forked from wownero/wownero
Merge pull request #168 from wowario/cnr
Cryptonight variant 4 aka CryptonightR #5126release-v0.5.0
commit
ca9ae12d5b
@ -0,0 +1,412 @@
|
||||
#ifndef VARIANT4_RANDOM_MATH_H
|
||||
#define VARIANT4_RANDOM_MATH_H
|
||||
|
||||
// Register size can be configured to either 32 bit (uint32_t) or 64 bit (uint64_t)
|
||||
typedef uint32_t v4_reg;
|
||||
|
||||
enum V4_Settings
|
||||
{
|
||||
// Generate code with minimal theoretical latency = 45 cycles, which is equivalent to 15 multiplications
|
||||
TOTAL_LATENCY = 15 * 3,
|
||||
|
||||
// Always generate at least 60 instructions
|
||||
NUM_INSTRUCTIONS = 60,
|
||||
|
||||
// Available ALUs for MUL
|
||||
// Modern CPUs typically have only 1 ALU which can do multiplications
|
||||
ALU_COUNT_MUL = 1,
|
||||
|
||||
// Total available ALUs
|
||||
// Modern CPUs have 4 ALUs, but we use only 3 because random math executes together with other main loop code
|
||||
ALU_COUNT = 3,
|
||||
};
|
||||
|
||||
enum V4_InstructionList
|
||||
{
|
||||
MUL, // a*b
|
||||
ADD, // a+b + C, -128 <= C <= 127
|
||||
SUB, // a-b
|
||||
ROR, // rotate right "a" by "b & 31" bits
|
||||
ROL, // rotate left "a" by "b & 31" bits
|
||||
XOR, // a^b
|
||||
RET, // finish execution
|
||||
V4_INSTRUCTION_COUNT = RET,
|
||||
};
|
||||
|
||||
// V4_InstructionDefinition is used to generate code from random data
|
||||
// Every random sequence of bytes is a valid code
|
||||
//
|
||||
// There are 8 registers in total:
|
||||
// - 4 variable registers
|
||||
// - 4 constant registers initialized from loop variables
|
||||
//
|
||||
// This is why dst_index is 2 bits
|
||||
enum V4_InstructionDefinition
|
||||
{
|
||||
V4_OPCODE_BITS = 3,
|
||||
V4_DST_INDEX_BITS = 2,
|
||||
V4_SRC_INDEX_BITS = 3,
|
||||
};
|
||||
|
||||
struct V4_Instruction
|
||||
{
|
||||
uint8_t opcode;
|
||||
uint8_t dst_index;
|
||||
uint8_t src_index;
|
||||
uint32_t C;
|
||||
};
|
||||
|
||||
#ifndef FORCEINLINE
|
||||
#ifdef __GNUC__
|
||||
#define FORCEINLINE __attribute__((always_inline)) inline
|
||||
#elif _MSC_VER
|
||||
#define FORCEINLINE __forceinline
|
||||
#else
|
||||
#define FORCEINLINE inline
|
||||
#endif
|
||||
#endif
|
||||
|
||||
#ifndef UNREACHABLE_CODE
|
||||
#ifdef __GNUC__
|
||||
#define UNREACHABLE_CODE __builtin_unreachable()
|
||||
#elif _MSC_VER
|
||||
#define UNREACHABLE_CODE __assume(false)
|
||||
#else
|
||||
#define UNREACHABLE_CODE
|
||||
#endif
|
||||
#endif
|
||||
|
||||
// Random math interpreter's loop is fully unrolled and inlined to achieve 100% branch prediction on CPU:
|
||||
// every switch-case will point to the same destination on every iteration of Cryptonight main loop
|
||||
//
|
||||
// This is about as fast as it can get without using low-level machine code generation
|
||||
static FORCEINLINE void v4_random_math(const struct V4_Instruction* code, v4_reg* r)
|
||||
{
|
||||
enum
|
||||
{
|
||||
REG_BITS = sizeof(v4_reg) * 8,
|
||||
};
|
||||
|
||||
#define V4_EXEC(i) \
|
||||
{ \
|
||||
const struct V4_Instruction* op = code + i; \
|
||||
const v4_reg src = r[op->src_index]; \
|
||||
v4_reg* dst = r + op->dst_index; \
|
||||
switch (op->opcode) \
|
||||
{ \
|
||||
case MUL: \
|
||||
*dst *= src; \
|
||||
break; \
|
||||
case ADD: \
|
||||
*dst += src + op->C; \
|
