adds dynamic cycle estimation

This commit is contained in:
Eyck Jentzsch 2023-10-25 17:13:52 +02:00
parent b7478965ab
commit b86d7a517d
4 changed files with 685 additions and 53 deletions

View File

@ -27,84 +27,84 @@ RV32I:
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,1]
delay: [1,2]
BEQ:
index: 4
encoding: 0b00000000000000000000000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,1]
delay: [1,2]
BNE:
index: 5
encoding: 0b00000000000000000001000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,1]
delay: [1,2]
BLT:
index: 6
encoding: 0b00000000000000000100000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,1]
delay: [1,2]
BGE:
index: 7
encoding: 0b00000000000000000101000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,1]
delay: [1,2]
BLTU:
index: 8
encoding: 0b00000000000000000110000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,1]
delay: [1,2]
BGEU:
index: 9
encoding: 0b00000000000000000111000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,1]
delay: [1,2]
LB:
index: 10
encoding: 0b00000000000000000000000000000011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
delay: 2
LH:
index: 11
encoding: 0b00000000000000000001000000000011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
delay: 2
LW:
index: 12
encoding: 0b00000000000000000010000000000011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
delay: 2
LBU:
index: 13
encoding: 0b00000000000000000100000000000011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
delay: 2
LHU:
index: 14
encoding: 0b00000000000000000101000000000011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
delay: 2
SB:
index: 15
encoding: 0b00000000000000000000000000100011
@ -356,56 +356,56 @@ RV32M:
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
delay: 2
MULH:
index: 50
encoding: 0b00000010000000000001000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
delay: 2
MULHSU:
index: 51
encoding: 0b00000010000000000010000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
delay: 2
MULHU:
index: 52
encoding: 0b00000010000000000011000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
delay: 2
DIV:
index: 53
encoding: 0b00000010000000000100000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
delay: 17
DIVU:
index: 54
encoding: 0b00000010000000000101000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
delay: 17
REM:
index: 55
encoding: 0b00000010000000000110000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
delay: 17
REMU:
index: 56
encoding: 0b00000010000000000111000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
delay: 17
Zca:
C__ADDI4SPN:
index: 57
@ -420,7 +420,7 @@ Zca:
mask: 0b1110000000000011
size: 16
branch: false
delay: 1
delay: 2
C__SW:
index: 59
encoding: 0b1100000000000000
@ -542,14 +542,14 @@ Zca:
mask: 0b1110000000000011
size: 16
branch: true
delay: [1,1]
delay: [1,2]
C__BNEZ:
index: 76
encoding: 0b1110000000000001
mask: 0b1110000000000011
size: 16
branch: true
delay: [1,1]
delay: [1,2]
C__SLLI:
index: 77
encoding: 0b0000000000000010
@ -564,7 +564,7 @@ Zca:
mask: 0b1110000000000011
size: 16
branch: false
delay: 1
delay: 2
C__MV:
index: 79
encoding: 0b1000000000000010

