bones-compiler 1.1.0 → 1.3.1

Sign up to get free protection for your applications and to get access to all the features.
Files changed (211) hide show
  1. checksums.yaml +15 -0
  2. data/CHANGELOG +37 -0
  3. data/LICENSE +1 -1
  4. data/README.rdoc +95 -70
  5. data/Rakefile +78 -3
  6. data/VERSION +1 -1
  7. data/bin/adarwin +17 -0
  8. data/examples/benchmarks/PolyBench/2mm.c +104 -0
  9. data/examples/benchmarks/{3mm.c → PolyBench/3mm.c} +5 -2
  10. data/examples/benchmarks/{adi.c → PolyBench/adi.c} +6 -3
  11. data/examples/benchmarks/{atax.c → PolyBench/atax.c} +5 -2
  12. data/examples/benchmarks/{bicg.c → PolyBench/bicg.c} +5 -2
  13. data/examples/benchmarks/{cholesky.c → PolyBench/cholesky.c} +3 -0
  14. data/examples/benchmarks/{common.h → PolyBench/common.h} +2 -2
  15. data/examples/benchmarks/{correlation.c → PolyBench/correlation.c} +16 -7
  16. data/examples/benchmarks/{covariance.c → PolyBench/covariance.c} +7 -2
  17. data/examples/benchmarks/{doitgen.c → PolyBench/doitgen.c} +5 -2
  18. data/examples/benchmarks/{durbin.c → PolyBench/durbin.c} +3 -0
  19. data/examples/benchmarks/{dynprog.c → PolyBench/dynprog.c} +3 -0
  20. data/examples/benchmarks/{fdtd-2d-apml.c → PolyBench/fdtd-2d-apml.c} +3 -0
  21. data/examples/benchmarks/{fdtd-2d.c → PolyBench/fdtd-2d.c} +5 -2
  22. data/examples/benchmarks/{floyd-warshall.c → PolyBench/floyd-warshall.c} +3 -0
  23. data/examples/benchmarks/{gemm.c → PolyBench/gemm.c} +5 -2
  24. data/examples/benchmarks/{gemver.c → PolyBench/gemver.c} +5 -2
  25. data/examples/benchmarks/{gesummv.c → PolyBench/gesummv.c} +5 -2
  26. data/examples/benchmarks/{gramschmidt.c → PolyBench/gramschmidt.c} +3 -0
  27. data/examples/benchmarks/{jacobi-1d-imper.c → PolyBench/jacobi-1d-imper.c} +10 -2
  28. data/examples/benchmarks/{jacobi-2d-imper.c → PolyBench/jacobi-2d-imper.c} +8 -3
  29. data/examples/benchmarks/{lu.c → PolyBench/lu.c} +3 -0
  30. data/examples/benchmarks/{ludcmp.c → PolyBench/ludcmp.c} +3 -0
  31. data/examples/benchmarks/{mvt.c → PolyBench/mvt.c} +6 -2
  32. data/examples/benchmarks/{reg_detect.c → PolyBench/reg_detect.c} +3 -0
  33. data/examples/benchmarks/{seidel-2d.c → PolyBench/seidel-2d.c} +3 -0
  34. data/examples/benchmarks/{symm.c → PolyBench/symm.c} +3 -0
  35. data/examples/benchmarks/{syr2k.c → PolyBench/syr2k.c} +5 -2
  36. data/examples/benchmarks/{syrk.c → PolyBench/syrk.c} +7 -4
  37. data/examples/benchmarks/{trisolv.c → PolyBench/trisolv.c} +3 -0
  38. data/examples/benchmarks/{trmm.c → PolyBench/trmm.c} +3 -0
  39. data/examples/benchmarks/Rodinia/cfd.c +180 -0
  40. data/examples/benchmarks/Rodinia/hotspot.c +228 -0
  41. data/examples/benchmarks/Rodinia/kmeans.c +164 -0
  42. data/examples/benchmarks/Rodinia/srad.c +188 -0
  43. data/examples/benchmarks/other/common.h +0 -0
  44. data/examples/benchmarks/other/dct.c +58 -0
  45. data/examples/benchmarks/other/mm.c +50 -0
  46. data/examples/benchmarks/{saxpy.c → other/saxpy.c} +11 -7
  47. data/examples/chunk/{example1.c → example01.c} +0 -0
  48. data/examples/chunk/{example2.c → example02.c} +0 -0
  49. data/examples/chunk/{example3.c → example03.c} +0 -0
  50. data/examples/chunk/{example4.c → example04.c} +0 -0
  51. data/examples/chunk/{example5.c → example05.c} +0 -0
