Update phatk kernel to one with new parameters for slightly less overhead again.

Make the queue kernel parameters call a function pointer to select phatk or poclbm.

Makefile.am

@@ -15,7 +15,7 @@ INCLUDES	= $(PTHREAD_FLAGS) -fno-strict-aliasing $(JANSSON_INCLUDES)
 
 bin_PROGRAMS	= cgminer
 
-bin_SCRIPTS	= phatk110714.cl poclbm110717.cl
+bin_SCRIPTS	= phatk110722.cl poclbm110717.cl
 
 cgminer_SOURCES	= elist.h miner.h compat.h			\
 		  main.c util.c					\
@@ -23,7 +23,7 @@ cgminer_SOURCES	= elist.h miner.h compat.h			\
 		  sha256_generic.c sha256_4way.c sha256_via.c	\
 		  sha256_cryptopp.c sha256_sse2_amd64.c		\
 		  sha256_sse4_amd64.c \
-		  phatk110714.cl poclbm110717.cl
+		  phatk110722.cl poclbm110717.cl
 
 cgminer_LDFLAGS	= $(PTHREAD_FLAGS)
 cgminer_LDADD	= @LIBCURL@ @JANSSON_LIBS@ @PTHREAD_LIBS@ @OPENCL_LIBS@ @NCURSES_LIBS@ @PDCURSES_LIBS@ lib/libgnu.a ccan/libccan.a

findnonce.c

@@ -72,7 +72,13 @@ void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data) {
 	blk->cty_a = A;
 	blk->cty_b = B;
 	blk->cty_c = C;
+
+	blk->C1addK5 = C + 0x59f111f1;
+
 	blk->cty_d = D;
+
+	blk->D1A = D + 0xb956c25b;
+
 	blk->cty_e = E;
 	blk->cty_f = F;
 	blk->cty_g = G;
@@ -94,6 +100,10 @@ void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data) {
 	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->W2 = data[2];
+
+	blk->W2A = blk->W2 + (rotr(blk->W16, 19) ^ rotr(blk->W16, 17) ^ (blk->W16 >> 10));
+	blk->W17_2 = 0x11002000 + (rotr(blk->W17, 19) ^ rotr(blk->W17, 17) ^ (blk->W17 >> 10));
+
 	blk->fW2 = data[2] + (rotr(blk->fW0, 17) ^ rotr(blk->fW0, 19) ^ (blk->fW0 >> 10));
 	blk->fW3 = 0x11002000 + (rotr(blk->fW1, 17) ^ rotr(blk->fW1, 19) ^ (blk->fW1 >> 10));
 	blk->fW15 = 0x00000280 + (rotr(blk->fW0, 7) ^ rotr(blk->fW0, 18) ^ (blk->fW0 >> 3));
@@ -101,6 +111,9 @@ void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data) {
 
 	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;
 }
 
 #define P(t) (W[(t)&0xF] = W[(t-16)&0xF] + (rotate(W[(t-15)&0xF], 25) ^ rotate(W[(t-15)&0xF], 14) ^ (W[(t-15)&0xF] >> 3)) + W[(t-7)&0xF] + (rotate(W[(t-2)&0xF], 15) ^ rotate(W[(t-2)&0xF], 13) ^ (W[(t-2)&0xF] >> 10)))

main.c

@@ -2373,7 +2373,7 @@ enum {
 #ifdef HAVE_OPENCL
 static _clState *clStates[16];
 
-static inline cl_int queue_kernel_parameters(_clState *clState, dev_blk_ctx *blk)
+static cl_int queue_poclbm_kernel(_clState *clState, dev_blk_ctx *blk)
 {
 	cl_kernel *kernel = &clState->kernel;
 	cl_int status = 0;
@@ -2395,15 +2395,6 @@ static inline cl_int queue_kernel_parameters(_clState *clState, dev_blk_ctx *blk
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->cty_h);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->nonce);
 
