myr-gr: remove unused allocated memory + pascal tweak

+ cleanup...

cuda_myriadgroestl.cu

@@ -5,6 +5,11 @@
 
 #include "cuda_helper.h"
 
+#ifdef __INTELLISENSE__
+#define __CUDA_ARCH__ 500
+#define __funnelshift_r(x,y,n) (x >> n)
+#endif
+
 #if __CUDA_ARCH__ >= 300
 // 64 Registers Variant for Compute 3.0
 #include "quark/groestl_functions_quad.h"
@@ -21,11 +26,8 @@ __constant__ uint32_t myriadgroestl_gpu_msg[32];
 // muss expandiert werden
 __constant__ uint32_t myr_sha256_gpu_constantTable[64];
 __constant__ uint32_t myr_sha256_gpu_constantTable2[64];
-__constant__ uint32_t myr_sha256_gpu_hashTable[8];
 
-uint32_t myr_sha256_cpu_hashTable[] = {
-    0x6a09e667, 0xbb67ae85, 0x3c6ef372, 0xa54ff53a, 0x510e527f, 0x9b05688c, 0x1f83d9ab, 0x5be0cd19 };
-uint32_t myr_sha256_cpu_constantTable[] = {
+const uint32_t myr_sha256_cpu_constantTable[] = {
 	0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5, 0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
 	0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3, 0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
 	0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc, 0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
@@ -36,7 +38,7 @@ uint32_t myr_sha256_cpu_constantTable[] = {
 	0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208, 0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2,
 };
 
-uint32_t myr_sha256_cpu_w2Table[] = {
+const uint32_t myr_sha256_cpu_w2Table[] = {
 	0x80000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
 	0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000200,
 	0x80000000, 0x01400000, 0x00205000, 0x00005088, 0x22000800, 0x22550014, 0x05089742, 0xa0000020,
@@ -44,9 +46,10 @@ uint32_t myr_sha256_cpu_w2Table[] = {
 	0x08b2b050, 0x9d7c4c27, 0x0ce2a393, 0x88e6e1ea, 0xa52b4335, 0x67a16f49, 0xd732016f, 0x4eeb2e91,
 	0x5dbf55e5, 0x8eee2335, 0xe2bc5ec2, 0xa83f4394, 0x45ad78f7, 0x36f3d0cd, 0xd99c05e8, 0xb0511dc7,
 	0x69bc7ac4, 0xbd11375b, 0xe3ba71e5, 0x3b209ff2, 0x18feee17, 0xe25ad9e7, 0x13375046, 0x0515089d,
-    0x4f0d0f04, 0x2627484e, 0x310128d2, 0xc668b434, 0x420841cc, 0x62d311b8, 0xe59ba771, 0x85a7a484 };
+	0x4f0d0f04, 0x2627484e, 0x310128d2, 0xc668b434, 0x420841cc, 0x62d311b8, 0xe59ba771, 0x85a7a484
+};
 
-#define SWAB32(x)        ( ((x & 0x000000FF) << 24) | ((x & 0x0000FF00) << 8) | ((x & 0x00FF0000) >> 8) | ((x & 0xFF000000) >> 24) )
+#define SWAB32(x) cuda_swab32(x)
 
 #if __CUDA_ARCH__ < 320
 	// Kepler (Compute 3.0)
@@ -55,6 +58,7 @@ uint32_t myr_sha256_cpu_w2Table[] = {
 	// Kepler (Compute 3.5)
 	#define ROTR32(x, n) __funnelshift_r( (x), (x), (n) )
 #endif
+
 #define R(x, n)         ((x) >> (n))
 #define Ch(x, y, z)     ((x & (y ^ z)) ^ z)
 #define Maj(x, y, z)    ((x & (y | z)) | (y & z))
@@ -65,28 +69,27 @@ uint32_t myr_sha256_cpu_w2Table[] = {
 
 __device__ void myriadgroestl_gpu_sha256(uint32_t *message)
 {
-    uint32_t W1[16];
-    uint32_t W2[16];
-
-    // Initialisiere die register a bis h mit der Hash-Tabelle
-    uint32_t regs[8];
-    uint32_t hash[8];
+	uint32_t regs[8], hash[8];
+	const uint32_t myr_sha256_gpu_hashTable[8] = {
+		0x6a09e667, 0xbb67ae85, 0x3c6ef372, 0xa54ff53a, 0x510e527f, 0x9b05688c, 0x1f83d9ab, 0x5be0cd19
+	};
 
