OpenCL GPU miner
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

679 lines
20 KiB

/*
* Copyright 2012 Andrew Smith
* Copyright 2012 Luke Dashjr
*
* This program is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License as published by the Free
* Software Foundation; either version 3 of the License, or (at your option)
* any later version. See COPYING for more details.
*/
#include "config.h"
#include <stdarg.h>
#include <stdio.h>
#include <unistd.h>
#include <math.h>
#include "logging.h"
#include "miner.h"
#include "fpgautils.h"
#include "util.h"
#define BITSTREAM_FILENAME "fpgaminer_top_fixed7_197MHz.ncd"
#define BISTREAM_USER_ID "\2\4$B"
#define MODMINER_CUTOFF_TEMP 60.0
#define MODMINER_OVERHEAT_TEMP 50.0
#define MODMINER_OVERHEAT_CLOCK -10
#define MODMINER_HW_ERROR_PERCENT 0.75
#define MODMINER_MAX_CLOCK 220
#define MODMINER_DEF_CLOCK 200
#define MODMINER_MIN_CLOCK 160
#define MODMINER_CLOCK_DOWN -2
#define MODMINER_CLOCK_SET 0
#define MODMINER_CLOCK_UP 2
// Maximum how many good shares in a row means clock up
// 96 is ~34m22s at 200MH/s
#define MODMINER_TRY_UP 96
// Initially how many good shares in a row means clock up
// This is doubled each down clock until it reaches MODMINER_TRY_UP
// 6 is ~2m9s at 200MH/s
#define MODMINER_EARLY_UP 6
struct device_api modminer_api;
static inline bool _bailout(int fd, struct cgpu_info *modminer, int prio, const char *fmt, ...)
{
if (fd != -1)
serial_close(fd);
if (modminer) {
modminer->device_fd = -1;
mutex_unlock(&modminer->device_mutex);
}
va_list ap;
va_start(ap, fmt);
vapplog(prio, fmt, ap);
va_end(ap);
return false;
}
// 45 noops sent when detecting, in case the device was left in "start job" reading
static const char NOOP[] = "\0\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff";
static bool modminer_detect_one(const char *devpath)
{
char buf[0x100];
char *devname;
ssize_t len;
int fd;
#ifdef WIN32
fd = serial_open(devpath, 0, 10, true);
if (fd < 0) {
applog(LOG_ERR, "ModMiner detect: failed to open %s", devpath);
return false;
}
(void)(write(fd, NOOP, sizeof(NOOP)-1) ?:0);
while (serial_read(fd, buf, sizeof(buf)) > 0)
;
// Version
if (1 != write(fd, "\x01", 1)) {
applog(LOG_ERR, "ModMiner detect: version request failed on %s (%d)", devpath, errno);
goto shin;
}
len = serial_read(fd, buf, sizeof(buf)-1);
if (len < 1) {
applog(LOG_ERR, "ModMiner detect: no version reply on %s (%d)", devpath, errno);
goto shin;
}
buf[len] = '\0';
devname = strdup(buf);
applog(LOG_DEBUG, "ModMiner identified as: %s", devname);
// FPGA count
if (1 != write(fd, "\x02", 1)) {
applog(LOG_ERR, "ModMiner detect: FPGA count request failed on %s (%d)", devpath, errno);
goto shin;
}
len = read(fd, buf, 1);
if (len < 1) {
applog(LOG_ERR, "ModMiner detect: timeout waiting for FPGA count from %s (%d)", devpath, errno);
goto shin;
}
serial_close(fd);
#else
fd = select_open(devpath);
if (fd < 0) {
applog(LOG_ERR, "ModMiner detect: failed to open %s", devpath);
return false;
}
// Don't care if they fail
select_write(fd, (char *)NOOP, sizeof(NOOP)-1);
// Will clear up to a max of sizeof(buf)-1 chars
select_read(fd, buf, sizeof(buf)-1);
// Version
if (select_write(fd, "\x01", 1) < 1) {
