Added alt kernel file option

algorithm.c

@@ -567,6 +567,7 @@ static cl_int queue_fresh_kernel(struct __clState *clState, struct _dev_blk_ctx
 typedef struct _algorithm_settings_t {
   const char *name; /* Human-readable identifier */
   algorithm_type_t type; //common algorithm type
+  const char *kernelfile; /* alternate kernel file */
   double   diff_multiplier1;
   double   diff_multiplier2;
   double   share_diff_multiplier;
@@ -587,7 +588,7 @@ typedef struct _algorithm_settings_t {
 static algorithm_settings_t algos[] = {
   // kernels starting from this will have difficulty calculated by using litecoin algorithm
 #define A_SCRYPT(a) \
-    { a, ALGO_SCRYPT, 1, 65536, 65536, 0, 0, 0xFF, 0xFFFFFFFFULL, 0x0000ffffUL, 0, -1, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, scrypt_regenhash, queue_scrypt_kernel, gen_hash, append_scrypt_compiler_options}
+  { a, ALGO_SCRYPT, "", 1, 65536, 65536, 0, 0, 0xFF, 0xFFFFFFFFULL, 0x0000ffffUL, 0, -1, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, scrypt_regenhash, queue_scrypt_kernel, gen_hash, append_scrypt_compiler_options}
   A_SCRYPT( "ckolivas" ),
   A_SCRYPT( "alexkarnew" ),
   A_SCRYPT( "alexkarnold" ),
@@ -598,7 +599,7 @@ static algorithm_settings_t algos[] = {
 
   // kernels starting from this will have difficulty calculated by using quarkcoin algorithm
 #define A_QUARK(a, b) \
-    { a, ALGO_QUARK, 256, 256, 256, 0, 0, 0xFF, 0xFFFFFFULL, 0x0000ffffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, b, queue_sph_kernel, gen_hash, append_x11_compiler_options}
+  { a, ALGO_QUARK, "", 256, 256, 256, 0, 0, 0xFF, 0xFFFFFFULL, 0x0000ffffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, b, queue_sph_kernel, gen_hash, append_x11_compiler_options }
   A_QUARK( "quarkcoin", quarkcoin_regenhash),
   A_QUARK( "qubitcoin", qubitcoin_regenhash),
   A_QUARK( "animecoin", animecoin_regenhash),
@@ -607,40 +608,40 @@ static algorithm_settings_t algos[] = {
 
   // kernels starting from this will have difficulty calculated by using bitcoin algorithm
 #define A_DARK(a, b) \
-    { a, ALGO_X11, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, b, queue_sph_kernel, gen_hash, append_x11_compiler_options}
+  { a, ALGO_X11, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, b, queue_sph_kernel, gen_hash, append_x11_compiler_options }
   A_DARK( "darkcoin",           darkcoin_regenhash),
   A_DARK( "inkcoin",            inkcoin_regenhash),
   A_DARK( "myriadcoin-groestl", myriadcoin_groestl_regenhash),
 #undef A_DARK
 
-  { "twecoin", ALGO_TWE, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, twecoin_regenhash, queue_sph_kernel, sha256, NULL},
-  { "maxcoin", ALGO_KECCAK, 1, 256, 1, 4, 15, 0x0F, 0xFFFFULL, 0x000000ffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, maxcoin_regenhash, queue_maxcoin_kernel, sha256, NULL},
+  { "twecoin", ALGO_TWE, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, twecoin_regenhash, queue_sph_kernel, sha256, NULL},
+  { "maxcoin", ALGO_KECCAK, "", 1, 256, 1, 4, 15, 0x0F, 0xFFFFULL, 0x000000ffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, maxcoin_regenhash, queue_maxcoin_kernel, sha256, NULL },
 
-  { "darkcoin-mod", ALGO_X11, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 10, 8 * 16 * 4194304, 0, darkcoin_regenhash, queue_darkcoin_mod_kernel, gen_hash, append_x11_compiler_options},
+  { "darkcoin-mod", ALGO_X11, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 10, 8 * 16 * 4194304, 0, darkcoin_regenhash, queue_darkcoin_mod_kernel, gen_hash, append_x11_compiler_options },
 
