Mercurial > hg > truffle
annotate src/gpu/ptx/vm/gpu_ptx.cpp @ 11954:905c26108117
removed warning
author | Doug Simon <doug.simon@oracle.com> |
---|---|
date | Thu, 10 Oct 2013 12:05:39 +0200 |
parents | 67a1e27a8dbb |
children | f020e149c1b6 |
rev | line source |
---|---|
10879 | 1 /* |
2 * Copyright (c) 2013, Oracle and/or its affiliates. All rights reserved. | |
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA | |
20 * or visit www.oracle.com if you need additional information or have any | |
21 * questions. | |
22 * | |
23 */ | |
24 | |
25 #include "precompiled.hpp" | |
11284
d876002b98e6
ptx: fix include for non-precompiledheader build in gpu_ptx.cpp
Bernhard Urban <bernhard.urban@jku.at>
parents:
11283
diff
changeset
|
26 #include "runtime/javaCalls.hpp" |
10879 | 27 #include "runtime/gpu.hpp" |
28 #include "utilities/globalDefinitions.hpp" | |
29 #include "utilities/ostream.hpp" | |
10883
5fcb30bcb90a
Fix compilation without precompiled headers.
Roland Schatz <roland.schatz@oracle.com>
parents:
10882
diff
changeset
|
30 #include "memory/allocation.hpp" |
5fcb30bcb90a
Fix compilation without precompiled headers.
Roland Schatz <roland.schatz@oracle.com>
parents:
10882
diff
changeset
|
31 #include "memory/allocation.inline.hpp" |
11596
91e5f927af63
Initial implementation of PTXRuntime (RegisterConfig, PTX description etc); guarded with new flag UseGPU. Specify -XX:+UseGPU to exercise this new implementation.
bharadwaj
parents:
11527
diff
changeset
|
32 #include "ptxKernelArguments.hpp" |
10879 | 33 |
34 void * gpu::Ptx::_device_context; | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
35 int gpu::Ptx::_cu_device = 0; |
10879 | 36 |
37 gpu::Ptx::cuda_cu_init_func_t gpu::Ptx::_cuda_cu_init; | |
38 gpu::Ptx::cuda_cu_ctx_create_func_t gpu::Ptx::_cuda_cu_ctx_create; | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
39 gpu::Ptx::cuda_cu_ctx_destroy_func_t gpu::Ptx::_cuda_cu_ctx_destroy; |
10879 | 40 gpu::Ptx::cuda_cu_ctx_synchronize_func_t gpu::Ptx::_cuda_cu_ctx_synchronize; |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
41 gpu::Ptx::cuda_cu_ctx_set_current_func_t gpu::Ptx::_cuda_cu_ctx_set_current; |
10879 | 42 gpu::Ptx::cuda_cu_device_get_count_func_t gpu::Ptx::_cuda_cu_device_get_count; |
43 gpu::Ptx::cuda_cu_device_get_name_func_t gpu::Ptx::_cuda_cu_device_get_name; | |
44 gpu::Ptx::cuda_cu_device_get_func_t gpu::Ptx::_cuda_cu_device_get; | |
45 gpu::Ptx::cuda_cu_device_compute_capability_func_t gpu::Ptx::_cuda_cu_device_compute_capability; | |
46 gpu::Ptx::cuda_cu_device_get_attribute_func_t gpu::Ptx::_cuda_cu_device_get_attribute; | |
47 gpu::Ptx::cuda_cu_launch_kernel_func_t gpu::Ptx::_cuda_cu_launch_kernel; | |
48 gpu::Ptx::cuda_cu_module_get_function_func_t gpu::Ptx::_cuda_cu_module_get_function; | |
49 gpu::Ptx::cuda_cu_module_load_data_ex_func_t gpu::Ptx::_cuda_cu_module_load_data_ex; | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
50 gpu::Ptx::cuda_cu_memcpy_dtoh_func_t gpu::Ptx::_cuda_cu_memcpy_dtoh; |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
51 gpu::Ptx::cuda_cu_memfree_func_t gpu::Ptx::_cuda_cu_memfree; |
10879 | 52 |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
53 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
54 /* |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
55 * see http://en.wikipedia.org/wiki/CUDA#Supported_GPUs |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
56 */ |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
57 int ncores(int major, int minor) { |
11843
372bacc13022
gtx_ptx.cpp: fix switch default label and operator precedence in ncores().
