Improve block find performance (llvm#412)

Author: b-sumner

amd/device-libs/ockl/src/dm.cl

@@ -51,26 +51,25 @@ struct kind_info_s {
     uint first_unusable;
     uint gap_unusable;
     uint pattern_unusable;
-    uint spread_factor;
 };
 
 static const __constant struct kind_info_s kinfo[NUM_KINDS] = {
-    { /*  0:   16 */ 130054, 129546, 110114, 16288,    6, 256, 0x00000000, 4195 },
-    { /*  1:   24 */  86927,  86758,  73744, 10904,  399, 512, 0x00000000, 2804 },
-    { /*  2:   32 */  65280,  64770,  55054,  8192,    0, 128, 0x00000000, 2107 },
-    { /*  3:   48 */  43576,  43406,  36895,  5504,   56, 256, 0x00000000, 1405 },
-    { /*  4:   64 */  32703,  32193,  27364,  4160,   63,  64, 0x00000000, 1054 },
-    { /*  5:   96 */  21816,  21646,  18399,  2816,   56, 128, 0x00000000,  703 },
-    { /*  6:  128 */  16367,  15856,  13477,  2176,   15,  32, 0x00008000,  527 },
-    { /*  7:  192 */  10915,  10745,   9133,  1472,   35,  64, 0x00000000,  352 },
-    { /*  8:  256 */   8187,   7676,   6524,  1280,   11,  16, 0x08000800,  265 },
-    { /*  9:  384 */   5459,   5289,   4495,   896,   19,  32, 0x00080000,  176 },
-    { /* 10:  512 */   4094,   3583,   3045,  1024,    6,   8, 0x40404040,  133 },
-    { /* 11:  768 */   2730,   2560,   2176,   512,   10,  16, 0x04000400,   89 },
-    { /* 12: 1024 */   2047,   1536,   1305,  1024,    3,   4, 0x88888888,   66 },
-    { /* 13: 1536 */   1365,   1195,   1015,   512,    5,   8, 0x20202020,   44 },
-    { /* 14: 2048 */   1023,    512,    435,  2048,    1,   2, 0xaaaaaaaa,   34 },
-    { /* 15: 3072 */    682,    512,    435,  2048,    2,   4, 0x44444444,   35 },
+    { /*  0:   16 */ 130054, 129546, 110114, 16288,    6, 256, 0x00000000 },
+    { /*  1:   24 */  86927,  86758,  73744, 10904,  399, 512, 0x00000000 },
+    { /*  2:   32 */  65280,  64770,  55054,  8192,    0, 128, 0x00000000 },
+    { /*  3:   48 */  43576,  43406,  36895,  5504,   56, 256, 0x00000000 },
+    { /*  4:   64 */  32703,  32193,  27364,  4160,   63,  64, 0x00000000 },
+    { /*  5:   96 */  21816,  21646,  18399,  2816,   56, 128, 0x00000000 },
+    { /*  6:  128 */  16367,  15856,  13477,  2176,   15,  32, 0x00008000 },
+    { /*  7:  192 */  10915,  10745,   9133,  1472,   35,  64, 0x00000000 },
+    { /*  8:  256 */   8187,   7676,   6524,  1280,   11,  16, 0x08000800 },
+    { /*  9:  384 */   5459,   5289,   4495,   896,   19,  32, 0x00080000 },
+    { /* 10:  512 */   4094,   3583,   3045,  1024,    6,   8, 0x40404040 },
+    { /* 11:  768 */   2730,   2560,   2176,   512,   10,  16, 0x04000400 },
+    { /* 12: 1024 */   2047,   1536,   1305,  1024,    3,   4, 0x88888888 },
+    { /* 13: 1536 */   1365,   1195,   1015,   512,    5,   8, 0x20202020 },
+    { /* 14: 2048 */   1023,    512,    435,  2048,    1,   2, 0xaaaaaaaa },
+    { /* 15: 3072 */    682,    512,    435,  2048,    2,   4, 0x44444444 },
 };
 
 // A slab is a chunk of memory used to provide "block"s whose addresses are
@@ -241,13 +240,6 @@ pattern_unusable(kind_t k)
     return kinfo[k].pattern_unusable;
 }
 
-// The multiplier used to spread out the probes of individual lanes while searching a slab of kind k
-static uint
-spread_factor(kind_t k)
-{
-    return kinfo[k].spread_factor;
-}
-
 // The number of active lanes at this point
 static uint
 active_lane_count(void)
@@ -849,7 +841,7 @@ block_find(__global sdata_t *sdp)
     uint i = 0;
     if (aid == 0)
         i = AFA(&sp->start, nactive, memory_order_relaxed);
-    i = ((first(i) + aid) * spread_factor(k) % num_blocks(k)) >> 5;
+    i = (((first(i) + aid) << 5) % num_blocks(k)) >> 5;
 
     uint n = (num_blocks(k) + 31) >> 5;