applog(LOG_ERR, "ModMiner detect: version request failed on %s (%d)", devpath, errno);
goto shin;
}
if ((len = select_read(fd, buf, sizeof(buf)-1)) < 1) {
applog(LOG_ERR, "ModMiner detect: no version reply on %s (%d)", devpath, errno);
goto shin;
}
buf[len] = '\0';
devname = strdup(buf);
applog(LOG_DEBUG, "ModMiner identified as: %s", devname);
// FPGA count
if (select_write(fd, "\x02", 1) < 1) {
applog(LOG_ERR, "ModMiner detect: FPGA count request failed on %s (%d)", devpath, errno);
goto shin;
}
if ((len = select_read(fd, buf, 1)) < 1) {
applog(LOG_ERR, "ModMiner detect: no FPGA count reply on %s (%d)", devpath, errno);
goto shin;
}
select_close(fd);
#endif
// TODO: check if it supports 2 byte temperatures and if not
// add a flag and set it use 1 byte and code to use the flag
if (buf[0] == 0) {
applog(LOG_ERR, "ModMiner detect: zero FPGA count from %s", devpath);
goto shin;
}
if (buf[0] < 1 || buf[0] > 4) {
applog(LOG_ERR, "ModMiner detect: invalid FPGA count (%u) from %s", buf[0], devpath);
goto shin;
}
applog(LOG_DEBUG, "ModMiner %s has %u FPGAs", devname, buf[0]);
struct cgpu_info *modminer;
modminer = calloc(1, sizeof(*modminer));
modminer->api = &modminer_api;
mutex_init(&modminer->device_mutex);
modminer->device_path = strdup(devpath);
modminer->device_fd = -1;
modminer->deven = DEV_ENABLED;
modminer->threads = buf[0];
modminer->name = devname;
return add_cgpu(modminer);
shin:
#ifdef WIN32
serial_close(fd);
#else
select_close(fd);
#endif
return false;
}
static int modminer_detect_auto()
{
return
serial_autodetect_udev (modminer_detect_one, "*ModMiner*") ?:
serial_autodetect_devserial(modminer_detect_one, "BTCFPGA_ModMiner") ?:
0;
}
static void modminer_detect()
{
serial_detect_auto(&modminer_api, modminer_detect_one, modminer_detect_auto);
}
#define bailout(...) return _bailout(-1, modminer, __VA_ARGS__);
#define bailout2(...) return _bailout(fd, modminer, __VA_ARGS__);
#define check_magic(L) do { \
if (1 != fread(buf, 1, 1, f)) \
bailout(LOG_ERR, "Error reading ModMiner firmware ('%c')", L); \
if (buf[0] != L) \
bailout(LOG_ERR, "ModMiner firmware has wrong magic ('%c')", L); \
} while(0)
#define read_str(eng) do { \
if (1 != fread(buf, 2, 1, f)) \
bailout(LOG_ERR, "Error reading ModMiner firmware (" eng " len)"); \
len = (ubuf[0] << 8) | ubuf[1]; \
if (len >= sizeof(buf)) \
bailout(LOG_ERR, "ModMiner firmware " eng " too long"); \
if (1 != fread(buf, len, 1, f)) \
bailout(LOG_ERR, "Error reading ModMiner firmware (" eng ")"); \
buf[len] = '\0'; \
} while(0)
#define status_read(eng) do { \
FD_ZERO(&fds); \
FD_SET(fd, &fds); \
select(fd+1, &fds, NULL, NULL, NULL); \
if (1 != read(fd, buf, 1)) \
bailout2(LOG_ERR, "%s %u: Error programming %s (" eng ")", modminer->api->name, modminer->device_id, modminer->device_path); \
if (buf[0] != 1) \
bailout2(LOG_ERR, "%s %u: Wrong " eng " programming %s", modminer->api->name, modminer->device_id, modminer->device_path); \
} while(0)
static bool modminer_fpga_upload_bitstream(struct cgpu_info *modminer)
{
fd_set fds;
char buf[0x100];
unsigned char *ubuf = (unsigned char *)buf;
unsigned long len;
char *p;
const char *fwfile = BITSTREAM_FILENAME;
