranch_cdbgsys_and_user ||
+ instr->opcode == aco_opcode::s_subvector_loop_begin ||
+ instr->opcode == aco_opcode::s_subvector_loop_end ||
+ instr->opcode == aco_opcode::s_setpc_b64 ||
+ instr->opcode == aco_opcode::s_swappc_b64 ||
+ instr->opcode == aco_opcode::s_getpc_b64 ||
+ instr->opcode == aco_opcode::s_call_b64;
+}
+
bool regs_intersect(PhysReg a_reg, unsigned a_size, PhysReg b_reg, unsigned b_size)
{
return a_reg > b_reg ?
return 0;
}
-int handle_instruction(NOP_ctx& ctx, aco_ptr<Instruction>& instr,
- std::vector<aco_ptr<Instruction>>& old_instructions,
- std::vector<aco_ptr<Instruction>>& new_instructions)
+int handle_instruction_gfx8_9(NOP_ctx_gfx8_9& ctx, aco_ptr<Instruction>& instr,
+
std::vector<aco_ptr<Instruction>>& old_instructions,
+
std::vector<aco_ptr<Instruction>>& new_instructions)
{
int new_idx = new_instructions.size();
return 0;
}
-std::pair<int, int> handle_instruction_gfx10(NOP_ctx& ctx, aco_ptr<Instruction>& instr,
- std::vector<aco_ptr<Instruction>>& old_instructions,
- std::vector<aco_ptr<Instruction>>& new_instructions)
+void handle_block_gfx8_9(NOP_ctx_gfx8_9& ctx, Block& block)
{
- int new_idx = new_instructions.size();
- unsigned vNOPs = 0;
- unsigned sNOPs = 0;
-
- /* break off from prevous SMEM group ("clause" seems to mean something different in RDNA) if needed */
- if (instr->format == Format::SMEM)
- sNOPs = std::max(sNOPs, handle_SMEM_clause(instr, new_idx, new_instructions));
-
- /* handle EXEC/M0/SGPR write following a VMEM instruction without a VALU or "waitcnt vmcnt(0)" in-between */
- if (instr->isSALU() || instr->format == Format::SMEM) {
- if (!instr->definitions.empty() && ctx.last_VMEM_since_scalar_write != -1) {
- ctx.last_VMEM_since_scalar_write = -1;
- vNOPs = 1;
+ std::vector<aco_ptr<Instruction>> instructions;
+ instructions.reserve(block.instructions.size());
+ for (unsigned i = 0; i < block.instructions.size(); i++) {
+ aco_ptr<Instruction>& instr = block.instructions[i];
+ unsigned NOPs = handle_instruction_gfx8_9(ctx, instr, block.instructions, instructions);
+ if (NOPs) {
+ // TODO: try to move the instruction down
+ /* create NOP */
+ aco_ptr<SOPP_instruction> nop{create_instruction<SOPP_instruction>(aco_opcode::s_nop, Format::SOPP, 0, 0)};
+ nop->imm = NOPs - 1;
+ nop->block = -1;
+ instructions.emplace_back(std::move(nop));
+ }
+
+ instructions.emplace_back(std::move(instr));
+ }
+
+ ctx.VALU_wrvcc -= instructions.size();
+ ctx.VALU_wrexec -= instructions.size();
+ ctx.VALU_wrsgpr -= instructions.size();
+ block.instructions = std::move(instructions);
+}
+
+void insert_NOPs_gfx8_9(Program* program)
+{
+ NOP_ctx_gfx8_9 ctx(program);
+
+ for (Block& block : program->blocks) {
+ if (block.instructions.empty())
+ continue;
+
+ handle_block_gfx8_9(ctx, block);
+ }
+}
+
+void handle_instruction_gfx10(NOP_ctx_gfx10 &ctx, aco_ptr<Instruction>& instr,
+ std::vector<aco_ptr<Instruction>>& old_instructions,
+ std::vector<aco_ptr<Instruction>>& new_instructions)
+{
+ /* VMEMtoScalarWriteHazard
+ * Handle EXEC/M0/SGPR write following a VMEM instruction without a VALU or "waitcnt vmcnt(0)" in-between.
