improve perf by 5% (OPTIM_FOR_FGLRX)

README.md

@@ -102,8 +102,14 @@ Troubleshooting performance issues:
   all the GPUs in the `--use` option, for example `silentarmy --use 0,1,2`
   if the host has three devices with IDs 0, 1, and 2.
 * If some GPUs have less than ~2.4 GB of GPU memory, run
-  `silentarmy --instances 1 --use ...` (2 instances use ~2.4 GB of GPU memory,
+  `silentarmy --instances 1` (2 instances use ~2.4 GB of GPU memory,
   1 instance uses ~1.2 GB of GPU memory.)
+* If you are using an AMD GPU with the **Radeon Software Crimson Edition**
+  driver, as opposed to the **AMDGPU-PRO** driver, then edit param.h and set
+  `OPTIM_FOR_FGLRX` to 1. This will improve performance by +5% and reduce
+  GPU memory usage from 1.2 GB per instance to 805 MB per instance. But do
+  **not** set it if you are using the AMDGPU-PRO driver or else it will
+  degrade performance by -15% or more.
 * If 1 instance still requires too much memory, edit `param.h` and set
   `NR_ROWS_LOG` to `19` (this reduces the per-instance memory usage to ~670 MB)
   and run with `--instances 1`.

input.cl

@@ -532,12 +532,15 @@ void equihash_round(uint round, __global char *ht_src, __global char *ht_dst,
     if (!cnt)
 	// no elements in row, no collisions
 	return ;
+#if NR_ROWS_LOG != 20 || !OPTIM_FOR_FGLRX
     p += xi_offset;
     for (i = 0; i < cnt; i++, p += SLOT_LEN)
         first_words[i] = *(__global uchar *)p;
+#endif
     // find collisions
     for (i = 0; i < cnt; i++)
         for (j = i + 1; j < cnt; j++)
+#if NR_ROWS_LOG != 20 || !OPTIM_FOR_FGLRX
             if ((first_words[i] & mask) ==
 		    (first_words[j] & mask))
               {
@@ -558,6 +561,9 @@ void equihash_round(uint round, __global char *ht_src, __global char *ht_dst,
       {
         i = collisions[n] & 0xff;
         j = collisions[n] >> 8;
+#else
+      {
+#endif
         a = (__global ulong *)
             (ht_src + tid * NR_SLOTS * SLOT_LEN + i * SLOT_LEN + xi_offset);
         b = (__global ulong *)

param.h

@@ -8,6 +8,10 @@
 // but occasionally misses ~1% of solutions.
 #define NR_ROWS_LOG                     20
 
+// Set this to 1 if you are using an AMD GPU with the Radeon Software Crimson
+// Edition driver (fglrx.ko), see README.md.
+#define OPTIM_FOR_FGLRX			0
+
 // Make hash tables OVERHEAD times larger than necessary to store the average
 // number of elements per row. The ideal value is as small as possible to
 // reduce memory usage, but not too small or else elements are dropped from the
@@ -25,6 +29,8 @@
 #define OVERHEAD                        3
 #elif NR_ROWS_LOG == 19
 #define OVERHEAD                        5
+#elif NR_ROWS_LOG == 20 && OPTIM_FOR_FGLRX
+#define OVERHEAD                        6
 #elif NR_ROWS_LOG == 20
 #define OVERHEAD                        9
 #endif