Andreas Woess <andreas.woess@jku.at>
parents:
11842
diff
changeset
|
58 int device_type = (major << 4) + minor; |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
59 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
60 switch (device_type) { |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
61 case 0x10: return 8; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
62 case 0x11: return 8; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
63 case 0x12: return 8; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
64 case 0x13: return 8; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
65 case 0x20: return 32; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
66 case 0x21: return 48; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
67 case 0x30: return 192; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
68 case 0x35: return 192; |
11843
372bacc13022
gtx_ptx.cpp: fix switch default label and operator precedence in ncores().
Andreas Woess <andreas.woess@jku.at>
parents:
11842
diff
changeset
|
69 default: |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
70 tty->print_cr("[CUDA] Warning: Unhandled device %x", device_type); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
71 return 0; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
72 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
73 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
74 |
10879 | 75 bool gpu::Ptx::initialize_gpu() { |
76 | |
77 /* Initialize CUDA driver API */ | |
78 int status = _cuda_cu_init(0); | |
79 if (status != GRAAL_CUDA_SUCCESS) { | |
80 tty->print_cr("Failed to initialize CUDA device"); | |
81 return false; | |
82 } | |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
83 |
10879 | 84 if (TraceGPUInteraction) { |
85 tty->print_cr("CUDA driver initialization: Success"); | |
86 } | |
87 | |
88 /* Get the number of compute-capable device count */ | |
89 int device_count = 0; | |
90 status = _cuda_cu_device_get_count(&device_count); | |
91 if (status != GRAAL_CUDA_SUCCESS) { | |
92 tty->print_cr("[CUDA] Failed to get compute-capable device count"); | |
93 return false; | |
94 } | |
95 | |
96 if (device_count == 0) { | |
97 tty->print_cr("[CUDA] Found no device supporting CUDA"); | |
98 return false; | |
99 } | |
100 | |
101 if (TraceGPUInteraction) { | |
102 tty->print_cr("[CUDA] Number of compute-capable devices found: %d", device_count); | |
103 } | |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
104 |
10879 | 105 /* Get the handle to the first compute device */ |
106 int device_id = 0; | |
107 /* Compute-capable device handle */ | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
108 status = _cuda_cu_device_get(&_cu_device, device_id); |
10879 | 109 |
110 if (status != GRAAL_CUDA_SUCCESS) { | |
111 tty->print_cr("[CUDA] Failed to get handle of first compute-capable device i.e., the one at ordinal: %d", device_id); | |
112 return false; | |
113 } | |
114 | |
115 if (TraceGPUInteraction) { | |
116 tty->print_cr("[CUDA] Got the handle of first compute-device"); | |
117 } | |
118 | |
119 /* Get device attributes */ | |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
120 int unified_addressing; |
10879 | 121 |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
122 status = _cuda_cu_device_get_attribute(&unified_addressing, GRAAL_CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING, _cu_device); |
11283
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
123 |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
124 if (status != GRAAL_CUDA_SUCCESS) { |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
125 tty->print_cr("[CUDA] Failed to query unified addressing mode of device: %d", _cu_device); |
11283
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
126 return false; |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
127 } |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
128 |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
129 if (TraceGPUInteraction) { |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
130 tty->print_cr("[CUDA] Unified addressing support on device %d: %d", _cu_device, unified_addressing); |
11283
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
131 } |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
132 |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
133 |
10879 | 134 /* Get device name */ |
135 char device_name[256]; | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
136 status = _cuda_cu_device_get_name(device_name, 256, _cu_device); |
10879 | 137 |
138 if (status != GRAAL_CUDA_SUCCESS) { | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
139 tty->print_cr("[CUDA] Failed to get name of device: %d", _cu_device); |
10879 | 140 return false; |
141 } | |
142 | |
143 if (TraceGPUInteraction) { | |
144 tty->print_cr("[CUDA] Using %s", device_name); | |
145 } | |
146 | |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
147 |
10879 | 148 return true; |
149 } | |
150 | |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
151 unsigned int gpu::Ptx::total_cores() { |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
152 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
153 int minor, major, nmp; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
154 int status = _cuda_cu_device_get_attribute(&minor, |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
155 GRAAL_CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
156 _cu_device); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