+ */
+ if (instr->isVMEM() || instr->format == Format::FLAT || instr->format == Format::GLOBAL ||
+ instr->format == Format::SCRATCH || instr->format == Format::DS) {
+ /* Remember all SGPRs that are read by the VMEM instruction */
+ mark_read_regs(instr, ctx.sgprs_read_by_VMEM);
+ } else if (instr->isSALU() || instr->format == Format::SMEM) {
+ /* Check if SALU writes an SGPR that was previously read by the VALU */
+ if (check_written_regs(instr, ctx.sgprs_read_by_VMEM)) {
+ ctx.sgprs_read_by_VMEM.reset();
+
+ /* Insert v_nop to mitigate the problem */
+ aco_ptr<VOP1_instruction> nop{create_instruction<VOP1_instruction>(aco_opcode::v_nop, Format::VOP1, 0, 0)};
+ new_instructions.emplace_back(std::move(nop));
}
- } else if (instr->isVMEM() || instr->isFlatOrGlobal()) {
- ctx.last_VMEM_since_scalar_write = new_idx;
} else if (instr->opcode == aco_opcode::s_waitcnt) {
+ /* Hazard is mitigated by "s_waitcnt vmcnt(0)" */
uint16_t imm = static_cast<SOPP_instruction*>(instr.get())->imm;
unsigned vmcnt = (imm & 0xF) | ((imm & (0x3 << 14)) >> 10);
if (vmcnt == 0)
- ctx.last_VMEM_since_scalar_write = -1;
+ ctx.sgprs_read_by_VMEM.reset();
} else if (instr->isVALU()) {
- ctx.last_VMEM_since_scalar_write = -1;
+ /* Hazard is mitigated by any VALU instruction */
+ ctx.sgprs_read_by_VMEM.reset();
}
/* VcmpxPermlaneHazard
ctx.has_VOPC = false;
}
- return std::make_pair(sNOPs, vNOPs);
-}
+ /* VcmpxExecWARHazard
+ * Handle any VALU instruction writing the exec mask after it was read by a non-VALU instruction.
+ */
+ if (!instr->isVALU() && instr_reads_exec(instr)) {
+ ctx.has_nonVALU_exec_read = true;
+ } else if (instr->isVALU()) {
+ if (instr_writes_exec(instr)) {
+ ctx.has_nonVALU_exec_read = false;
+
+ /* Insert s_waitcnt_depctr instruction with magic imm to mitigate the problem */
+ aco_ptr<SOPP_instruction> depctr{create_instruction<SOPP_instruction>(aco_opcode::s_waitcnt_depctr, Format::SOPP, 0, 1)};
+ depctr->imm = 0xfffe;
+ depctr->definitions[0] = Definition(sgpr_null, s1);
+ new_instructions.emplace_back(std::move(depctr));
+ } else if (instr_writes_sgpr(instr)) {
+ /* Any VALU instruction that writes an SGPR mitigates the problem */
+ ctx.has_nonVALU_exec_read = false;
+ }
+ } else if (instr->opcode == aco_opcode::s_waitcnt_depctr) {
+ /* s_waitcnt_depctr can mitigate the problem if it has a magic imm */
+ const SOPP_instruction *sopp = static_cast<const SOPP_instruction *>(instr.get());
+ if ((sopp->imm & 0xfffe) == 0xfffe)
+ ctx.has_nonVALU_exec_read = false;
+ }
+ /* SMEMtoVectorWriteHazard
+ * Handle any VALU instruction writing an SGPR after an SMEM reads it.