char fpgaid = 4; // "all FPGAs"
FILE *f = open_bitstream("modminer", fwfile);
if (!f)
bailout(LOG_ERR, "Error opening ModMiner firmware file %s", fwfile);
if (1 != fread(buf, 2, 1, f))
bailout(LOG_ERR, "Error reading ModMiner firmware (magic)");
if (buf[0] || buf[1] != 9)
bailout(LOG_ERR, "ModMiner firmware has wrong magic (9)");
if (-1 == fseek(f, 11, SEEK_CUR))
bailout(LOG_ERR, "ModMiner firmware seek failed");
check_magic('a');
read_str("design name");
applog(LOG_DEBUG, "ModMiner firmware file %s info:", fwfile);
applog(LOG_DEBUG, " Design name: %s", buf);
p = strrchr(buf, ';') ?: buf;
p = strrchr(buf, '=') ?: p;
if (p[0] == '=')
++p;
unsigned long fwusercode = (unsigned long)strtoll(p, &p, 16);
if (p[0] != '\0')
bailout(LOG_ERR, "Bad usercode in ModMiner firmware file");
if (fwusercode == 0xffffffff)
bailout(LOG_ERR, "ModMiner firmware doesn't support user code");
applog(LOG_DEBUG, " Version: %u, build %u", (fwusercode >> 8) & 0xff, fwusercode & 0xff);
check_magic('b');
read_str("part number");
applog(LOG_DEBUG, " Part number: %s", buf);
check_magic('c');
read_str("build date");
applog(LOG_DEBUG, " Build date: %s", buf);
check_magic('d');
read_str("build time");
applog(LOG_DEBUG, " Build time: %s", buf);
check_magic('e');
if (1 != fread(buf, 4, 1, f))
bailout(LOG_ERR, "Error reading ModMiner firmware (data len)");
len = ((unsigned long)ubuf[0] << 24) | ((unsigned long)ubuf[1] << 16) | (ubuf[2] << 8) | ubuf[3];
applog(LOG_DEBUG, " Bitstream size: %lu", len);
SOCKETTYPE fd = modminer->device_fd;
applog(LOG_WARNING, "%s %u: Programming %s... DO NOT EXIT CGMINER UNTIL COMPLETE", modminer->api->name, modminer->device_id, modminer->device_path);
buf[0] = '\x05'; // Program Bitstream
buf[1] = fpgaid;
buf[2] = (len >> 0) & 0xff;
buf[3] = (len >> 8) & 0xff;
buf[4] = (len >> 16) & 0xff;
buf[5] = (len >> 24) & 0xff;
if (6 != write(fd, buf, 6))
bailout2(LOG_ERR, "%s %u: Error programming %s (cmd)", modminer->api->name, modminer->device_id, modminer->device_path);
status_read("cmd reply");
ssize_t buflen;
while (len) {
buflen = len < 32 ? len : 32;
if (fread(buf, buflen, 1, f) != 1)
bailout2(LOG_ERR, "%s %u: File underrun programming %s (%d bytes left)", modminer->api->name, modminer->device_id, modminer->device_path, len);
if (write(fd, buf, buflen) != buflen)
bailout2(LOG_ERR, "%s %u: Error programming %s (data)", modminer->api->name, modminer->device_id, modminer->device_path);
status_read("status");
len -= buflen;
}
status_read("final status");
applog(LOG_WARNING, "%s %u: Done programming %s", modminer->api->name, modminer->device_id, modminer->device_path);
return true;
}
static bool modminer_device_prepare(struct cgpu_info *modminer)
{
int fd = serial_open(modminer->device_path, 0, 10, true);
if (unlikely(-1 == fd))
bailout(LOG_ERR, "%s %u: Failed to open %s", modminer->api->name, modminer->device_id, modminer->device_path);
modminer->device_fd = fd;
applog(LOG_INFO, "%s %u: Opened %s", modminer->api->name, modminer->device_id, modminer->device_path);
struct timeval now;
gettimeofday(&now, NULL);
get_datestamp(modminer->init, &now);
return true;
}
#undef bailout
static bool modminer_fpga_prepare(struct thr_info *thr)
