summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--doc/compile_Windows.md2
-rw-r--r--xmrstak/backend/amd/amd_gpu/opencl/wolf-skein.cl76
-rw-r--r--xmrstak/backend/amd/autoAdjust.hpp30
-rw-r--r--xmrstak/backend/cpu/minethd.cpp7
-rw-r--r--xmrstak/backend/nvidia/nvcc_code/cuda_extra.cu6
5 files changed, 84 insertions, 37 deletions
diff --git a/doc/compile_Windows.md b/doc/compile_Windows.md
index e32906c..a512373 100644
--- a/doc/compile_Windows.md
+++ b/doc/compile_Windows.md
@@ -25,7 +25,7 @@
- Download and install https://developer.nvidia.com/cuda-downloads
- For minimal install choose `Custom installation options` during the install and select
- - CUDA/Develpment
+ - CUDA/Development
- CUDA/Visual Studio Integration (ignore the warning during the install that VS2017 is not supported)
- CUDA/Runtime
- Driver components
diff --git a/xmrstak/backend/amd/amd_gpu/opencl/wolf-skein.cl b/xmrstak/backend/amd/amd_gpu/opencl/wolf-skein.cl
index 868757b..279b652 100644
--- a/xmrstak/backend/amd/amd_gpu/opencl/wolf-skein.cl
+++ b/xmrstak/backend/amd/amd_gpu/opencl/wolf-skein.cl
@@ -3,6 +3,7 @@ R"===(
#define WOLF_SKEIN_CL
// Vectorized Skein implementation macros and functions by Wolf
+// Updated by taisel
#define SKEIN_KS_PARITY 0x1BD11BDAA9FC1A22
@@ -22,11 +23,11 @@ static const __constant ulong SKEIN512_256_IV[8] =
0xC36FBAF9393AD185UL, 0x3EEDBA1833EDFC13UL
};
-#define SKEIN_INJECT_KEY(p, s) do { \
+#define SKEIN_INJECT_KEY(p, s, q) do { \
p += h; \
- p.s5 += t[s % 3]; \
- p.s6 += t[(s + 1) % 3]; \
- p.s7 += s; \
+ p.s5 += t[s]; \
+ p.s6 += t[select(s + 1U, 0U, s == 2U)]; \
+ p.s7 += q; \
} while(0)
ulong SKEIN_ROT(const uint2 x, const uint y)
@@ -35,55 +36,55 @@ ulong SKEIN_ROT(const uint2 x, const uint y)
else return(as_ulong(amd_bitalign(x.s10, x, 32 - (y - 32))));
}
-void SkeinMix8(ulong4 *pv0, ulong4 *pv1, const uint rc0, const uint rc1, const uint rc2, const uint rc3)
+void SkeinMix8(ulong4 *pv0, ulong4 *pv1, const ulong4 rc)
{
*pv0 += *pv1;
- (*pv1).s0 = SKEIN_ROT(as_uint2((*pv1).s0), rc0);
- (*pv1).s1 = SKEIN_ROT(as_uint2((*pv1).s1), rc1);
- (*pv1).s2 = SKEIN_ROT(as_uint2((*pv1).s2), rc2);
- (*pv1).s3 = SKEIN_ROT(as_uint2((*pv1).s3), rc3);
+ (*pv1).s0 = SKEIN_ROT(as_uint2((*pv1).s0), rc.s0);
+ (*pv1).s1 = SKEIN_ROT(as_uint2((*pv1).s1), rc.s1);
+ (*pv1).s2 = SKEIN_ROT(as_uint2((*pv1).s2), rc.s2);
+ (*pv1).s3 = SKEIN_ROT(as_uint2((*pv1).s3), rc.s3);
*pv1 ^= *pv0;
}
-ulong8 SkeinEvenRound(ulong8 p, const ulong8 h, const ulong *t, const uint s)
+ulong8 SkeinEvenRound(ulong8 p, const ulong8 h, const ulong *t, const uint s, const uint q)
{
- SKEIN_INJECT_KEY(p, s);
+ SKEIN_INJECT_KEY(p, s, q);
ulong4 pv0 = p.even, pv1 = p.odd;
- SkeinMix8(&pv0, &pv1, 46, 36, 19, 37);
+ SkeinMix8(&pv0, &pv1, (ulong4)(46, 36, 19, 37));
pv0 = shuffle(pv0, (ulong4)(1, 2, 3, 0));
pv1 = shuffle(pv1, (ulong4)(0, 3, 2, 1));
- SkeinMix8(&pv0, &pv1, 33, 27, 14, 42);
+ SkeinMix8(&pv0, &pv1, (ulong4)(33, 27, 14, 42));
pv0 = shuffle(pv0, (ulong4)(1, 2, 3, 0));
pv1 = shuffle(pv1, (ulong4)(0, 3, 2, 1));
- SkeinMix8(&pv0, &pv1, 17, 49, 36, 39);
+ SkeinMix8(&pv0, &pv1, (ulong4)(17, 49, 36, 39));
pv0 = shuffle(pv0, (ulong4)(1, 2, 3, 0));
pv1 = shuffle(pv1, (ulong4)(0, 3, 2, 1));