View File

@ -0,0 +1,624 @@
RV32I:
LUI:
index: 0
encoding: 0b00000000000000000000000000110111
mask: 0b00000000000000000000000001111111
size: 32
branch: false
delay: 1
AUIPC:
index: 1
encoding: 0b00000000000000000000000000010111
mask: 0b00000000000000000000000001111111
size: 32
branch: false
delay: 1
JAL:
index: 2
encoding: 0b00000000000000000000000001101111
mask: 0b00000000000000000000000001111111
size: 32
branch: true
delay: 1
JALR:
index: 3
encoding: 0b00000000000000000000000001100111
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,2]
BEQ:
index: 4
encoding: 0b00000000000000000000000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,2]
BNE:
index: 5
encoding: 0b00000000000000000001000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,2]
BLT:
index: 6
encoding: 0b00000000000000000100000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,2]
BGE:
index: 7
encoding: 0b00000000000000000101000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,2]
BLTU:
index: 8
encoding: 0b00000000000000000110000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,2]
BGEU:
index: 9
encoding: 0b00000000000000000111000001100011
mask: 0b00000000000000000111000001111111
size: 32
branch: true
delay: [1,2]
LB:
index: 10
encoding: 0b00000000000000000000000000000011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 2
LH:
index: 11
encoding: 0b00000000000000000001000000000011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 2
LW:
index: 12
encoding: 0b00000000000000000010000000000011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 2
LBU:
index: 13
encoding: 0b00000000000000000100000000000011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 2
LHU:
index: 14
encoding: 0b00000000000000000101000000000011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 2
SB:
index: 15
encoding: 0b00000000000000000000000000100011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
SH:
index: 16
encoding: 0b00000000000000000001000000100011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
SW:
index: 17
encoding: 0b00000000000000000010000000100011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
ADDI:
index: 18
encoding: 0b00000000000000000000000000010011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
SLTI:
index: 19
encoding: 0b00000000000000000010000000010011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
SLTIU:
index: 20
encoding: 0b00000000000000000011000000010011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
XORI:
index: 21
encoding: 0b00000000000000000100000000010011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
ORI:
index: 22
encoding: 0b00000000000000000110000000010011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
ANDI:
index: 23
encoding: 0b00000000000000000111000000010011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
SLLI:
index: 24
encoding: 0b00000000000000000001000000010011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: "u_24:20"
SRLI:
index: 25
encoding: 0b00000000000000000101000000010011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: "u_24:20"
SRAI:
index: 26
encoding: 0b01000000000000000101000000010011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: "u_24:20"
ADD:
index: 27
encoding: 0b00000000000000000000000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
SUB:
index: 28
encoding: 0b01000000000000000000000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
SLL:
index: 29
encoding: 0b00000000000000000001000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: "X_24:20"
SLT:
index: 30
encoding: 0b00000000000000000010000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
SLTU:
index: 31
encoding: 0b00000000000000000011000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
XOR:
index: 32
encoding: 0b00000000000000000100000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
SRL:
index: 33
encoding: 0b00000000000000000101000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: "X_24:20"
SRA:
index: 34
encoding: 0b01000000000000000101000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: "X_24:20"
OR:
index: 35
encoding: 0b00000000000000000110000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
AND:
index: 36
encoding: 0b00000000000000000111000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 1
FENCE:
index: 37
encoding: 0b00000000000000000000000000001111
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
ECALL:
index: 38
encoding: 0b00000000000000000000000001110011
mask: 0b11111111111111111111111111111111
attributes: [[name:no_cont]]
size: 32
branch: false
delay: 1
EBREAK:
index: 39
encoding: 0b00000000000100000000000001110011
mask: 0b11111111111111111111111111111111
attributes: [[name:no_cont]]
size: 32
branch: false
delay: 1
MRET:
index: 40
encoding: 0b00110000001000000000000001110011
mask: 0b11111111111111111111111111111111
attributes: [[name:no_cont]]
size: 32
branch: false
delay: 1
WFI:
index: 41
encoding: 0b00010000010100000000000001110011
mask: 0b11111111111111111111111111111111
size: 32
branch: false
delay: 1
Zicsr:
CSRRW:
index: 42
encoding: 0b00000000000000000001000001110011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