  52. data/examples/chunk/example06.c +45 -0
  53. data/examples/chunk/example07.c +49 -0
  54. data/examples/dependences/example01.c +42 -0
  55. data/examples/dependences/example02.c +40 -0
  56. data/examples/dependences/example03.c +43 -0
  57. data/examples/dependences/example04.c +44 -0
  58. data/examples/dependences/example05.c +42 -0
  59. data/examples/element/{example1.c → example01.c} +0 -0
  60. data/examples/element/{example2.c → example02.c} +2 -2
  61. data/examples/element/{example3.c → example03.c} +0 -0
  62. data/examples/element/{example4.c → example04.c} +0 -0
  63. data/examples/element/{example5.c → example05.c} +0 -0
  64. data/examples/element/{example6.c → example06.c} +0 -0
  65. data/examples/element/{example7.c → example07.c} +0 -0
  66. data/examples/element/{example8.c → example08.c} +0 -0
  67. data/examples/element/{example9.c → example09.c} +0 -0
  68. data/examples/element/example13.c +73 -0
  69. data/examples/fusion/example01.c +68 -0
  70. data/examples/fusion/example02.c +73 -0
  71. data/examples/fusion/example03.c +72 -0
  72. data/examples/fusion/example04.c +61 -0
  73. data/examples/fusion/example05.c +55 -0
  74. data/examples/neighbourhood/{example1.c → example01.c} +0 -0
  75. data/examples/neighbourhood/{example2.c → example02.c} +0 -0
  76. data/examples/neighbourhood/{example3.c → example03.c} +0 -0
  77. data/examples/neighbourhood/{example4.c → example04.c} +0 -0
  78. data/examples/neighbourhood/example05.c +44 -0
  79. data/examples/shared/{example1.c → example01.c} +0 -0
  80. data/examples/shared/{example2.c → example02.c} +0 -0
  81. data/examples/shared/{example3.c → example03.c} +0 -0
  82. data/examples/shared/{example4.c → example04.c} +0 -0
  83. data/examples/shared/{example5.c → example05.c} +0 -0
  84. data/lib/adarwin.rb +62 -0
  85. data/lib/adarwin/dependences.rb +268 -0
  86. data/lib/adarwin/engine.rb +277 -0
  87. data/lib/adarwin/fusion.rb +174 -0
  88. data/lib/adarwin/interval.rb +57 -0
  89. data/lib/adarwin/memorycopies.rb +153 -0
  90. data/lib/adarwin/nest.rb +225 -0
  91. data/lib/adarwin/preprocessor.rb +76 -0
  92. data/lib/adarwin/reference.rb +261 -0
  93. data/lib/bones.rb +4 -55
  94. data/lib/bones/algorithm.rb +77 -40
  95. data/lib/bones/copy.rb +26 -0
  96. data/lib/bones/engine.rb +147 -31
  97. data/lib/bones/preprocessor.rb +92 -12
  98. data/lib/bones/species.rb +4 -3
  99. data/lib/bones/structure.rb +14 -4
  100. data/lib/castaddon.rb +11 -6
  101. data/lib/castaddon/node_adarwin.rb +245 -0
  102. data/lib/castaddon/node_bones.rb +316 -0
  103. data/lib/castaddon/node_common.rb +289 -0
  104. data/lib/castaddon/transformations.rb +236 -0
  105. data/lib/common.rb +216 -0
  106. data/skeletons/CPU-C/common/header.c +3 -0
  107. data/skeletons/CPU-C/common/mem_global.c +0 -0
  108. data/skeletons/CPU-C/common/timer_2_start.c +11 -13
  109. data/skeletons/CPU-C/common/timer_2_stop.c +1 -1
  110. data/skeletons/CPU-C/common/timer_globals.c +29 -0
  111. data/skeletons/CPU-OPENCL-INTEL/common/globals.c +1 -1
  112. data/skeletons/CPU-OPENCL-INTEL/common/header.c +3 -0
  113. data/skeletons/CPU-OPENCL-INTEL/common/mem_copy_D2H.c +7 -2
  114. data/skeletons/CPU-OPENCL-INTEL/common/mem_copy_H2D.c +4 -2
  115. data/skeletons/CPU-OPENCL-INTEL/common/mem_global.c +0 -0