{
struct cgpu_info *modminer = thr->cgpu;
// Don't need to lock the mutex here,
// since prepare runs from the main thread before the miner threads start
if (modminer->device_fd == -1 && !modminer_device_prepare(modminer))
return false;
struct modminer_fpga_state *state;
state = thr->cgpu_data = calloc(1, sizeof(struct modminer_fpga_state));
state->next_work_cmd[0] = '\x08'; // Send Job
state->next_work_cmd[1] = thr->device_thread; // FPGA id
state->shares_to_good = MODMINER_EARLY_UP;
return true;
}
/*
* Clocking rules:
* If device exceeds cutoff temp - shut down - and decrease the clock by
* MODMINER_OVERHEAT_CLOCK for when it restarts
*
* When to clock down:
* If device overheats
* or
* If device gets MODMINER_HW_ERROR_PERCENT errors since last clock up or down
* if clock is <= default it requires 2 HW to do this test
* if clock is > default it only requires 1 HW to do this test
*
* When to clock up:
* If device gets shares_to_good good shares in a row
*
* N.B. clock must always be a multiple of 2
*/
static bool modminer_delta_clock(struct thr_info *thr, bool needlock, int delta, bool temp)
{
struct cgpu_info *modminer = thr->cgpu;
struct modminer_fpga_state *state = thr->cgpu_data;
char fpgaid = thr->device_thread;
int fd = modminer->device_fd;
unsigned char cmd[6], buf[1];
struct timeval now;
gettimeofday(&now, NULL);
// Only do once if multiple shares per work or multiple reasons
// Since the temperature down clock test is first in the code this is OK
if (tdiff(&now, &(state->last_changed)) < 0.5)
return false;
// Update before possibly aborting to avoid repeating unnecessarily
memcpy(&(state->last_changed), &now, sizeof(struct timeval));
state->shares = 0;
state->shares_last_hw = 0;
state->hw_errors = 0;
// If drop requested due to temperature, clock drop is always allowed
if (!temp && delta < 0 && state->clock <= MODMINER_MIN_CLOCK)
return false;
if (delta > 0 && state->clock >= MODMINER_MAX_CLOCK)
return false;
if (delta < 0) {
if ((state->shares_to_good * 2) < MODMINER_TRY_UP)
state->shares_to_good *= 2;
else
state->shares_to_good = MODMINER_TRY_UP;
}
state->clock += delta;
cmd[0] = '\x06'; // set clock speed
cmd[1] = fpgaid;
cmd[2] = state->clock;
cmd[3] = cmd[4] = cmd[5] = '\0';
if (needlock)
mutex_lock(&modminer->device_mutex);
if (6 != write(fd, cmd, 6))
bailout2(LOG_ERR, "%s%u.%u: Error writing (set clock speed)", modminer->api->name, modminer->device_id, fpgaid);
if (serial_read(fd, &buf, 1) != 1)
bailout2(LOG_ERR, "%s%u.%u: Error reading (set clock speed)", modminer->api->name, modminer->device_id, fpgaid);
if (needlock)
mutex_unlock(&modminer->device_mutex);
applog(LOG_WARNING, "%s%u.%u: Set clock speed %sto %u", modminer->api->name, modminer->device_id, fpgaid, (delta < 0) ? "down " : (delta > 0 ? "up " : ""), state->clock);
return true;
}
static bool modminer_fpga_init(struct thr_info *thr)
{
struct cgpu_info *modminer = thr->cgpu;
struct modminer_fpga_state *state = thr->cgpu_data;
int fd;
char fpgaid = thr->device_thread;
unsigned char cmd[2], buf[4];
mutex_lock(&modminer->device_mutex);
fd = modminer->device_fd;
if (fd == -1) {
// Died in another thread...
