update c11 like tribus + 2.2.1 readme

README.txt

@@ -1,5 +1,5 @@
 
-ccminer 2.2 (August 2017) "Equihash, tribus and optimized skunk"
+ccminer 2.2.1 (Sept. 2017) "optimized tribus kernel (Maxwell+)"
 ---------------------------------------------------------------
 
 ***************************************************************
@@ -277,6 +277,8 @@ so we can more efficiently implement new algorithms using the latest hardware
 features.
 
 >>> RELEASE HISTORY <<<
+  Sep. 01st 2017  v2.2.1
+                  Improve tribus algo on recent cards (up to +10%)
 
   Aug. 13th 2017  v2.2
                   New skunk algo, using the heavy streebog algorithm

compat/ccminer-config.h

@@ -164,7 +164,7 @@
 #define PACKAGE_URL "http://github.com/tpruvot/ccminer"
 
 /* Define to the version of this package. */
-#define PACKAGE_VERSION "2.2"
+#define PACKAGE_VERSION "2.2.1"
 
 /* If using the C implementation of alloca, define if you know the
    direction of stack growth for your system; otherwise it will be

tribus/tribus.cu

@@ -146,6 +146,7 @@ extern "C" int scanhash_tribus(int thr_id, struct work *work, uint32_t max_nonce
 				gpu_increment_reject(thr_id);
 				if (!opt_quiet)
 					gpulog(LOG_WARNING, thr_id, "result for %08x does not validate on CPU!", work->nonces[0]);
+				cudaMemset(d_resNonce[thr_id], 0xFF, 2 * sizeof(uint32_t));
 				pdata[19] = work->nonces[0] + 1;
 				continue;
 			}

x11/c11.cu

@@ -18,10 +18,13 @@ extern "C"
 #include "cuda_helper.h"
 #include "cuda_x11.h"
 
+void tribus_echo512_final(int thr_id, uint32_t threads, uint32_t *d_hash, uint32_t *d_resNonce, const uint64_t target);
+
 #include <stdio.h>
 #include <memory.h>
 
 static uint32_t *d_hash[MAX_GPUS];
+static uint32_t *d_resNonce[MAX_GPUS];
 
 // Flax/Chaincoin C11 CPU Hash
 extern "C" void c11hash(void *output, const void *input)
@@ -103,6 +106,7 @@ extern "C" void c11hash(void *output, const void *input)
 #endif
 
 static bool init[MAX_GPUS] = { 0 };
+static bool use_compat_kernels[MAX_GPUS] = { 0 };
 
 extern "C" int scanhash_c11(int thr_id, struct work* work, uint32_t max_nonce, unsigned long *hashes_done)
 {
@@ -118,7 +122,8 @@ extern "C" int scanhash_c11(int thr_id, struct work* work, uint32_t max_nonce, u
 
 	if (!init[thr_id])
 	{
-		cudaSetDevice(device_map[thr_id]);
+		int dev_id = device_map[thr_id];
+		cudaSetDevice(dev_id);
 		if (opt_cudaschedule == -1 && gpu_threads == 1) {
 			cudaDeviceReset();
 			// reduce cpu usage
@@ -127,6 +132,9 @@ extern "C" int scanhash_c11(int thr_id, struct work* work, uint32_t max_nonce, u
 		}
 		gpulog(LOG_INFO, thr_id, "Intensity set to %g, %u cuda threads", throughput2intensity(throughput), throughput);
 
+		cuda_get_arch(thr_id);
+		use_compat_kernels[thr_id] = (cuda_arch[dev_id] < 500);
+
 		quark_blake512_cpu_init(thr_id, throughput);
 		quark_bmw512_cpu_init(thr_id, throughput);
 		quark_groestl512_cpu_init(thr_id, throughput);
@@ -135,11 +143,13 @@ extern "C" int scanhash_c11(int thr_id, struct work* work, uint32_t max_nonce, u
 		quark_jh512_cpu_init(thr_id, throughput);
 		x11_luffaCubehash512_cpu_init(thr_id, throughput);
 		x11_shavite512_cpu_init(thr_id, throughput);
+		if (use_compat_kernels[thr_id])
 			x11_echo512_cpu_init(thr_id, throughput);
 		if (x11_simd512_cpu_init(thr_id, throughput) != 0) {
 			return 0;
 		}
-		CUDA_CALL_OR_RET_X(cudaMalloc(&d_hash[thr_id], 64 * throughput), 0); // why 64 ?
+		CUDA_CALL_OR_RET_X(cudaMalloc(&d_hash[thr_id], 64 * throughput), 0);
+		CUDA_SAFE_CALL(cudaMalloc(&d_resNonce[thr_id], 2 * sizeof(uint32_t)));
 
 		cuda_check_cpu_init(thr_id, throughput);
 
