Mercurial > hg > graal-compiler
annotate src/gpu/ptx/vm/gpu_ptx.cpp @ 12653:1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
* Change dynamic loading of CUDA driver API functions to load 32-bit or 64-bit versions of depending on the the host architecture.
* Add ability to generate PTX kernels to be launched both on 32-bit and 64-bit hosts.
* Use Unified Virtual Memory APIs to perform array argument marshalling.
* PTX array storage test runs on the device and returns correct results.
* More integer test failures on GPU fixed.
| author | S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> |
|---|---|
| date | Fri, 01 Nov 2013 18:34:03 -0400 |
| parents | f020e149c1b6 |
| children | 220ed109bf77 |
| 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; |
|
12653
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
52 gpu::Ptx::cuda_cu_mem_host_register_func_t gpu::Ptx::_cuda_cu_mem_host_register; |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
53 gpu::Ptx::cuda_cu_mem_host_get_device_pointer_func_t gpu::Ptx::_cuda_cu_mem_host_get_device_pointer; |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
54 gpu::Ptx::cuda_cu_mem_host_unregister_func_t gpu::Ptx::_cuda_cu_mem_host_unregister; |
| 10879 | 55 |
|
12653
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
56 #define STRINGIFY(x) #x |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
57 |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
58 #define LOOKUP_CUDA_FUNCTION(name, alias) \ |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
59 _##alias = \ |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
60 CAST_TO_FN_PTR(alias##_func_t, os::dll_lookup(handle, STRINGIFY(name))); \ |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
61 if (_##alias == NULL) { \ |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
62 tty->print_cr("[CUDA] ***** Error: Failed to lookup %s", STRINGIFY(name)); \ |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
63 return 0; \ |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
64 } \ |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
65 |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
66 #define LOOKUP_CUDA_V2_FUNCTION(name, alias) LOOKUP_CUDA_FUNCTION(name##_v2, alias) |
|
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
67 |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
68 /* |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
69 * 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
|
70 */ |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
71 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
|
72 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
|
73 |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
74 switch (device_type) { |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
75 case 0x10: return 8; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
76 case 0x11: return 8; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
77 case 0x12: return 8; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
78 case 0x13: return 8; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
79 case 0x20: return 32; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
80 case 0x21: return 48; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
81 case 0x30: return 192; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
82 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
|
83 default: |
|
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
84 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
|
85 return 0; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
86 } |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
87 } |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
88 |
| 10879 | 89 bool gpu::Ptx::initialize_gpu() { |
| 90 | |
| 91 /* Initialize CUDA driver API */ | |
| 92 int status = _cuda_cu_init(0); | |
| 93 if (status != GRAAL_CUDA_SUCCESS) { | |
| 94 tty->print_cr("Failed to initialize CUDA device"); | |
| 95 return false; | |
| 96 } | |
|
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
97 |
| 10879 | 98 if (TraceGPUInteraction) { |
| 99 tty->print_cr("CUDA driver initialization: Success"); | |
| 100 } | |
| 101 | |
| 102 /* Get the number of compute-capable device count */ | |
| 103 int device_count = 0; | |
| 104 status = _cuda_cu_device_get_count(&device_count); | |
| 105 if (status != GRAAL_CUDA_SUCCESS) { | |
| 106 tty->print_cr("[CUDA] Failed to get compute-capable device count"); | |
| 107 return false; | |
| 108 } | |
| 109 | |
| 110 if (device_count == 0) { | |
| 111 tty->print_cr("[CUDA] Found no device supporting CUDA"); | |
| 112 return false; | |
| 113 } | |
| 114 | |
| 115 if (TraceGPUInteraction) { | |
| 116 tty->print_cr("[CUDA] Number of compute-capable devices found: %d", device_count); | |
| 117 } | |
|
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
118 |
| 10879 | 119 /* Get the handle to the first compute device */ |
| 120 int device_id = 0; | |
| 121 /* 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
|
122 status = _cuda_cu_device_get(&_cu_device, device_id); |
| 10879 | 123 |
| 124 if (status != GRAAL_CUDA_SUCCESS) { | |
| 125 tty->print_cr("[CUDA] Failed to get handle of first compute-capable device i.e., the one at ordinal: %d", device_id); | |
| 126 return false; | |
| 127 } | |
| 128 | |
| 129 if (TraceGPUInteraction) { | |
| 130 tty->print_cr("[CUDA] Got the handle of first compute-device"); | |
| 131 } | |
| 132 | |
| 133 /* Get device attributes */ | |
|
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
134 int unified_addressing; |
| 10879 | 135 |
|
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_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
|
137 |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
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 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
|
140 return false; |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
141 } |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
142 |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
143 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
|
144 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
|
145 } |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
146 |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
147 |
| 10879 | 148 /* Get device name */ |
| 149 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
|
150 status = _cuda_cu_device_get_name(device_name, 256, _cu_device); |
| 10879 | 151 |
| 152 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
|
153 tty->print_cr("[CUDA] Failed to get name of device: %d", _cu_device); |
| 10879 | 154 return false; |
| 155 } | |
| 156 | |
| 157 if (TraceGPUInteraction) { | |
| 158 tty->print_cr("[CUDA] Using %s", device_name); | |
| 159 } | |
| 160 | |
|
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
161 |
| 10879 | 162 return true; |
| 163 } | |
| 164 | |
|
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
165 unsigned int gpu::Ptx::total_cores() { |
|
