fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-01 08:08:06 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

aco: Add helper to get cycle info for an instruction.

Browse source 
For use in s_delay_alu tracking

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19743>
This commit is contained in:
Bas Nieuwenhuizen 2022-11-14 18:58:32 +00:00 • committed by Marge Bot
parent 352e492c7b
commit cd3bf56ace
2 changed files with 30 additions and 12 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

11
src/amd/compiler/aco_ir.h
View file

@ -2313,6 +2313,17 @@ void collect_presched_stats(Program* program);
void collect_preasm_stats(Program* program);
void collect_postasm_stats(Program* program, const std::vector<uint32_t>& code);
struct Instruction_cycle_info {
/* Latency until the result is ready (if not needing a waitcnt) */
unsigned latency;
/* How many cycles issuing this instruction takes (i.e. cycles till the next instruction can be
* issued)*/
unsigned issue_cycles;
};
Instruction_cycle_info get_cycle_info(const Program& program, const Instruction& instr);
enum print_flags {
print_no_ssa = 0x1,
print_perf_info = 0x2,

31
src/amd/compiler/aco_statistics.cpp
View file

@ -132,14 +132,14 @@ is_dual_issue_capable(const Program& program, const Instruction& instruction)
}
static perf_info
get_perf_info(Program* program, aco_ptr<Instruction>& instr)
get_perf_info(const Program& program, const Instruction& instr)
{
instr_class cls = instr_info.classes[(int)instr->opcode];
instr_class cls = instr_info.classes[(int)instr.opcode];
#define WAIT(res) BlockCycleEstimator::res, 0
#define WAIT_USE(res, cnt) BlockCycleEstimator::res, cnt
if (program->gfx_level >= GFX10) {
if (program.gfx_level >= GFX10) {
/* fp64 might be incorrect */
switch (cls) {
case instr_class::valu32:
@ -162,8 +162,8 @@ get_perf_info(Program* program, aco_ptr<Instruction>& instr)
case instr_class::branch:
case instr_class::sendmsg: return {0, WAIT_USE(branch_sendmsg, 1)};
case instr_class::ds:
return instr->isDS() && instr->ds().gds ? perf_info{0, WAIT_USE(export_gds, 1)}
: perf_info{0, WAIT_USE(lds, 1)};
return instr.isDS() && instr.ds().gds ? perf_info{0, WAIT_USE(export_gds, 1)}
: perf_info{0, WAIT_USE(lds, 1)};
case instr_class::exp: return {0, WAIT_USE(export_gds, 1)};
case instr_class::vmem: return {0, WAIT_USE(vmem, 1)};
case instr_class::barrier:
@ -178,8 +178,8 @@ get_perf_info(Program* program, aco_ptr<Instruction>& instr)
case instr_class::valu64: return {8, WAIT_USE(valu, 8)};
case instr_class::valu_quarter_rate32: return {16, WAIT_USE(valu, 16)};
case instr_class::valu_fma:
return program->dev.has_fast_fma32 ? perf_info{4, WAIT_USE(valu, 4)}
: perf_info{16, WAIT_USE(valu, 16)};
return program.dev.has_fast_fma32 ? perf_info{4, WAIT_USE(valu, 4)}
: perf_info{16, WAIT_USE(valu, 16)};
case instr_class::valu_transcendental32: return {16, WAIT_USE(valu, 16)};
case instr_class::valu_double: return {64, WAIT_USE(valu, 64)};
case instr_class::valu_double_add: return {32, WAIT_USE(valu, 32)};
@ -191,8 +191,8 @@ get_perf_info(Program* program, aco_ptr<Instruction>& instr)
return {8, WAIT_USE(branch_sendmsg, 8)};
return {4, WAIT_USE(branch_sendmsg, 4)};
case instr_class::ds:
return instr->isDS() && instr->ds().gds ? perf_info{4, WAIT_USE(export_gds, 4)}
: perf_info{4, WAIT_USE(lds, 4)};
return instr.isDS() && instr.ds().gds ? perf_info{4, WAIT_USE(export_gds, 4)}
: perf_info{4, WAIT_USE(lds, 4)};
case instr_class::exp: return {16, WAIT_USE(export_gds, 16)};
case instr_class::vmem: return {4, WAIT_USE(vmem, 4)};
case instr_class::barrier:
@ -209,7 +209,7 @@ get_perf_info(Program* program, aco_ptr<Instruction>& instr)
void
BlockCycleEstimator::use_resources(aco_ptr<Instruction>& instr)
{
perf_info perf = get_perf_info(program, instr);
perf_info perf = get_perf_info(*program, *instr);
if (perf.rsrc0 != resource_count) {
res_available[(int)perf.rsrc0] = cur_cycle + perf.cost0;
@ -225,7 +225,7 @@ BlockCycleEstimator::use_resources(aco_ptr<Instruction>& instr)
int32_t
BlockCycleEstimator::cycles_until_res_available(aco_ptr<Instruction>& instr)
{
perf_info perf = get_perf_info(program, instr);
perf_info perf = get_perf_info(*program, *instr);
int32_t cost = 0;
if (perf.rsrc0 != resource_count)
@ -379,7 +379,7 @@ is_vector(aco_opcode op)
void
BlockCycleEstimator::add(aco_ptr<Instruction>& instr)
{
perf_info perf = get_perf_info(program, instr);
perf_info perf = get_perf_info(*program, *instr);
cur_cycle += get_dependency_cost(instr);
@ -602,4 +602,11 @@ collect_postasm_stats(Program* program, const std::vector<uint32_t>& code)
program->statistics[aco_statistic_hash] = util_hash_crc32(code.data(), code.size() * 4);
}
Instruction_cycle_info
get_cycle_info(const Program& program, const Instruction& instr)
{
perf_info info = get_perf_info(program, instr);
return Instruction_cycle_info{(unsigned)info.latency, std::max(info.cost0, info.cost1)};
}
} // namespace aco