-	if (clState->hasBitAlign == true) {
-		/* Parameters for phatk kernel */
-		status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->W2);
-		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->PreVal4);
-		status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->T1);
-	} else {
-		/* Parameters for poclbm kernel */
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->fW0);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->fW1);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->fW2);
@@ -2412,7 +2403,43 @@ static inline cl_int queue_kernel_parameters(_clState *clState, dev_blk_ctx *blk
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->fW01r);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->fcty_e);
 	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->fcty_e2);
+
+	status |= clSetKernelArg(*kernel, num++, sizeof(clState->outputBuffer),
+				 (void *)&clState->outputBuffer);
+
+	return status;
 }
+
+static cl_int queue_phatk_kernel(_clState *clState, dev_blk_ctx *blk)
+{
+	cl_kernel *kernel = &clState->kernel;
+	cl_int status = 0;
+	int num = 0;
+
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->ctx_a);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->ctx_b);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->ctx_c);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->ctx_d);
+	status |= clSetKernelArg(*kernel, num++, sizeof(uint), (void *)&blk->ctx_e);
+	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_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);
+
+	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(clState->outputBuffer),
 				 (void *)&clState->outputBuffer);
 
@@ -2432,6 +2459,8 @@ static void set_threads_hashes(unsigned int vectors, unsigned int *threads,
 
 static void *gpuminer_thread(void *userdata)
 {
+	cl_int (*queue_kernel_parameters)(_clState *, dev_blk_ctx *);
+
 	const unsigned long cycle = opt_log_interval / 5 ? : 1;
 	struct timeval tv_start, tv_end, diff, tv_workstart;
 	struct thr_info *mythr = userdata;
@@ -2459,6 +2488,11 @@ static void *gpuminer_thread(void *userdata)
 	bool requested = true;
 	uint32_t total_hashes = 0, hash_div = 1;
 
+	if (clState->hasBitAlign)
+		queue_kernel_parameters = &queue_phatk_kernel;
+	else
+		queue_kernel_parameters = &queue_poclbm_kernel;
+
 	if (opt_dynamic) {
 		/* Minimise impact on desktop if we want dynamic mode */
 		setpriority(PRIO_PROCESS, 0, 19);

miner.h

@@ -258,6 +258,8 @@ typedef struct {
 	cl_uint fW01r; cl_uint fcty_e; cl_uint fcty_e2;
 	cl_uint W16; cl_uint W17; cl_uint W2;
 	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;
 } dev_blk_ctx;
 #else
 typedef struct {

ocl.c

@@ -341,7 +341,7 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 	char numbuf[10];
 	char filename[16];
 	if (clState->hasBitAlign)
-		strcpy(filename, "phatk110714.cl");
+		strcpy(filename, "phatk110722.cl");
 	else
 		strcpy(filename, "poclbm110717.cl");
 	FILE *binaryfile;
@@ -370,7 +370,7 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 
 	strcpy(binaryfilename, name);
 	if (clState->hasBitAlign) {
-		strcat(binaryfilename, "phatk110714");
+		strcat(binaryfilename, "phatk110722");
 		strcat(binaryfilename, "bitalign");
 	} else
 		strcat(binaryfilename, "poclbm110717");

phatk110722.cl

@@ -1,7 +1,7 @@
 // This file is taken and modified from the public-domain poclbm project, and
 // we have therefore decided to keep it public-domain in Phoenix.
 
-// 2011-07-11: further modified by Diapolo and still public-domain
+// 2011-07-12: further modified by Diapolo and still public-domain
 // -ck version to be compatible with cgminer
 // 2011-07-14: shorter code
 
@@ -38,6 +38,8 @@ __constant ulong L = 0x198c7e2a2;
 #define BFI_INTX
 #define BITALIGNX
 