 	// pre
-#pragma unroll 8
+	#pragma unroll 8
 	for (int k=0; k < 8; k++)
 	{
 		regs[k] = myr_sha256_gpu_hashTable[k];
 		hash[k] = regs[k];
 	}
 
-#pragma unroll 16
-    for(int k=0;k<16;k++)
+	uint32_t W1[16];
+	#pragma unroll 16
+	for(int k=0; k<16; k++)
 		W1[k] = SWAB32(message[k]);
 
-// Progress W1
-#pragma unroll 16
-    for(int j=0;j<16;j++)
+	// Progress W1
+	#pragma unroll 16
+	for(int j=0; j<16; j++)
 	{
 		uint32_t T1, T2;
 		T1 = regs[7] + S1(regs[4]) + Ch(regs[4], regs[5], regs[6]) + myr_sha256_gpu_constantTable[j] + W1[j];
@@ -98,24 +101,27 @@ __device__ void myriadgroestl_gpu_sha256(uint32_t *message)
 		regs[4] += T1;
 	}
 
-// Progress W2...W3
-////// PART 1
-#pragma unroll 2
-    for(int j=0;j<2;j++)
+	// Progress W2...W3
+	uint32_t W2[16];
+
+	////// PART 1
+	#pragma unroll 2
+	for(int j=0; j<2; j++)
 		W2[j] = s1(W1[14+j]) + W1[9+j] + s0(W1[1+j]) + W1[j];
-#pragma unroll 5
+
+	#pragma unroll 5
 	for(int j=2;j<7;j++)
 		W2[j] = s1(W2[j-2]) + W1[9+j] + s0(W1[1+j]) + W1[j];
 
-#pragma unroll 8
-    for(int j=7;j<15;j++)
+	#pragma unroll 8
+	for(int j=7; j<15; j++)
 		W2[j] = s1(W2[j-2]) + W2[j-7] + s0(W1[1+j]) + W1[j];
 
 	W2[15] = s1(W2[13]) + W2[8] + s0(W2[0]) + W1[15];
 
-    // Rundenfunktion
-#pragma unroll 16
-    for(int j=0;j<16;j++)
+	// Round function
+	#pragma unroll 16
+	for(int j=0; j<16; j++)
 	{
 		uint32_t T1, T2;
 		T1 = regs[7] + S1(regs[4]) + Ch(regs[4], regs[5], regs[6]) + myr_sha256_gpu_constantTable[j + 16] + W2[j];
@@ -127,23 +133,23 @@ __device__ void myriadgroestl_gpu_sha256(uint32_t *message)
 		regs[4] += T1;
 	}
 
-////// PART 2
-#pragma unroll 2
-    for(int j=0;j<2;j++)
+	////// PART 2
+	#pragma unroll 2
+	for(int j=0; j<2; j++)
 		W1[j] = s1(W2[14+j]) + W2[9+j] + s0(W2[1+j]) + W2[j];
-#pragma unroll 5
-    for(int j=2;j<7;j++)
+	#pragma unroll 5
+	for(int j=2; j<7; j++)
 		W1[j] = s1(W1[j-2]) + W2[9+j] + s0(W2[1+j]) + W2[j];
 
-#pragma unroll 8
-    for(int j=7;j<15;j++)
+	#pragma unroll 8
+	for(int j=7; j<15; j++)
 		W1[j] = s1(W1[j-2]) + W1[j-7] + s0(W2[1+j]) + W2[j];
 
 	W1[15] = s1(W1[13]) + W1[8] + s0(W1[0]) + W2[15];
 
-    // Rundenfunktion
-#pragma unroll 16
-    for(int j=0;j<16;j++)
+	// Round function
+	#pragma unroll 16
+	for(int j=0; j<16; j++)
 	{
 		uint32_t T1, T2;
 		T1 = regs[7] + S1(regs[4]) + Ch(regs[4], regs[5], regs[6]) + myr_sha256_gpu_constantTable[j + 32] + W1[j];
@@ -155,23 +161,23 @@ __device__ void myriadgroestl_gpu_sha256(uint32_t *message)
 		regs[4] += T1;
 	}
 
-////// PART 3
-#pragma unroll 2
-    for(int j=0;j<2;j++)
+	////// PART 3
+	#pragma unroll 2
+	for(int j=0; j<2; j++)
 		W2[j] = s1(W1[14+j]) + W1[9+j] + s0(W1[1+j]) + W1[j];
-#pragma unroll 5
-    for(int j=2;j<7;j++)
+	#pragma unroll 5
+	for(int j=2; j<7; j++)
 		W2[j] = s1(W2[j-2]) + W1[9+j] + s0(W1[1+j]) + W1[j];
 
