Added support for CUDA compute capability 3.0. Multiple CUDA architecture compilation support.

CMakeLists.txt

@@ -394,7 +394,13 @@ if(WITH_CUDA)
     set(
             CUDA_NVCC_FLAGS
             ${CUDA_NVCC_FLAGS};
-            -O3 -arch=compute_35 -std=c++11
+            -O3 -std=c++11 --ptxas-options=-v
+            -gencode=arch=compute_75,code="sm_75,compute_75"
+            -gencode=arch=compute_61,code="sm_61,compute_61"
+            -gencode=arch=compute_52,code="sm_52,compute_52"
+            -gencode=arch=compute_50,code="sm_50,compute_50"
+            -gencode=arch=compute_35,code="sm_35,compute_35"
+            -gencode=arch=compute_30,code="sm_30,compute_30"
     )
     cuda_add_library(cuda_hasher MODULE ${SOURCE_CUDA_HASHER})
     set_target_properties(cuda_hasher

src/crypto/argon2_hasher/hash/gpu/cuda/cuda_kernel.cu

@@ -19,6 +19,11 @@
 
 #include "blake2b.cu"
 
+#ifndef __CUDA_ARCH__
+#define __CUDA_ARCH__ 0
+#endif
+
+#if (__CUDA_ARCH__ >= 350)
     #define COMPUTE	\
         asm ("{"	\
             ".reg .u32 s1, s2, s3, s4;\n\t"	\
@@ -70,6 +75,39 @@
             "shf.r.wrap.b32 %3, s3, s4, 31;\n\t"	\
             "shf.r.wrap.b32 %2, s4, s3, 31;\n\t"	\
         "}" : "+r"(tmp_a.x), "+r"(tmp_a.y), "+r"(tmp_a.z), "+r"(tmp_a.w), "+r"(tmp_b.x), "+r"(tmp_b.y), "+r"(tmp_b.z), "+r"(tmp_b.w));
+#else
+    #define downsample(x, lo, hi) \
+    { \
+        lo = (uint32_t)x; \
+        hi = (uint32_t)(x >> 32); \
+    }
+
+    #define upsample(lo, hi) (((uint64_t)(hi) << 32) | (uint64_t)(lo))
+
+    #define rotate(x, n) (((x) >> (64-n)) | ((x) << n))
+
+    #define fBlaMka(x, y) ((x) + (y) + 2 * upsample((uint32_t)(x) * (uint32_t)y, __umulhi((uint32_t)(x), (uint32_t)(y))))
+
+    #define COMPUTE \
+    { \
+        uint64_t a64 = upsample(tmp_a.x, tmp_a.y); \
+        uint64_t b64 = upsample(tmp_a.z, tmp_a.w); \
+        uint64_t c64 = upsample(tmp_b.x, tmp_b.y); \
+        uint64_t d64 = upsample(tmp_b.z, tmp_b.w); \
+        a64 = fBlaMka(a64, b64);          \
+        d64 = rotate(d64 ^ a64, 32);      \
+        c64 = fBlaMka(c64, d64);          \
+        b64 = rotate(b64 ^ c64, 40);      \
+        a64 = fBlaMka(a64, b64);          \
+        d64 = rotate(d64 ^ a64, 48);      \
+        c64 = fBlaMka(c64, d64);          \
+        b64 = rotate(b64 ^ c64, 1);       \
+        downsample(a64, tmp_a.x, tmp_a.y); \
+        downsample(b64, tmp_a.z, tmp_a.w); \
+        downsample(c64, tmp_b.x, tmp_b.y); \
+        downsample(d64, tmp_b.z, tmp_b.w); \
+    }
+#endif // __CUDA_ARCH__
 
 #define G1(data)           \
 {                           \