157 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
158 if (status != GRAAL_CUDA_SUCCESS) { |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
159 tty->print_cr("[CUDA] Failed to get minor attribute of device: %d", _cu_device); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
160 return 0; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
161 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
162 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
163 status = _cuda_cu_device_get_attribute(&major, |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
164 GRAAL_CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
165 _cu_device); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
166 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
167 if (status != GRAAL_CUDA_SUCCESS) { |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
168 tty->print_cr("[CUDA] Failed to get major attribute of device: %d", _cu_device); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
169 return 0; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
170 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
171 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
172 status = _cuda_cu_device_get_attribute(&nmp, |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
173 GRAAL_CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
174 _cu_device); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
175 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
176 if (status != GRAAL_CUDA_SUCCESS) { |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
177 tty->print_cr("[CUDA] Failed to get numberof MPs on device: %d", _cu_device); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
178 return 0; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
179 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
180 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
181 int total = nmp * ncores(major, minor); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
182 |
11894
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
183 int max_threads_per_block, warp_size, async_engines, can_map_host_memory, concurrent_kernels; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
184 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
185 status = _cuda_cu_device_get_attribute(&max_threads_per_block, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
186 GRAAL_CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
187 _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
188 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
189 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
190 tty->print_cr("[CUDA] Failed to get GRAAL_CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK: %d", _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
191 return 0; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
192 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
193 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
194 status = _cuda_cu_device_get_attribute(&warp_size, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
195 GRAAL_CU_DEVICE_ATTRIBUTE_WARP_SIZE, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
196 _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
197 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
198 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
199 tty->print_cr("[CUDA] Failed to get GRAAL_CU_DEVICE_ATTRIBUTE_WARP_SIZE: %d", _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
200 return 0; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
201 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
202 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
203 status = _cuda_cu_device_get_attribute(&async_engines, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
204 GRAAL_CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
205 _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
206 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
207 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
208 tty->print_cr("[CUDA] Failed to get GRAAL_CU_DEVICE_ATTRIBUTE_WARP_SIZE: %d", _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
209 return 0; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
210 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
211 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
212 status = _cuda_cu_device_get_attribute(&can_map_host_memory, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
213 GRAAL_CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
214 _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
215 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
216 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
217 tty->print_cr("[CUDA] Failed to get GRAAL_CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY: %d", _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
218 return 0; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
219 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
220 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
221 status = _cuda_cu_device_get_attribute(&concurrent_kernels, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
222 GRAAL_CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
223 _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
224 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
225 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
226 tty->print_cr("[CUDA] Failed to get GRAAL_CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS: %d", _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