-  { "marucoin", ALGO_X13, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, marucoin_regenhash, queue_sph_kernel, gen_hash, append_x13_compiler_options},
-  { "marucoin-mod", ALGO_X13, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 12, 8 * 16 * 4194304, 0, marucoin_regenhash, queue_marucoin_mod_kernel, gen_hash, append_x13_compiler_options},
-  { "marucoin-modold", ALGO_X13, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 10, 8 * 16 * 4194304, 0, marucoin_regenhash, queue_marucoin_mod_old_kernel, gen_hash, append_x13_compiler_options},
+  { "marucoin", ALGO_X13, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, marucoin_regenhash, queue_sph_kernel, gen_hash, append_x13_compiler_options },
+  { "marucoin-mod", ALGO_X13, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 12, 8 * 16 * 4194304, 0, marucoin_regenhash, queue_marucoin_mod_kernel, gen_hash, append_x13_compiler_options },
+  { "marucoin-modold", ALGO_X13, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 10, 8 * 16 * 4194304, 0, marucoin_regenhash, queue_marucoin_mod_old_kernel, gen_hash, append_x13_compiler_options },
 
-  { "x14", ALGO_X14, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 13, 8 * 16 * 4194304, 0, x14_regenhash, queue_x14_kernel, gen_hash, append_x13_compiler_options},
-  { "x14old", ALGO_X14, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 10, 8 * 16 * 4194304, 0, x14_regenhash, queue_x14_old_kernel, gen_hash, append_x13_compiler_options},
+  { "x14", ALGO_X14, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 13, 8 * 16 * 4194304, 0, x14_regenhash, queue_x14_kernel, gen_hash, append_x13_compiler_options },
+  { "x14old", ALGO_X14, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 10, 8 * 16 * 4194304, 0, x14_regenhash, queue_x14_old_kernel, gen_hash, append_x13_compiler_options },
 
-  { "bitblock", ALGO_X15, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 14, 4 * 16 * 4194304, 0, bitblock_regenhash, queue_bitblock_kernel, gen_hash, append_x13_compiler_options},
-  { "bitblockold", ALGO_X15, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 10, 4 * 16 * 4194304, 0, bitblock_regenhash, queue_bitblockold_kernel, gen_hash, append_x13_compiler_options},
+  { "bitblock", ALGO_X15, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 14, 4 * 16 * 4194304, 0, bitblock_regenhash, queue_bitblock_kernel, gen_hash, append_x13_compiler_options },
+  { "bitblockold", ALGO_X15, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 10, 4 * 16 * 4194304, 0, bitblock_regenhash, queue_bitblockold_kernel, gen_hash, append_x13_compiler_options },
 
-  { "talkcoin-mod", ALGO_NIST, 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 4,  8 * 16 * 4194304, 0, talkcoin_regenhash, queue_talkcoin_mod_kernel, gen_hash, append_x11_compiler_options},
+  { "talkcoin-mod", ALGO_NIST, "", 1, 1, 1, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 4, 8 * 16 * 4194304, 0, talkcoin_regenhash, queue_talkcoin_mod_kernel, gen_hash, append_x11_compiler_options },
 
-  { "fresh", ALGO_FRESH, 1, 256, 256, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 4, 4 * 16 * 4194304, 0, fresh_regenhash, queue_fresh_kernel, gen_hash, NULL},
+  { "fresh", ALGO_FRESH, "", 1, 256, 256, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 4, 4 * 16 * 4194304, 0, fresh_regenhash, queue_fresh_kernel, gen_hash, NULL },
 
   // kernels starting from this will have difficulty calculated by using fuguecoin algorithm
 #define A_FUGUE(a, b) \
-    { a, ALGO_FUGUE, 1, 256, 256, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, b, queue_sph_kernel, sha256, NULL}
+  { a, ALGO_FUGUE, "", 1, 256, 256, 0, 0, 0xFF, 0xFFFFULL, 0x0000ffffUL, 0, 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, b, queue_sph_kernel, sha256, NULL }
   A_FUGUE( "fuguecoin",   fuguecoin_regenhash),
   A_FUGUE( "groestlcoin", groestlcoin_regenhash),
 #undef A_FUGUE
 