+#define O 15
+
 #ifdef BITALIGN
 	#pragma OPENCL EXTENSION cl_amd_media_ops : enable
 	#define rot(x, y) amd_bitalign(x, x, (u)(32 - y))
@@ -59,99 +61,105 @@ __constant ulong L = 0x198c7e2a2;
 #define s1(n) (rot(Vals[(132 - n) % 8], 26) ^ rot(Vals[(132 - n) % 8], 21) ^ rot(Vals[(132 - n) % 8], 7))
 #define ch(n) (Ch(Vals[(132 - n) % 8], Vals[(133 - n) % 8], Vals[(134 - n) % 8]))
 #define ma(n) (Ma(Vals[(129 - n) % 8], Vals[(130 - n) % 8], Vals[(128 - n) % 8]))
-#define t1(n) (K[n % 64] + Vals[(135 - n) % 8] + W[n] + s1(n) + ch(n))
+#define t1(n) (K[n % 64] + Vals[(135 - n) % 8] + W[n - O] + s1(n) + ch(n))
+#define t1_no_W(n) (K[n % 64] + Vals[(135 - n) % 8] + s1(n) + ch(n))
 
 // intermediate W calculations
-#define P1(x) (rot(W[x - 2], 15) ^ rot(W[x - 2], 13) ^ (W[x - 2] >> 10U))
-#define P2(x) (rot(W[x - 15], 25) ^ rot(W[x - 15], 14) ^ (W[x - 15] >> 3U))
-#define P3(x) W[x - 7]
-#define P4(x) W[x - 16]
+#define P1(x) (rot(W[x - 2 - O], 15) ^ rot(W[x - 2 - O], 13) ^ (W[x - 2 - O] >> 10U))
+#define P2(x) (rot(W[x - 15 - O], 25) ^ rot(W[x - 15 - O], 14) ^ (W[x - 15 - O] >> 3U))
+#define P3(x) W[x - 7 - O]
+#define P4(x) W[x - 16 - O]
 
 // full W calculation
-#define W(x) (W[x] = P4(x) + P3(x) + P2(x) + P1(x))
+#define W(x) (W[x - O] = P4(x) + P3(x) + P2(x) + P1(x))
 
 // SHA round without W calc
 #define sharound(n) { Vals[(131 - n) % 8] += t1(n); Vals[(135 - n) % 8] = t1(n) + s0(n) + ma(n); }
+#define sharound_no_W(n) { Vals[(131 - n) % 8] += t1_no_W(n); Vals[(135 - n) % 8] = t1_no_W(n) + s0(n) + ma(n); }
 
 __kernel void search(	const uint state0, const uint state1, const uint state2, const uint state3,
 						const uint state4, const uint state5, const uint state6, const uint state7,
-						const uint B1, const uint C1, const uint D1,
+						const uint B1, const uint C1, const uint C1addK5, const uint D1,
 						const uint F1, const uint G1, const uint H1,
 						const uint base,
 						const uint W2,
-						const uint W16, const uint W17,
-						const uint PreVal4, const uint T1,
+						const uint W16, const uint W17, const uint W17_2,
+						const uint PreVal4addT1, const uint T1substate0,
 						__global uint * output)
 {
-	u W[124];
+	u W[124 - O];
 	u Vals[8];
+#ifdef VECTORS4
+        u W_3 = base + (get_global_id(0) << 2) + (uint4)(0, 1, 2, 3);
+#elif defined VECTORS2
+        u W_3 = base + (get_global_id(0) << 1) + (uint2)(0, 1);
+#else
+        u W_3 = base + get_global_id(0);
+#endif
+	u Temp;
 	
+	Vals[0] = W_3 + PreVal4addT1 + T1substate0;
 	Vals[1] = B1;
 	Vals[2] = C1;
+
+	Vals[4] = W_3 + PreVal4addT1;
 	Vals[5] = F1;
 	Vals[6] = G1;
 	
