Make postcalc_hash asynchronous as well.

cpu-miner.c

@@ -277,20 +277,6 @@ static struct option options[] = {
 	{ "userpass", 1, NULL, 1002 },
 };
 
-struct work {
-	unsigned char	data[128];
-	unsigned char	hash1[64];
-	unsigned char	midstate[32];
-	unsigned char	target[32];
-
-	unsigned char	hash[32];
-
-	uint32_t		output[1];
-	uint32_t		res_nonce;
-	uint32_t		valid;
-	dev_blk_ctx		blk;
-};
-
 static bool jobj_binary(const json_t *obj, const char *key,
 			void *buf, size_t buflen)
 {
@@ -964,7 +950,7 @@ static void *gpuminer_thread(void *userdata)
 			for (i = 0; i < 127; i++) {
 				if (res[i]) {
 					applog(LOG_INFO, "GPU %d found something?", gpu_from_thr_id(thr_id));
-					postcalc_hash(mythr, &work->blk, work, res[i]);
+					postcalc_hash_async(mythr, work, res[i]);
 				} else
 					break;
 			}

findnonce.c

@@ -9,9 +9,12 @@
 
 #include <stdio.h>
 #include <inttypes.h>
+#include <pthread.h>
+#include <string.h>
 
 #include "ocl.h"
 #include "findnonce.h"
+#include "miner.h"
 
 const uint32_t SHA256_K[64] = {
 	0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5,
@@ -131,8 +134,21 @@ void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data) {
   R(E, F, G, H, A, B, C, D, P(u+4), SHA256_K[u+4]); \
   R(D, E, F, G, H, A, B, C, P(u+5), SHA256_K[u+5])
 
-void postcalc_hash(struct thr_info *thr, dev_blk_ctx *blk, struct work *work, uint32_t start)
+struct pc_data {
+	struct thr_info *thr;
+	struct work work;
+	uint32_t start;
+	pthread_t pth;
+};
+
+static void *postcalc_hash(void *userdata)
 {
+	struct pc_data *pcd = (struct pc_data *)userdata;
+	struct thr_info *thr = pcd->thr;
+	dev_blk_ctx *blk = &pcd->work.blk;
+	struct work *work = &pcd->work;
+	uint32_t start = pcd->start;
+
 	cl_uint A, B, C, D, E, F, G, H;
 	cl_uint W[16];
 	cl_uint nonce;
@@ -185,4 +201,24 @@ void postcalc_hash(struct thr_info *thr, dev_blk_ctx *blk, struct work *work, ui
 out:
 	if (unlikely(best_g == ~0))
 		applog(LOG_ERR, "No best_g found! Error in OpenCL code?");
+	free(pcd);
+}
+
+void postcalc_hash_async(struct thr_info *thr, struct work *work, uint32_t start)
+{
+	struct pc_data *pcd = malloc(sizeof(struct pc_data));
+	if (unlikely(!pcd)) {
+		applog(LOG_ERR, "Failed to malloc pc_data in postcalc_hash_async");
+		return;
+	}
+
+	pcd->thr = thr;
+	memcpy(&pcd->work, work, sizeof(struct work));
+	pcd->start = start;
+
+	if (pthread_create(&pcd->pth, NULL, postcalc_hash, (void *)pcd)) {
+		applog(LOG_ERR, "Failed to create postcalc_hash thread");
+		return;
+	}
+	pthread_detach(pcd->pth);
 }

findnonce.h

@@ -1,22 +1,10 @@
-#ifdef __APPLE_CC__
+#ifndef __FINDNONCE_H__
-#include <OpenCL/opencl.h>
+#define __FINDNONCE_H__
-#else
-#include <CL/cl.h>
-#endif
 #include "miner.h"
 
 #define MAXTHREADS (0xFFFFFFFF)
 #define BUFFERSIZE (sizeof(uint32_t) * 128)
 
-typedef struct {
-    cl_uint ctx_a; cl_uint ctx_b; cl_uint ctx_c; cl_uint ctx_d;
-    cl_uint ctx_e; cl_uint ctx_f; cl_uint ctx_g; cl_uint ctx_h;
-    cl_uint cty_a; cl_uint cty_b; cl_uint cty_c; cl_uint cty_d;
-    cl_uint cty_e; cl_uint cty_f; cl_uint cty_g; cl_uint cty_h;
-    cl_uint merkle; cl_uint ntime; cl_uint nbits; cl_uint nonce;
-	cl_uint fW0; cl_uint fW1; cl_uint fW2; cl_uint fW3; cl_uint fW15;
-	cl_uint fW01r; cl_uint fcty_e; cl_uint fcty_e2;
-} dev_blk_ctx;
-
 extern void precalc_hash(dev_blk_ctx *blk, uint32_t *state, uint32_t *data);
-extern void postcalc_hash(struct thr_info *thr, dev_blk_ctx *blk, struct work *work, uint32_t start);
+extern void postcalc_hash_async(struct thr_info *thr, struct work *work, uint32_t start);
+#endif /*__FINDNONCE_H__*/

miner.h

@@ -9,6 +9,11 @@
 #include <pthread.h>
 #include <jansson.h>
 #include <curl/curl.h>
+#ifdef __APPLE_CC__
+#include <OpenCL/opencl.h>
+#else
+#include <CL/cl.h>
+#endif
 
 #ifdef STDC_HEADERS
 # include <stdlib.h>
@@ -194,7 +199,31 @@ extern bool use_syslog;
 extern struct thr_info *thr_info;
 extern int longpoll_thr_id;
 extern struct work_restart *work_restart;
-struct work;
+
+typedef struct {
+    cl_uint ctx_a; cl_uint ctx_b; cl_uint ctx_c; cl_uint ctx_d;
+    cl_uint ctx_e; cl_uint ctx_f; cl_uint ctx_g; cl_uint ctx_h;
+    cl_uint cty_a; cl_uint cty_b; cl_uint cty_c; cl_uint cty_d;
+    cl_uint cty_e; cl_uint cty_f; cl_uint cty_g; cl_uint cty_h;
+    cl_uint merkle; cl_uint ntime; cl_uint nbits; cl_uint nonce;
+	cl_uint fW0; cl_uint fW1; cl_uint fW2; cl_uint fW3; cl_uint fW15;
+	cl_uint fW01r; cl_uint fcty_e; cl_uint fcty_e2;
+} dev_blk_ctx;
+
+struct work {
+	unsigned char	data[128];
+	unsigned char	hash1[64];
+	unsigned char	midstate[32];
+	unsigned char	target[32];
+
+	unsigned char	hash[32];
+
+	uint32_t		output[1];
+	uint32_t		res_nonce;
+	uint32_t		valid;
+	dev_blk_ctx		blk;
+};
+
 bool submit_nonce(struct thr_info *thr, struct work *work, uint32_t nonce);
 
 extern void applog(int prio, const char *fmt, ...);