   // Terminator (do not remove)
-  { NULL, ALGO_UNK, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, NULL, NULL, NULL, NULL}
+  { NULL, ALGO_UNK, "", 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, NULL, NULL, NULL, NULL }
 };
 
 void copy_algorithm_settings(algorithm_t* dest, const char* algo)
@@ -653,6 +654,7 @@ void copy_algorithm_settings(algorithm_t* dest, const char* algo)
     if (strcmp(src->name, algo) == 0)
     {
       strcpy(dest->name, src->name);
+	  dest->kernelfile = src->kernelfile;
       dest->type = src->type;
 
       dest->diff_multiplier1 = src->diff_multiplier1;
@@ -720,6 +722,8 @@ void set_algorithm(algorithm_t* algo, const char* newname_alias)
   const char* newname;
   //load previous algorithm nfactor in case nfactor was applied before algorithm... or default to 10
   uint8_t old_nfactor = ((algo->nfactor)?algo->nfactor:0);
+  //load previous kernel file name if was applied before algorithm...
+  const char *kernelfile = algo->kernelfile;
   uint8_t nfactor = 10;
 
   if (!(newname = lookup_algorithm_alias(newname_alias, &nfactor)))
@@ -732,6 +736,11 @@ void set_algorithm(algorithm_t* algo, const char* newname_alias)
     nfactor = old_nfactor;
 
   set_algorithm_nfactor(algo, nfactor);
+
+  //reapply kernelfile if was set
+  if (!empty_string(kernelfile)) {
+    algo->kernelfile = kernelfile;
+  }
 }
 
 void set_algorithm_nfactor(algorithm_t* algo, const uint8_t nfactor)
@@ -761,6 +770,5 @@ void set_algorithm_nfactor(algorithm_t* algo, const uint8_t nfactor)
 
 bool cmp_algorithm(algorithm_t* algo1, algorithm_t* algo2)
 {
-  return (strcmp(algo1->name, algo2->name) == 0) &&
-         (algo1->nfactor == algo2->nfactor);
+  return (!safe_cmp(algo1->name, algo2->name) && !safe_cmp(algo1->kernelfile, algo2->kernelfile) && (algo1->nfactor == algo2->nfactor));
 }

algorithm.h

@@ -42,6 +42,7 @@ struct work;
 typedef struct _algorithm_t {
   char     name[20]; /* Human-readable identifier */
   algorithm_type_t type; //algorithm type
+  const char *kernelfile; /* alternate kernel file */
   uint32_t n;        /* N (CPU/Memory tradeoff parameter) */
   uint8_t  nfactor;  /* Factor of N above (n = 2^nfactor) */
   double   diff_multiplier1;

config_parser.c

@@ -173,6 +173,14 @@ char *set_default_devices(const char *arg)
   return NULL;
 }
 
+char *set_default_kernelfile(const char *arg)
+{
+  applog(LOG_INFO, "Set default kernel file to %s", arg);
+  default_profile.algorithm.kernelfile = arg;
+  
+  return NULL;
+}
+
 char *set_default_lookup_gap(const char *arg)
 {
   default_profile.lookup_gap = arg;
@@ -276,7 +284,7 @@ char *set_profile_algorithm(const char *arg)
 {
   struct profile *profile = get_current_profile();
 
-  //applog(LOG_DEBUG, "Setting profile %s algorithm to %s", profile->name, arg);
+  applog(LOG_DEBUG, "Setting profile %s algorithm to %s", profile->name, arg);
   set_algorithm(&profile->algorithm, arg);
 
   return NULL;
@@ -289,6 +297,16 @@ char *set_profile_devices(const char *arg)
   return NULL;
 }
 
+char *set_profile_kernelfile(const char *arg)
+{
+  struct profile *profile = get_current_profile();
+  
+  applog(LOG_DEBUG, "Setting profile %s algorithm kernel file to %s", profile->name, arg);
+  profile->algorithm.kernelfile = arg;
+  
+  return NULL;
+}
+
 char *set_profile_lookup_gap(const char *arg)
 {
   struct profile *profile = get_current_profile();
@@ -946,6 +964,20 @@ void apply_pool_profile(struct pool *pool)
   }
   applog(LOG_DEBUG, "Pool %i Algorithm set to \"%s\"", pool->pool_no, pool->algorithm.name);
 