- SkeinMix8(&pv0, &pv1, 44, 9, 54, 56);
+ SkeinMix8(&pv0, &pv1, (ulong4)(44, 9, 54, 56));
return(shuffle2(pv0, pv1, (ulong8)(1, 4, 2, 7, 3, 6, 0, 5)));
}
-ulong8 SkeinOddRound(ulong8 p, const ulong8 h, const ulong *t, const uint s)
+ulong8 SkeinOddRound(ulong8 p, const ulong8 h, const ulong *t, const uint s, const uint q)
{
- SKEIN_INJECT_KEY(p, s);
+ SKEIN_INJECT_KEY(p, s, q);
ulong4 pv0 = p.even, pv1 = p.odd;
- SkeinMix8(&pv0, &pv1, 39, 30, 34, 24);
+ SkeinMix8(&pv0, &pv1, (ulong4)(39, 30, 34, 24));
pv0 = shuffle(pv0, (ulong4)(1, 2, 3, 0));
pv1 = shuffle(pv1, (ulong4)(0, 3, 2, 1));
- SkeinMix8(&pv0, &pv1, 13, 50, 10, 17);
+ SkeinMix8(&pv0, &pv1, (ulong4)(13, 50, 10, 17));
pv0 = shuffle(pv0, (ulong4)(1, 2, 3, 0));
pv1 = shuffle(pv1, (ulong4)(0, 3, 2, 1));
- SkeinMix8(&pv0, &pv1, 25, 29, 39, 43);
+ SkeinMix8(&pv0, &pv1, (ulong4)(25, 29, 39, 43));
pv0 = shuffle(pv0, (ulong4)(1, 2, 3, 0));
pv1 = shuffle(pv1, (ulong4)(0, 3, 2, 1));
- SkeinMix8(&pv0, &pv1, 8, 35, 56, 22);
+ SkeinMix8(&pv0, &pv1, (ulong4)(8, 35, 56, 22));
return(shuffle2(pv0, pv1, (ulong8)(1, 4, 2, 7, 3, 6, 0, 5)));
}
@@ -92,20 +93,47 @@ ulong8 Skein512Block(ulong8 p, ulong8 h, ulong h8, const ulong *t)
#pragma unroll
for(int i = 0; i < 18; ++i)
{
- p = SkeinEvenRound(p, h, t, i);
+ p = SkeinEvenRound(p, h, t, 0U, i);
++i;
ulong tmp = h.s0;
h = shuffle(h, (ulong8)(1, 2, 3, 4, 5, 6, 7, 0));
h.s7 = h8;
h8 = tmp;
- p = SkeinOddRound(p, h, t, i);
+ p = SkeinOddRound(p, h, t, 1U, i);
+ ++i;
+ tmp = h.s0;
+ h = shuffle(h, (ulong8)(1, 2, 3, 4, 5, 6, 7, 0));
+ h.s7 = h8;
+ h8 = tmp;
+ p = SkeinEvenRound(p, h, t, 2U, i);
+ ++i;
+ tmp = h.s0;
+ h = shuffle(h, (ulong8)(1, 2, 3, 4, 5, 6, 7, 0));
+ h.s7 = h8;
+ h8 = tmp;
+ p = SkeinOddRound(p, h, t, 0U, i);
+ ++i;
+ tmp = h.s0;
+ h = shuffle(h, (ulong8)(1, 2, 3, 4, 5, 6, 7, 0));
+ h.s7 = h8;
+ h8 = tmp;
+ p = SkeinEvenRound(p, h, t, 1U, i);
+ ++i;
+ tmp = h.s0;
+ h = shuffle(h, (ulong8)(1, 2, 3, 4, 5, 6, 7, 0));
+ h.s7 = h8;
+ h8 = tmp;
+ p = SkeinOddRound(p, h, t, 2U, i);
tmp = h.s0;
h = shuffle(h, (ulong8)(1, 2, 3, 4, 5, 6, 7, 0));
h.s7 = h8;
h8 = tmp;
}
- SKEIN_INJECT_KEY(p, 18);
+ p += h;
+ p.s5 += t[0];
+ p.s6 += t[1];
+ p.s7 += 18;
return(p);
}
diff --git a/xmrstak/backend/amd/autoAdjust.hpp b/xmrstak/backend/amd/autoAdjust.hpp
index 511a712..93b71ba 100644
--- a/xmrstak/backend/amd/autoAdjust.hpp
+++ b/xmrstak/backend/amd/autoAdjust.hpp
@@ -94,7 +94,6 @@ private:
}
std::string conf;
- int i = 0;
for(auto& ctx : devVec)
{
/* 1000 is a magic selected limit, the reason is that more than 2GiB memory
@@ -127,14 +126,27 @@ private:
size_t possibleIntensity = std::min( maxThreads , maxIntensity );
// map intensity to a multiple of the compute unit count, 8 is the number of threads per work group
size_t intensity = (possibleIntensity / (8 * ctx.computeUnits)) * ctx.computeUnits * 8;
- conf += std::string(" // gpu: ") + ctx.name + " memory:" + std::to_string(availableMem / byteToMiB) + "\n";
- conf += std::string(" // compute units: ") + std::to_string(ctx.computeUnits) + "\n";
- // set 8 threads per block (this is a good value for the most gpus)
- conf += std::string(" { \"index\" : ") + std::to_string(ctx.deviceIdx) + ",\n" +