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 int minor, major, nmp; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
168 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
|
169 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
|
170 _cu_device); |
|
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 if (status != GRAAL_CUDA_SUCCESS) { |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
173 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
|
174 return 0; |
|
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 |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
177 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
|
178 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
|
179 _cu_device); |
|
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 if (status != GRAAL_CUDA_SUCCESS) { |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
182 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
|
183 return 0; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
184 } |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
185 |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
186 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
|
187 GRAAL_CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
188 _cu_device); |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
189 |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
190 if (status != GRAAL_CUDA_SUCCESS) { |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
191 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
|
192 return 0; |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
193 } |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
194 |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
195 int total = nmp * ncores(major, minor); |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
196 |
|
11894
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
197 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
|
198 |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
199 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
|
200 GRAAL_CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
201 _cu_device); |
|
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 if (status != GRAAL_CUDA_SUCCESS) { |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
204 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
|
205 return 0; |
|
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 |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
208 status = _cuda_cu_device_get_attribute(&warp_size, |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
209 GRAAL_CU_DEVICE_ATTRIBUTE_WARP_SIZE, |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
210 _cu_device); |
|
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 if (status != GRAAL_CUDA_SUCCESS) { |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
213 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
|
214 return 0; |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
215 } |
|
12653
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
216 |
|
11894
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
217 status = _cuda_cu_device_get_attribute(&async_engines, |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
218 GRAAL_CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT, |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
219 _cu_device); |
|
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 if (status != GRAAL_CUDA_SUCCESS) { |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
222 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
|
223 return 0; |
|
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 |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
226 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
|
227 GRAAL_CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY, |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
228 _cu_device); |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
229 |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
230 if (status != GRAAL_CUDA_SUCCESS) { |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
231 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
|
232 return 0; |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
233 } |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
234 |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
235 status = _cuda_cu_device_get_attribute(&concurrent_kernels, |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
236 GRAAL_CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS, |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
237 _cu_device); |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
238 |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
239 if (status != GRAAL_CUDA_SUCCESS) { |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
240 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
|
241 return 0; |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
242 } |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
243 |
|
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
244 if (TraceGPUInteraction) { |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
245 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
|
246 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
|
247 total, async_engines, can_map_host_memory, concurrent_kernels); |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
248 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
|
249 } |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
250 return (total); |
|
12653
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
251 |
|
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
252 } |
|
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
253 |
| 10879 | 254 void *gpu::Ptx::generate_kernel(unsigned char *code, int code_len, const char *name) { |
| 255 | |
| 256 struct CUmod_st * cu_module; | |
| 257 // Use three JIT compiler options | |
| 258 const unsigned int jit_num_options = 3; | |
| 259 int *jit_options = NEW_C_HEAP_ARRAY(int, jit_num_options, mtCompiler); | |
| 260 void **jit_option_values = NEW_C_HEAP_ARRAY(void *, jit_num_options, mtCompiler); | |
| 261 | |
| 262 // Set up PTX JIT compiler options | |
| 263 // 1. set size of compilation log buffer | |
| 264 int jit_log_buffer_size = 1024; | |
| 265 jit_options[0] = GRAAL_CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES; | |
| 266 jit_option_values[0] = (void *)(size_t)jit_log_buffer_size; | |
| 267 | |
| 268 // 2. set pointer to compilation log buffer | |
| 269 char *jit_log_buffer = NEW_C_HEAP_ARRAY(char, jit_log_buffer_size, mtCompiler); | |
| 270 jit_options[1] = GRAAL_CU_JIT_INFO_LOG_BUFFER; | |
| 271 jit_option_values[1] = jit_log_buffer; | |
| 272 | |
| 273 // 3. set pointer to set the Maximum # of registers (32) for the kernel | |
| 274 int jit_register_count = 32; | |
| 275 jit_options[2] = GRAAL_CU_JIT_MAX_REGISTERS; | |
| 276 jit_option_values[2] = (void *)(size_t)jit_register_count; | |
| 277 | |
|
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
|