+  // if the pool doesn't have a specific kernel file...
+  if (empty_string(pool->algorithm.kernelfile)) {
+    // ...but profile does, apply it to the pool
+    if (!empty_string(profile->algorithm.kernelfile)) {
+      pool->algorithm.kernelfile = profile->algorithm.kernelfile;
+      applog(LOG_DEBUG, "Pool %i Kernel File set to \"%s\"", pool->pool_no, pool->algorithm.kernelfile);
+      // ...or default profile does, apply it to the pool
+    }
+    else if (!empty_string(default_profile.algorithm.kernelfile)) {
+      pool->algorithm.kernelfile = default_profile.algorithm.kernelfile;
+      applog(LOG_DEBUG, "Pool %i Kernel File set to \"%s\"", pool->pool_no, pool->algorithm.kernelfile);
+    }
+  }
+
   if(pool_cmp(pool->devices, default_profile.devices))
   {
     if(!empty_string(profile->devices))
@@ -1233,6 +1265,10 @@ static json_t *build_pool_json()
     // devices
     if (!build_pool_json_add(obj, "device", pool->devices, profile->devices, default_profile.devices, pool->pool_no))
       return NULL;
+    
+    // kernelfile
+    if (!build_pool_json_add(obj, "kernelfile", pool->algorithm.kernelfile, profile->algorithm.kernelfile, default_profile.algorithm.kernelfile, pool->pool_no))
+      return NULL;
 
     // lookup-gap
     if (!build_pool_json_add(obj, "lookup-gap", pool->lookup_gap, profile->lookup_gap, default_profile.lookup_gap, pool->pool_no))
@@ -1338,6 +1374,10 @@ static json_t *build_profile_settings_json(json_t *object, struct profile *profi
   if (!build_profile_json_add(object, "device", profile->devices, default_profile.devices, isdefault, parentkey, profile->profile_no))
     return NULL;
 
+  // kernelfile
+  if (!build_profile_json_add(object, "kernelfile", profile->algorithm.kernelfile, default_profile.algorithm.kernelfile, isdefault, parentkey, profile->profile_no))
+    return NULL;
+
   // lookup-gap
   if (!build_profile_json_add(object, "lookup-gap", profile->lookup_gap, default_profile.lookup_gap, isdefault, parentkey, profile->profile_no))
     return NULL;

config_parser.h

@@ -47,6 +47,7 @@ extern int total_profiles;
 extern char *set_default_algorithm(const char *arg);
 extern char *set_default_nfactor(const char *arg);
 extern char *set_default_devices(const char *arg);
+extern char *set_default_kernelfile(const char *arg);
 extern char *set_default_lookup_gap(const char *arg);
 extern char *set_default_intensity(const char *arg);
 extern char *set_default_xintensity(const char *arg);
@@ -67,6 +68,7 @@ extern char *set_default_worksize(const char *arg);
 extern char *set_profile_name(const char *arg);
 extern char *set_profile_algorithm(const char *arg);
 extern char *set_profile_devices(const char *arg);
+extern char *set_profile_kernelfile(const char *arg);
 extern char *set_profile_lookup_gap(const char *arg);
 extern char *set_profile_intensity(const char *arg);
 extern char *set_profile_xintensity(const char *arg);

ocl.c

@@ -298,8 +298,9 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *alg
   char filename[255];
   char strbuf[32];
 
-  sprintf(strbuf, "%s.cl", cgpu->algorithm.name);
+  sprintf(strbuf, "%s.cl", (!empty_string(cgpu->algorithm.kernelfile) ? cgpu->algorithm.kernelfile : cgpu->algorithm.name));
   strcpy(filename, strbuf);
+  applog(LOG_DEBUG, "Using source file %s", filename);
 
   /* For some reason 2 vectors is still better even if the card says
    * otherwise, and many cards lie about their max so use 256 as max
@@ -382,7 +383,7 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *alg
   build_data->opencl_version = get_opencl_version(devices[gpu]);
   build_data->patch_bfi = needs_bfi_patch(build_data);
 