  116. data/skeletons/CPU-OPENCL-INTEL/common/mem_prologue.c +6 -3
  117. data/skeletons/CPU-OPENCL-INTEL/common/timer_2_stop.c +1 -1
  118. data/skeletons/CPU-OPENCL-INTEL/common/timer_globals.c +24 -0
  119. data/skeletons/CPU-OPENMP/common/globals.c +1 -0
  120. data/skeletons/CPU-OPENMP/common/header.c +3 -0
  121. data/skeletons/CPU-OPENMP/common/mem_global.c +0 -0
  122. data/skeletons/CPU-OPENMP/common/timer_1_start.c +0 -12
  123. data/skeletons/CPU-OPENMP/common/timer_2_stop.c +1 -1
  124. data/skeletons/CPU-OPENMP/common/timer_globals.c +33 -0
  125. data/skeletons/GPU-CUDA/common/globals.c +27 -3
  126. data/skeletons/GPU-CUDA/common/header.c +2 -0
  127. data/skeletons/GPU-CUDA/common/mem_async_alloc.c +6 -0
  128. data/skeletons/GPU-CUDA/common/mem_async_copyin.c +6 -0
  129. data/skeletons/GPU-CUDA/common/mem_async_copyout.c +6 -0
  130. data/skeletons/GPU-CUDA/common/mem_async_free.c +6 -0
  131. data/skeletons/GPU-CUDA/common/mem_copy_D2H.c +2 -1
  132. data/skeletons/GPU-CUDA/common/mem_copy_H2D.c +2 -1
  133. data/skeletons/GPU-CUDA/common/mem_global.c +1 -0
  134. data/skeletons/GPU-CUDA/common/mem_prologue.c +1 -2
  135. data/skeletons/GPU-CUDA/common/scheduler.c +86 -0
  136. data/skeletons/GPU-CUDA/common/timer_2_start.c +2 -4
  137. data/skeletons/GPU-CUDA/common/timer_2_stop.c +3 -5
  138. data/skeletons/GPU-CUDA/common/timer_globals.c +26 -0
  139. data/skeletons/GPU-CUDA/kernel/2xN-N-chunk-1-N-to-D-element.kernel.cu +5 -7
  140. data/skeletons/GPU-CUDA/kernel/N-N-chunk-1-N-to-D-element.kernel.cu +4 -6
  141. data/skeletons/GPU-CUDA/kernel/default.host.c +1 -1
  142. data/skeletons/GPU-CUDA/kernel/default.kernel.cu +6 -8
  143. data/skeletons/GPU-CUDA/skeletons.txt +6 -5
  144. data/{examples/benchmarks/2mm.c → test/examples/benchmarks/PolyBench/2mm_species.c} +19 -15
  145. data/test/examples/benchmarks/PolyBench/3mm_species.c +82 -0
  146. data/test/examples/benchmarks/PolyBench/adi_species.c +89 -0
  147. data/test/examples/benchmarks/PolyBench/atax_species.c +69 -0
  148. data/test/examples/benchmarks/PolyBench/bicg_species.c +71 -0
  149. data/test/examples/benchmarks/PolyBench/cholesky_species.c +68 -0
  150. data/test/examples/benchmarks/PolyBench/correlation_species.c +97 -0
  151. data/test/examples/benchmarks/PolyBench/covariance_species.c +78 -0
  152. data/test/examples/benchmarks/PolyBench/doitgen_species.c +67 -0
  153. data/test/examples/benchmarks/PolyBench/durbin_species.c +80 -0
  154. data/test/examples/benchmarks/PolyBench/dynprog_species.c +71 -0
  155. data/test/examples/benchmarks/PolyBench/fdtd-2d-apml_species.c +112 -0
  156. data/test/examples/benchmarks/PolyBench/fdtd-2d_species.c +78 -0
  157. data/test/examples/benchmarks/PolyBench/floyd-warshall_species.c +54 -0
  158. data/test/examples/benchmarks/PolyBench/gemm_species.c +73 -0
  159. data/test/examples/benchmarks/PolyBench/gemver_species.c +93 -0
  160. data/test/examples/benchmarks/PolyBench/gesummv_species.c +68 -0
  161. data/test/examples/benchmarks/PolyBench/gramschmidt_species.c +78 -0
  162. data/test/examples/benchmarks/PolyBench/jacobi-1d-imper_species.c +59 -0
  163. data/test/examples/benchmarks/PolyBench/jacobi-2d-imper_species.c +65 -0
  164. data/test/examples/benchmarks/PolyBench/lu_species.c +57 -0
