From 11c5ec810d5dcafb0fcf2f3a6b6e0eaca7b152cf Mon Sep 17 00:00:00 2001
From: Tanguy Pruvot <tanguy.pruvot@gmail.com>
Date: Sun, 9 Nov 2014 21:55:19 +0100
Subject: [PATCH] Handle intensity param in all algos

and add a check related to start/max nounce params
---
 JHA/jackpotcoin.cu  |  6 +++---
 blake32.cu          |  7 +++----
 cuda_nist5.cu       |  6 +++---
 heavy/heavy.cu      | 10 +++++-----
 keccak/keccak256.cu |  8 +++-----
 quark/animecoin.cu  |  3 ++-
 quark/quarkcoin.cu  |  4 +++-
 qubit/deep.cu       |  3 ++-
 qubit/doom.cu       |  3 ++-
 qubit/qubit.cu      |  5 +++--
 x11/fresh.cu        |  4 +++-
 x11/s3.cu           |  9 +++++----
 x11/x11.cu          |  3 ++-
 x13/x13.cu          |  3 ++-
 x15/whirlpool.cu    |  8 ++++----
 x15/x14.cu          |  8 +++++---
 x15/x15.cu          |  8 +++++---
 x17/x17.cu          |  7 ++++---
 18 files changed, 59 insertions(+), 46 deletions(-)

diff --git a/JHA/jackpotcoin.cu b/JHA/jackpotcoin.cu
index 804e609..9050edb 100644
--- a/JHA/jackpotcoin.cu
+++ b/JHA/jackpotcoin.cu
@@ -97,9 +97,8 @@ extern "C" int scanhash_jackpot(int thr_id, uint32_t *pdata,
 	if (opt_benchmark)
 		((uint32_t*)ptarget)[7] = 0x0000ff;
 
-	const uint32_t Htarg = ptarget[7];
-
-	const int throughput = 256*4096*4; // 100;
+	int throughput = opt_work_size ? opt_work_size : (1 << 22); // 256*4096*4
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	if (!init[thr_id])
@@ -212,6 +211,7 @@ extern "C" int scanhash_jackpot(int thr_id, uint32_t *pdata,
 		{
 			unsigned int rounds;
 			uint32_t vhash64[8];
+			uint32_t Htarg = ptarget[7];
 			be32enc(&endiandata[19], foundNonce);
 
 			// diese jackpothash Funktion gibt die Zahl der Runden zurück
diff --git a/blake32.cu b/blake32.cu
index 1fb763a..c5a2197 100644
--- a/blake32.cu
+++ b/blake32.cu
@@ -16,7 +16,6 @@ extern "C" {
 
 /* threads per block and throughput (intensity) */
 #define TPB 128
-#define INTENSITY (1 << 20) // = 1048576 nonces per call
 
 /* added in sph_blake.c */
 extern "C" int blake256_rounds = 14;
@@ -393,15 +392,15 @@ extern "C" int scanhash_blake256(int thr_id, uint32_t *pdata, const uint32_t *pt
 {
 	const uint32_t first_nonce = pdata[19];
 	static bool init[8] = { 0, 0, 0, 0, 0, 0, 0, 0 };
-	uint64_t targetHigh = ((uint64_t*)ptarget)[3]; // 0x00000000.0fffffff
+	uint64_t targetHigh = ((uint64_t*)ptarget)[3];
 	uint32_t _ALIGN(64) endiandata[20];
 #if PRECALC64
 	uint32_t _ALIGN(64) midstate[8];
 #else
 	uint32_t crcsum;
 #endif
-	/* todo: -i param */
-	uint32_t throughput = min(INTENSITY, max_nonce - first_nonce);
+	uint32_t throughput = opt_work_size ? opt_work_size : (1 << 20); // 1048576 nonces per call
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	int rc = 0;
 
diff --git a/cuda_nist5.cu b/cuda_nist5.cu
index 1189f18..5a41b0d 100644
--- a/cuda_nist5.cu
+++ b/cuda_nist5.cu
@@ -77,9 +77,8 @@ extern "C" int scanhash_nist5(int thr_id, uint32_t *pdata,
 	if (opt_benchmark)
 		((uint32_t*)ptarget)[7] = 0x00FF;
 
