More docs, and add '--ryzen' as a hack for Ryzen folks to get nice perf.

README.md

@@ -42,7 +42,7 @@ _OR_
  * make
 
 #### Note for Debian/Ubuntu users:
- * apt-get install automake autoconf pkg-config libcurl4-openssl-dev libjansson-dev libssl-dev libgmp-dev zlib1g-dev
+ * `apt-get install automake autoconf pkg-config libcurl4-openssl-dev libjansson-dev libssl-dev libgmp-dev zlib1g-dev`
 
 #### Notes for AIX users:
  * To build a 64-bit binary, export OBJECT_MODE=64
@@ -119,6 +119,23 @@ If for some reason you want to remove HugePages (or adjust the size):
 
 `sudo nano /etc/sysctl.conf`, scroll to the bottom, and remove / edit the line `vm.nr_hugepages=size`, `Ctrl+O`, `[Enter]`, `Ctrl+X`.  Then, like before, `sudo sysctl -p`.  Note that you can also reboot and this will cause HugePages to allocate / deallocate.
 
+### Oneways, cpu affinity.
+
+You can now specify a number of 'oneway' threads to acompany your 'default way' threads.  Default way is determined by your CPU instruction set.
+
+'-1 n' will specify the number of oneway threads to spawn.  You can also use `--oneways n`.  Some folks (on arm especially) see perf gains due to the implementation.  Additionally, there are new options to help control affinity for these threads and 'default' way threads too.  `--cpu-affinity-stride N`, `--cpu-affinity-default-index N`, `--cpu-affinity-oneway-index N`, and `--cpu-priority-oneway 0-5`
+
+Affinity stride works by saying how many cpus should be skipped before assigning the thread to it.  So, if you have 8 cpu's and set 'stride' to 3, you will start with CPU (0 + 3 * 0), then CPU (0 + 3 * 1), then CPU (0 + 3 * 2),  aka  CPU's 0, 3, 6.  'default' index is the starting index ('0' in the example) for the number of `-t` threads.  'oneway-index' is the starting index for oneway `-1` threads.
+
+Play around with them, and pass `-D` to get some debug output.
+
+e.g, 'Bind oneways to odd threads, 'defaults' to even threads: `./cpuminer ... -t 4 -1 4 --cpu-affinity-stride 2 --cpu-affinity-default-index 0 --cpu-affinity-oneway-index 1`
+e.g.2, 'Bind oneways to the last 4 cpu's after the defaults' `./cpuminer ... -t 4 -1 4 --cpu-affinity-stride 1 --cpu-affinity-default-index 0 --cpu-affinity-oneway-index 4` (edited)
+
+# RYZEN USERS pass `--ryzen`.
+
+Ryzen's implementation of AVX2 is ... subpar.  Please pass `--ryzen` on the commandline to default to the AVX implementation.  Users reported ~25% gains.
+
 ### Connecting through a proxy
 
 Use the --proxy option.
@@ -131,17 +148,16 @@ When the --proxy option is not used, the program honors the http_proxy and all_p
 
 GCC
 =======
-Some people have reported increases by using GCC 7.
+Some people have reported increases by using GCC 7.2.  Please note, this will replace your existing GCC installation, which may become unrecoverable if any errors occur in the `make` process.
 
 To build and install GCC 7.2 on Ubuntu do the following:
 * `apt-get -y install unzip flex`
 * `wget https://github.com/gcc-mirror/gcc/archive/gcc-7_2_0-release.zip`
 * `unzip gcc-7_2_0-release.zip`
-* `cd gcc-7_2_0-release`
+* `cd gcc-gcc-7_2_0-release`
 * `sudo ./contrib/download_prerequisites`
-* `mkdir build`
-* `cd build`
-* `../configure `
+* `mkdir build && cd build`
+* `../configure --enable-languages=c,c++ --disable-multilib`
 * `make -j 8`
 * `make install`
 

algo/scrypt.c

@@ -1771,7 +1771,13 @@ static inline void scrypt_core(uint32_t *X, uint32_t *V, int N)
 bool printed = false;
 unsigned char *scrypt_buffer_alloc(int N, int forceThroughput)
 {
-	uint32_t size = (forceThroughput == -1 ? scrypt_best_throughput() : forceThroughput) * 32 * (N + 1) * sizeof(uint32_t);
+	uint32_t throughput = (forceThroughput == -1 ? scrypt_best_throughput() : forceThroughput);
+	if (opt_ryzen_1x) {
+		// force throughput to be 3 (aka AVX) instead of AVX2.
+		throughput = 3;
+	}
+
+	uint32_t size = throughput * 32 * (N + 1) * sizeof(uint32_t);
 