227 return 0; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
228 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
229 |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
230 if (TraceGPUInteraction) { |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
231 tty->print_cr("[CUDA] Compatibility version of device %d: %d.%d", _cu_device, major, minor); |
11894
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
232 tty->print_cr("[CUDA] Number of cores: %d async engines: %d can map host mem: %d concurrent kernels: %d", |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
233 total, async_engines, can_map_host_memory, concurrent_kernels); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
234 tty->print_cr("[CUDA] Max threads per block: %d warp size: %d", max_threads_per_block, warp_size); |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
235 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
236 return (total); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
237 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
238 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
239 |
10879 | 240 void *gpu::Ptx::generate_kernel(unsigned char *code, int code_len, const char *name) { |
241 | |
242 struct CUmod_st * cu_module; | |
243 // Use three JIT compiler options | |
244 const unsigned int jit_num_options = 3; | |
245 int *jit_options = NEW_C_HEAP_ARRAY(int, jit_num_options, mtCompiler); | |
246 void **jit_option_values = NEW_C_HEAP_ARRAY(void *, jit_num_options, mtCompiler); | |
247 | |
248 // Set up PTX JIT compiler options | |
249 // 1. set size of compilation log buffer | |
250 int jit_log_buffer_size = 1024; | |
251 jit_options[0] = GRAAL_CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES; | |
252 jit_option_values[0] = (void *)(size_t)jit_log_buffer_size; | |
253 | |
254 // 2. set pointer to compilation log buffer | |
255 char *jit_log_buffer = NEW_C_HEAP_ARRAY(char, jit_log_buffer_size, mtCompiler); | |
256 jit_options[1] = GRAAL_CU_JIT_INFO_LOG_BUFFER; | |
257 jit_option_values[1] = jit_log_buffer; | |
258 | |
259 // 3. set pointer to set the Maximum # of registers (32) for the kernel | |
260 int jit_register_count = 32; | |
261 jit_options[2] = GRAAL_CU_JIT_MAX_REGISTERS; | |
262 jit_option_values[2] = (void *)(size_t)jit_register_count; | |
263 | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
264 /* Create CUDA context to compile and execute the kernel */ |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
265 int status = _cuda_cu_ctx_create(&_device_context, 0, _cu_device); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
266 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
267 if (status != GRAAL_CUDA_SUCCESS) { |
11634
bff2b88444f5
Start of PTX array passing
Morris Meyer <morris.meyer@oracle.com>
parents:
11606
diff
changeset
|
268 tty->print_cr("[CUDA] Failed to create CUDA context for device(%d): %d", _cu_device, status); |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
269 return NULL; |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
270 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
271 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
272 if (TraceGPUInteraction) { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
273 tty->print_cr("[CUDA] Success: Created context for device: %d", _cu_device); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
274 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
275 |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
276 status = _cuda_cu_ctx_set_current(_device_context); |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
277 |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
278 if (status != GRAAL_CUDA_SUCCESS) { |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
279 tty->print_cr("[CUDA] Failed to set current context for device: %d", _cu_device); |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
280 return NULL; |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
281 } |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
282 |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
283 if (TraceGPUInteraction) { |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
284 tty->print_cr("[CUDA] Success: Set current context for device: %d", _cu_device); |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
285 } |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
286 |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
287 if (TraceGPUInteraction) { |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
288 tty->print_cr("[CUDA] PTX Kernel\n%s", code); |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
289 tty->print_cr("[CUDA] Function name : %s", name); |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
290 |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
291 } |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
292 |
10879 | 293 /* Load module's data with compiler options */ |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
294 status = _cuda_cu_module_load_data_ex(&cu_module, (void*) code, jit_num_options, |
10879 | 295 jit_options, (void **)jit_option_values); |
296 if (status != GRAAL_CUDA_SUCCESS) { | |
297 if (status == GRAAL_CUDA_ERROR_NO_BINARY_FOR_GPU) { | |
298 tty->print_cr("[CUDA] Check for malformed PTX kernel or incorrect PTX compilation options"); | |
299 } | |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
300 tty->print_cr("[CUDA] *** Error (%d) Failed to load module data with online compiler options for method %s", |