mutex_unlock(&modminer->device_mutex);
return false;
}
cmd[0] = '\x04'; // Read USER code (bitstream id)
cmd[1] = fpgaid;
if (write(fd, cmd, 2) != 2)
bailout2(LOG_ERR, "%s%u.%u: Error writing (read USER code)", modminer->api->name, modminer->device_id, fpgaid);
if (serial_read(fd, buf, 4) != 4)
bailout2(LOG_ERR, "%s%u.%u: Error reading (read USER code)", modminer->api->name, modminer->device_id, fpgaid);
if (memcmp(buf, BISTREAM_USER_ID, 4)) {
applog(LOG_ERR, "%s%u.%u: FPGA not programmed", modminer->api->name, modminer->device_id, fpgaid);
if (!modminer_fpga_upload_bitstream(modminer))
return false;
}
else
applog(LOG_DEBUG, "%s%u.%u: FPGA is already programmed :)", modminer->api->name, modminer->device_id, fpgaid);
state->clock = MODMINER_DEF_CLOCK;
modminer_delta_clock(thr, false, MODMINER_CLOCK_SET, false);
mutex_unlock(&modminer->device_mutex);
thr->primary_thread = true;
return true;
}
static void get_modminer_statline_before(char *buf, struct cgpu_info *modminer)
{
char info[18] = " | ";
int tc = modminer->threads;
bool havetemp = false;
int i;
if (tc > 4)
tc = 4;
for (i = tc - 1; i >= 0; --i) {
struct thr_info *thr = modminer->thr[i];
struct modminer_fpga_state *state = thr->cgpu_data;
float temp = state->temp;
info[i*3+2] = '/';
if (temp) {
havetemp = true;
if (temp > 9)
info[i*3+0] = 0x30 + (temp / 10);
info[i*3+1] = 0x30 + ((int)temp % 10);
}
}
if (havetemp) {
info[tc*3-1] = ' ';
info[tc*3] = 'C';
strcat(buf, info);
}
else
strcat(buf, " | ");
}
static bool modminer_prepare_next_work(struct modminer_fpga_state *state, struct work *work)
{
char *midstate = state->next_work_cmd + 2;
char *taildata = midstate + 32;
if (!(memcmp(midstate, work->midstate, 32) || memcmp(taildata, work->data + 64, 12)))
return false;
memcpy(midstate, work->midstate, 32);
memcpy(taildata, work->data + 64, 12);
return true;
}
static bool modminer_start_work(struct thr_info *thr)
{
fd_set fds;
struct cgpu_info *modminer = thr->cgpu;
struct modminer_fpga_state *state = thr->cgpu_data;
char fpgaid = thr->device_thread;
SOCKETTYPE fd = modminer->device_fd;
char buf[1];
mutex_lock(&modminer->device_mutex);
if (46 != write(fd, state->next_work_cmd, 46))
bailout2(LOG_ERR, "%s%u.%u: Error writing (start work)", modminer->api->name, modminer->device_id, fpgaid);
gettimeofday(&state->tv_workstart, NULL);
state->hashes = 0;
status_read("start work");
mutex_unlock(&modminer->device_mutex);
return true;
}
#define work_restart(thr) thr->work_restart
static uint64_t modminer_process_results(struct thr_info *thr)
{
struct cgpu_info *modminer = thr->cgpu;
struct modminer_fpga_state *state = thr->cgpu_data;
char fpgaid = thr->device_thread;
int fd = modminer->device_fd;
struct work *work = &state->running_work;
char cmd[2], temperature[2];
uint32_t nonce;
long iter;
uint32_t curr_hw_errors;
// \x0a is 1 byte temperature
// \x0d is 2 byte temperature
cmd[0] = '\x0d';
cmd[1] = fpgaid;
mutex_lock(&modminer->device_mutex);
if (2 == write(fd, cmd, 2) && read(fd, &temperature, 2) == 2)
{
// Only accurate to 2 and a bit places
state->temp = roundf((temperature[1] * 256.0 + temperature[0]) / 0.128) / 1000.0;
if (!fpgaid)
modminer->temp = state->temp;
if (state->temp >= MODMINER_OVERHEAT_TEMP) {
if (state->temp >= MODMINER_CUTOFF_TEMP) {