CSRRS:
index: 43
encoding: 0b00000000000000000010000001110011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
CSRRC:
index: 44
encoding: 0b00000000000000000011000001110011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
CSRRWI:
index: 45
encoding: 0b00000000000000000101000001110011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
CSRRSI:
index: 46
encoding: 0b00000000000000000110000001110011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
CSRRCI:
index: 47
encoding: 0b00000000000000000111000001110011
mask: 0b00000000000000000111000001111111
size: 32
branch: false
delay: 1
Zifencei:
FENCE_I:
index: 48
encoding: 0b00000000000000000001000000001111
mask: 0b00000000000000000111000001111111
attributes: [[name:flush]]
size: 32
branch: false
delay: 1
RV32M:
MUL:
index: 49
encoding: 0b00000010000000000000000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 32
MULH:
index: 50
encoding: 0b00000010000000000001000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 32
MULHSU:
index: 51
encoding: 0b00000010000000000010000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 32
MULHU:
index: 52
encoding: 0b00000010000000000011000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 32
DIV:
index: 53
encoding: 0b00000010000000000100000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 33
DIVU:
index: 54
encoding: 0b00000010000000000101000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 33
REM:
index: 55
encoding: 0b00000010000000000110000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 33
REMU:
index: 56
encoding: 0b00000010000000000111000000110011
mask: 0b11111110000000000111000001111111
size: 32
branch: false
delay: 33
Zca:
C__ADDI4SPN:
index: 57
encoding: 0b0000000000000000
mask: 0b1110000000000011
size: 16
branch: false
delay: 1
C__LW:
index: 58
encoding: 0b0100000000000000
mask: 0b1110000000000011
size: 16
branch: false
delay: 2
C__SW:
index: 59
encoding: 0b1100000000000000
mask: 0b1110000000000011
size: 16
branch: false
delay: 1
C__ADDI:
index: 60
encoding: 0b0000000000000001
mask: 0b1110000000000011
size: 16
branch: false
delay: 1
C__NOP:
index: 61
encoding: 0b0000000000000001
mask: 0b1110111110000011
size: 16
branch: false
delay: 1
C__JAL:
index: 62
encoding: 0b0010000000000001
mask: 0b1110000000000011
attributes: [[name:enable, value:1]]
size: 16
branch: true
delay: 1
C__LI:
index: 63
encoding: 0b0100000000000001
mask: 0b1110000000000011
size: 16
branch: false
delay: 1
C__LUI:
index: 64
encoding: 0b0110000000000001
mask: 0b1110000000000011
size: 16
branch: false
delay: 1
C__ADDI16SP:
index: 65
encoding: 0b0110000100000001
mask: 0b1110111110000011
size: 16
branch: false
delay: 1
__reserved_clui:
index: 66
encoding: 0b0110000000000001
mask: 0b1111000001111111
size: 16
branch: false
delay: 1
C__SRLI:
index: 67
encoding: 0b1000000000000001
mask: 0b1111110000000011
attributes: [[name:enable, value:1]]
size: 16
branch: false
delay: "u_12:12*16+u_6:2"
C__SRAI:
index: 68
encoding: 0b1000010000000001
mask: 0b1111110000000011
attributes: [[name:enable, value:1]]
size: 16
branch: false
delay: "u_12:12*16+u_6:2"
C__ANDI:
index: 69
encoding: 0b1000100000000001
mask: 0b1110110000000011
size: 16
branch: false
delay: 1
C__SUB:
index: 70
encoding: 0b1000110000000001
mask: 0b1111110001100011
size: 16
branch: false
delay: 1
C__XOR:
index: 71
encoding: 0b1000110000100001
mask: 0b1111110001100011
size: 16
branch: false
delay: 1
C__OR:
index: 72
encoding: 0b1000110001000001
mask: 0b1111110001100011
size: 16
branch: false
delay: 1
C__AND:
index: 73
encoding: 0b1000110001100001
mask: 0b1111110001100011
size: 16
branch: false
delay: 1
C__J:
index: 74
encoding: 0b1010000000000001
mask: 0b1110000000000011
size: 16
branch: true
delay: 1
C__BEQZ:
index: 75
encoding: 0b1100000000000001
mask: 0b1110000000000011
size: 16
branch: true
delay: [1,2]
C__BNEZ:
index: 76
encoding: 0b1110000000000001
mask: 0b1110000000000011
size: 16
branch: true
delay: [1,2]
C__SLLI:
index: 77
encoding: 0b0000000000000010
mask: 0b1111000000000011
attributes: [[name:enable, value:1]]
size: 16
branch: false
delay: "u_12:12*16+u_6:2"
C__LWSP:
index: 78
encoding: 0b0100000000000010
mask: 0b1110000000000011
size: 16
branch: false
delay: 2
C__MV:
index: 79
encoding: 0b1000000000000010
mask: 0b1111000000000011
size: 16
branch: false
delay: 1
C__JR:
index: 80
encoding: 0b1000000000000010
mask: 0b1111000001111111
size: 16
branch: true
delay: 1
__reserved_cmv:
index: 81
encoding: 0b1000000000000010
mask: 0b1111111111111111
size: 16
branch: false
delay: 1
C__ADD:
index: 82
encoding: 0b1001000000000010
mask: 0b1111000000000011
size: 16
branch: false
delay: 1
C__JALR:
index: 83
encoding: 0b1001000000000010
mask: 0b1111000001111111
size: 16
branch: true
delay: 1
C__EBREAK:
index: 84
encoding: 0b1001000000000010
mask: 0b1111111111111111
size: 16
branch: false
delay: 1
C__SWSP:
index: 85
encoding: 0b1100000000000010
mask: 0b1110000000000011
size: 16
branch: false
delay: 1
DII:
index: 86
encoding: 0b0000000000000000
mask: 0b1111111111111111
size: 16
branch: false
delay: 1