10879 | 301 status, name); |
302 return NULL; | |
303 } | |
304 | |
305 if (TraceGPUInteraction) { | |
306 tty->print_cr("[CUDA] Loaded data for PTX Kernel"); | |
307 } | |
308 | |
309 struct CUfunc_st * cu_function; | |
310 | |
311 status = _cuda_cu_module_get_function(&cu_function, cu_module, name); | |
312 | |
313 if (status != GRAAL_CUDA_SUCCESS) { | |
11283
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
314 tty->print_cr("[CUDA] *** Error: Failed to get function %s", name); |
10879 | 315 return NULL; |
316 } | |
317 | |
318 if (TraceGPUInteraction) { | |
319 tty->print_cr("[CUDA] Got function handle for %s", name); | |
320 } | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
321 |
10879 | 322 return cu_function; |
323 } | |
324 | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
325 bool gpu::Ptx::execute_kernel(address kernel, PTXKernelArguments &ptxka, JavaValue &ret) { |
11822
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
326 return gpu::Ptx::execute_warp(1, 1, 1, kernel, ptxka, ret); |
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
327 } |
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
328 |
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
329 bool gpu::Ptx::execute_warp(int dimX, int dimY, int dimZ, |
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
330 address kernel, PTXKernelArguments &ptxka, JavaValue &ret) { |
10879 | 331 // grid dimensionality |
332 unsigned int gridX = 1; | |
333 unsigned int gridY = 1; | |
334 unsigned int gridZ = 1; | |
335 | |
336 // thread dimensionality | |
11822
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
337 unsigned int blockX = dimX; |
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
338 unsigned int blockY = dimY; |
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
339 unsigned int blockZ = dimZ; |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
340 |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
341 struct CUfunc_st* cu_function = (struct CUfunc_st*) kernel; |
10879 | 342 |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
343 void * config[5] = { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
344 GRAAL_CU_LAUNCH_PARAM_BUFFER_POINTER, ptxka._kernelArgBuffer, |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
345 GRAAL_CU_LAUNCH_PARAM_BUFFER_SIZE, &(ptxka._bufferOffset), |
11283
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
346 GRAAL_CU_LAUNCH_PARAM_END |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
347 }; |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
348 |
10879 | 349 if (kernel == NULL) { |
350 return false; | |
351 } | |
352 | |
353 if (TraceGPUInteraction) { | |
354 tty->print_cr("[CUDA] launching kernel"); | |
355 } | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
356 |
10879 | 357 int status = _cuda_cu_launch_kernel(cu_function, |
358 gridX, gridY, gridZ, | |
359 blockX, blockY, blockZ, | |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
360 0, NULL, NULL, (void **) &config); |
10879 | 361 if (status != GRAAL_CUDA_SUCCESS) { |
362 tty->print_cr("[CUDA] Failed to launch kernel"); | |
363 return false; | |
364 } | |
365 | |
366 if (TraceGPUInteraction) { | |
11822
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
367 tty->print_cr("[CUDA] Success: Kernel Launch: X: %d Y: %d Z: %d", blockX, blockY, blockZ); |
10879 | 368 } |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
369 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
370 status = _cuda_cu_ctx_synchronize(); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
371 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
372 if (status != GRAAL_CUDA_SUCCESS) { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
373 tty->print_cr("[CUDA] Failed to synchronize launched kernel (%d)", status); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
374 return false; |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
375 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
376 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
377 if (TraceGPUInteraction) { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
378 tty->print_cr("[CUDA] Success: Synchronized launch kernel"); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
379 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
380 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
381 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
382 // Get the result. TODO: Move this code to get_return_oop() |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
383 BasicType return_type = ptxka.get_ret_type(); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
384 switch (return_type) { |
11821
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
385 case T_INT: |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
386 { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
387 int return_val; |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
388 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._return_value_ptr, T_INT_BYTE_SIZE); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
389 if (status != GRAAL_CUDA_SUCCESS) { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
390 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
391 return false; |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
392 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
393 ret.set_jint(return_val); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
394 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
395 break; |
11901