@@ -151,7 +161,10 @@ extern "C" int scanhash_c11(int thr_id, struct work* work, uint32_t max_nonce, u
 		be32enc(&endiandata[k], pdata[k]);
 
 	quark_blake512_cpu_setBlock_80(thr_id, endiandata);
+	if (use_compat_kernels[thr_id])
 		cuda_check_cpu_setTarget(ptarget);
+	else
+		cudaMemset(d_resNonce[thr_id], 0xFF, 2 * sizeof(uint32_t));
 
 	do {
 		int order = 0;
@@ -175,24 +188,32 @@ extern "C" int scanhash_c11(int thr_id, struct work* work, uint32_t max_nonce, u
 		TRACE("shavite:");
 		x11_simd512_cpu_hash_64(thr_id, throughput, pdata[19], NULL, d_hash[thr_id], order++);
 		TRACE("simd   :");
+
+		if (use_compat_kernels[thr_id]) {
 			x11_echo512_cpu_hash_64(thr_id, throughput, pdata[19], NULL, d_hash[thr_id], order++);
-		TRACE("echo => ");
+			work->nonces[0] = cuda_check_hash(thr_id, throughput, pdata[19], d_hash[thr_id]);
+			work->nonces[1] = UINT32_MAX;
+		} else {
+			tribus_echo512_final(thr_id, throughput, d_hash[thr_id], d_resNonce[thr_id], AS_U64(&ptarget[6]));
+			cudaMemcpy(&work->nonces[0], d_resNonce[thr_id], 2 * sizeof(uint32_t), cudaMemcpyDeviceToHost);
+		}
 
 		*hashes_done = pdata[19] - first_nonce + throughput;
 
-		work->nonces[0] = cuda_check_hash(thr_id, throughput, pdata[19], d_hash[thr_id]);
 		if (work->nonces[0] != UINT32_MAX)
 		{
-			const uint32_t Htarg = ptarget[7];
 			uint32_t _ALIGN(64) vhash[8];
+			const uint32_t Htarg = ptarget[7];
+			const uint32_t startNounce = pdata[19];
+			if (!use_compat_kernels[thr_id]) work->nonces[0] += startNounce;
 			be32enc(&endiandata[19], work->nonces[0]);
 			c11hash(vhash, endiandata);
 
 			if (vhash[7] <= Htarg && fulltest(vhash, ptarget)) {
 				work->valid_nonces = 1;
 				work_set_target_ratio(work, vhash);
-				work->nonces[1] = cuda_check_hash_suppl(thr_id, throughput, pdata[19], d_hash[thr_id], 1);
+				if (work->nonces[1] != UINT32_MAX) {
-				if (work->nonces[1] != 0) {
+					work->nonces[1] += startNounce;
 					be32enc(&endiandata[19], work->nonces[1]);
 					c11hash(vhash, endiandata);
 					bn_set_target_ratio(work, vhash, 1);
@@ -207,6 +228,7 @@ extern "C" int scanhash_c11(int thr_id, struct work* work, uint32_t max_nonce, u
 				gpu_increment_reject(thr_id);
 				if (!opt_quiet)
 					gpulog(LOG_WARNING, thr_id, "result for %08x does not validate on CPU!", work->nonces[0]);
+				cudaMemset(d_resNonce[thr_id], 0xFF, 2 * sizeof(uint32_t));
 				pdata[19] = work->nonces[0] + 1;
 				continue;
 			}
@@ -234,6 +256,8 @@ extern "C" void free_c11(int thr_id)
 	cudaThreadSynchronize();
 
 	cudaFree(d_hash[thr_id]);
+	cudaFree(d_resNonce[thr_id]);
+
 	quark_blake512_cpu_free(thr_id);
 	quark_groestl512_cpu_free(thr_id);
 	x11_simd512_cpu_free(thr_id);