||||
break; \
|
||||
case SUB: \
|
||||
*dst -= src; \
|
||||
break; \
|
||||
case ROR: \
|
||||
{ \
|
||||
const uint32_t shift = src % REG_BITS; \
|
||||
*dst = (*dst >> shift) | (*dst << (REG_BITS - shift)); \
|
||||
} \
|
||||
break; \
|
||||
case ROL: \
|
||||
{ \
|
||||
const uint32_t shift = src % REG_BITS; \
|
||||
*dst = (*dst << shift) | (*dst >> (REG_BITS - shift)); \
|
||||
} \
|
||||
break; \
|
||||
case XOR: \
|
||||
*dst ^= src; \
|
||||
break; \
|
||||
case RET: \
|
||||
return; \
|
||||
default: \
|
||||
UNREACHABLE_CODE; \
|
||||
break; \
|
||||
} \
|
||||
}
|
||||
|
||||
#define V4_EXEC_10(j) \
|
||||
V4_EXEC(j + 0) \
|
||||
V4_EXEC(j + 1) \
|
||||
V4_EXEC(j + 2) \
|
||||
V4_EXEC(j + 3) \
|
||||
V4_EXEC(j + 4) \
|
||||
V4_EXEC(j + 5) \
|
||||
V4_EXEC(j + 6) \
|
||||
V4_EXEC(j + 7) \
|
||||
V4_EXEC(j + 8) \
|
||||
V4_EXEC(j + 9)
|
||||
|
||||
// Generated program can have 60 + a few more (usually 2-3) instructions to achieve required latency
|
||||
// I've checked all block heights < 10,000,000 and here is the distribution of program sizes:
|
||||
//
|
||||
// 60 28495
|
||||
// 61 106077
|
||||
// 62 2455855
|
||||
// 63 5114930
|
||||
// 64 1020868
|
||||
// 65 1109026
|
||||
// 66 151756
|
||||
// 67 8429
|
||||
// 68 4477
|
||||
// 69 87
|
||||
|
||||
// Unroll 70 instructions here
|
||||
V4_EXEC_10(0); // instructions 0-9
|
||||
V4_EXEC_10(10); // instructions 10-19
|
||||
V4_EXEC_10(20); // instructions 20-29
|
||||
V4_EXEC_10(30); // instructions 30-39
|
||||
V4_EXEC_10(40); // instructions 40-49
|
||||
V4_EXEC_10(50); // instructions 50-59
|
||||
V4_EXEC_10(60); // instructions 60-69
|
||||
|
||||
#undef V4_EXEC_10
|
||||
#undef V4_EXEC
|
||||
}
|
||||
|
||||
// If we don't have enough data available, generate more
|
||||
static FORCEINLINE void check_data(size_t* data_index, const size_t bytes_needed, char* data, const size_t data_size)
|
||||
{
|
||||
if (*data_index + bytes_needed > data_size)
|
||||
{
|
||||
hash_extra_blake(data, data_size, data);
|
||||
*data_index = 0;
|
||||
}
|
||||
}
|
||||
|
||||
// Generates as many random math operations as possible with given latency and ALU restrictions
|
||||
static inline int v4_random_math_init(struct V4_Instruction* code, const uint64_t height)
|
||||
{
|
||||
// MUL is 3 cycles, 3-way addition and rotations are 2 cycles, SUB/XOR are 1 cycle
|
||||
// These latencies match real-life instruction latencies for Intel CPUs starting from Sandy Bridge and up to Skylake/Coffee lake
|
||||
//
|
||||
// AMD Ryzen has the same latencies except 1-cycle ROR/ROL, so it'll be a bit faster than Intel Sandy Bridge and newer processors
|
||||
// Surprisingly, Intel Nehalem also has 1-cycle ROR/ROL, so it'll also be faster than Intel Sandy Bridge and newer processors
|
||||
// AMD Bulldozer has 4 cycles latency for MUL (slower than Intel) and 1 cycle for ROR/ROL (faster than Intel), so average performance will be the same
|
||||
// Source: https://www.agner.org/optimize/instruction_tables.pdf
|
||||
const int op_latency[V4_INSTRUCTION_COUNT] = { 3, 2, 1, 2, 2, 1 };
|
||||
|
||||
// Instruction latencies for theoretical ASIC implementation
|
||||
const int asic_op_latency[V4_INSTRUCTION_COUNT] = { 3, 1, 1, 1, 1, 1 };
|
||||
|
||||
// Available ALUs for each instruction