-	W[2] = W2;
-#ifdef VECTORS4
-        Vals[4] = (W[3] = base + (get_global_id(0) << 2) + (uint4)(0, 1, 2, 3)) + PreVal4;
-#elif defined VECTORS2
-        Vals[4] = (W[3] = base + (get_global_id(0) << 1) + (uint2)(0, 1)) + PreVal4;
-#else
-        Vals[4] = (W[3] = base + get_global_id(0)) + PreVal4;
-#endif
 	// used in: P2(19) == 285220864 (0x11002000), P4(20)
-	W[4] = 0x80000000U;
+	// W[4] = 0x80000000U;
 	// P1(x) is 0 for x == 7, 8, 9, 10, 11, 12, 13, 14, 15, 16
 	// P2(x) is 0 for x == 20, 21, 22, 23, 24, 25, 26, 27, 28, 29
 	// P3(x) is 0 for x == 12, 13, 14, 15, 16, 17, 18, 19, 20, 21
 	// P4(x) is 0 for x == 21, 22, 23, 24, 25, 26, 27, 28, 29, 30
 	// W[x] in sharound(x) is 0 for x == 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
-	W[14] = W[13] = W[12] = W[11] = W[10] = W[9] = W[8] = W[7] = W[6] = W[5] = 0x00000000U;
+	// W[14] = W[13] = W[12] = W[11] = W[10] = W[9] = W[8] = W[7] = W[6] = W[5] = 0x00000000U;
 	// used in: P2(30) == 10485845 (0xA00055), P3(22), P4(31)
 	// K[15] + W[15] == 0xc19bf174 + 0x00000280U = 0xc19bf3f4
-	W[15] = 0x00000280U;
-
-	W[16] = W16;
-	W[17] = W17;
-	// removed P3(18) from add because it is == 0
-	W[18] = P1(18) + P4(18) + P2(18);
-	// removed P3(19) from add because it is == 0
-	W[19] = (u)0x11002000 + P1(19) + P4(19);
-	// removed P2(20), P3(20) from add because it is == 0
-	W[20] = P1(20) + P4(20);
-	W[21] = P1(21);
-	W[22] = P1(22) + P3(22);
-	W[23] = P1(23) + P3(23);
-	W[24] = P1(24) + P3(24);
-	W[25] = P1(25) + P3(25);
-	W[26] = P1(26) + P3(26);
-	W[27] = P1(27) + P3(27);
-	W[28] = P1(28) + P3(28);
-	W[29] = P1(29) + P3(29);
-	W[30] = (u)0xA00055 + P1(30) + P3(30);
-	
-	// Round 3
-	Vals[0] = state0 + Vals[4];
-	Vals[4] += T1;
+	W[15 - O] = 0x00000280U;
+	W[16 - O] = W16;
+	W[17 - O] = W17;
+	W[18 - O] = W2 + (rot(W_3, 25) ^ rot(W_3, 14) ^ (W_3 >> 3U));
+	W[19 - O] = W_3 + W17_2;
+	W[20 - O] = (u)0x80000000U + P1(20);
+	W[21 - O] = P1(21);
+	W[22 - O] = P1(22) + P3(22);
+	W[23 - O] = P1(23) + P3(23);
+	W[24 - O] = P1(24) + P3(24);
+	W[25 - O] = P1(25) + P3(25);
+	W[26 - O] = P1(26) + P3(26);
+	W[27 - O] = P1(27) + P3(27);
+	W[28 - O] = P1(28) + P3(28);
+	W[29 - O] = P1(29) + P3(29);
+	W[30 - O] = (u)0xA00055 + P1(30) + P3(30);
 	
 	// Round 4
-	// K[4] + W[4] == 0x3956c25b + 0x80000000U = 0xb956c25b
-	Vals[7] = (Vals[3] = (u)0xb956c25b + D1 + s1(4) + ch(4)) + H1;
-	Vals[3] += s0(4) + ma(4);
+	Temp = D1 + ch(4) + s1(4);
+	Vals[7] = Temp + H1;
+	Vals[3] = Temp + ma(4) + s0(4);	
 
