From a2728af4f793dcf1a1074b784ddf8ac33401c7c7 Mon Sep 17 00:00:00 2001
From: SChernykh <sergey.v.chernykh@gmail.com>
Date: Thu, 10 Mar 2022 00:24:49 +0100
Subject: [PATCH] Dero HE (astrobwt/v2) CUDA config generator

---
 src/backend/cuda/CudaConfig_gen.h               |  9 ++++++++-
 src/backend/cuda/runners/CudaAstroBWTRunner.cpp | 16 +++++++++++++---
 src/backend/cuda/runners/CudaAstroBWTRunner.h   |  2 ++
 3 files changed, 23 insertions(+), 4 deletions(-)

diff --git a/src/backend/cuda/CudaConfig_gen.h b/src/backend/cuda/CudaConfig_gen.h
index d92b05260..73c77e48d 100644
--- a/src/backend/cuda/CudaConfig_gen.h
+++ b/src/backend/cuda/CudaConfig_gen.h
@@ -139,7 +139,14 @@ size_t inline generate<Algorithm::RANDOM_X>(Threads<CudaThreads> &threads, const
 template<>
 size_t inline generate<Algorithm::ASTROBWT>(Threads<CudaThreads> &threads, const std::vector<CudaDevice> &devices)
 {
-    return generate(Algorithm::kASTROBWT, threads, Algorithm::ASTROBWT_DERO, devices);
+    size_t count = 0;
+
+    if (!threads.isExist(Algorithm::ASTROBWT_DERO_2)) {
+        count += threads.move(Algorithm::kASTROBWT_DERO_2, CudaThreads(devices, Algorithm::ASTROBWT_DERO_2));
+    }
+
+    count += generate(Algorithm::kASTROBWT, threads, Algorithm::ASTROBWT_DERO, devices);
+    return count;
 }
 #endif
 
diff --git a/src/backend/cuda/runners/CudaAstroBWTRunner.cpp b/src/backend/cuda/runners/CudaAstroBWTRunner.cpp
index 43431aac4..6b078e4f8 100644
--- a/src/backend/cuda/runners/CudaAstroBWTRunner.cpp
+++ b/src/backend/cuda/runners/CudaAstroBWTRunner.cpp
@@ -32,11 +32,17 @@
 constexpr uint32_t xmrig::CudaAstroBWTRunner::BWT_DATA_STRIDE;
 
 
-xmrig::CudaAstroBWTRunner::CudaAstroBWTRunner(size_t index, const CudaLaunchData &data) :
-    CudaBaseRunner(index, data)
+xmrig::CudaAstroBWTRunner::CudaAstroBWTRunner(size_t index, const CudaLaunchData &data)
+    : CudaBaseRunner(index, data)
+    , m_algorithm(data.algorithm)
 {
     m_intensity = m_data.thread.threads() * m_data.thread.blocks();
     m_intensity -= m_intensity % 32;
+
+    // Dero HE has very fast blocks, so we can't use high intensity
+    if ((m_algorithm == Algorithm::ASTROBWT_DERO_2) && (m_intensity > 4096)) {
+        m_intensity = 4096;
+    }
 }
 
 
@@ -58,10 +64,14 @@ bool xmrig::CudaAstroBWTRunner::set(const Job &job, uint8_t *blob)
 
 size_t xmrig::CudaAstroBWTRunner::roundSize() const
 {
+    if (m_algorithm == Algorithm::ASTROBWT_DERO_2) {
+        return m_intensity;
+    }
+
     constexpr uint32_t STAGE1_SIZE = 147253;
     constexpr uint32_t STAGE1_DATA_STRIDE = (STAGE1_SIZE + 256 + 255) & ~255U;
 
-    const uint32_t BATCH2_SIZE = m_intensity;
+    const uint32_t BATCH2_SIZE = static_cast<uint32_t>(m_intensity);
     const uint32_t BWT_ALLOCATION_SIZE = BATCH2_SIZE * BWT_DATA_STRIDE;
     const uint32_t BATCH1_SIZE = (BWT_ALLOCATION_SIZE / STAGE1_DATA_STRIDE) & ~255U;
 
diff --git a/src/backend/cuda/runners/CudaAstroBWTRunner.h b/src/backend/cuda/runners/CudaAstroBWTRunner.h
index 0afee8c1c..d217d65a5 100644
--- a/src/backend/cuda/runners/CudaAstroBWTRunner.h
+++ b/src/backend/cuda/runners/CudaAstroBWTRunner.h
@@ -27,6 +27,7 @@
 
 
 #include "backend/cuda/runners/CudaBaseRunner.h"
+#include "base/crypto/Algorithm.h"
 
 
 namespace xmrig {
@@ -50,6 +51,7 @@ protected:
 
 private:
     size_t m_intensity  = 0;
+    Algorithm m_algorithm;
 };