xmr: link the --bfactor setting (0-11)

README.txt

@@ -57,7 +57,7 @@ Vertcoin Lyra2RE
 Ziftrcoin (ZR5)
 Boolberry (Wild Keccak)
 Monero (Cryptonight)
-Aeon (Cryptonight-light)
+Aeon (Cryptonight-lite)
 
 where some of these coins have a VERY NOTABLE nVidia advantage
 over competing AMD (OpenCL Only) implementations.
@@ -193,11 +193,16 @@ Scrypt specific options:
                         cache for mining. Kepler devices may profit.
       --no-autotune     disable auto-tuning of kernel launch parameters
 
-
-XMR and Wildkeccak specific:
+CryptoNight specific options:
   -l, --launch-config   gives the launch configuration for each kernel
                         in a comma separated list, one per device.
+      --bfactor=[0-12]  Run Cryptonight core kernel in smaller pieces,
+                        From 0 (ui freeze) to 12 (smooth), win default is 11
+                        This is a per-device setting like the launch config.
+
 Wildkeccak specific:
+  -l, --launch-config   gives the launch configuration for each kernel
+                        in a comma separated list, one per device.
   -k, --scratchpad url  Url used to download the scratchpad cache.
 
 

algos.h

@@ -127,6 +127,8 @@ static inline int algo_to_int(char* arg)
 			i = ALGO_AUTO;
 		else if (!strcasecmp("cryptonight-light", arg))
 			i = ALGO_CRYPTOLIGHT;
+		else if (!strcasecmp("cryptonight-lite", arg))
+			i = ALGO_CRYPTOLIGHT;
 		else if (!strcasecmp("flax", arg))
 			i = ALGO_C11;
 		else if (!strcasecmp("diamond", arg))

ccminer.cpp

@@ -146,7 +146,8 @@ int device_singlememory[MAX_GPUS] = { 0 };
 // implemented scrypt options
 int parallel = 2; // All should be made on GPU
 char *device_config[MAX_GPUS] = { 0 };
-int device_backoff[MAX_GPUS] = { 0 };
+int device_backoff[MAX_GPUS] = { 0 }; // scrypt
+int device_bfactor[MAX_GPUS] = { 0 }; // cryptonight
 int device_lookup_gap[MAX_GPUS] = { 0 };
 int device_interactive[MAX_GPUS] = { 0 };
 int opt_nfactor = 0;
@@ -371,8 +372,9 @@ struct option options[] = {
 	{ "interactive", 1, NULL, 1050 },  // scrypt
 	{ "lookup-gap", 1, NULL, 'L' },    // scrypt
 	{ "texture-cache", 1, NULL, 1051 },// scrypt
-	{ "launch-config", 1, NULL, 'l' }, // scrypt & bbr
+	{ "launch-config", 1, NULL, 'l' }, // scrypt bbr xmr
 	{ "scratchpad", 1, NULL, 'k' },    // bbr
+	{ "bfactor", 1, NULL, 1055 },      // xmr
 	{ "max-temp", 1, NULL, 1060 },
 	{ "max-diff", 1, NULL, 1061 },
 	{ "max-rate", 1, NULL, 1062 },
@@ -442,7 +444,16 @@ Scrypt specific options:\n\
 ";
 
 static char const xmr_usage[] = "\n\
-CryptoNote specific options:\n\
+CryptoNight specific options:\n\
+  -l, --launch-config   gives the launch configuration for each kernel\n\
+                        in a comma separated list, one per device.\n\
+      --bfactor=[0-12]  Run Cryptonight core kernel in smaller pieces,\n\
+                        From 0 (ui freeze) to 12 (smooth), win default is 11\n\
+                        This is a per-device setting like the launch config.\n\
+";
+
+static char const bbr_usage[] = "\n\
+Boolberry specific options:\n\
   -l, --launch-config   gives the launch configuration for each kernel\n\
                         in a comma separated list, one per device.\n\
   -k, --scratchpad url  Url used to download the scratchpad cache.\n\
@@ -2441,6 +2452,7 @@ static void *miner_thread(void *userdata)
 			if (opt_led_mode == LED_MODE_SHARES)
 				gpu_led_percent(dev_id, 50);
 
+			work.submit_nonce_id = 0;
 			nonceptr[0] = work.nonces[0];
 			if (!submit_work(mythr, &work))
 				break;
@@ -2468,6 +2480,7 @@ static void *miner_thread(void *userdata)
 				if (!submit_work(mythr, &work))
 					break;
 				nonceptr[0] = curnonce;
+				work.nonces[1] = 0; // reset
 			}
 		}
 	}
@@ -2852,12 +2865,16 @@ static void show_usage_and_exit(int status)
 		fprintf(stderr, "Try `" PROGRAM_NAME " --help' for more information.\n");
 	else
 		printf(usage);
+
 	if (opt_algo == ALGO_SCRYPT || opt_algo == ALGO_SCRYPT_JANE) {
 		printf(scrypt_usage);
 	}
-	if (opt_algo == ALGO_CRYPTONIGHT || opt_algo == ALGO_CRYPTOLIGHT || opt_algo == ALGO_WILDKECCAK) {
+	else if (opt_algo == ALGO_CRYPTONIGHT || opt_algo == ALGO_CRYPTOLIGHT) {
 		printf(xmr_usage);
 	}
+	else if (opt_algo == ALGO_WILDKECCAK) {
+		printf(bbr_usage);
+	}
 	proper_exit(status);
 }
 