-  strcpy(build_data->binary_filename, cgpu->algorithm.name);
+  strcpy(build_data->binary_filename, (!empty_string(cgpu->algorithm.kernelfile) ? cgpu->algorithm.kernelfile : cgpu->algorithm.name));
   strcat(build_data->binary_filename, name);
   if (clState->goffset)
     strcat(build_data->binary_filename, "g");
@@ -392,6 +393,7 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *alg
     algorithm->set_compile_options(build_data, cgpu, algorithm);
 
   strcat(build_data->binary_filename, ".bin");
+  applog(LOG_DEBUG, "Using binary file %s", build_data->binary_filename);
 
   // Load program from file or build it if it doesn't exist
   if (!(clState->program = load_opencl_binary_kernel(build_data))) {

sgminer.c

@@ -797,6 +797,16 @@ static char *set_pool_devices(const char *arg)
   return NULL;
 }
 
+static char *set_pool_kernelfile(const char *arg)
+{
+  struct pool *pool = get_current_pool();
+  
+  applog(LOG_DEBUG, "Setting pool %i algorithm kernel file to %s", pool->pool_no, arg);
+  pool->algorithm.kernelfile = arg;
+  
+  return NULL;
+}
+
 static char *set_pool_lookup_gap(const char *arg)
 {
   struct pool *pool = get_current_pool();
@@ -1421,6 +1431,9 @@ struct opt_table opt_config_table[] = {
   OPT_WITH_ARG("--keccak-unroll",
       set_int_0_to_9999, opt_show_intval, &opt_keccak_unroll,
       "Set SPH_KECCAK_UNROLL for Xn derived algorithms (Default: 0)"),
+  OPT_WITH_ARG("--kernelfile",
+      set_default_kernelfile, NULL, NULL,
+      "Set the algorithm kernel source file (without file extension)."),
   OPT_WITH_ARG("--lookup-gap",
       set_default_lookup_gap, NULL, NULL,
       "Set GPU lookup gap for scrypt mining, comma separated"),
@@ -1523,6 +1536,9 @@ struct opt_table opt_config_table[] = {
   OPT_WITH_ARG("--pool-device",
       set_pool_devices, NULL, NULL,
       "Select devices to use with pool, one value, range and/or comma separated (e.g. 0-2,4) default: all"),
+  OPT_WITH_ARG("--pool-kernelfile",
+      set_pool_kernelfile, NULL, NULL,
+      "Set the pool's algorithm kernel source file (without file extension)."),
   OPT_WITH_ARG("--pool-lookup-gap",
       set_pool_lookup_gap, NULL, NULL,
       "Set Pool GPU lookup gap for scrypt mining, comma separated"),
@@ -1581,6 +1597,9 @@ struct opt_table opt_config_table[] = {
   OPT_WITH_ARG("--profile-device",
       set_profile_devices, NULL, NULL,
       "Select devices to use with profile, one value, range and/or comma separated (e.g. 0-2,4) default: all"),
+  OPT_WITH_ARG("--profile-kernelfile",
+      set_profile_kernelfile, NULL, NULL,
+      "Set the profile's algorithm kernel source file (without file extension)."),
   OPT_WITH_ARG("--profile-lookup-gap",
       set_profile_lookup_gap, NULL, NULL,
       "Set Profile GPU lookup gap for scrypt mining, comma separated"),
@@ -2865,9 +2884,10 @@ static void show_hash(struct work *work, char *hashshow)
      work->block? " BLOCK!" : "");
   } else {
     swab256(rhash, work->hash);
-    for (ofs = 0; ofs <= 28; ofs ++) {
-      if (rhash[ofs])
+    for (ofs = 0; ofs <= 28; ++ofs) {
+      if (rhash[ofs]) {
         break;
+      }
     }
     hash32 = (uint32_t *)(rhash + ofs);
     h32 = be32toh(*hash32);
@@ -3277,9 +3297,9 @@ static void calc_diff(struct work *work, double known)
   struct sgminer_pool_stats *pool_stats = &(work->pool->sgminer_pool_stats);
   double difficulty;
 
-  if (known)
+  if (known) {
     work->work_difficulty = known;
-  else {
+  } else {
     double d64, dcut64;
 
     d64 = work->pool->algorithm.diff_multiplier2 * truediffone;