-#pragma unroll 8
-    for(int j=7;j<15;j++)
+	#pragma unroll 8
+	for(int j=7; j<15; j++)
 		W2[j] = s1(W2[j-2]) + W2[j-7] + s0(W1[1+j]) + W1[j];
 
 	W2[15] = s1(W2[13]) + W2[8] + s0(W2[0]) + W1[15];
 
-    // Rundenfunktion
-#pragma unroll 16
-    for(int j=0;j<16;j++)
+	// Round function
+	#pragma unroll 16
+	for(int j=0; j<16; j++)
 	{
 		uint32_t T1, T2;
 		T1 = regs[7] + S1(regs[4]) + Ch(regs[4], regs[5], regs[6]) + myr_sha256_gpu_constantTable[j + 48] + W2[j];
@@ -183,20 +189,20 @@ __device__ void myriadgroestl_gpu_sha256(uint32_t *message)
 		regs[4] += T1;
 	}
 
-#pragma unroll 8
-    for(int k=0;k<8;k++)
+	#pragma unroll 8
+	for(int k=0; k<8; k++)
 		hash[k] += regs[k];
 
 	/////
-    ///// Zweite Runde (wegen Msg-Padding)
+	///// 2nd Round (wegen Msg-Padding)
 	/////
-#pragma unroll 8
-    for(int k=0;k<8;k++)
+	#pragma unroll 8
+	for(int k=0; k<8; k++)
 		regs[k] = hash[k];
 
-// Progress W1
-#pragma unroll 64
-    for(int j=0;j<64;j++)
+	// Progress W1
+	#pragma unroll 64
+	for(int j=0; j<64; j++)
 	{
 		uint32_t T1, T2;
 		T1 = regs[7] + S1(regs[4]) + Ch(regs[4], regs[5], regs[6]) + myr_sha256_gpu_constantTable2[j];
@@ -208,14 +214,14 @@ __device__ void myriadgroestl_gpu_sha256(uint32_t *message)
 		regs[4] += T1;
 	}
 
-#pragma unroll 8
-    for(int k=0;k<8;k++)
+	#pragma unroll 8
+	for(int k=0; k<8; k++)
 		hash[k] += regs[k];
 
-    //// FERTIG
+	//// Close
 
-#pragma unroll 8
-    for(int k=0;k<8;k++)
+	#pragma unroll 8
+	for(int k=0; k<8; k++)
 		message[k] = SWAB32(hash[k]);
 }
 
@@ -229,8 +235,9 @@ __global__ void __launch_bounds__(256, 4)
 	{
 		// GROESTL
 		uint32_t paddedInput[8];
-#pragma unroll 8
-        for(int k=0;k<8;k++) paddedInput[k] = myriadgroestl_gpu_msg[4*k+threadIdx.x%4];
+		#pragma unroll 8
+		for(int k=0; k<8; k++)
+			paddedInput[k] = myriadgroestl_gpu_msg[4*k+threadIdx.x%4];
 
 		uint32_t nounce = startNounce + thread;
 		if ((threadIdx.x % 4) == 3)
@@ -249,8 +256,8 @@ __global__ void __launch_bounds__(256, 4)
 		if ((threadIdx.x & 0x03) == 0)
 		{
 			uint32_t *outpHash = &hashBuffer[16 * thread];
-#pragma unroll 16
-            for(int k=0;k<16;k++) outpHash[k] = out_state[k];
+			#pragma unroll 16
+			for(int k=0; k<16; k++) outpHash[k] = out_state[k];
 		}
 	}
 #endif
@@ -267,7 +274,8 @@ __global__ void
 
 		uint32_t out_state[16];
 		uint32_t *inpHash = &hashBuffer[16 * thread];
-#pragma unroll 16
+
+		#pragma unroll 16
 		for (int i=0; i < 16; i++)
 			out_state[i] = inpHash[i];
 
@@ -276,7 +284,7 @@ __global__ void
 		int i, position = -1;
 		bool rc = true;
 
-#pragma unroll 8
+		#pragma unroll 8
 		for (i = 7; i >= 0; i--) {
 			if (out_state[i] > pTarget[i]) {
 				if(position < i) {
@@ -292,8 +300,7 @@ __global__ void
 			 }
 		}
 