 	// Round 5
-	Vals[2] = K[5] + C1 + s1(5) + ch(5) + s0(5) + ma(5);
-	Vals[6] = K[5] + C1 + G1 + s1(5) + ch(5);
-
-	sharound(6);
-	sharound(7);
-	sharound(8);
-	sharound(9);
-	sharound(10);
-	sharound(11);
-	sharound(12);
-	sharound(13);
-	sharound(14);
+	Temp = C1addK5 + ch(5) + s1(5);
+	Vals[6] = Temp + G1;
+	Vals[2] = Temp + ma(5) + s0(5);
+
+	// W[6] to W[14] are 0, so no need to add them!
+	sharound_no_W(6);
+	sharound_no_W(7);
+	sharound_no_W(8);
+	sharound_no_W(9);
+	sharound_no_W(10);
+	sharound_no_W(11);
+	sharound_no_W(12);
+	sharound_no_W(13);
+	sharound_no_W(14);
+
+//	#define sharound(n) { Vals[(131 - n) % 8] += t1(n); Vals[(135 - n) % 8] = t1(n) + s0(n) + ma(n); }
+//	#define t1(n) (K[n % 64] + Vals[(135 - n) % 8] + W[n] + s1(n) + ch(n))
+
+//	Vals[(131 - 15) % 8] += (Vals[(135 - 15) % 8] = (u)0xc19bf3f4 + Vals[(135 - 15) % 8] + s1(15) + ch(15));
+//	Vals[(135 - 15) % 8] += s0(15) + ma(15);
+
 	sharound(15);
 	sharound(16);
 	sharound(17);
@@ -236,35 +244,35 @@ __kernel void search(	const uint state0, const uint state1, const uint state2, c
 	W(63);
 	sharound(63);
 
-	W[64] = state0 + Vals[0];
-	W[65] = state1 + Vals[1];
-	W[66] = state2 + Vals[2];
-	W[67] = state3 + Vals[3];
-	W[68] = state4 + Vals[4];
-	W[69] = state5 + Vals[5];
-	W[70] = state6 + Vals[6];
-	W[71] = state7 + Vals[7];
+	W[64 - O] = state0 + Vals[0];
+	W[65 - O] = state1 + Vals[1];
+	W[66 - O] = state2 + Vals[2];
+	W[67 - O] = state3 + Vals[3];
+	W[68 - O] = state4 + Vals[4];
+	W[69 - O] = state5 + Vals[5];
+	W[70 - O] = state6 + Vals[6];
+	W[71 - O] = state7 + Vals[7];
 	// used in: P2(87) = 285220864 (0x11002000), P4(88)
 	// K[72] + W[72] ==
-	W[72] = 0x80000000U;
+	W[72 - O] = 0x80000000U;
 	// P1(x) is 0 for x == 75, 76, 77, 78, 79, 80
 	// P2(x) is 0 for x == 88, 89, 90, 91, 92, 93
 	// P3(x) is 0 for x == 80, 81, 82, 83, 84, 85
 	// P4(x) is 0 for x == 89, 90, 91, 92, 93, 94
 	// W[x] in sharound(x) is 0 for x == 73, 74, 75, 76, 77, 78
-	W[78] = W[77] = W[76] = W[75] = W[74] = W[73] = 0x00000000U;
+	// W[78] = W[77] = W[76] = W[75] = W[74] = W[73] = 0x00000000U;
 	// used in: P1(81) = 10485760 (0xA00000), P2(94) = 4194338 (0x400022), P3(86), P4(95)
 	// K[79] + W[79] ==
-	W[79] = 0x00000100U;
+	W[79 - O] = 0x00000100U;
 
 	Vals[0] = H[0];
 	Vals[1] = H[1];
 	Vals[2] = H[2];
-	Vals[3] = (u)L + W[64];
+	Vals[3] = (u)L + W[64 - O];
 	Vals[4] = H[3];
 	Vals[5] = H[4];
 	Vals[6] = H[5];
-	Vals[7] = H[6] + W[64];
+	Vals[7] = H[6] + W[64 - O];
 	
 	sharound(65);
 	sharound(66);
@@ -274,21 +282,23 @@ __kernel void search(	const uint state0, const uint state1, const uint state2, c
 	sharound(70);
 	sharound(71);
 	sharound(72);
-	sharound(73);
-	sharound(74);
-	sharound(75);
-	sharound(76);
-	sharound(77);
-	sharound(78);
+
+	// W is also zero for these rounds
+	sharound_no_W(73);
+	sharound_no_W(74);
+	sharound_no_W(75);
+	sharound_no_W(76);
+	sharound_no_W(77);
+	sharound_no_W(78);
+
 	sharound(79);
 	