-	const uint32_t Htarg = ptarget[7];
-
-	const int throughput = 256*4096; // 100;
+	int throughput = opt_work_size ? opt_work_size : (1 << 20); // 256*4096
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	if (!init[thr_id])
@@ -119,6 +118,7 @@ extern "C" int scanhash_nist5(int thr_id, uint32_t *pdata,
 		if  (foundNonce != 0xffffffff)
 		{
 			uint32_t vhash64[8];
+			uint32_t Htarg = ptarget[7];
 			be32enc(&endiandata[19], foundNonce);
 			nist5hash(vhash64, endiandata);
 
diff --git a/heavy/heavy.cu b/heavy/heavy.cu
index f5a71f2..3221e77 100644
--- a/heavy/heavy.cu
+++ b/heavy/heavy.cu
@@ -281,8 +281,10 @@ int scanhash_heavy_cpp(int thr_id, uint32_t *pdata,
  const uint32_t *ptarget, uint32_t max_nonce,
  unsigned long *hashes_done, uint32_t maxvote, int blocklen)
 {
+    const uint32_t first_nonce = pdata[19]; /* to check */
     // CUDA will process thousands of threads.
-    const int throughput = 4096 * 128;
+    int throughput = opt_work_size ? opt_work_size : (1 << 19); // 128*4096
+    throughput = min(throughput, max_nonce - first_nonce);
 
     if (opt_benchmark)
         ((uint32_t*)ptarget)[7] = 0x000000ff;
@@ -296,8 +298,6 @@ int scanhash_heavy_cpp(int thr_id, uint32_t *pdata,
     int nrmCalls[6];
     memset(nrmCalls, 0, sizeof(int) * 6);
 
-    uint32_t start_nonce = pdata[19];    
-
     // für jeden Hash ein individuelles Target erstellen basierend
     // auf dem höchsten Bit, das in ptarget gesetzt ist.
     int highbit = findhighbit(ptarget, 8);
@@ -418,7 +418,7 @@ int scanhash_heavy_cpp(int thr_id, uint32_t *pdata,
                         }
                         else
                         {
-                            *hashes_done = pdata[19] - start_nonce;
+                            *hashes_done = pdata[19] - first_nonce;
                             rc = 1;
                             goto exit;
                         }
@@ -432,7 +432,7 @@ emptyNonceVector:
         pdata[19] += throughput;
 
     } while (pdata[19] < max_nonce && !work_restart[thr_id].restart);
-    *hashes_done = pdata[19] - start_nonce;
+    *hashes_done = pdata[19] - first_nonce;
 
 exit:
     cudaFreeHost(cpu_nonceVector);
diff --git a/keccak/keccak256.cu b/keccak/keccak256.cu
index 52108d0..3d4c8cc 100644
--- a/keccak/keccak256.cu
+++ b/keccak/keccak256.cu
@@ -46,9 +46,8 @@ extern "C" int scanhash_keccak256(int thr_id, uint32_t *pdata,
 	if (opt_benchmark)
 		((uint32_t*)ptarget)[7] = 0x000f;
 
-	const uint32_t Htarg = ptarget[7];
-
-	const int throughput = 256*256*8*8;
+	int throughput = opt_work_size ? opt_work_size : (1 << 22); // 256*256*8*8
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	if (!init[thr_id]) {
@@ -72,10 +71,9 @@ extern "C" int scanhash_keccak256(int thr_id, uint32_t *pdata,
 		uint32_t foundNonce = keccak256_cpu_hash_80(thr_id, throughput, pdata[19], d_hash[thr_id], order++);
 		if  (foundNonce != 0xffffffff)
 		{
-
 			uint32_t vhash64[8];
+			uint32_t Htarg = ptarget[7];
 			be32enc(&endiandata[19], foundNonce);
-
 			keccak256_hash(vhash64, endiandata);
 