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
396 case T_BOOLEAN: |
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
397 { |
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
398 int return_val; |
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
399 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._return_value_ptr, T_INT_BYTE_SIZE); |
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
400 if (status != GRAAL_CUDA_SUCCESS) { |
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
401 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status); |
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
402 return false; |
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
403 } |
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
404 ret.set_jint(return_val); |
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
405 } |
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
406 break; |
11894
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
407 case T_FLOAT: |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
408 { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
409 float return_val; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
410 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._return_value_ptr, T_FLOAT_BYTE_SIZE); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
411 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
412 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
413 return false; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
414 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
415 ret.set_jfloat(return_val); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
416 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
417 break; |
11902
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
418 case T_DOUBLE: |
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
419 { |
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
420 double return_val; |
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
421 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._return_value_ptr, T_DOUBLE_BYTE_SIZE); |
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
422 if (status != GRAAL_CUDA_SUCCESS) { |
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
423 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status); |
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
424 return false; |
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
425 } |
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
426 ret.set_jdouble(return_val); |
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
427 } |
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
428 break; |
11821
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
429 case T_LONG: |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
430 { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
431 long return_val; |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
432 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._return_value_ptr, T_LONG_BYTE_SIZE); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
433 if (status != GRAAL_CUDA_SUCCESS) { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
434 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
435 return false; |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
436 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
437 ret.set_jlong(return_val); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
438 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
439 break; |
11821
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
440 case T_VOID: |
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
441 break; |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
442 default: |
11821
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
443 tty->print_cr("[CUDA] TODO *** Unhandled return type: %d", return_type); |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
444 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
445 |
11821
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
446 // handle post-invocation object and array arguemtn |
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
447 ptxka.reiterate(); |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
448 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
449 // Free device memory allocated for result |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
450 status = gpu::Ptx::_cuda_cu_memfree(ptxka._return_value_ptr); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
451 if (status != GRAAL_CUDA_SUCCESS) { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
452 tty->print_cr("[CUDA] *** Error (%d) Failed to free device memory of return value", status); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
453 return false; |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
454 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
455 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
456 if (TraceGPUInteraction) { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
457 tty->print_cr("[CUDA] Success: Freed device memory of return value"); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
458 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
459 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
460 // Destroy context |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
461 status = gpu::Ptx::_cuda_cu_ctx_destroy(_device_context); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