@@ -3187,6 +3204,20 @@ void parse_arg(int key, char *arg)
 				device_texturecache[n++] = last;
 		}
 		break;
+	case 1055: /* cryptonight --bfactor */
+		{
+			char *pch = strtok(arg, ",");
+			int n = 0, last = atoi(arg);
+			while (pch != NULL) {
+				last = atoi(pch);
+				if (last > 15) last = 15;
+				device_bfactor[n++] = last;
+				pch = strtok(NULL, ",");
+			}
+			while (n < MAX_GPUS)
+				device_bfactor[n++] = last;
+		}
+		break;
 	case 1070: /* --gpu-clock */
 		{
 			char *pch = strtok(arg,",");
@@ -3659,6 +3690,7 @@ int main(int argc, char *argv[])
 		device_name[i] = NULL;
 		device_config[i] = NULL;
 		device_backoff[i] = is_windows() ? 12 : 2;
+		device_bfactor[i] = is_windows() ? 11 : 0;
 		device_lookup_gap[i] = 1;
 		device_batchsize[i] = 1024;
 		device_interactive[i] = -1;

crypto/cryptolight-core.cu

@@ -4,20 +4,11 @@
 #include <sys/time.h>
 #include <unistd.h>
 
-#include <cuda.h>
-#include <cuda_runtime.h>
-
 #include "cryptolight.h"
 #define LONG_SHL_IDX 18
 #define LONG_LOOPS32 0x40000
 
-#ifdef WIN32
-static __thread int cn_bfactor = 11;
-static __thread int cn_bsleep = 100;
-#else
-static __thread int cn_bfactor = 0;
-static __thread int cn_bsleep = 0;
-#endif
+extern int device_backoff[MAX_GPUS];
 
 #include "cn_aes.cuh"
 
@@ -259,6 +250,8 @@ void cryptolight_core_gpu_phase3(int threads, const uint32_t * __restrict__ long
 	}
 }
 
+extern int device_bfactor[MAX_GPUS];
+
 __host__
 void cryptolight_core_cpu_hash(int thr_id, int blocks, int threads, uint32_t *d_long_state, uint32_t *d_ctx_state, uint32_t *d_ctx_a, uint32_t *d_ctx_b, uint32_t *d_ctx_key1, uint32_t *d_ctx_key2)
 {
@@ -267,8 +260,8 @@ void cryptolight_core_cpu_hash(int thr_id, int blocks, int threads, uint32_t *d_
 	dim3 block4(threads << 2);
 	dim3 block8(threads << 3);
 
-	const int bfactor = cn_bfactor; // device_bfactor[thr_id];
-	const int bsleep = cn_bsleep; //device_bsleep[thr_id];
+	const int bfactor = device_bfactor[thr_id];
+	const int bsleep = bfactor ? 100 : 0;
 
 	int i, partcount = 1 << bfactor;
 	int dev_id = device_map[thr_id];

crypto/cryptonight-core.cu

@@ -4,22 +4,10 @@
 #include <sys/time.h>
 #include <unistd.h>
 
-#include <cuda.h>
-#include <cuda_runtime.h>
-
 #include "cryptonight.h"
 #define LONG_SHL_IDX 19
 #define LONG_LOOPS32 0x80000
 
-#ifdef WIN32
-// to prevent ui freeze
-static __thread int cn_bfactor = 11;
-static __thread int cn_bsleep = 100;
-#else
-static __thread int cn_bfactor = 0;
-static __thread int cn_bsleep = 0;
-#endif
-
 #include "cn_aes.cuh"
 
 #define MUL_SUM_XOR_DST(a,c,dst) { \
@@ -264,6 +252,8 @@ void cryptonight_core_gpu_phase3(int threads, const uint32_t * __restrict__ long
 	}
 }
 
+extern int device_bfactor[MAX_GPUS];
+
 __host__
 void cryptonight_core_cpu_hash(int thr_id, int blocks, int threads, uint32_t *d_long_state, uint32_t *d_ctx_state, uint32_t *d_ctx_a, uint32_t *d_ctx_b, uint32_t *d_ctx_key1, uint32_t *d_ctx_key2)
 {
@@ -272,8 +262,8 @@ void cryptonight_core_cpu_hash(int thr_id, int blocks, int threads, uint32_t *d_
 	dim3 block4(threads << 2);
 	dim3 block8(threads << 3);
 
-	const int bfactor = cn_bfactor; // device_bfactor[thr_id];
-	const int bsleep = cn_bsleep; //device_bsleep[thr_id];
+	const int bfactor = device_bfactor[thr_id];
+	const int bsleep = bfactor ? 100 : 0;
 
 	int i, partcount = 1 << bfactor;
 	int dev_id = device_map[thr_id];