 			if (vhash64[7] <= Htarg && fulltest(vhash64, ptarget)) {
diff --git a/quark/animecoin.cu b/quark/animecoin.cu
index bb4f7da..b461e92 100644
--- a/quark/animecoin.cu
+++ b/quark/animecoin.cu
@@ -170,7 +170,8 @@ extern "C" int scanhash_anime(int thr_id, uint32_t *pdata,
 		((uint32_t*)ptarget)[7] = 0x00000f;
 
 	const uint32_t Htarg = ptarget[7];
-	const int throughput = 256*2048; // 100;
+	int throughput = opt_work_size ? opt_work_size : (1 << 20); // 256*2048*2
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	if (!init[thr_id])
diff --git a/quark/quarkcoin.cu b/quark/quarkcoin.cu
index 1adac57..2604461 100644
--- a/quark/quarkcoin.cu
+++ b/quark/quarkcoin.cu
@@ -135,9 +135,11 @@ extern "C" int scanhash_quark(int thr_id, uint32_t *pdata,
     unsigned long *hashes_done)
 {
 	const uint32_t first_nonce = pdata[19];
-	const int throughput = 256*4096; // 100;
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 
+	int throughput = opt_work_size ? opt_work_size : (1 << 20); // 256*4096
+	throughput = min(throughput, max_nonce - first_nonce);
+
 	if (opt_benchmark)
 		((uint32_t*)ptarget)[7] = 0x00FF;
 
diff --git a/qubit/deep.cu b/qubit/deep.cu
index 8e9a8be..538a851 100644
--- a/qubit/deep.cu
+++ b/qubit/deep.cu
@@ -60,9 +60,10 @@ extern "C" int scanhash_deep(int thr_id, uint32_t *pdata,
 	unsigned long *hashes_done)
 {
 	const uint32_t first_nonce = pdata[19];
-	const int throughput = 256*256*8*8;
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	uint32_t endiandata[20];
+	int throughput = opt_work_size ? opt_work_size : (1 << 22); // 256*256*8*8
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	if (opt_benchmark)
 		((uint32_t*)ptarget)[7] = 0x0000ff;
diff --git a/qubit/doom.cu b/qubit/doom.cu
index 8f93ac4..f8254b3 100644
--- a/qubit/doom.cu
+++ b/qubit/doom.cu
@@ -40,9 +40,10 @@ extern "C" int scanhash_doom(int thr_id, uint32_t *pdata,
 	unsigned long *hashes_done)
 {
 	const uint32_t first_nonce = pdata[19];
-	const int throughput = 256*256*8*8;
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	uint32_t endiandata[20];
+	int throughput = opt_work_size ? opt_work_size : (1 << 22); // 256*256*8*8
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	if (opt_benchmark)
 		((uint32_t*)ptarget)[7] = 0x0000ff;
diff --git a/qubit/qubit.cu b/qubit/qubit.cu
index 11eb0d1..45b0d14 100644
--- a/qubit/qubit.cu
+++ b/qubit/qubit.cu
@@ -78,10 +78,11 @@ extern "C" int scanhash_qubit(int thr_id, uint32_t *pdata,
 	const uint32_t *ptarget, uint32_t max_nonce,
 	unsigned long *hashes_done)
 {
-	const uint32_t first_nonce = pdata[19];
-	const int throughput = 256*256*8;
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	uint32_t endiandata[20];
+	const uint32_t first_nonce = pdata[19];
+	int throughput = opt_work_size ? opt_work_size : (1 << 19); // 256*256*8
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	if (opt_benchmark)
 		((uint32_t*)ptarget)[7] = 0x0000ff;
diff --git a/x11/fresh.cu b/x11/fresh.cu
index 0c5179a..8aa1214 100644
--- a/x11/fresh.cu
+++ b/x11/fresh.cu
@@ -75,10 +75,12 @@ extern "C" int scanhash_fresh(int thr_id, uint32_t *pdata,
 	unsigned long *hashes_done)
 {
 	const uint32_t first_nonce = pdata[19];
-	const int throughput = 256*256*8;
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	uint32_t endiandata[20];
 
