Some more ROCm performance tuning for -m 77xx

OpenCL/m07700_a0-optimized.cl

@@ -17,8 +17,13 @@
 #include "inc_hash_md5.cl"
 #endif
 
+#ifdef IS_AMD
+#define GETCHAR(a,p)  (((a)[(p) / 4] >> (((p) & 3) * 8)) & 0xff)
+#define PUTCHAR(a,p,c) ((a)[(p) / 4] = (((a)[(p) / 4] & ~(0xff << (((p) & 3) * 8))) | ((c) << (((p) & 3) * 8))))
+#else
 #define GETCHAR(a,p)   ((u8 *)(a))[(p)]
 #define PUTCHAR(a,p,c) ((u8 *)(a))[(p)] = (u8) (c)
+#endif
 
 #define SETSHIFTEDINT(a,n,v)        \
 {                                   \

OpenCL/m07700_a1-optimized.cl

@@ -15,8 +15,13 @@
 #include "inc_hash_md5.cl"
 #endif
 
+#ifdef IS_AMD
+#define GETCHAR(a,p)  (((a)[(p) / 4] >> (((p) & 3) * 8)) & 0xff)
+#define PUTCHAR(a,p,c) ((a)[(p) / 4] = (((a)[(p) / 4] & ~(0xff << (((p) & 3) * 8))) | ((c) << (((p) & 3) * 8))))
+#else
 #define GETCHAR(a,p)   ((u8 *)(a))[(p)]
 #define PUTCHAR(a,p,c) ((u8 *)(a))[(p)] = (u8) (c)
+#endif
 
 #define SETSHIFTEDINT(a,n,v)        \
 {                                   \

OpenCL/m07700_a3-optimized.cl

@@ -15,8 +15,13 @@
 #include "inc_hash_md5.cl"
 #endif
 
+#ifdef IS_AMD
+#define GETCHAR(a,p)  (((a)[(p) / 4] >> (((p) & 3) * 8)) & 0xff)
+#define PUTCHAR(a,p,c) ((a)[(p) / 4] = (((a)[(p) / 4] & ~(0xff << (((p) & 3) * 8))) | ((c) << (((p) & 3) * 8))))
+#else
 #define GETCHAR(a,p)   ((u8 *)(a))[(p)]
 #define PUTCHAR(a,p,c) ((u8 *)(a))[(p)] = (u8) (c)
+#endif
 
 CONSTANT_VK u32a sapb_trans_tbl[256] =
 {

OpenCL/m07701_a0-optimized.cl

@@ -17,8 +17,13 @@
 #include "inc_hash_md5.cl"
 #endif
 
+#ifdef IS_AMD
+#define GETCHAR(a,p)  (((a)[(p) / 4] >> (((p) & 3) * 8)) & 0xff)
+#define PUTCHAR(a,p,c) ((a)[(p) / 4] = (((a)[(p) / 4] & ~(0xff << (((p) & 3) * 8))) | ((c) << (((p) & 3) * 8))))
+#else
 #define GETCHAR(a,p)   ((u8 *)(a))[(p)]
 #define PUTCHAR(a,p,c) ((u8 *)(a))[(p)] = (u8) (c)
+#endif
 
 #define SETSHIFTEDINT(a,n,v)        \
 {                                   \

OpenCL/m07701_a1-optimized.cl

@@ -15,8 +15,13 @@
 #include "inc_hash_md5.cl"
 #endif
 
+#ifdef IS_AMD
+#define GETCHAR(a,p)  (((a)[(p) / 4] >> (((p) & 3) * 8)) & 0xff)
+#define PUTCHAR(a,p,c) ((a)[(p) / 4] = (((a)[(p) / 4] & ~(0xff << (((p) & 3) * 8))) | ((c) << (((p) & 3) * 8))))
+#else
 #define GETCHAR(a,p)   ((u8 *)(a))[(p)]
 #define PUTCHAR(a,p,c) ((u8 *)(a))[(p)] = (u8) (c)
+#endif
 
 #define SETSHIFTEDINT(a,n,v)        \
 {                                   \

OpenCL/m07701_a3-optimized.cl

@@ -15,8 +15,13 @@
 #include "inc_hash_md5.cl"
 #endif
 
+#ifdef IS_AMD
+#define GETCHAR(a,p)  (((a)[(p) / 4] >> (((p) & 3) * 8)) & 0xff)
+#define PUTCHAR(a,p,c) ((a)[(p) / 4] = (((a)[(p) / 4] & ~(0xff << (((p) & 3) * 8))) | ((c) << (((p) & 3) * 8))))
+#else
 #define GETCHAR(a,p)   ((u8 *)(a))[(p)]
 #define PUTCHAR(a,p,c) ((u8 *)(a))[(p)] = (u8) (c)
+#endif
 
 CONSTANT_VK u32a sapb_trans_tbl[256] =
 {