From 3234e9d6b5195789f3c8deb8513998d3917772a5 Mon Sep 17 00:00:00 2001 From: Jens Steube Date: Thu, 20 Jun 2019 16:16:56 +0200 Subject: [PATCH] Some more ROCm performance tuning for -m 77xx --- OpenCL/m07700_a0-optimized.cl | 5 +++++ OpenCL/m07700_a1-optimized.cl | 5 +++++ OpenCL/m07700_a3-optimized.cl | 5 +++++ OpenCL/m07701_a0-optimized.cl | 5 +++++ OpenCL/m07701_a1-optimized.cl | 5 +++++ OpenCL/m07701_a3-optimized.cl | 5 +++++ 6 files changed, 30 insertions(+) diff --git a/OpenCL/m07700_a0-optimized.cl b/OpenCL/m07700_a0-optimized.cl index b62bf6ee6..5902ecca4 100644 --- a/OpenCL/m07700_a0-optimized.cl +++ b/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) \ { \ diff --git a/OpenCL/m07700_a1-optimized.cl b/OpenCL/m07700_a1-optimized.cl index 216c29116..dec9f47d9 100644 --- a/OpenCL/m07700_a1-optimized.cl +++ b/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) \ { \ diff --git a/OpenCL/m07700_a3-optimized.cl b/OpenCL/m07700_a3-optimized.cl index 211ec9f2f..0e4175444 100644 --- a/OpenCL/m07700_a3-optimized.cl +++ b/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] = { diff --git a/OpenCL/m07701_a0-optimized.cl b/OpenCL/m07701_a0-optimized.cl index c816f9bf3..3207e1f86 100644 --- a/OpenCL/m07701_a0-optimized.cl +++ b/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) \ { \ diff --git a/OpenCL/m07701_a1-optimized.cl b/OpenCL/m07701_a1-optimized.cl index a83fbddda..32812f0ca 100644 --- a/OpenCL/m07701_a1-optimized.cl +++ b/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) \ { \ diff --git a/OpenCL/m07701_a3-optimized.cl b/OpenCL/m07701_a3-optimized.cl index 4b0aa394e..2e314990a 100644 --- a/OpenCL/m07701_a3-optimized.cl +++ b/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] = {