applog(LOG_WARNING, "%s%u.%u: Hit thermal cutoff limit (%f) at %f, disabling device!", modminer->api->name, modminer->device_id, fpgaid, MODMINER_CUTOFF_TEMP, state->temp);
modminer_delta_clock(thr, true, MODMINER_OVERHEAT_CLOCK, true);
dev_error(modminer, REASON_DEV_THERMAL_CUTOFF);
modminer->deven = DEV_RECOVER;
} else {
applog(LOG_WARNING, "%s%u.%u Overheat limit (%f) reached %f", modminer->api->name, modminer->device_id, fpgaid, MODMINER_OVERHEAT_TEMP, state->temp);
modminer_delta_clock(thr, true, MODMINER_CLOCK_DOWN, true);
dev_error(modminer, REASON_DEV_OVER_HEAT);
}
}
}
cmd[0] = '\x09';
iter = 200;
while (1) {
if (write(fd, cmd, 2) != 2)
bailout2(LOG_ERR, "%s%u.%u: Error reading (get nonce)", modminer->api->name, modminer->device_id, fpgaid);
serial_read(fd, &nonce, 4);
mutex_unlock(&modminer->device_mutex);
if (memcmp(&nonce, "\xff\xff\xff\xff", 4)) {
state->shares++;
state->no_nonce_counter = 0;
curr_hw_errors = state->hw_errors;
submit_nonce(thr, work, nonce);
if (state->hw_errors > curr_hw_errors) {
state->shares_last_hw = state->shares;
if (state->clock > MODMINER_DEF_CLOCK || state->hw_errors > 1) {
float pct = (state->hw_errors * 100.0 / (state->shares ? : 1.0));
if (pct >= MODMINER_HW_ERROR_PERCENT)
modminer_delta_clock(thr, true, MODMINER_CLOCK_DOWN, false);
}
} else {
// If we've reached the required good shares in a row then clock up
if ((state->shares - state->shares_last_hw) >= state->shares_to_good)
modminer_delta_clock(thr, true, MODMINER_CLOCK_UP, false);
}
} else if (++state->no_nonce_counter > 18000) {
// TODO: NFI what this is - but will be gone
// when the threading rewrite is done
state->no_nonce_counter = 0;
modminer_delta_clock(thr, true, MODMINER_CLOCK_DOWN, false);
}
if (work_restart(thr))
break;
usleep(10000);
if (work_restart(thr) || !--iter)
break;
mutex_lock(&modminer->device_mutex);
}
struct timeval tv_workend, elapsed;
gettimeofday(&tv_workend, NULL);
timersub(&tv_workend, &state->tv_workstart, &elapsed);
uint64_t hashes = (uint64_t)state->clock * (((uint64_t)elapsed.tv_sec * 1000000) + elapsed.tv_usec);
if (hashes > 0xffffffff)
hashes = 0xffffffff;
else
if (hashes <= state->hashes)
hashes = 1;
else
hashes -= state->hashes;
state->hashes += hashes;
return hashes;
}
static int64_t modminer_scanhash(struct thr_info *thr, struct work *work, int64_t __maybe_unused max_nonce)
{
struct modminer_fpga_state *state = thr->cgpu_data;
int64_t hashes = 0;
bool startwork;
startwork = modminer_prepare_next_work(state, work);
if (state->work_running) {
hashes = modminer_process_results(thr);
if (work_restart(thr)) {
state->work_running = false;
return 0;
}
} else
state->work_running = true;
if (startwork) {
if (!modminer_start_work(thr))
return -1;
memcpy(&state->running_work, work, sizeof(state->running_work));
}
// This is intentionally early
work->blk.nonce += hashes;
return hashes;
}
static void modminer_hw_error(struct thr_info *thr)
{
struct modminer_fpga_state *state = thr->cgpu_data;
state->hw_errors++;
}
static void modminer_fpga_shutdown(struct thr_info *thr)
{
free(thr->cgpu_data);
}
struct device_api modminer_api = {
.dname = "modminer",
.name = "MMQ",
.api_detect = modminer_detect,
.get_statline_before = get_modminer_statline_before,
.thread_prepare = modminer_fpga_prepare,
.thread_init = modminer_fpga_init,
.scanhash = modminer_scanhash,
.hw_error = modminer_hw_error,
.thread_shutdown = modminer_fpga_shutdown,
};