278 /* Create CUDA context to compile and execute the kernel */ |
|
12653
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
279 int status = _cuda_cu_ctx_create(&_device_context, GRAAL_CU_CTX_MAP_HOST, _cu_device); |
|
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
|
280 |
|
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
|
281 if (status != GRAAL_CUDA_SUCCESS) { |
|
11634
bff2b88444f5
Start of PTX array passing
Morris Meyer <morris.meyer@oracle.com>
parents:
11606
diff
changeset
|
282 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
|
283 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
|
284 } |
|
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
|
285 |
|
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
|
286 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
|
287 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
|
288 } |
|
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
|
289 |
|
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
290 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
|
291 |
|
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
292 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
|
293 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
|
294 return NULL; |
|
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
295 } |
|
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
296 |
|
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
297 if (TraceGPUInteraction) { |
|
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
298 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
|
299 } |
|
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
300 |
|
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
301 if (TraceGPUInteraction) { |
|
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
302 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
|
303 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
|
304 |
|
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
305 } |
|
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
306 |
| 10879 | 307 /* 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
|
308 status = _cuda_cu_module_load_data_ex(&cu_module, (void*) code, jit_num_options, |
| 10879 | 309 jit_options, (void **)jit_option_values); |
| 310 if (status != GRAAL_CUDA_SUCCESS) { | |
| 311 if (status == GRAAL_CUDA_ERROR_NO_BINARY_FOR_GPU) { | |
| 312 tty->print_cr("[CUDA] Check for malformed PTX kernel or incorrect PTX compilation options"); | |
| 313 } | |
|
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
314 tty->print_cr("[CUDA] *** Error (%d) Failed to load module data with online compiler options for method %s", |
| 10879 | 315 status, name); |
| 316 return NULL; | |
| 317 } | |
| 318 | |
| 319 if (TraceGPUInteraction) { | |
| 320 tty->print_cr("[CUDA] Loaded data for PTX Kernel"); | |
| 321 } | |
| 322 | |
| 323 struct CUfunc_st * cu_function; | |
| 324 | |
| 325 status = _cuda_cu_module_get_function(&cu_function, cu_module, name); | |
| 326 | |
| 327 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
|
328 tty->print_cr("[CUDA] *** Error: Failed to get function %s", name); |
| 10879 | 329 return NULL; |
| 330 } | |
| 331 | |
| 332 if (TraceGPUInteraction) { | |
| 333 tty->print_cr("[CUDA] Got function handle for %s", name); | |
| 334 } | |
|
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
|
335 |
| 10879 | 336 return cu_function; |
| 337 } | |
| 338 | |
|
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
|
339 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
|
340 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
|
341 } |
|
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
342 |
|
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
343 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
|
344 address kernel, PTXKernelArguments &ptxka, JavaValue &ret) { |
| 10879 | 345 // grid dimensionality |
| 346 unsigned int gridX = 1; | |
| 347 unsigned int gridY = 1; | |
| 348 unsigned int gridZ = 1; | |
| 349 | |
| 350 // thread dimensionality | |
|
11822
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
351 unsigned int blockX = dimX; |
|
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
352 unsigned int blockY = dimY; |
|
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
353 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
|
354 |
|
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
|
355 struct CUfunc_st* cu_function = (struct CUfunc_st*) kernel; |
| 10879 | 356 |
|
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
|
357 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
|
358 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
|
359 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
|
360 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
|
361 }; |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
362 |
| 10879 | 363 if (kernel == NULL) { |
| 364 return false; | |
| 365 } | |
| 366 | |
| 367 if (TraceGPUInteraction) { | |
| 368 tty->print_cr("[CUDA] launching kernel"); | |
| 369 } | |
|
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
|
370 |
| 10879 | 371 int status = _cuda_cu_launch_kernel(cu_function, |
| 372 gridX, gridY, gridZ, | |
| 373 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
|
374 0, NULL, NULL, (void **) &config); |
| 10879 | 375 if (status != GRAAL_CUDA_SUCCESS) { |
| 376 tty->print_cr("[CUDA] Failed to launch kernel"); | |
| 377 return false; | |
| 378 } | |
| 379 | |
| 380 if (TraceGPUInteraction) { | |
|
11822
365d8f385fb5
PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11821
diff
changeset
|
381 tty->print_cr("[CUDA] Success: Kernel Launch: X: %d Y: %d Z: %d", blockX, blockY, blockZ); |
| 10879 | 382 } |
|
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
|
383 |
|
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 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
|
385 |
|
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 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
|
387 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
|
388 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
|
389 } |
|
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 |
|
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 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
|
392 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
|
393 } |
|
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 |
|
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
|
396 // 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
|
397 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
|
398 switch (return_type) { |
|
11821
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
399 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
|
400 { |
|
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
|
401 int return_val; |
|
12519
f020e149c1b6
PTX codegen enhancements; fixes to PTX test regressions.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
11902
diff
changeset
|
402 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_INT_BYTE_SIZE); |