View File

@ -1,4 +1,3 @@
/*******************************************************************************
* Copyright (C) 2017 - 2023, MINRES Technologies GmbH
* All rights reserved.
@ -34,6 +33,7 @@
******************************************************************************/
#include "cycle_estimate.h"
#include <iss/plugin/calculator.h>
#include <yaml-cpp/yaml.h>
#include <iss/arch_if.h>
@ -48,11 +48,12 @@ iss::plugin::cycle_estimate::cycle_estimate(string const& config_file_name)
{
}
iss::plugin::cycle_estimate::~cycle_estimate() {
}
iss::plugin::cycle_estimate::~cycle_estimate() = default;
bool iss::plugin::cycle_estimate::registration(const char* const version, vm_if& vm) {
instr_if = vm.get_arch()->get_instrumentation_if();
assert(instr_if && "No instrumentation interface available but callback executed");
reg_base_ptr = reinterpret_cast<uint32_t*>(vm.get_arch()->get_regs_base_ptr());
if(!instr_if) return false;
const string core_name = instr_if->core_type_name();
if (config_file_name.length() > 0) {
@ -67,17 +68,23 @@ bool iss::plugin::cycle_estimate::registration(const char* const version, vm_if&
auto isa_subset = p.first;
auto instructions = p.second;
for (auto const& instr : instructions) {
instr_desc res;
auto idx = instr.second["index"].as<unsigned>();
if(delays.size()<=idx)
delays.resize(idx+1);
auto& res = delays[idx];
res.is_branch = instr.second["branch"].as<bool>();
auto delay = instr.second["delay"];
if(delay.IsSequence()) {
res.not_taken = delay[0].as<uint64_t>();
res.taken = delay[1].as<uint64_t>();
} else {
res.not_taken = delay.as<uint64_t>();
res.taken = res.not_taken;
try {
res.not_taken = delay.as<uint64_t>();
res.taken = res.not_taken;
} catch (const YAML::BadConversion& e) {
res.f = iss::plugin::calculator(reg_base_ptr, delay.as<std::string>());
}
}
delays.push_back(std::move(res));
}
}
} catch (YAML::ParserException &e) {
@ -90,15 +97,18 @@ bool iss::plugin::cycle_estimate::registration(const char* const version, vm_if&
}
}
return true;
}
void iss::plugin::cycle_estimate::callback(instr_info_t instr_info) {
assert(instr_if && "No instrumentation interface available but callback executed");
auto entry = delays[instr_info.instr_id];
bool taken = instr_if->is_branch_taken();
if (taken && (entry.taken > 1))
instr_if->update_last_instr_cycles(entry.taken);
else if (entry.not_taken > 1)
instr_if->update_last_instr_cycles(entry.not_taken);
size_t instr_id = instr_info.instr_id;
auto entry = delays[instr_id];
if(instr_info.phase_id==PRE_SYNC) {
if(entry.f)
current_delay = entry.f(instr_if->get_instr_word());
} else {
if(!entry.f)
current_delay = instr_if->is_branch_taken()? entry.taken: entry.not_taken;
if(current_delay>1)
instr_if->update_last_instr_cycles(current_delay);
}
}

View File

@ -40,24 +40,20 @@
#include <string>
#include <unordered_map>
#include <vector>
#include <functional>
namespace iss {
namespace plugin {
class cycle_estimate: public vm_plugin {
BEGIN_BF_DECL(instr_desc, uint32_t)
BF_FIELD(taken, 24, 8)
BF_FIELD(not_taken, 16, 8)
BF_FIELD(is_branch, 8, 8)
BF_FIELD(size, 0, 8)
instr_desc(uint32_t size, uint32_t taken, uint32_t not_taken, bool branch): instr_desc() {
this->size=size;
this->taken=taken;
this->not_taken=not_taken;
this->is_branch=branch;
}
END_BF_DECL();
struct instr_desc {
size_t size;
bool is_branch;
unsigned not_taken;
unsigned taken;
std::function<unsigned(uint64_t)> f;
};
public:
cycle_estimate() = delete;
@ -76,13 +72,15 @@ public:
bool registration(const char *const version, vm_if &arch) override;
sync_type get_sync() override { return POST_SYNC; };
sync_type get_sync() override { return ALL_SYNC; };
void callback(instr_info_t instr_info) override;
private:
iss::instrumentation_if *instr_if;
iss::instrumentation_if *instr_if{nullptr};
uint32_t* reg_base_ptr {nullptr};
std::vector<instr_desc> delays;
unsigned current_delay{0};
struct pair_hash {
size_t operator()(const std::pair<uint64_t, uint64_t> &p) const {
std::hash<uint64_t> hash;