first commit

2019-08-25 16:12:39 +08:00 · 2019-08-25 16:12:39 +08:00 · 604f1cb18a
commit 604f1cb18a
9 changed files with 1683 additions and 0 deletions
--- a/gpu.h
+++ b/gpu.h
@ -0,0 +1,80 @@
+//#define GPUTYPE 8800		// MP=16 
+//#define NJBL_value 32
+
+//#define GPUTYPE C1060		// MP=30 
+//#define NJBL_value 30
+
+//#define GPUTYPE M2070		// MP=14 
+//#define NJBL_value 28
+
+//#define GPUTYPE GF460		// MP=7	 
+//#define NJBL_value 14
+
+//#define GPUTYPE GF470		// MP=14 
+//#define NJBL_value 28
+
+//#define GPUTYPE GF480		// MP=15 
+//#define NJBL_value 30
+
+//#define GPUTYPE GF570		// MP=15 
+//#define NJBL_value 30
+
+#define GPUTYPE K20m		// MP=13 
+#define NJBL_value 26
+
+//#define GPUTYPE GF660		// MP=5 
+//#define NJBL_value 10
+
+//#define GPUTYPE GF780		// MP=15
+//#define NJBL_value 30
+
+//#define GPUTYPE GF1080	// MP=20 
+//#define NJBL_value 20
+
+//#define GPUTYPE GFTITANX	// MP=24
+//#define NJBL_value 24
+
+
+//	NJBLOCKS = 1 or 2 x MP 
+//	NXREDUCE is a power of 2 and >= NJBLOCKS ~32 is good for all
+//	NYREDUCE is a power of 2 < NXREDUCE ~4 is good for all
+
+enum{
+	SM_VER   = 35,
+	NTHREADS = 128,
+	NTHSCAT  = 64,
+	NIBLOCKS = 16,
+	NJPSHRE  = 32,
+	NIMAX    = (NTHREADS * NIBLOCKS),
+	NJBLOCKS = NJBL_value,
+	NXREDUCE = 32,
+	NYREDUCE = 4,
+};
+
+
+/*
+----------------------------------------------------------------------------
+type	 MP x Cores  x GPU clock rate 	SP Gflop/s	DP Gflop/s 
+	      per MP	   (GHz)       (nbody n=200k)  (nbody n=200k)
+
+8800 GT  14 x  8     x 1.62 =  181.44	 		  - 
+8800 512 16 x  8     x 1.80 =  230.40	 298.266	  - 
+9800 GTX 16 x  8     x 1.85 =  236.80	 331.754	  -
+250 GTS  16 x  8     x 1.73 =  221.44	 
+
+C1060    30 x  8     x 1.30 =  312.00	 450.861	 55.412
+
+M2070	 14 x  32    x 1.15 =  515.20	 546.416	250.454
+
+460 v2    7 x  48    x 1.56 =  524.16    421.335	 67.939
+470      14 x  32    x 1.22 =  546.56    578.572	105.796
+480      15 x  32    x 1.40 =  672.00	 705.514	145.502
+570	 15 x  32    x 1.46 =  700.80	 736.882	151.960
+
+670	  7 x 192    x 0.71 =  954.24	1151.030        100.092
+680	  8 x 192    x 0.71 = 1090.56	1187.710	101.211
+K20m     13 x 192    x 0.71 = 1772.16	1322.374	576.523
+TITAN    14 x 192    x 0.88 = 2365.44	2026.287	751.671
+780 Ti	 15 x 192    x 0.93 = 2678.40	2691.603	198.987
+----------------------------------------------------------------------------
+*/