thodg/cgminer

diff --git a/findnonce.c b/findnonce.c
index 660971c..229b3e6 100644
--- a/findnonce.c
+++ b/findnonce.c
@@ -55,7 +55,7 @@ inline uint32_t ByteReverse(uint32_t value)
 
 void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data) {
 	cl_uint A, B, C, D, E, F, G, H;
-
+	
 	A = state[0];
 	B = state[1];
 	C = state[2];
@@ -96,9 +96,19 @@ void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data) {
 	blk->merkle = data[0];
 	blk->ntime = data[1];
 	blk->nbits = data[2];
-
+	
 	blk->W16 = blk->fW0 = data[0] + (rotr(data[1], 7) ^ rotr(data[1], 18) ^ (data[1] >> 3));
 	blk->W17 = blk->fW1 = data[1] + (rotr(data[2], 7) ^ rotr(data[2], 18) ^ (data[2] >> 3)) + 0x01100000;
+	blk->PreVal4 = blk->fcty_e = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + (D ^ (B & (C ^ D))) + 0xe9b5dba5;
+	blk->T1 = blk->fcty_e2 = (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + ((F & G) | (H & (F | G)));
+	blk->PreVal4_2 = blk->PreVal4 + blk->T1;
+	blk->PreVal0 = blk->PreVal4 + state[0];
+	blk->PreW31 = 0x00000280 + (rotr(blk->W16,  7) ^ rotr(blk->W16, 18) ^ (blk->W16 >> 3));
+	blk->PreW32 = blk->W16 + ((rotr(blk->W17, 7) ^ rotr(blk->W17, 18) ^ (blk->W17 >> 3)));
+	blk->PreW18 = data[2] + (rotr(blk->W16, 17) ^ rotr(blk->W16, 19) ^ (blk->W16 >> 10));
+	blk->PreW19 = 0x11002000 + (rotr(blk->W17, 17) ^ rotr(blk->W17, 19) ^ (blk->W17 >> 10));
+	
+	
 	blk->W2 = data[2];
 
 	blk->W2A = blk->W2 + (rotr(blk->W16, 19) ^ rotr(blk->W16, 17) ^ (blk->W16 >> 10));
@@ -109,9 +119,7 @@ void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data) {
 	blk->fW15 = 0x00000280 + (rotr(blk->fW0, 7) ^ rotr(blk->fW0, 18) ^ (blk->fW0 >> 3));
 	blk->fW01r = blk->fW0 + (rotr(blk->fW1, 7) ^ rotr(blk->fW1, 18) ^ (blk->fW1 >> 3));
 
-	blk->PreVal4 = blk->fcty_e = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + (D ^ (B & (C ^ D))) + 0xe9b5dba5;
-	blk->T1 = blk->fcty_e2 = (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + ((F & G) | (H & (F | G)));
-
+	
 	blk->PreVal4addT1 = blk->PreVal4 + blk->T1;
 	blk->T1substate0 = state[0] - blk->T1;
 }
diff --git a/main.c b/main.c
index 8f2a3ab..7f1165a 100644
--- a/main.c
+++ b/main.c
@@ -3411,7 +3411,7 @@ static cl_int queue_poclbm_kernel(_clState *clState, dev_blk_ctx *blk)
 
 static cl_int queue_phatk_kernel(_clState *clState, dev_blk_ctx *blk)
 {
-	cl_kernel *kernel = &clState->kernel;
+		cl_kernel *kernel = &clState->kernel;
 	cl_int status = 0;
 	int num = 0;
 
@@ -3423,21 +3423,27 @@ static cl_int queue_phatk_kernel(_clState *clState, dev_blk_ctx *blk)
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->ctx_f);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->ctx_g);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->ctx_h);
+	
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->cty_b);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->cty_c);
-	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->C1addK5);
-	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->D1A);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->cty_d);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->cty_f);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->cty_g);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->cty_h);
-	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->nonce);
+	
+	uint nonces[2];
+	nonces[0] = blk->nonce;
+	nonces[1] = (blk->nonce)+1;
+	status |= clSetKernelArg(*kernel, num++, 2 * sizeof(uint), (void *)nonces);
 
-	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->W2A);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->W16);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->W17);
-	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->W17_2);
-	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->PreVal4addT1);
-	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->T1substate0);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->PreVal4_2);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->PreVal0);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->PreW18);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->PreW19);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->PreW31);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->PreW32);
 
 	status |= clSetKernelArg(*kernel, num++, sizeof(clState->outputBuffer),
 				 (void *)&clState->outputBuffer);
diff --git a/miner.h b/miner.h
index 402ef9f..910693c 100644
--- a/miner.h
+++ b/miner.h
@@ -288,6 +288,12 @@ typedef struct {
 	cl_uint PreVal4; cl_uint T1;
 	cl_uint C1addK5; cl_uint D1A; cl_uint W2A; cl_uint W17_2;
 	cl_uint PreVal4addT1; cl_uint T1substate0;
+	cl_uint PreVal4_2;
+	cl_uint PreVal0;
+	cl_uint PreW18;
+	cl_uint PreW19;
+	cl_uint PreW31;
+	cl_uint PreW32;
 } dev_blk_ctx;
 #else
 typedef struct {
diff --git a/ocl.c b/ocl.c
index 45f7095..15c2517 100644
--- a/ocl.c
+++ b/ocl.c
@@ -355,8 +355,8 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 			break;
 		case KL_NONE: /* Shouldn't happen */
 		case KL_PHATK:
-			strcpy(filename, "phatk110722.cl");
-			strcpy(binaryfilename, "phatk110722");
+			strcpy(filename, "phatk2_2.cl");
+			strcpy(binaryfilename, "phatk2_2");
 			break;
 	}
 
@@ -501,7 +501,14 @@ build:
 	}
 
 	/* create a cl program executable for all the devices specified */
-	status = clBuildProgram(clState->program, 1, &devices[gpu], NULL, NULL, NULL);
+	char CompilerOptions[256];
+	sprintf(CompilerOptions, "%s%i", "-DWORKSIZE=", clState->work_size);
+	//int n = 1000;
+	//while(n--)
+	//	printf("%s", CompilerOptions);
+	//return 1;
+	status = clBuildProgram(clState->program, 1, &devices[gpu], CompilerOptions , NULL, NULL);
+
 	if (status != CL_SUCCESS)
 	{
 		applog(LOG_ERR, "Error: Building Program (clBuildProgram)");