- " \"intensity\" : " + std::to_string(intensity) + ", \"worksize\" : " + std::to_string(8) + ",\n" +
- " \"affine_to_cpu\" : false, \"strided_index\" : true\n"
- " },\n";
- ++i;
+ //If the intensity is 0, then it's because the multiple of the unit count is greater than intensity
+ if (intensity == 0)
+ {
+ printer::inst()->print_msg(L0, "WARNING: Auto detected intensity unexpectedly low. Try to set the environment variable GPU_SINGLE_ALLOC_PERCENT.");
+ intensity = possibleIntensity;
+
+ }
+ if (intensity != 0)
+ {
+ conf += std::string(" // gpu: ") + ctx.name + " memory:" + std::to_string(availableMem / byteToMiB) + "\n";
+ conf += std::string(" // compute units: ") + std::to_string(ctx.computeUnits) + "\n";
+ // set 8 threads per block (this is a good value for the most gpus)
+ conf += std::string(" { \"index\" : ") + std::to_string(ctx.deviceIdx) + ",\n" +
+ " \"intensity\" : " + std::to_string(intensity) + ", \"worksize\" : " + std::to_string(8) + ",\n" +
+ " \"affine_to_cpu\" : false, \"strided_index\" : true\n"
+ " },\n";
+ }
+ else
+ {
+ printer::inst()->print_msg(L0, "WARNING: Ignore gpu %s, %s MiB free memory is not enough to suggest settings.", ctx.name.c_str(), std::to_string(availableMem / byteToMiB).c_str());
+ }
}
configTpl.replace("PLATFORMINDEX",std::to_string(platformIndex));
diff --git a/xmrstak/backend/cpu/minethd.cpp b/xmrstak/backend/cpu/minethd.cpp
index 1d9165e..cef4f8e 100644
--- a/xmrstak/backend/cpu/minethd.cpp
+++ b/xmrstak/backend/cpu/minethd.cpp
@@ -448,12 +448,13 @@ void minethd::work_main()
globalStates::inst().calc_start_nonce(result.iNonce, oWork.bNiceHash, nonce_chunk);
}
- *piNonce = ++result.iNonce;
+ *piNonce = result.iNonce;
hash_fun(oWork.bWorkBlob, oWork.iWorkSize, result.bResult, ctx);
if (*piHashVal < oWork.iTarget)
executor::inst()->push_event(ex_event(result, oWork.iPoolId));
+ result.iNonce++;
std::this_thread::yield();
}
@@ -637,7 +638,7 @@ void minethd::multiway_work_main(cn_hash_fun_multi hash_fun_multi)
}
for (size_t i = 0; i < N; i++)
- *piNonce[i] = ++iNonce;
+ *piNonce[i] = iNonce++;
hash_fun_multi(bWorkBlob, oWork.iWorkSize, bHashOut, ctx);
@@ -645,7 +646,7 @@ void minethd::multiway_work_main(cn_hash_fun_multi hash_fun_multi)
{
if (*piHashVal[i] < oWork.iTarget)
{
- executor::inst()->push_event(ex_event(job_result(oWork.sJobID, iNonce - N + 1 + i, bHashOut + 32 * i, iThreadNo), oWork.iPoolId));
+ executor::inst()->push_event(ex_event(job_result(oWork.sJobID, iNonce - N + i, bHashOut + 32 * i, iThreadNo), oWork.iPoolId));
}
}
diff --git a/xmrstak/backend/nvidia/nvcc_code/cuda_extra.cu b/xmrstak/backend/nvidia/nvcc_code/cuda_extra.cu
index d865e13..d5d0039 100644
--- a/xmrstak/backend/nvidia/nvcc_code/cuda_extra.cu
+++ b/xmrstak/backend/nvidia/nvcc_code/cuda_extra.cu
@@ -441,6 +441,12 @@ extern "C" int cuda_get_deviceinfo(nvid_ctx* ctx)
maxMemUsage = size_t(1024u) * byteToMiB;
}
+ if(props.multiProcessorCount <= 6)
+ {
+ // limit memory usage for low end devices to reduce the number of threads
+ maxMemUsage = size_t(1024u) * byteToMiB;
+ }
+
int* tmp;
cudaError_t err;
// a device must be selected to get the right memory usage later on
OpenPOWER on IntegriCloud