  165. data/test/examples/benchmarks/PolyBench/ludcmp_species.c +89 -0
  166. data/test/examples/benchmarks/PolyBench/mvt_species.c +69 -0
  167. data/test/examples/benchmarks/PolyBench/reg_detect_species.c +86 -0
  168. data/test/examples/benchmarks/PolyBench/seidel-2d_species.c +53 -0
  169. data/test/examples/benchmarks/PolyBench/symm_species.c +74 -0
  170. data/test/examples/benchmarks/PolyBench/syr2k_species.c +69 -0
  171. data/test/examples/benchmarks/PolyBench/syrk_species.c +66 -0
  172. data/test/examples/benchmarks/PolyBench/trisolv_species.c +61 -0
  173. data/test/examples/benchmarks/PolyBench/trmm_species.c +61 -0
  174. data/test/examples/chunk/example01_species.c +58 -0
  175. data/test/examples/chunk/example02_species.c +48 -0
  176. data/test/examples/chunk/example03_species.c +63 -0
  177. data/test/examples/chunk/example04_species.c +58 -0
  178. data/test/examples/chunk/example05_species.c +56 -0
  179. data/test/examples/chunk/example06_species.c +49 -0
  180. data/test/examples/chunk/example07_species.c +53 -0
  181. data/test/examples/dependences/example01_species.c +46 -0
  182. data/test/examples/dependences/example02_species.c +44 -0
  183. data/test/examples/dependences/example03_species.c +47 -0
  184. data/test/examples/dependences/example04_species.c +48 -0
  185. data/test/examples/dependences/example05_species.c +46 -0
  186. data/test/examples/element/example01_species.c +50 -0
  187. data/test/examples/element/example02_species.c +50 -0
  188. data/test/examples/element/example03_species.c +62 -0
  189. data/test/examples/element/example04_species.c +53 -0
  190. data/test/examples/element/example05_species.c +59 -0
  191. data/test/examples/element/example06_species.c +50 -0
  192. data/test/examples/element/example07_species.c +58 -0
  193. data/test/examples/element/example08_species.c +49 -0
  194. data/test/examples/element/example09_species.c +52 -0
  195. data/test/examples/element/example10_species.c +54 -0
  196. data/test/examples/element/example11_species.c +51 -0
  197. data/test/examples/element/example12_species.c +60 -0
  198. data/test/examples/element/example13_species.c +77 -0
  199. data/test/examples/neighbourhood/example01_species.c +57 -0
  200. data/test/examples/neighbourhood/example02_species.c +56 -0
  201. data/test/examples/neighbourhood/example03_species.c +83 -0
  202. data/test/examples/neighbourhood/example04_species.c +55 -0
  203. data/test/examples/neighbourhood/example05_species.c +48 -0
  204. data/test/examples/shared/example01_species.c +49 -0
  205. data/test/examples/shared/example02_species.c +55 -0
  206. data/test/examples/shared/example03_species.c +59 -0
  207. data/test/examples/shared/example04_species.c +56 -0
  208. data/test/examples/shared/example05_species.c +52 -0
  209. metadata +193 -73
  210. data/examples/benchmarks/overview.txt +0 -38
  211. data/lib/castaddon/node.rb +0 -753
@@ -0,0 +1,66 @@
1
+ //
2
+ // This file is part of the Bones source-to-source compiler examples. The C-code
3
+ // is largely identical in terms of functionality and variable naming to the code
4
+ // found in PolyBench/C version 3.2. For more information on PolyBench/C or Bones
5
+ // please use the contact information below.
