From 332396a0036ef1e3407462acb815e8a21fc3dc52 Mon Sep 17 00:00:00 2001
From: jsteube <jens.steube@gmail.com>
Date: Fri, 28 Jul 2017 02:28:52 +0200
Subject: [PATCH] Fix SCRYPT on ROCm
---
OpenCL/m08900.cl | 26 ++++++++++++--------------
OpenCL/m15700.cl | 26 ++++++++++++--------------
src/shared.c | 6 ------
3 files changed, 24 insertions(+), 34 deletions(-)
diff --git a/OpenCL/m08900.cl b/OpenCL/m08900.cl
index 402c2e0c0..9d65e1a71 100644
--- a/OpenCL/m08900.cl
+++ b/OpenCL/m08900.cl
@@ -138,6 +138,16 @@ void scrypt_smix (uint4 *X, uint4 *T, __global uint4 *V0, __global uint4 *V1, __
const u32 xd4 = x / 4;
const u32 xm4 = x & 3;
+ __global uint4 *V;
+
+ switch (xm4)
+ {
+ case 0: V = V0; break;
+ case 1: V = V1; break;
+ case 2: V = V2; break;
+ case 3: V = V3; break;
+ }
+
#ifdef _unroll
#pragma unroll
#endif
@@ -156,13 +166,7 @@ void scrypt_smix (uint4 *X, uint4 *T, __global uint4 *V0, __global uint4 *V1, __
for (u32 y = 0; y < ySIZE; y++)
{
- switch (xm4)
- {
- case 0: for (u32 z = 0; z < zSIZE; z++) V0[CO] = X[z]; break;
- case 1: for (u32 z = 0; z < zSIZE; z++) V1[CO] = X[z]; break;
- case 2: for (u32 z = 0; z < zSIZE; z++) V2[CO] = X[z]; break;
- case 3: for (u32 z = 0; z < zSIZE; z++) V3[CO] = X[z]; break;
- }
+ for (u32 z = 0; z < zSIZE; z++) V[CO] = X[z];
for (u32 i = 0; i < SCRYPT_TMTO; i++) salsa_r (X);
}
@@ -175,13 +179,7 @@ void scrypt_smix (uint4 *X, uint4 *T, __global uint4 *V0, __global uint4 *V1, __
const u32 km = k - (y * SCRYPT_TMTO);
- switch (xm4)
- {
- case 0: for (u32 z = 0; z < zSIZE; z++) T[z] = V0[CO]; break;
- case 1: for (u32 z = 0; z < zSIZE; z++) T[z] = V1[CO]; break;
- case 2: for (u32 z = 0; z < zSIZE; z++) T[z] = V2[CO]; break;
- case 3: for (u32 z = 0; z < zSIZE; z++) T[z] = V3[CO]; break;
- }
+ for (u32 z = 0; z < zSIZE; z++) T[z] = V[CO];
for (u32 i = 0; i < km; i++) salsa_r (T);
diff --git a/OpenCL/m15700.cl b/OpenCL/m15700.cl
index 57a33dc1b..a1971093f 100644
--- a/OpenCL/m15700.cl
+++ b/OpenCL/m15700.cl
@@ -138,6 +138,16 @@ void scrypt_smix (uint4 *X, uint4 *T, __global uint4 *V0, __global uint4 *V1, __
const u32 xd4 = x / 4;
const u32 xm4 = x & 3;
+ __global uint4 *V;
+
+ switch (xm4)
+ {
+ case 0: V = V0; break;
+ case 1: V = V1; break;
+ case 2: V = V2; break;
+ case 3: V = V3; break;
+ }
+
#ifdef _unroll
#pragma unroll
#endif
@@ -156,13 +166,7 @@ void scrypt_smix (uint4 *X, uint4 *T, __global uint4 *V0, __global uint4 *V1, __
for (u32 y = 0; y < ySIZE; y++)
{
- switch (xm4)
- {
- case 0: for (u32 z = 0; z < zSIZE; z++) V0[CO] = X[z]; break;
- case 1: for (u32 z = 0; z < zSIZE; z++) V1[CO] = X[z]; break;
- case 2: for (u32 z = 0; z < zSIZE; z++) V2[CO] = X[z]; break;
- case 3: for (u32 z = 0; z < zSIZE; z++) V3[CO] = X[z]; break;
- }
+ for (u32 z = 0; z < zSIZE; z++) V[CO] = X[z];
for (u32 i = 0; i < SCRYPT_TMTO; i++) salsa_r (X);
}
@@ -175,13 +179,7 @@ void scrypt_smix (uint4 *X, uint4 *T, __global uint4 *V0, __global uint4 *V1, __
const u32 km = k - (y * SCRYPT_TMTO);
- switch (xm4)
- {
- case 0: for (u32 z = 0; z < zSIZE; z++) T[z] = V0[CO]; break;
- case 1: for (u32 z = 0; z < zSIZE; z++) T[z] = V1[CO]; break;
- case 2: for (u32 z = 0; z < zSIZE; z++) T[z] = V2[CO]; break;
- case 3: for (u32 z = 0; z < zSIZE; z++) T[z] = V3[CO]; break;
- }
+ for (u32 z = 0; z < zSIZE; z++) T[z] = V[CO];
for (u32 i = 0; i < km; i++) salsa_r (T);
diff --git a/src/shared.c b/src/shared.c
index 83442586d..9ea3d575b 100644
--- a/src/shared.c
+++ b/src/shared.c
@@ -334,12 +334,6 @@ void setup_environment_variables ()
putenv ((char *) "DISPLAY=:0");
}
- if (getenv ("GPU_FORCE_64BIT_PTR") == NULL)
- putenv ((char *) "GPU_FORCE_64BIT_PTR=0");
-
- if (getenv ("GPU_USE_SYNC_OBJECTS") == NULL)
- putenv ((char *) "GPU_USE_SYNC_OBJECTS=1");
-
if (getenv ("OCL_CODE_CACHE_ENABLE") == NULL)
putenv ((char *) "OCL_CODE_CACHE_ENABLE=0");