From 3f6bd5a25e5b6c0e22a99f5d0b296be61b766a63 Mon Sep 17 00:00:00 2001
From: psychocrypt <psychocryptHPC@gmail.com>
Date: Fri, 1 Feb 2019 22:30:33 +0100
Subject: [PATCH] CUDA: optimze cn_gpu auto suggestion
optimize the algorithm for cryptonight_gpu autosuggestion
---
.../backend/nvidia/nvcc_code/cuda_extra.cu | 45 +++++++++++++++----
1 file changed, 37 insertions(+), 8 deletions(-)
diff --git a/xmrstak/backend/nvidia/nvcc_code/cuda_extra.cu b/xmrstak/backend/nvidia/nvcc_code/cuda_extra.cu
index e4574e2..a37ecc8 100644
--- a/xmrstak/backend/nvidia/nvcc_code/cuda_extra.cu
+++ b/xmrstak/backend/nvidia/nvcc_code/cuda_extra.cu
@@ -593,6 +593,10 @@ extern "C" int cuda_get_deviceinfo(nvid_ctx* ctx)
}
}
+ auto neededAlgorithms = ::jconf::inst()->GetCurrentCoinSelection().GetAllAlgorithms();
+ bool useCryptonight_gpu = std::find(neededAlgorithms.begin(), neededAlgorithms.end(), cryptonight_gpu) != neededAlgorithms.end();
+
+
// set all device option those marked as auto (-1) to a valid value
if(ctx->device_blocks == -1)
{
@@ -600,8 +604,11 @@ extern "C" int cuda_get_deviceinfo(nvid_ctx* ctx)
* - 3 * SMX count for >=sm_30
* - 2 * SMX count for <sm_30
*/
- ctx->device_blocks = props.multiProcessorCount *
- ( props.major < 3 ? 2 : 3 );
+ ctx->device_blocks = props.multiProcessorCount * (props.major < 3 ? 2 : 3);
+
+ // use 6 blocks per SM for sm_2X else 8 blocks
+ if(useCryptonight_gpu)
+ ctx->device_blocks = props.multiProcessorCount * (props.major < 3 ? 6 : 8);
// increase bfactor for low end devices to avoid that the miner is killed by the OS
if(props.multiProcessorCount <= 6)
@@ -613,7 +620,16 @@ extern "C" int cuda_get_deviceinfo(nvid_ctx* ctx)
* `cryptonight_core_gpu_phase1` and `cryptonight_core_gpu_phase3` starts
* `8 * ctx->device_threads` threads per block
*/
- ctx->device_threads = 64;
+ const uint32_t maxThreadsPerBlock = props.major < 3 ? 512 : 1024;
+
+ // for the most algorithms we are using 8 threads per hash
+ uint32_t threadsPerHash = 8;
+
+ // phase2_gpu uses 16 threads per hash
+ if(useCryptonight_gpu)
+ threadsPerHash = 16;
+
+ ctx->device_threads = maxThreadsPerBlock / threadsPerHash;
constexpr size_t byteToMiB = 1024u * 1024u;
// no limit by default 1TiB
@@ -678,8 +694,6 @@ extern "C" int cuda_get_deviceinfo(nvid_ctx* ctx)
ctx->total_device_memory = totalMemory;
ctx->free_device_memory = freeMemory;
- auto neededAlgorithms = ::jconf::inst()->GetCurrentCoinSelection().GetAllAlgorithms();
-
size_t hashMemSize = 0;
for(const auto algo : neededAlgorithms)
{
@@ -725,10 +739,9 @@ extern "C" int cuda_get_deviceinfo(nvid_ctx* ctx)
// use only odd number of threads
ctx->device_threads = ctx->device_threads & 0xFFFFFFFE;
- if(props.major == 2 && ctx->device_threads > 64)
+ if(ctx->device_threads > maxThreadsPerBlock / threadsPerHash)
{
- // Fermi gpus only support 512 threads per block (we need start 4 * configured threads)
- ctx->device_threads = 64;
+ ctx->device_threads = maxThreadsPerBlock / threadsPerHash;
}
// check if cryptonight_monero_v8 is selected for the user pool
@@ -749,6 +762,22 @@ extern "C" int cuda_get_deviceinfo(nvid_ctx* ctx)
ctx->device_blocks = blockOptimal;
}
}
+ else if(useCryptonight_gpu)
+ {
+ // 8 based on my profiling sessions maybe it must be adjusted later
+ size_t threads = 8;
+ // 8 is chosen by checking the occupancy calculator
+ size_t blockOptimal = 8 * ctx->device_mpcount;
+ if(gpuArch >= 70)
+ blockOptimal = 5 * ctx->device_mpcount;
+
+ if(blockOptimal * threads * hashMemSize < limitedMemory)
+ {
+ ctx->device_threads = threads;
+ ctx->device_blocks = blockOptimal;
+ }
+
+ }
}
printf("device init succeeded\n");
--
GitLab