6
+ //
7
+ // == More information on PolyBench/C
8
+ // Contact............Louis-Noel Pouchet <pouchet@cse.ohio-state.edu>
9
+ // Web address........http://polybench.sourceforge.net/
10
+ //
11
+ // == More information on Bones
12
+ // Contact............Cedric Nugteren <c.nugteren@tue.nl>
13
+ // Web address........http://parse.ele.tue.nl/bones/
14
+ //
15
+ // == File information
16
+ // Filename...........benchmark/syrk.c
17
+ // Author.............Cedric Nugteren
18
+ // Last modified on...07-May-2013
19
+ //
20
+
21
+ #include "common.h"
22
+
23
+ // This is 'syrk', an algorithm for symmetric rank-k operations
24
+ int main(void) {
25
+ int i,j,k;
26
+
27
+ // Declare arrays on the stack
28
+ float A[NI][NJ];
29
+ float C[NI][NI];
30
+
31
+ // Set the constants
32
+ float alpha = 32412;
33
+ float beta = 2123;
34
+
35
+ // Set the input data
36
+ for (i=0; i<NI; i++) {
37
+ for (j=0; j<NJ; j++) {
38
+ A[i][j] = ((float) i*j) / NI;
39
+ }
40
+ }
41
+ for (i=0; i<NI; i++) {
42
+ for (j=0; j<NI; j++) {
43
+ C[i][j] = ((float) i*j) / NI;
44
+ }
45
+ }
46
+
47
+ // Perform the computation (C := alpha*A*A' + beta*C)
48
+ #pragma scop
49
+ {
50
+ #pragma species kernel C[0:NI-1,0:NI-1]|element ^ A[0:NI-1,0:NJ-1]|chunk(0:0,0:NJ-1) -> C[0:NI-1,0:NI-1]|element
51
+ for (i = 0; i < NI; i++) {
52
+ for (j = 0; j < NI; j++) {
53
+ C[i][j] *= beta;
54
+ for (k = 0; k < NJ; k++) {
55
+ C[i][j] += alpha * A[i][k] * A[j][k];
56
+ }
57
+ }
58
+ }
59
+ #pragma species endkernel syrk_k1
60
+ }
61
+ #pragma endscop
62
+
63
+ // Clean-up and exit the function
64
+ fflush(stdout);
65
+ return 0;
66
+ }
@@ -0,0 +1,61 @@
1
+ //
2
+ // This file is part of the Bones source-to-source compiler examples. The C-code
3
+ // is largely identical in terms of functionality and variable naming to the code
4
+ // found in PolyBench/C version 3.2. For more information on PolyBench/C or Bones
5
+ // please use the contact information below.
6
+ //
7
+ // == More information on PolyBench/C
8
+ // Contact............Louis-Noel Pouchet <pouchet@cse.ohio-state.edu>
9
+ // Web address........http://polybench.sourceforge.net/
10
+ //
11
+ // == More information on Bones
12
+ // Contact............Cedric Nugteren <c.nugteren@tue.nl>
13
+ // Web address........http://parse.ele.tue.nl/bones/
14
+ //
15
+ // == File information
16
+ // Filename...........benchmark/trisolv.c
17
+ // Author.............Cedric Nugteren
18
+ // Last modified on...03-Jul-2012
19
+ //
20
+
21
+ #include "common.h"
22
+
23
+ // This is 'trisolv', a solver for linear systems with triangular matrices
24
+ int main(void) {
25
+ int i,j;
26
+ float A_i_i;
27
+
28
+ // Declare arrays on the stack
29
+ float A[NX][NX];
30
+ float c[NX];
31
+ float x[NX];
32
+
33
+ // Set the input data
34
+ for (i=0; i<NX; i++) {
35
+ c[i] = ((float) i) / NX;
36
+ x[i] = ((float) i) / NX;
37
+ for (j=0; j<NX; j++) {
38
+ A[i][j] = ((float) i*j) / NX;
39
+ }
40
+ }
41
+
42
+ // Perform the computation
43
+ #pragma scop
44
+ {
45
+ for (i = 0; i < NX; i++) {
46
+ x[i] = c[i];
47
+ A_i_i = A[i][i];
48
+ #pragma species kernel x[i:i]|full ^ A[i:i,0:i-1]|element ^ x[0:i-1]|element -> x[i:i]|shared
49
+ for (j = 0; j <= i - 1; j++) {
50
+ x[i] = x[i] - A[i][j] * x[j];
51
+ x[i] = x[i] / A_i_i;
52
+ }
53
+ #pragma species endkernel trisolv_k4
54
+ }
55
+ }
56
+ #pragma endscop
57
+
58
+ // Clean-up and exit the function
59
+ fflush(stdout);
60
+ return 0;
61
+ }
@@ -0,0 +1,61 @@
1
+ //
2
+ // This file is part of the Bones source-to-source compiler examples. The C-code
3
+ // is largely identical in terms of functionality and variable naming to the code
4
+ // found in PolyBench/C version 3.2. For more information on PolyBench/C or Bones
5
+ // please use the contact information below.