+ */
+ if (instr->format == Format::SMEM) {
+ /* Remember all SGPRs that are read by the SMEM instruction */
+ mark_read_regs(instr, ctx.sgprs_read_by_SMEM);
+ } else if (VALU_writes_sgpr(instr)) {
+ /* Check if VALU writes an SGPR that was previously read by SMEM */
+ if (check_written_regs(instr, ctx.sgprs_read_by_SMEM)) {
+ ctx.sgprs_read_by_SMEM.reset();
+
+ /* Insert s_mov to mitigate the problem */
+ aco_ptr<SOP1_instruction> s_mov{create_instruction<SOP1_instruction>(aco_opcode::s_mov_b32, Format::SOP1, 1, 1)};
+ s_mov->definitions[0] = Definition(sgpr_null, s1);
+ s_mov->operands[0] = Operand(0u);
+ new_instructions.emplace_back(std::move(s_mov));
+ }
+ } else if (instr->isSALU()) {
+ if (instr->format != Format::SOPP) {
+ /* SALU can mitigate the hazard */
+ ctx.sgprs_read_by_SMEM.reset();
+ } else {
+ /* Reducing lgkmcnt count to 0 always mitigates the hazard. */
+ const SOPP_instruction *sopp = static_cast<const SOPP_instruction *>(instr.get());
+ if (sopp->opcode == aco_opcode::s_waitcnt_lgkmcnt) {
+ if (sopp->imm == 0 && sopp->definitions[0].physReg() == sgpr_null)
+ ctx.sgprs_read_by_SMEM.reset();
+ } else if (sopp->opcode == aco_opcode::s_waitcnt) {
+ unsigned lgkm = (sopp->imm >> 8) & 0x3f;
+ if (lgkm == 0)
+ ctx.sgprs_read_by_SMEM.reset();
+ }
+ }
+ }
+
+ /* LdsBranchVmemWARHazard
+ * Handle VMEM/GLOBAL/SCRATCH->branch->DS and DS->branch->VMEM/GLOBAL/SCRATCH patterns.
+ */
+ if (instr->isVMEM() || instr->format == Format::GLOBAL || instr->format == Format::SCRATCH) {
+ ctx.has_VMEM = true;
+ ctx.has_branch_after_VMEM = false;
+ /* Mitigation for DS is needed only if there was already a branch after */
+ ctx.has_DS = ctx.has_branch_after_DS;
+ } else if (instr->format == Format::DS) {
+ ctx.has_DS = true;
+ ctx.has_branch_after_DS = false;
+ /* Mitigation for VMEM is needed only if there was already a branch after */
+ ctx.has_VMEM = ctx.has_branch_after_VMEM;
+ } else if (instr_is_branch(instr)) {
+ ctx.has_branch_after_VMEM = ctx.has_VMEM;
+ ctx.has_branch_after_DS = ctx.has_DS;
+ } else if (instr->opcode == aco_opcode::s_waitcnt_vscnt) {
+ /* Only s_waitcnt_vscnt can mitigate the hazard */
+ const SOPK_instruction *sopk = static_cast<const SOPK_instruction *>(instr.get());
+ if (sopk->definitions[0].physReg() == sgpr_null && sopk->imm == 0)
+ ctx.has_VMEM = ctx.has_branch_after_VMEM = ctx.has_DS = ctx.has_branch_after_DS = false;
+ }
+ if ((ctx.has_VMEM && ctx.has_branch_after_DS) || (ctx.has_DS && ctx.has_branch_after_VMEM)) {
+ ctx.has_VMEM = ctx.has_branch_after_VMEM = ctx.has_DS = ctx.has_branch_after_DS = false;
+
+ /* Insert s_waitcnt_vscnt to mitigate the problem */
+ aco_ptr<SOPK_instruction> wait{create_instruction<SOPK_instruction>(aco_opcode::s_waitcnt_vscnt, Format::SOPK, 0, 1)};
+ wait->definitions[0] = Definition(sgpr_null, s1);
+ wait->imm = 0;
+ new_instructions.emplace_back(std::move(wait));
+ }
+}
-void handle_block(NOP_ctx& ctx, Block& block)
+void handle_block_gfx10(NOP_ctx_gfx10& ctx, Block& block)
{
+ if (block.instructions.empty())
+ return;
+
std::vector<aco_ptr<Instruction>> instructions;
instructions.reserve(block.instructions.size());
- for (unsigned i = 0; i < block.instructions.size(); i++) {