462 if (status != GRAAL_CUDA_SUCCESS) { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
463 tty->print_cr("[CUDA] *** Error (%d) Failed to destroy context", status); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
464 return false; |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
465 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
466 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
467 if (TraceGPUInteraction) { |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
468 tty->print_cr("[CUDA] Success: Destroy context"); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
469 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
470 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
471 return (status == GRAAL_CUDA_SUCCESS); |
10879 | 472 } |
473 | |
474 #if defined(LINUX) | |
475 static const char cuda_library_name[] = "libcuda.so"; | |
476 #elif defined(__APPLE__) | |
477 static char const cuda_library_name[] = "/usr/local/cuda/lib/libcuda.dylib"; | |
478 #else | |
479 static char const cuda_library_name[] = ""; | |
480 #endif | |
481 | |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
482 #define STD_BUFFER_SIZE 1024 |
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
483 |
10879 | 484 bool gpu::Ptx::probe_linkage() { |
485 if (cuda_library_name != NULL) { | |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
486 char *buffer = (char*)malloc(STD_BUFFER_SIZE); |
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
487 void *handle = os::dll_load(cuda_library_name, buffer, STD_BUFFER_SIZE); |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
488 free(buffer); |
10879 | 489 if (handle != NULL) { |
490 _cuda_cu_init = | |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
491 CAST_TO_FN_PTR(cuda_cu_init_func_t, os::dll_lookup(handle, "cuInit")); |
10879 | 492 _cuda_cu_ctx_create = |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
493 CAST_TO_FN_PTR(cuda_cu_ctx_create_func_t, os::dll_lookup(handle, "cuCtxCreate")); |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
494 _cuda_cu_ctx_destroy = |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
495 CAST_TO_FN_PTR(cuda_cu_ctx_destroy_func_t, os::dll_lookup(handle, "cuCtxDestroy")); |
10879 | 496 _cuda_cu_ctx_synchronize = |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
497 CAST_TO_FN_PTR(cuda_cu_ctx_synchronize_func_t, os::dll_lookup(handle, "cuCtxSynchronize")); |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
498 _cuda_cu_ctx_set_current = |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
499 CAST_TO_FN_PTR(cuda_cu_ctx_set_current_func_t, os::dll_lookup(handle, "cuCtxSetCurrent")); |
10879 | 500 _cuda_cu_device_get_count = |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
501 CAST_TO_FN_PTR(cuda_cu_device_get_count_func_t, os::dll_lookup(handle, "cuDeviceGetCount")); |
10879 | 502 _cuda_cu_device_get_name = |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
503 CAST_TO_FN_PTR(cuda_cu_device_get_name_func_t, os::dll_lookup(handle, "cuDeviceGetName")); |
10879 | 504 _cuda_cu_device_get = |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
505 CAST_TO_FN_PTR(cuda_cu_device_get_func_t, os::dll_lookup(handle, "cuDeviceGet")); |
10879 | 506 _cuda_cu_device_compute_capability = |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
507 CAST_TO_FN_PTR(cuda_cu_device_compute_capability_func_t, os::dll_lookup(handle, "cuDeviceComputeCapability")); |
10879 | 508 _cuda_cu_device_get_attribute = |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
509 CAST_TO_FN_PTR(cuda_cu_device_get_attribute_func_t, os::dll_lookup(handle, "cuDeviceGetAttribute")); |
10879 | 510 _cuda_cu_module_get_function = |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
511 CAST_TO_FN_PTR(cuda_cu_module_get_function_func_t, os::dll_lookup(handle, "cuModuleGetFunction")); |
10879 | 512 _cuda_cu_module_load_data_ex = |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
513 CAST_TO_FN_PTR(cuda_cu_module_load_data_ex_func_t, os::dll_lookup(handle, "cuModuleLoadDataEx")); |
10879 | 514 _cuda_cu_launch_kernel = |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
515 CAST_TO_FN_PTR(cuda_cu_launch_kernel_func_t, os::dll_lookup(handle, "cuLaunchKernel")); |
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
516 _cuda_cu_memalloc = |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
517 CAST_TO_FN_PTR(cuda_cu_memalloc_func_t, os::dll_lookup(handle, "cuMemAlloc")); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
518 _cuda_cu_memfree = |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
519 CAST_TO_FN_PTR(cuda_cu_memfree_func_t, os::dll_lookup(handle, "cuMemFree")); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
520 _cuda_cu_memcpy_htod = |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
521 CAST_TO_FN_PTR(cuda_cu_memcpy_htod_func_t, os::dll_lookup(handle, "cuMemcpyHtoD")); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
522 _cuda_cu_memcpy_dtoh = |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
523 CAST_TO_FN_PTR(cuda_cu_memcpy_dtoh_func_t, os::dll_lookup(handle, "cuMemcpyDtoH")); |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
524 |
10879 | 525 if (TraceGPUInteraction) { |
526 tty->print_cr("[CUDA] Success: library linkage"); | |
527 } | |
528 return true; | |
529 } else { | |
530 // Unable to dlopen libcuda | |
531 return false; | |
532 } | |
533 } else { | |
534 tty->print_cr("Unsupported CUDA platform"); | |
535 return false; | |
536 } | |
537 tty->print_cr("Failed to find CUDA linkage"); | |
538 return false; | |
539 } |