-        if(rc == true)
-            if(resNounce[0] > nounce)
+		if(rc && resNounce[0] > nounce)
 			resNounce[0] = nounce;
 	}
 #endif
@@ -303,44 +310,34 @@ __global__ void
 __host__
 void myriadgroestl_cpu_init(int thr_id, uint32_t threads)
 {
-    cudaMemcpyToSymbol( myr_sha256_gpu_hashTable,
-                        myr_sha256_cpu_hashTable,
-                        sizeof(uint32_t) * 8 );
-
-    cudaMemcpyToSymbol( myr_sha256_gpu_constantTable,
-                        myr_sha256_cpu_constantTable,
-                        sizeof(uint32_t) * 64 );
-
-    // zweite CPU-Tabelle bauen und auf die GPU laden
 	uint32_t temp[64];
-    for(int i=0;i<64;i++)
+	for(int i=0; i<64; i++)
 		temp[i] = myr_sha256_cpu_w2Table[i] + myr_sha256_cpu_constantTable[i];
 
 	cudaMemcpyToSymbol( myr_sha256_gpu_constantTable2,
 						temp,
 						sizeof(uint32_t) * 64 );
 
-    // Speicher für Gewinner-Nonce belegen
-    cudaMalloc(&d_resultNonce[thr_id], sizeof(uint32_t));
+	cudaMemcpyToSymbol( myr_sha256_gpu_constantTable,
+						myr_sha256_cpu_constantTable,
+						sizeof(uint32_t) * 64 );
 
-    // Speicher für temporäreHashes
-    cudaMalloc(&d_outputHashes[thr_id], 16*sizeof(uint32_t)*threads);
+	cudaMalloc(&d_outputHashes[thr_id], (size_t) 64 * threads);
+	cudaMalloc(&d_resultNonce[thr_id], sizeof(uint32_t));
 }
 
 __host__
 void myriadgroestl_cpu_free(int thr_id)
 {
-    cudaFree(d_resultNonce[thr_id]);
 	cudaFree(d_outputHashes[thr_id]);
+	cudaFree(d_resultNonce[thr_id]);
 }
 
 __host__
 void myriadgroestl_cpu_setBlock(int thr_id, void *data, void *pTargetIn)
 {
 	// Nachricht expandieren und setzen
-    uint32_t msgBlock[32];
-
-    memset(msgBlock, 0, sizeof(uint32_t) * 32);
+	uint32_t msgBlock[32] = { 0 };
 	memcpy(&msgBlock[0], data, 80);
 
 	// Erweitere die Nachricht auf den Nachrichtenblock (padding)
@@ -352,27 +349,20 @@ void myriadgroestl_cpu_setBlock(int thr_id, void *data, void *pTargetIn)
 	// auf der GPU ausgeführt)
 
 	// Blockheader setzen (korrekte Nonce und Hefty Hash fehlen da drin noch)
-    cudaMemcpyToSymbol( myriadgroestl_gpu_msg,
-                        msgBlock,
-                        128);
+	cudaMemcpyToSymbol(myriadgroestl_gpu_msg, msgBlock, 128);
 
 	cudaMemset(d_resultNonce[thr_id], 0xFF, sizeof(uint32_t));
-    cudaMemcpyToSymbol( pTarget,
-                        pTargetIn,
-                        sizeof(uint32_t) * 8 );
+	cudaMemcpyToSymbol(pTarget, pTargetIn, 32);
 }
 
 __host__
-void myriadgroestl_cpu_hash(int thr_id, uint32_t threads, uint32_t startNounce, void *outputHashes, uint32_t *nounce)
+void myriadgroestl_cpu_hash(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *resNounce)
 {
 	uint32_t threadsperblock = 256;
 
 	// Compute 3.0 benutzt die registeroptimierte Quad Variante mit Warp Shuffle
 	// mit den Quad Funktionen brauchen wir jetzt 4 threads pro Hash, daher Faktor 4 bei der Blockzahl
-    const int factor=4;
-
-    // Größe des dynamischen Shared Memory Bereichs
-    size_t shared_size = 0;
+	const int factor = 4;
 
 	cudaMemset(d_resultNonce[thr_id], 0xFF, sizeof(uint32_t));
 	// berechne wie viele Thread Blocks wir brauchen
@@ -384,12 +374,12 @@ void myriadgroestl_cpu_hash(int thr_id, uint32_t threads, uint32_t startNounce,
 		return;
 	}
 