|
||||
const int op_ALUs[V4_INSTRUCTION_COUNT] = { ALU_COUNT_MUL, ALU_COUNT, ALU_COUNT, ALU_COUNT, ALU_COUNT, ALU_COUNT };
|
||||
|
||||
char data[32];
|
||||
memset(data, 0, sizeof(data));
|
||||
*((uint64_t*)data) = height;
|
||||
|
||||
size_t data_index = sizeof(data);
|
||||
|
||||
int code_size;
|
||||
do {
|
||||
int latency[8];
|
||||
int asic_latency[8];
|
||||
|
||||
// Tracks previous instruction and value of the source operand for registers R0-R3 throughout code execution
|
||||
// byte 0: current value of the destination register
|
||||
// byte 1: instruction opcode
|
||||
// byte 2: current value of the source register
|
||||
//
|
||||
// Registers R4-R7 are constant and are threatened as having the same value because when we do
|
||||
// the same operation twice with two constant source registers, it can be optimized into a single operation
|
||||
int inst_data[8] = { 0, 1, 2, 3, -1, -1, -1, -1 };
|
||||
|
||||
bool alu_busy[TOTAL_LATENCY + 1][ALU_COUNT];
|
||||
bool is_rotation[V4_INSTRUCTION_COUNT];
|
||||
bool rotated[4];
|
||||
int rotate_count = 0;
|
||||
|
||||
memset(latency, 0, sizeof(latency));
|
||||
memset(asic_latency, 0, sizeof(asic_latency));
|
||||
memset(alu_busy, 0, sizeof(alu_busy));
|
||||
memset(is_rotation, 0, sizeof(is_rotation));
|
||||
memset(rotated, 0, sizeof(rotated));
|
||||
is_rotation[ROR] = true;
|
||||
is_rotation[ROL] = true;
|
||||
|
||||
int num_retries = 0;
|
||||
code_size = 0;
|
||||
|
||||
// Generate random code to achieve minimal required latency for our abstract CPU
|
||||
// Try to get this latency for all 4 registers
|
||||
while (((latency[0] < TOTAL_LATENCY) || (latency[1] < TOTAL_LATENCY) || (latency[2] < TOTAL_LATENCY) || (latency[3] < TOTAL_LATENCY)) && (num_retries < 64))
|
||||
{
|
||||
check_data(&data_index, 1, data, sizeof(data));
|
||||
|
||||
const uint8_t c = ((uint8_t*)data)[data_index++];
|
||||
|
||||
// MUL = opcodes 0-2
|
||||
// ADD = opcode 3
|
||||
// SUB = opcode 4
|
||||
// ROR/ROL = opcode 5, shift direction is selected randomly
|
||||
// XOR = opcodes 6-7
|
||||
uint8_t opcode = c & ((1 << V4_OPCODE_BITS) - 1);
|
||||
if (opcode == 5)
|
||||
{
|
||||
check_data(&data_index, 1, data, sizeof(data));
|
||||
opcode = (data[data_index++] >= 0) ? ROR : ROL;
|
||||
}
|
||||
else if (opcode >= 6)
|
||||
{
|
||||
opcode = XOR;
|
||||
}
|
||||
else
|
||||
{
|
||||
opcode = (opcode <= 2) ? MUL : (opcode - 2);
|
||||
}
|
||||
|
||||
uint8_t dst_index = (c >> V4_OPCODE_BITS) & ((1 << V4_DST_INDEX_BITS) - 1);
|
||||
uint8_t src_index = (c >> (V4_OPCODE_BITS + V4_DST_INDEX_BITS)) & ((1 << V4_SRC_INDEX_BITS) - 1);
|
||||
|
||||
const int a = dst_index;
|
||||
int b = src_index;
|
||||
|
||||
// Don't do ADD/SUB/XOR with the same register
|
||||
if (((opcode == ADD) || (opcode == SUB) || (opcode == XOR)) && (a == b))
|
||||
{
|
||||
// a is always < 4, so we don't need to check bounds here
|
||||
b = a + 4;
|
||||
src_index = b;
|
||||
}
|
||||
|
||||
// Don't do rotation with the same destination twice because it's equal to a single rotation
|
||||
if (is_rotation[opcode] && rotated[a])
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
// Don't do the same instruction (except MUL) with the same source value twice because all other cases can be optimized:
|
||||
// 2xADD(a, b, C) = ADD(a, b*2, C1+C2), same for SUB and rotations
|
||||
// 2xXOR(a, b) = NOP
|
||||
if ((opcode != MUL) && ((inst_data[a] & 0xFFFF00) == (opcode << 8) + ((inst_data[b] & 255) << 16)))
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
// Find which ALU is available (and when) for this instruction
|
||||
int next_latency = (latency[a] > latency[b]) ? latency[a] : latency[b];
|
||||
int alu_index = -1;
|
||||
while (next_latency < TOTAL_LATENCY)
|
||||
{
|
||||
for (int i = op_ALUs[opcode] - 1; i >= 0; --i)
|
||||
{
|
||||
if (!alu_busy[next_latency][i])
|
||||
{
|
||||
// ADD is implemented as two 1-cycle instructions on a real CPU, so do an additional availability check
|
||||
if ((opcode == ADD) && alu_busy[next_latency + 1][i])
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
// Rotation can only start when previous rotation is finished, so do an additional availability check
|
||||
if (is_rotation[opcode] && (next_latency < rotate_count * op_latency[opcode]))
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
alu_index = i;
|
||||
break;
|
||||
}
|
||||
}
|
||||
if (alu_index >= 0)
|
||||
{
|
||||
break;
|
||||
}
|
||||
++next_latency;
|
||||
}
|
||||
|
||||
// Don't generate instructions that leave some register unchanged for more than 7 cycles
|
||||
if (next_latency > latency[a] + 7)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
next_latency += op_latency[opcode];
|
||||
|
||||
if (next_latency <= TOTAL_LATENCY)
|
||||
{
|
||||
if (is_rotation[opcode])
|
||||
{
|
||||
++rotate_count;
|
||||
}
|
||||
|
||||
// Mark ALU as busy only for the first cycle when it starts executing the instruction because ALUs are fully pipelined
|
||||
alu_busy[next_latency - op_latency[opcode]][alu_index] = true;
|
||||
latency[a] = next_latency;
|
||||
|
||||
// ASIC is supposed to have enough ALUs to run as many independent instructions per cycle as possible, so latency calculation for ASIC is simple
|
||||
asic_latency[a] = ((asic_latency[a] > asic_latency[b]) ? asic_latency[a] : asic_latency[b]) + asic_op_latency[opcode];
|
||||
|
||||
rotated[a] = is_rotation[opcode];
|
||||
|
||||
inst_data[a] = code_size + (opcode << 8) + ((inst_data[b] & 255) << 16);
|
||||
|
||||
code[code_size].opcode = opcode;
|
||||
code[code_size].dst_index = dst_index;
|
||||
code[code_size].src_index = src_index;
|
||||
code[code_size].C = 0;
|
||||
|
||||
if (opcode == ADD)
|
||||
{
|
||||
// ADD instruction is implemented as two 1-cycle instructions on a real CPU, so mark ALU as busy for the next cycle too
|
||||
alu_busy[next_latency - op_latency[opcode] + 1][alu_index] = true;
|
||||
|
||||
// ADD instruction requires 4 more random bytes for 32-bit constant "C" in "a = a + b + C"
|
||||
check_data(&data_index, sizeof(uint32_t), data, sizeof(data));
|
||||
code[code_size].C = *((uint32_t*)&data[data_index]);
|
||||
data_index += sizeof(uint32_t);
|
||||
}
|
||||
|
||||
++code_size;
|
||||
if (code_size >= NUM_INSTRUCTIONS)
|
||||
{
|
||||
break;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
++num_retries;
|
||||
}
|
||||
}
|
||||
|
||||
// ASIC has more execution resources and can extract as much parallelism from the code as possible
|
||||
// We need to add a few more MUL and ROR instructions to achieve minimal required latency for ASIC