+	int throughput = opt_work_size ? opt_work_size : (1 << 19); // 256*256*8;
+	throughput = min(throughput, max_nonce - first_nonce);
+
 	if (opt_benchmark)
 		((uint32_t*)ptarget)[7] = 0x00ff;
 
diff --git a/x11/s3.cu b/x11/s3.cu
index 08f4018..24e73bb 100644
--- a/x11/s3.cu
+++ b/x11/s3.cu
@@ -57,14 +57,15 @@ extern "C" int scanhash_s3(int thr_id, uint32_t *pdata,
 	unsigned long *hashes_done)
 {
 	const uint32_t first_nonce = pdata[19];
+	static bool init[8] = { 0, 0, 0, 0, 0, 0, 0, 0 };
+
 #ifdef WIN32
 	// reduce a bit the intensity on windows
-	const int throughput = 256 * 256 * 8;
+	int throughput = opt_work_size ? opt_work_size : (1 << 19); // 256*256*8;
 #else
-	const int throughput = 256 * 256 * 8 * 2;
+	int throughput = opt_work_size ? opt_work_size : (1 << 20); // 256*256*8*2;
 #endif
-
-	static bool init[8] = { 0, 0, 0, 0, 0, 0, 0, 0 };
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	if (opt_benchmark)
 		((uint32_t*)ptarget)[7] = 0xF;
diff --git a/x11/x11.cu b/x11/x11.cu
index 4834130..6dad14c 100644
--- a/x11/x11.cu
+++ b/x11/x11.cu
@@ -138,8 +138,9 @@ extern "C" int scanhash_x11(int thr_id, uint32_t *pdata,
     unsigned long *hashes_done)
 {
 	const uint32_t first_nonce = pdata[19];
-	const int throughput = 256*256*8;
 	static bool init[8] = {0,0,0,0,0,0,0,0};
+	int throughput = opt_work_size ? opt_work_size : (1 << 19); // 256*256*8;
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	if (opt_benchmark)
 		((uint32_t*)ptarget)[7] = 0x0000f;
diff --git a/x13/x13.cu b/x13/x13.cu
index c751211..747ec92 100644
--- a/x13/x13.cu
+++ b/x13/x13.cu
@@ -162,7 +162,8 @@ extern "C" int scanhash_x13(int thr_id, uint32_t *pdata,
 
 	const uint32_t Htarg = ptarget[7];
 
-	const int throughput = 256*256*8;
+	int throughput = opt_work_size ? opt_work_size : (1 << 19); // 256*256*8;
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	if (!init[thr_id])
diff --git a/x15/whirlpool.cu b/x15/whirlpool.cu
index 192a71c..c9575ed 100644
--- a/x15/whirlpool.cu
+++ b/x15/whirlpool.cu
@@ -56,13 +56,13 @@ extern "C" int scanhash_whc(int thr_id, uint32_t *pdata,
     unsigned long *hashes_done)
 {
 	const uint32_t first_nonce = pdata[19];
-	const int throughput = 256*256*8;
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	uint32_t endiandata[20];
-	uint32_t Htarg = ptarget[7];
+	int throughput = opt_work_size ? opt_work_size : (1 << 19); // 256*256*8;
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	if (opt_benchmark)
-		((uint32_t*)ptarget)[7] = Htarg = 0x0000ff;
+		((uint32_t*)ptarget)[7] = 0x0000ff;
 
 	if (!init[thr_id]) {
 		cudaSetDevice(device_map[thr_id]);
@@ -91,8 +91,8 @@ extern "C" int scanhash_whc(int thr_id, uint32_t *pdata,
 		if (foundNonce != 0xffffffff)
 		{
 			uint32_t vhash64[8];
+			uint32_t Htarg = ptarget[7];
 			be32enc(&endiandata[19], foundNonce);
-
 			wcoinhash(vhash64, endiandata);
 