6
+ //
7
+ // == More information on PolyBench/C
8
+ // Contact............Louis-Noel Pouchet <pouchet@cse.ohio-state.edu>
9
+ // Web address........http://polybench.sourceforge.net/
10
+ //
11
+ // == More information on Bones
12
+ // Contact............Cedric Nugteren <c.nugteren@tue.nl>
13
+ // Web address........http://parse.ele.tue.nl/bones/
14
+ //
15
+ // == File information
16
+ // Filename...........benchmark/trmm.c
17
+ // Author.............Cedric Nugteren
18
+ // Last modified on...26-Jun-2012
19
+ //
20
+
21
+ #include "common.h"
22
+
23
+ // This is 'trmm', a triangular matrix multiplication kernel
24
+ int main(void) {
25
+ int i,j,k;
26
+
27
+ // Declare arrays on the stack
28
+ float A[NI][NI];
29
+ float B[NI][NI];
30
+
31
+ // Set the constants
32
+ int alpha = 32412;
33
+
34
+ // Set the input data
35
+ for (i=0; i<NI; i++) {
36
+ for (j=0; j<NI; j++) {
37
+ A[i][j] = ((float) i*j) / NI;
38
+ B[i][j] = ((float) i*j) / NI;
39
+ }
40
+ }
41
+
42
+ // Perform the computation (B := alpha*A'*B, with A triangular)
43
+ #pragma scop
44
+ {
45
+ for (i = 1; i < NI; i++) {
46
+ for (j = 0; j < NI; j++) {
47
+ #pragma species kernel B[i:i,j:j]|full ^ A[i:i,0:i-1]|element ^ B[j:j,0:i-1]|element -> B[i:i,j:j]|shared
48
+ for (k = 0; k < i; k++) {
49
+ B[i][j] += alpha * A[i][k] * B[j][k];
50
+ }
51
+ #pragma species endkernel trmm_k3
52
+ }
53
+ }
54
+ }
55
+ #pragma endscop
56
+
57
+ // Clean-up and exit the function
58
+ fflush(stdout);
59
+ return 0;
60
+ }
61
+
@@ -0,0 +1,58 @@
1
+ //
2
+ // This file is part of the Bones source-to-source compiler examples. This C-code
3
+ // example is meant to illustrate the use of Bones. For more information on Bones
4
+ // use the contact information below.
5
+ //
6
+ // == More information on Bones
7
+ // Contact............Cedric Nugteren <c.nugteren@tue.nl>
8
+ // Web address........http://parse.ele.tue.nl/bones/
9
+ //
10
+ // == File information
11
+ // Filename...........chunk/example1.c
12
+ // Author.............Cedric Nugteren
13
+ // Last modified on...16-April-2012
14
+ //
15
+
16
+ #include <stdio.h>
17
+
18
+ // This is 'example1', a basic chunk to element example using a 2D tile
19
+ int main(void) {
20
+ int i,j;
21
+ int i2,j2;
22
+ int result = 0;
23
+
24
+ // Declare input/output arrays
25
+ int A[560][32];
26
+ int B[56][16];
27
+
28
+ // Set the input data
29
+ for(i=0;i<560;i++) {
30
+ for(j=0;j<32;j++) {
31
+ A[i][j] = i+j;
32
+ }
33
+ }
34
+
35
+ // Perform the computation
36
+ #pragma scop
37
+ {
38
+ #pragma species kernel A[0:559,0:31]|chunk(0:9,0:1) -> B[0:55,0:15]|element
39
+ for (i = 0; i < 56; i++) {
40
+ for (j = 0; j < 16; j++) {
41
+ result = 0;
42
+ for (i2 = 0; i2 < 10; i2++) {
43
+ for (j2 = 0; j2 < 2; j2++) {
44
+ result = result + A[i * 10 + i2][j * 2 + j2];
45
+ }
46
+ }
47
+ B[i][j] = result;
48
+ }
49
+ }
50
+ #pragma species endkernel example01_k1
51
+ }
52
+ #pragma endscop
53
+
54
+ // Clean-up and exit the function
55
+ fflush(stdout);
56
+ return 0;
57
+ }
58
+
@@ -0,0 +1,48 @@
1
+ //
2
+ // This file is part of the Bones source-to-source compiler examples. This C-code
3
+ // example is meant to illustrate the use of Bones. For more information on Bones
4
+ // use the contact information below.