|
||||
// Get this latency for at least 1 of the 4 registers
|
||||
const int prev_code_size = code_size;
|
||||
while ((asic_latency[0] < TOTAL_LATENCY) && (asic_latency[1] < TOTAL_LATENCY) && (asic_latency[2] < TOTAL_LATENCY) && (asic_latency[3] < TOTAL_LATENCY))
|
||||
{
|
||||
int min_idx = 0;
|
||||
int max_idx = 0;
|
||||
for (int i = 1; i < 4; ++i)
|
||||
{
|
||||
if (asic_latency[i] < asic_latency[min_idx]) min_idx = i;
|
||||
if (asic_latency[i] > asic_latency[max_idx]) max_idx = i;
|
||||
}
|
||||
|
||||
const uint8_t pattern[3] = { ROR, MUL, MUL };
|
||||
const uint8_t opcode = pattern[(code_size - prev_code_size) % 3];
|
||||
latency[min_idx] = latency[max_idx] + op_latency[opcode];
|
||||
asic_latency[min_idx] = asic_latency[max_idx] + asic_op_latency[opcode];
|
||||
|
||||
code[code_size].opcode = opcode;
|
||||
code[code_size].dst_index = min_idx;
|
||||
code[code_size].src_index = max_idx;
|
||||
code[code_size].C = 0;
|
||||
++code_size;
|
||||
}
|
||||
|
||||
// There is ~99.8% chance that code_size >= NUM_INSTRUCTIONS here, so second iteration is required rarely
|
||||
} while (code_size < NUM_INSTRUCTIONS);
|
||||
|
||||
// Add final instruction to stop the interpreter
|
||||
code[code_size].opcode = RET;
|
||||
code[code_size].dst_index = 0;
|
||||
code[code_size].src_index = 0;
|
||||
code[code_size].C = 0;
|
||||
|
||||
return code_size;
|
||||
}
|
||||
|
||||
#endif
|
@ -0,0 +1,10 @@
|
||||
9d47bf4c41b7e8e727e681715acb47fa1677cdba9ca7bcb05ad8cc8abd5daa66 5468697320697320612074657374205468697320697320612074657374205468697320697320612074657374 1806260
|
||||
0d4a495cb844a3ca8ba4edb8e6bcf829ef1c06d9cdea2b62ca46c2a21b8b0a79 4c6f72656d20697073756d20646f6c6f722073697420616d65742c20636f6e73656374657475722061646970697363696e67 1806261
|
||||
a1d6d848b5c5915fccd2f64cf216c6b1a02cf7c77bc80d8d4e51b419e88ff0dd 656c69742c2073656420646f20656975736d6f642074656d706f7220696e6369646964756e74207574206c61626f7265 1806262
|
||||
af3a8544a0221a148c2ac90484b19861e3afca33fe17021efb8ad6496b567915 657420646f6c6f7265206d61676e6120616c697175612e20557420656e696d206164206d696e696d2076656e69616d2c 1806263
|
||||
313399e0963ae8a99dab8af66d343e097dae0c0feb08dbc43ccdafef5515f413 71756973206e6f737472756420657865726369746174696f6e20756c6c616d636f206c61626f726973206e697369 1806264
|
||||
6021c6ef90bff9ae94a7506d623d3a7a86c1756d655f50dd558f716d64622a34 757420616c697175697020657820656120636f6d6d6f646f20636f6e7365717561742e20447569732061757465 1806265
|
||||
2b13000535f3db5f9b9b84a65c4351f386cd2cdedebb8c3ad2eab086e6a3fee5 697275726520646f6c6f7220696e20726570726568656e646572697420696e20766f6c7570746174652076656c6974 1806266
|
||||
fc0e1dad8e895749dc90eb690bc1ba059a1cd772afaaf65a106bf9e5e6b80503 657373652063696c6c756d20646f6c6f726520657520667567696174206e756c6c612070617269617475722e 1806267
|
||||
b60b0afe144deff7d903ed2d5545e77ebe66a3c51fee7016eeb8fee9eb630c0f 4578636570746575722073696e74206f6363616563617420637570696461746174206e6f6e2070726f6964656e742c 1806268
|
||||
64774b27e7d5fec862fc4c0c13ac6bf09123b6f05bb0e4b75c97f379a2b3a679 73756e7420696e2063756c706120717569206f666669636961206465736572756e74206d6f6c6c697420616e696d20696420657374206c61626f72756d2e 1806269
|
Loading…
Reference in new issue