-	// removed P1(80), P3(80) from add because it is == 0
-	W[80] = P2(80) + P4(80);
-	W[81] = (u)0xA00000 + P4(81) + P2(81);
-	W[82] = P4(82) + P2(82) + P1(82);
-	W[83] = P4(83) + P2(83) + P1(83);
-	W[84] = P4(84) + P2(84) + P1(84);
-	W[85] = P4(85) + P2(85) + P1(85);
+	W[80 - O] = P2(80) + P4(80);
+	W[81 - O] = (u)0xA00000 + P4(81) + P2(81);
+	W[82 - O] = P4(82) + P2(82) + P1(82);
+	W[83 - O] = P4(83) + P2(83) + P1(83);
+	W[84 - O] = P4(84) + P2(84) + P1(84);
+	W[85 - O] = P4(85) + P2(85) + P1(85);
 	W(86);
 
 	sharound(80);
@@ -299,27 +309,25 @@ __kernel void search(	const uint state0, const uint state1, const uint state2, c
 	sharound(85);
 	sharound(86);
 
-	W[87] = (u)0x11002000 + P4(87) + P3(87) + P1(87);
+	W[87 - O] = (u)0x11002000 + P4(87) + P3(87) + P1(87);
 	sharound(87);
-	W[88] = P4(88) + P3(88) + P1(88);
+	W[88 - O] = (u)0x80000000U + P3(88) + P1(88);
 	sharound(88);
-	W[89] = P3(89) + P1(89);
+	W[89 - O] = P3(89) + P1(89);
 	sharound(89);
-	W[90] = P3(90) + P1(90);
+	W[90 - O] = P3(90) + P1(90);
 	sharound(90);
-	W[91] = P3(91) + P1(91);
+	W[91 - O] = P3(91) + P1(91);
 	sharound(91);
-	W[92] = P3(92) + P1(92);
+	W[92 - O] = P3(92) + P1(92);
 	sharound(92);
-	// removed P2(93), P4(93) from add because it is == 0
-	W[93] = P3(93) + P1(93);
+	W[93 - O] = P3(93) + P1(93);
 	sharound(93);
-	// removed P4(94) from add because it is == 0
-	W[94] = (u)0x400022 + P3(94) + P1(94);
+	W[94 - O] = (u)0x400022 + P3(94) + P1(94);
 	sharound(94);
-	
-	W(95);
+	W[95 - O] = (u)0x00000100U + P3(95) + P2(95) + P1(95);
 	sharound(95);
+
 	W(96);
 	sharound(96);
 	W(97);
@@ -386,26 +394,26 @@ __kernel void search(	const uint state0, const uint state1, const uint state2, c
 #if defined(VECTORS4) || defined(VECTORS2)
 	if (Vals[7].x == -H[7])
 	{
-		output[MAXBUFFERS] = output[NFLAG & W[3].x] =  W[3].x;
+		output[MAXBUFFERS] = output[NFLAG & (W[3].x >> 2)] =  W_3.x;
 	}
 	if (Vals[7].y == -H[7])
 	{
-		output[MAXBUFFERS] = output[NFLAG & W[3].y] =  W[3].y;
+		output[MAXBUFFERS] = output[NFLAG & (W[3].y >> 2)] =  W_3.y;
 	}
 #ifdef VECTORS4
 	if (Vals[7].z == -H[7])
 	{
-		output[MAXBUFFERS] = output[NFLAG & W[3].z] =  W[3].z;
+		output[MAXBUFFERS] = output[NFLAG & (W[3].z >> 2)] =  W_3.z;
 	}
 	if (Vals[7].w == -H[7])
 	{
-		output[MAXBUFFERS] = output[NFLAG & W[3].w] =  W[3].w;
+		output[MAXBUFFERS] = output[NFLAG & (W[3].w >> 2)] =  W_3.w;
 	}
 #endif
 #else
 	if (Vals[7] == -H[7])
 	{
-		output[MAXBUFFERS] = output[NFLAG & W[3]] =  W[3];
+		output[MAXBUFFERS] = output[NFLAG & (W[3] >> 2)] =  W_3;
 	}
 #endif
 }