5
+ //
6
+ // == More information on Bones
7
+ // Contact............Cedric Nugteren <c.nugteren@tue.nl>
8
+ // Web address........http://parse.ele.tue.nl/bones/
9
+ //
10
+ // == File information
11
+ // Filename...........chunk/example2.c
12
+ // Author.............Cedric Nugteren
13
+ // Last modified on...16-April-2012
14
+ //
15
+
16
+ #include <stdio.h>
17
+ #define SIZE 2048
18
+ #define HALFSIZE (SIZE/2)
19
+
20
+ // This is 'example2', demonstrating a chunk-example without an inner-loop, everything is unrolled manually
21
+ int main(void) {
22
+ int i;
23
+
24
+ // Declare input/output arrays
25
+ float A[SIZE];
26
+ float B[HALFSIZE];
27
+
28
+ // Set the input data
29
+ for(i=0;i<SIZE;i++) {
30
+ A[i] = i%6+i;
31
+ }
32
+
33
+ // Perform the computation
34
+ #pragma scop
35
+ {
36
+ #pragma species kernel A[0:2*HALFSIZE-1]|chunk(0:1) -> B[0:HALFSIZE-1]|element
37
+ for (i = 0; i < HALFSIZE; i++) {
38
+ B[i] = A[i * 2] + A[i * 2 + 1];
39
+ }
40
+ #pragma species endkernel example02_k1
41
+ }
42
+ #pragma endscop
43
+
44
+ // Clean-up and exit the function
45
+ fflush(stdout);
46
+ return 0;
47
+ }
48
+
@@ -0,0 +1,63 @@
1
+ //
2
+ // This file is part of the Bones source-to-source compiler examples. This C-code
3
+ // example is meant to illustrate the use of Bones. For more information on Bones
4
+ // use the contact information below.
5
+ //
6
+ // == More information on Bones
7
+ // Contact............Cedric Nugteren <c.nugteren@tue.nl>
8
+ // Web address........http://parse.ele.tue.nl/bones/
9
+ //
10
+ // == File information
11
+ // Filename...........chunk/example3.c
12
+ // Author.............Cedric Nugteren
13
+ // Last modified on...16-April-2012
14
+ //
15
+
16
+ #include <stdio.h>
17
+ #define BASE 2048
18
+ #define TILE 64
19
+ #define SIZE (BASE*TILE)
20
+
21
+ // This is 'example3', demonstrating a chunked input and a chunked output, and showing the importance of ordering (array referenced first should be placed first)
22
+ int main(void) {
23
+ int i;
24
+ int t = 0;
25
+ float result = 0;
26
+
27
+ // Declare input/output arrays
28
+ float A[BASE];
29
+ float B[SIZE];
30
+ float out1[SIZE];
31
+ float out2[SIZE];
32
+
33
+ // Set the input data
34
+ for(i=0;i<BASE;i++) {
35
+ A[i] = 0.6;
36
+ }
37
+ for(i=0;i<SIZE;i++) {
38
+ B[i] = i%6+i;
39
+ }
40
+
41
+ // Perform the computation
42
+ #pragma scop
43
+ {
44
+ #pragma species kernel A[0:BASE-1]|element ^ B[0:(BASE*TILE)-1]|chunk(0:TILE-1) -> out1[0:(BASE*TILE)-1]|chunk(0:TILE-1) ^ out2[0:(BASE*TILE)-1]|chunk(0:TILE-1)
45
+ for (i = 0; i < BASE; i++) {
46
+ result = A[i];
47
+ for (t = 0; t < TILE; t++) {
48
+ result = result + t * B[i * TILE + t];
49
+ }
50
+ for (t = 0; t < TILE; t++) {
51
+ out1[i * TILE + t] = result;
52
+ out2[i * TILE + t] = -result;
53
+ }
54
+ }
55
+ #pragma species endkernel example03_k1
56
+ }
57
+ #pragma endscop
58
+
59
+ // Clean-up and exit the function
60
+ fflush(stdout);
61
+ return 0;
62
+ }
63
+
@@ -0,0 +1,58 @@
1
+ //
2
+ // This file is part of the Bones source-to-source compiler examples. This C-code
3
+ // example is meant to illustrate the use of Bones. For more information on Bones
4
+ // use the contact information below.