 #ifdef __linux__
 	unsigned char* m_memory = (unsigned char*)(mmap(0, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB | MAP_POPULATE, 0, 0));
@@ -1987,6 +1993,11 @@ extern int scanhash_scrypt(int thr_id, struct work *work, uint32_t max_nonce, ui
 	uint32_t n = pdata[19] - 1;
 	const uint32_t Htarg = ptarget[7];
 	int throughput = scrypt_best_throughput();
+	if (opt_ryzen_1x) {
+		// force throughput to be 3 (aka AVX) instead of AVX2.
+		throughput = 3;
+	}
+
 	int i;
 
 #ifdef HAVE_SHA256_4WAY

cpu-miner.c

@@ -125,6 +125,7 @@ int opt_affinity_stride = 1;
 int opt_affinity_default_index = 0;
 int opt_affinity_oneway_index = 0;
 int opt_oneway_priority = 0;
+bool opt_ryzen_1x = false;
 int* thread_affinty_array = NULL;
 int num_cpus;
 char *rpc_url;
@@ -200,6 +201,7 @@ Options:\n\
       --cert=FILE       certificate for mining server using SSL\n\
   -x, --proxy=[PROTOCOL://]HOST[:PORT]  connect through a proxy\n\
   -t, --threads=N       number of miner threads (default: number of processors)\n\
+  -1, --oneways=N       number of miner threads that are forced to 'oneway' (default: 0)\n\
   -r, --retries=N       number of times to retry if a network call fails\n\
                           (default: retry indefinitely)\n\
   -R, --retry-pause=N   time to pause between retries, in seconds (default: 30)\n\
@@ -234,11 +236,23 @@ Options:\n\
       --cputest         debug hashes from cpu algorithms\n\
       --cpu-affinity    set process affinity to cpu core(s), mask 0x3 for cores 0 and 1\n\
       --cpu-priority    set process priority (default: 0 idle, 2 normal to 5 highest)\n\
+\n\
+      --cpu-affinity-stride N  \n\
+            how many processors to skip when assigining affinity based on indicies\n\
+             cannot be used with '--cpu-affinity'  (default: 1) See README.md for more details.\n\
+      --cpu-affinity-default-index N \n\
+            which cpu to start affinity for 'default' way threads (0-based). (default: 0) See README.md for more details.\n\
+      --cpu-affinity-oneway-index N \n\
+            which cpu to start affinity for 'default' way threads (0-based). (default: [After default threads]) See README.md for more details.\n\
+      --cpu-priority-oneway 0-5\n\
+            what priority oneway threads have (0 lowest, 5 highest) (default: 0)\n\
+\n\
   -b, --api-bind        IP/Port for the miner API (default: 127.0.0.1:4048)\n\
       --api-remote      Allow remote control\n\
       --max-temp=N      Only mine if cpu temp is less than specified value (linux)\n\
       --max-rate=N[KMG] Only mine if net hashrate is less than specified value\n\
       --max-diff=N      Only mine if net difficulty is less than specified value\n\
+      --ryzen           Force AVX, and disable AVX2.  Ryzen 1*** is much faster.\n\
   -c, --config=FILE     load a JSON-format configuration file\n\
   -V, --version         display version information and exit\n\
   -h, --help            display this help text and exit\n\
@@ -269,6 +283,7 @@ static struct option const options[] = {
 	{ "cpu-affinity-stride", 1, NULL, 1050 },
 	{ "cpu-affinity-default-index", 1, NULL, 1051 },
 	{ "cpu-affinity-oneway-index", 1, NULL, 1052 },
+	{ "ryzen", 0, NULL, 2000 },
 	{ "no-color", 0, NULL, 1002 },
 	{ "debug", 0, NULL, 'D' },
 	{ "diff-factor", 1, NULL, 'f' },
@@ -2583,6 +2598,9 @@ void parse_arg(int key, char *arg)
 		opt_affinity_oneway_index = v;
 		use_affinity_mask = -1;
 		break;	
+	case 2000: // "ryzen"
+		opt_ryzen_1x = true;
+		break;
 	case 'V':
 		show_version_and_exit();
 	case 'h':

miner.h

@@ -297,6 +297,7 @@ extern bool opt_stratum_stats;
 extern char *opt_cert;
 extern char *opt_proxy;
 extern long opt_proxy_type;
+extern bool opt_ryzen_1x;
 extern bool use_syslog;
 extern bool use_colors;
 extern pthread_mutex_t applog_lock;