- aco_ptr<Instruction>& instr = block.instructions[i];
- unsigned NOPs = handle_instruction(ctx, instr, block.instructions, instructions);
- if (NOPs) {
- // TODO: try to move the instruction down
- /* create NOP */
- aco_ptr<SOPP_instruction> nop{create_instruction<SOPP_instruction>(aco_opcode::s_nop, Format::SOPP, 0, 0)};
- nop->imm = NOPs - 1;
- nop->block = -1;
- instructions.emplace_back(std::move(nop));
- }
+ for (aco_ptr<Instruction>& instr : block.instructions) {
+ handle_instruction_gfx10(ctx, instr, block.instructions, instructions);
instructions.emplace_back(std::move(instr));
}
- ctx.VALU_wrvcc -= instructions.size();
- ctx.VALU_wrexec -= instructions.size();
- ctx.VALU_wrsgpr -= instructions.size();
block.instructions = std::move(instructions);
}
-void handle_block_gfx10(NOP_ctx& ctx, Block& block)
+void mitigate_hazards_gfx10(Program *program)
{
- std::vector<aco_ptr<Instruction>> instructions;
- instructions.reserve(block.instructions.size());
- for (unsigned i = 0; i < block.instructions.size(); i++) {
- aco_ptr<Instruction>& instr = block.instructions[i];
- std::pair<int, int> NOPs = handle_instruction_gfx10(ctx, instr, block.instructions, instructions);
- for (int i = 0; i < NOPs.second; i++) {
- // TODO: try to move the instruction down
- /* create NOP */
- aco_ptr<VOP1_instruction> nop{create_instruction<VOP1_instruction>(aco_opcode::v_nop, Format::VOP1, 0, 0)};
- instructions.emplace_back(std::move(nop));
- }
- if (NOPs.first) {
- // TODO: try to move the instruction down
- /* create NOP */
- aco_ptr<SOPP_instruction> nop{create_instruction<SOPP_instruction>(aco_opcode::s_nop, Format::SOPP, 0, 0)};
- nop->imm = NOPs.first - 1;
- nop->block = -1;
- instructions.emplace_back(std::move(nop));
+ NOP_ctx_gfx10 all_ctx[program->blocks.size()];
+ std::stack<unsigned> loop_header_indices;
+
+ for (unsigned i = 0; i < program->blocks.size(); i++) {
+ Block& block = program->blocks[i];
+ NOP_ctx_gfx10 &ctx = all_ctx[i];
+
+ if (block.kind & block_kind_loop_header) {
+ loop_header_indices.push(i);
+ } else if (block.kind & block_kind_loop_exit) {
+ /* Go through the whole loop again */
+ for (unsigned idx = loop_header_indices.top(); idx < i; idx++) {
+ NOP_ctx_gfx10 loop_block_ctx;
+ for (unsigned b : program->blocks[idx].linear_preds)
+ loop_block_ctx.join(all_ctx[b]);
+
+ handle_block_gfx10(loop_block_ctx, program->blocks[idx]);
+
+ /* We only need to continue if the loop header context changed */
+ if (idx == loop_header_indices.top() && loop_block_ctx == all_ctx[idx])
+ break;
+
+ all_ctx[idx] = loop_block_ctx;
+ }
+
+ loop_header_indices.pop();
}
- instructions.emplace_back(std::move(instr));
- }
+ for (unsigned b : block.linear_preds)
+ ctx.join(all_ctx[b]);
- block.instructions = std::move(instructions);
+ handle_block_gfx10(ctx, block);
+ }
}
} /* end namespace */
-
void insert_NOPs(Program* program)
{
- NOP_ctx ctx(program);
-
- for (Block& block : program->blocks) {
- if (block.instructions.empty())
- continue;
-
- if (ctx.chip_class >= GFX10)
- handle_block_gfx10(ctx, block);
- else
- handle_block(ctx, block);
- }
+ if (program->chip_class >= GFX10)
+ mitigate_hazards_gfx10(program);
+ else
+ insert_NOPs_gfx8_9(program);
}
}
projects / mesa.git / blobdiff