5
+ //
6
+ // == More information on Bones
7
+ // Contact............Cedric Nugteren <c.nugteren@tue.nl>
8
+ // Web address........http://parse.ele.tue.nl/bones/
9
+ //
10
+ // == File information
11
+ // Filename...........chunk/example4.c
12
+ // Author.............Cedric Nugteren
13
+ // Last modified on...16-April-2012
14
+ //
15
+
16
+ #include <stdio.h>
17
+ #define SIZE 1024
18
+
19
+ // This is 'example4', a basic element to chunk example with an if-statement in the body
20
+ int main(void) {
21
+ int i;
22
+ int threshold = 19;
23
+
24
+ // Declare input/output arrays
25
+ char A[SIZE];
26
+ char B[SIZE*2];
27
+
28
+ // Set the input data
29
+ for(i=0;i<SIZE;i++) {
30
+ A[i] = i%100;
31
+ }
32
+
33
+ // Set the output data to zero
34
+ for(i=0;i<SIZE*2;i++) {
35
+ B[i] = 0;
36
+ }
37
+
38
+ // Perform the computation
39
+ #pragma scop
40
+ {
41
+ #pragma species kernel A[0:SIZE-1]|element -> B[0:2*SIZE-1]|chunk(0:1)
42
+ for (i = 0; i < SIZE; i++) {
43
+ B[i * 2] = A[i];
44
+ if (A[i] > threshold) {
45
+ B[i * 2 + 1] = A[i];
46
+ } else {
47
+ B[i * 2 + 1] = 0;
48
+ }
49
+ }
50
+ #pragma species endkernel example04_k1
51
+ }
52
+ #pragma endscop
53
+
54
+ // Clean-up and exit the function
55
+ fflush(stdout);
56
+ return 0;
57
+ }
58
+
@@ -0,0 +1,56 @@
1
+ //
2
+ // This file is part of the Bones source-to-source compiler examples. This C-code
3
+ // example is meant to illustrate the use of Bones. For more information on Bones
4
+ // use the contact information below.
5
+ //
6
+ // == More information on Bones
7
+ // Contact............Cedric Nugteren <c.nugteren@tue.nl>
8
+ // Web address........http://parse.ele.tue.nl/bones/
9
+ //
10
+ // == File information
11
+ // Filename...........chunk/example5.c
12
+ // Author.............Cedric Nugteren
13
+ // Last modified on...16-April-2012
14
+ //
15
+
16
+ #include <stdio.h>
17
+
18
+ // This is 'example5', demonstrating a chunk to element example without a temporary variable but directly writing to the output
19
+ int main(void) {
20
+ int i,j;
21
+ int i2,j2;
22
+
23
+ // Declare input/output arrays
24
+ int A[560][32];
25
+ int B[56][16];
26
+
27
+ // Set the input data
28
+ for(i=0;i<560;i++) {
29
+ for(j=0;j<32;j++) {
30
+ A[i][j] = i+j;
31
+ }
32
+ }
33
+
34
+ // Perform the computation
35
+ #pragma scop
36
+ {
37
+ #pragma species kernel A[0:559,0:31]|chunk(0:9,0:1) -> B[0:55,0:15]|element
38
+ for (i = 0; i < 56; i++) {
39
+ for (j = 0; j < 16; j++) {
40
+ B[i][j] = 0;
41
+ for (i2 = 0; i2 < 10; i2++) {
42
+ for (j2 = 0; j2 < 2; j2++) {
43
+ B[i][j] = B[i][j] + A[i * 10 + i2][j * 2 + j2];
44
+ }
45
+ }
46
+ }
47
+ }
48
+ #pragma species endkernel example05_k1
49
+ }
50
+ #pragma endscop
51
+
52
+ // Clean-up and exit the function
53
+ fflush(stdout);
54
+ return 0;
55
+ }
56
+