Don't enqueuewrite buffer at all for pad8 and pass work details around for scrypt in dev_blk.

cgminer.c

@@ -3983,11 +3983,15 @@ bool hashtest(const struct work *work)
 
 bool test_nonce(struct work *work, uint32_t nonce)
 {
+	uint32_t *work_nonce = (uint32_t *)(work->data + 64 + 12);
+
+	*work_nonce = htobe32(nonce);
+#if 0
 	work->data[64 + 12 + 0] = (nonce >> 0) & 0xff;
 	work->data[64 + 12 + 1] = (nonce >> 8) & 0xff;
 	work->data[64 + 12 + 2] = (nonce >> 16) & 0xff;
 	work->data[64 + 12 + 3] = (nonce >> 24) & 0xff;
-
+#endif
 	if (opt_scrypt)
 		return true;
 

driver-opencl.c

@@ -993,14 +993,14 @@ static cl_int queue_diablo_kernel(_clState *clState, dev_blk_ctx *blk, cl_uint t
 }
 
 #ifdef USE_SCRYPT
-static cl_int queue_scrypt_kernel(_clState *clState, dev_blk_ctx *blk, cl_uint threads)
+static cl_int queue_scrypt_kernel(_clState *clState, dev_blk_ctx *blk, __maybe_unused cl_uint threads)
 {
-	cl_uint4 *midstate = (cl_uint4 *)blk->midstate;
+	cl_uint4 *midstate = (cl_uint4 *)blk->work->midstate;
 	cl_kernel *kernel = &clState->kernel;
 	unsigned int num = 0;
 	cl_int status = 0;
-	int i;
 
+	clState->cldata = blk->work->data;
 	status = clEnqueueWriteBuffer(clState->commandQueue, clState->CLbuffer0, true, 0, 80, clState->cldata, 0, NULL,NULL);
 
 	CL_SET_ARG(clState->CLbuffer0);
@@ -1349,13 +1349,6 @@ static bool opencl_thread_init(struct thr_info *thr)
 		return false;
 	}
 
-#ifdef USE_SCRYPT
-	if (opt_scrypt) {
-		if (clState->padbufsize > BUFFERSIZE)
-			blank_res = realloc(blank_res, clState->padbufsize);
-		status = clEnqueueWriteBuffer(clState->commandQueue, clState->padbuffer8, true, 0, clState->padbufsize, blank_res, 0, NULL,NULL);
-	}
-#endif
 	status |= clEnqueueWriteBuffer(clState->commandQueue, clState->outputBuffer, CL_TRUE, 0,
 			BUFFERSIZE, blank_res, 0, NULL, NULL);
 	if (unlikely(status != CL_SUCCESS)) {
@@ -1385,6 +1378,11 @@ static void opencl_free_work(struct thr_info *thr, struct work *work)
 
 static bool opencl_prepare_work(struct thr_info __maybe_unused *thr, struct work *work)
 {
+#ifdef USE_SCRYPT
+	if (opt_scrypt)
+		work->blk.work = work;
+	else
+#endif
 		precalc_hash(&work->blk, (uint32_t *)(work->midstate), (uint32_t *)(work->data + 64));
 	return true;
 }
@@ -1446,9 +1444,6 @@ static int64_t opencl_scanhash(struct thr_info *thr, struct work *work,
 	if (hashes > gpu->max_hashes)
 		gpu->max_hashes = hashes;
 
-#ifdef USE_SCRYPT
-	clState->cldata = work->data;
-#endif
 	status = thrdata->queue_kernel_parameters(clState, &work->blk, globalThreads[0]);
 	if (unlikely(status != CL_SUCCESS)) {
 		applog(LOG_ERR, "Error: clSetKernelArg of all params failed.");

findnonce.c

@@ -45,7 +45,8 @@ const uint32_t SHA256_K[64] = {
 	d = d + h; \
 	h = h + (rotate(a, 30) ^ rotate(a, 19) ^ rotate(a, 10)) + ((a & b) | (c & (a | b)))
 
-void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data) {
+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];
@@ -127,10 +128,6 @@ void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data) {
 	blk->fiveA = blk->ctx_f + SHA256_K[5];
 	blk->sixA = blk->ctx_g + SHA256_K[6];
 	blk->sevenA = blk->ctx_h + SHA256_K[7];
-
-#ifdef USE_SCRYPT
-	blk->midstate = (unsigned char *)state;
-#endif
 }
 
 #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)))

miner.h

@@ -671,7 +671,7 @@ typedef struct {
 	cl_uint zeroA, zeroB;
 	cl_uint oneA, twoA, threeA, fourA, fiveA, sixA, sevenA;
 #ifdef USE_SCRYPT
-	unsigned char *midstate;
+	struct work *work;
 #endif
 } dev_blk_ctx;
 #else