 			if (vhash64[7] <= Htarg && fulltest(vhash64, ptarget))
diff --git a/x15/x14.cu b/x15/x14.cu
index 7335a0d..b24a61d 100644
--- a/x15/x14.cu
+++ b/x15/x14.cu
@@ -168,13 +168,14 @@ extern "C" int scanhash_x14(int thr_id, uint32_t *pdata,
 	unsigned long *hashes_done)
 {
 	const uint32_t first_nonce = pdata[19];
-	const int throughput = 256*256*8;
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	uint32_t endiandata[20];
-	uint32_t Htarg = ptarget[7];
+
+	int throughput = opt_work_size ? opt_work_size : (1 << 19); // 256*256*8;
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	if (opt_benchmark)
-		((uint32_t*)ptarget)[7] = Htarg = 0xff;
+		((uint32_t*)ptarget)[7] = 0xff;
 
 	if (!init[thr_id])
 	{
@@ -230,6 +231,7 @@ extern "C" int scanhash_x14(int thr_id, uint32_t *pdata,
 			uint32_t vhash64[8];
 			be32enc(&endiandata[19], foundNonce);
 			x14hash(vhash64, endiandata);
+			uint32_t Htarg = ptarget[7];
 			if (vhash64[7] <= Htarg && fulltest(vhash64, ptarget)) {
 				pdata[19] = foundNonce;
 				*hashes_done = foundNonce - first_nonce + 1;
diff --git a/x15/x15.cu b/x15/x15.cu
index 012cafd..ec73805 100644
--- a/x15/x15.cu
+++ b/x15/x15.cu
@@ -177,13 +177,14 @@ extern "C" int scanhash_x15(int thr_id, uint32_t *pdata,
 	unsigned long *hashes_done)
 {
 	const uint32_t first_nonce = pdata[19];
-	const int throughput = 256*256*8;
 	static bool init[8] = {0,0,0,0,0,0,0,0};
 	uint32_t endiandata[20];
-	uint32_t Htarg = ptarget[7];
+
+	int throughput = opt_work_size ? opt_work_size : (1 << 19); // 256*256*8;
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	if (opt_benchmark)
-		((uint32_t*)ptarget)[7] = Htarg = 0x00FF;
+		((uint32_t*)ptarget)[7] = 0x00FF;
 
 	if (!init[thr_id])
 	{
@@ -241,6 +242,7 @@ extern "C" int scanhash_x15(int thr_id, uint32_t *pdata,
 		{
 			/* check now with the CPU to confirm */
 			uint32_t vhash64[8];
+			uint32_t Htarg = ptarget[7];
 			be32enc(&endiandata[19], foundNonce);
 			x15hash(vhash64, endiandata);
 			if (vhash64[7] <= Htarg && fulltest(vhash64, ptarget)) {
diff --git a/x17/x17.cu b/x17/x17.cu
index 608ce60..e4573a5 100644
--- a/x17/x17.cu
+++ b/x17/x17.cu
@@ -197,12 +197,12 @@ extern "C" int scanhash_x17(int thr_id, uint32_t *pdata,
 	unsigned long *hashes_done)
 {
 	const uint32_t first_nonce = pdata[19];
-	const int throughput = 256*256*8;
 	static bool init[8] = {0,0,0,0,0,0,0,0};
-	uint32_t Htarg = ptarget[7];
+	int throughput = opt_work_size ? opt_work_size : (1 << 19); // 256*256*8;
+	throughput = min(throughput, max_nonce - first_nonce);
 
 	if (opt_benchmark)
-		((uint32_t*)ptarget)[7] = Htarg = 0x00FF;
+		((uint32_t*)ptarget)[7] = 0x00FF;
 
 	if (!init[thr_id])
 	{
@@ -265,6 +265,7 @@ extern "C" int scanhash_x17(int thr_id, uint32_t *pdata,
 		if (foundNonce != 0xffffffff)
 		{
 			uint32_t vhash64[8];
+			uint32_t Htarg = ptarget[7];
 			be32enc(&endiandata[19], foundNonce);
 			x17hash(vhash64, endiandata);