Multiple miscellaneous changes:

- Attempting to integrate recent changes from multiple various versions of phatk kernel
  - reduces ALU OPs by 5 on VLIW4, and by 8 on VLIW5 hardware.
- Apply some of the phatk optimizations to the poclbm kernel as well.
- On the off-chance the compiler does not precalculate the Vals[... % 8], replace it with the equivelant Vals[... & 7]
- change patching the #defines to passing them via clBuildProgram
- Add compiler hints about the requested work group size and vectorization.

findnonce.c

@@ -113,7 +113,7 @@ void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data) {
 	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;
+	blk->T1substate0 = state[0] + blk->PreVal4;
 }
 
 #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)))

ocl.c

@@ -497,49 +497,38 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 build:
 	memcpy(source, rawsource, pl);
 
+	char cl_options[1024] = "";
 	/* Patch the source file with the preferred_vwidth */
 	if (clState->preferred_vwidth > 1) {
-		char *find = strstr(source, "VECTORSX");
-
-		if (unlikely(!find)) {
-			applog(LOG_ERR, "Unable to find VECTORSX in source");
-			return NULL;
-		}
-		find += 7; // "VECTORS"
 		if (clState->preferred_vwidth == 2)
-			strncpy(find, "2", 1);
-		else
-			strncpy(find, "4", 1);
+			strcat(cl_options, " -DVECTORS2 ");
+		if (clState->preferred_vwidth == 4)
+			strcat(cl_options, " -DVECTORS4 ");
+		if (opt_debug)
+			applog(LOG_DEBUG, "Appended -DVECTORS%d to build flags", clState->preferred_vwidth);
+	}
+
+	/* Provide compiler hint for requested worksize */
+	if (clState->work_size) {
+		sprintf(numbuf, "%d", (int)clState->work_size);
+		strcat(cl_options, " -DWORKSIZE=");
+		strcat(cl_options, numbuf);
 		if (opt_debug)
-			applog(LOG_DEBUG, "Patched source to suit %d vectors", clState->preferred_vwidth);
+			applog(LOG_DEBUG, "Appended -DWORKSIZE=%d to build flags", clState->work_size);
 	}
 
 	/* Patch the source file defining BITALIGN */
 	if (clState->hasBitAlign) {
-		char *find = strstr(source, "BITALIGNX");
-
-		if (unlikely(!find)) {
-			applog(LOG_ERR, "Unable to find BITALIGNX in source");
-			return NULL;
-		}
-		find += 8; // "BITALIGN"
-		strncpy(find, " ", 1);
+		strcat(cl_options, " -DBITALIGN");
 		if (opt_debug)
-			applog(LOG_DEBUG, "cl_amd_media_ops found, patched source with BITALIGN");
+			applog(LOG_DEBUG, "cl_amd_media_ops found, Appended -DBITALIGN to build flags");
 	} else if (opt_debug)
 		applog(LOG_DEBUG, "cl_amd_media_ops not found, will not BITALIGN patch");
 
 	if (patchbfi) {
-		char *find = strstr(source, "BFI_INTX");
-
-		if (unlikely(!find)) {
-			applog(LOG_ERR, "Unable to find BFI_INTX in source");
-			return NULL;
-		}
-		find += 7; // "BFI_INT"
-		strncpy(find, " ", 1);
+		strcat(cl_options, " -DBFI_INT");
 		if (opt_debug)
-			applog(LOG_DEBUG, "cl_amd_media_ops found, patched source with BFI_INT");
+			applog(LOG_DEBUG, "cl_amd_media_ops found, Appended -DBFI_INT to build flags");
 	} else if (opt_debug)
 		applog(LOG_DEBUG, "cl_amd_media_ops not found, will not BFI_INT patch");
 
@@ -551,7 +540,7 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 	}
 
 	/* create a cl program executable for all the devices specified */
-	status = clBuildProgram(clState->program, 1, &devices[gpu], NULL, NULL, NULL);
+	status = clBuildProgram(clState->program, 1, &devices[gpu], (const char *)&cl_options, NULL, NULL);
 	if (status != CL_SUCCESS)
 	{
 		applog(LOG_ERR, "Error: Building Program (clBuildProgram)");