-    myriadgroestl_gpu_hash_quad<<<grid, block, shared_size>>>(threads, startNounce, d_outputHashes[thr_id]);
+	myriadgroestl_gpu_hash_quad <<< grid, block >>> (threads, startNounce, d_outputHashes[thr_id]);
 	dim3 grid2((threads + threadsperblock-1)/threadsperblock);
-    myriadgroestl_gpu_hash_quad2<<<grid2, block, shared_size>>>(threads, startNounce, d_resultNonce[thr_id], d_outputHashes[thr_id]);
+	myriadgroestl_gpu_hash_quad2 <<< grid2, block >>> (threads, startNounce, d_resultNonce[thr_id], d_outputHashes[thr_id]);
 
 	// Strategisches Sleep Kommando zur Senkung der CPU Last
 	MyStreamSynchronize(NULL, 0, thr_id);
 
-    cudaMemcpy(nounce, d_resultNonce[thr_id], sizeof(uint32_t), cudaMemcpyDeviceToHost);
+	cudaMemcpy(resNounce, d_resultNonce[thr_id], sizeof(uint32_t), cudaMemcpyDeviceToHost);
 }

myriadgroestl.cpp

@@ -10,7 +10,7 @@
 void myriadgroestl_cpu_init(int thr_id, uint32_t threads);
 void myriadgroestl_cpu_free(int thr_id);
 void myriadgroestl_cpu_setBlock(int thr_id, void *data, void *pTargetIn);
-void myriadgroestl_cpu_hash(int thr_id, uint32_t threads, uint32_t startNounce, void *outputHashes, uint32_t *nounce);
+void myriadgroestl_cpu_hash(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *nounce);
 
 void myriadhash(void *state, const void *input)
 {
@@ -37,18 +37,18 @@ int scanhash_myriad(int thr_id, struct work *work, uint32_t max_nonce, unsigned
 	uint32_t *pdata = work->data;
 	uint32_t *ptarget = work->target;
 	uint32_t start_nonce = pdata[19];
-	uint32_t throughput = cuda_default_throughput(thr_id, 1U << 17);
+	int dev_id = device_map[thr_id];
+	int intensity = (device_sm[dev_id] >= 600) ? 20 : 18;
+	uint32_t throughput = cuda_default_throughput(thr_id, 1U << intensity);
 	if (init[thr_id]) throughput = min(throughput, max_nonce - start_nonce);
 
-	uint32_t *outputHash = (uint32_t*)malloc(throughput * 64);
-
 	if (opt_benchmark)
 		ptarget[7] = 0x0000ff;
 
 	// init
 	if(!init[thr_id])
 	{
-		cudaSetDevice(device_map[thr_id]);
+		cudaSetDevice(dev_id);
 		if (opt_cudaschedule == -1 && gpu_threads == 1) {
 			cudaDeviceReset();
 			// reduce cpu usage
@@ -62,14 +62,13 @@ int scanhash_myriad(int thr_id, struct work *work, uint32_t max_nonce, unsigned
 	for (int k=0; k < 20; k++)
 		be32enc(&endiandata[k], pdata[k]);
 
-	// Context mit dem Endian gedrehten Blockheader vorbereiten (Nonce wird später ersetzt)
 	myriadgroestl_cpu_setBlock(thr_id, endiandata, (void*)ptarget);
 
 	do {
 		// GPU
 		uint32_t foundNounce = UINT32_MAX;
 
-		myriadgroestl_cpu_hash(thr_id, throughput, pdata[19], outputHash, &foundNounce);
+		myriadgroestl_cpu_hash(thr_id, throughput, pdata[19], &foundNounce);
 
 		*hashes_done = pdata[19] - start_nonce + throughput;
 
@@ -81,9 +80,8 @@ int scanhash_myriad(int thr_id, struct work *work, uint32_t max_nonce, unsigned
 			if (vhash[7] <= ptarget[7] && fulltest(vhash, ptarget)) {
 				work_set_target_ratio(work, vhash);
 				pdata[19] = foundNounce;
-				free(outputHash);
 				return 1;
-			} else {
+			} else if (vhash[7] > ptarget[7]) {
 				gpulog(LOG_WARNING, thr_id, "result for %08x does not validate on CPU!", foundNounce);
 			}
 		}
@@ -98,7 +96,6 @@ int scanhash_myriad(int thr_id, struct work *work, uint32_t max_nonce, unsigned
 
 	*hashes_done = max_nonce - start_nonce;
 
-	free(outputHash);
 	return 0;
 }