|
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
|
403 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
|
404 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
|
405 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
|
406 } |
|
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
|
407 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
|
408 } |
|
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
|
409 break; |
|
11901
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
410 case T_BOOLEAN: |
|
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
411 { |
|
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
412 int return_val; |
|
12519
f020e149c1b6
PTX codegen enhancements; fixes to PTX test regressions.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
11902
diff
changeset
|
413 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_INT_BYTE_SIZE); |
|
11901
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
414 if (status != GRAAL_CUDA_SUCCESS) { |
|
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
415 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
|
416 return false; |
|
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
417 } |
|
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
418 ret.set_jint(return_val); |
|
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
419 } |
|
61767ccd4600
PTX boolean return value, emitIntegerTestMove, warnings
Morris Meyer <morris.meyer@oracle.com>
parents:
11894
diff
changeset
|
420 break; |
|
11894
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
421 case T_FLOAT: |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
422 { |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
423 float return_val; |
|
12519
f020e149c1b6
PTX codegen enhancements; fixes to PTX test regressions.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
11902
diff
changeset
|
424 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_FLOAT_BYTE_SIZE); |
|
11894
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
425 if (status != GRAAL_CUDA_SUCCESS) { |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
426 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
|
427 return false; |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
428 } |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
429 ret.set_jfloat(return_val); |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
430 } |
|
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
431 break; |
|
11902
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
432 case T_DOUBLE: |
|
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
433 { |
|
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
434 double return_val; |
|
12519
f020e149c1b6
PTX codegen enhancements; fixes to PTX test regressions.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
11902
diff
changeset
|
435 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_DOUBLE_BYTE_SIZE); |
|
11902
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
436 if (status != GRAAL_CUDA_SUCCESS) { |
|
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
437 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
|
438 return false; |
|
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
439 } |
|
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
440 ret.set_jdouble(return_val); |
|
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
441 } |
|
67a1e27a8dbb
PTX initial float and double
Morris Meyer <morris.meyer@oracle.com>
parents:
11901
diff
changeset
|
442 break; |
|
11821
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
443 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
|
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 long return_val; |
|
12519
f020e149c1b6
PTX codegen enhancements; fixes to PTX test regressions.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
11902
diff
changeset
|
446 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_LONG_BYTE_SIZE); |
|
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
|
447 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
|
448 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
|
449 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
|
450 } |
|
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 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
|
452 } |
|
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 break; |
|
11821
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
454 case T_VOID: |
|
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
455 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
|
456 default: |
|
11821
d8659ad83fcc
PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents:
11634
diff
changeset
|
457 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
|
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 // Free device memory allocated for result |
|
12519
f020e149c1b6
PTX codegen enhancements; fixes to PTX test regressions.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
11902
diff
changeset
|
461 status = gpu::Ptx::_cuda_cu_memfree(ptxka._dev_return_value); |
|
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
|
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 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
|
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: 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
|
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 // 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
|
472 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
|
473 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
|
474 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
|
475 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
|
476 } |
|
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
|
477 |
|
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
|
478 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
|
479 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
|
480 } |
|
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
|
481 |
|
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
|
482 return (status == GRAAL_CUDA_SUCCESS); |
| 10879 | 483 } |
| 484 | |
| 485 #if defined(LINUX) | |
| 486 static const char cuda_library_name[] = "libcuda.so"; | |
| 487 #elif defined(__APPLE__) | |
| 488 static char const cuda_library_name[] = "/usr/local/cuda/lib/libcuda.dylib"; | |
| 489 #else | |
| 490 static char const cuda_library_name[] = ""; | |
| 491 #endif | |
| 492 | |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
493 #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
|
494 |
| 10879 | 495 bool gpu::Ptx::probe_linkage() { |
| 496 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
|
497 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
|
498 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
|
499 free(buffer); |
| 10879 | 500 if (handle != NULL) { |
|
12653
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
501 LOOKUP_CUDA_FUNCTION(cuInit, cuda_cu_init); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
502 LOOKUP_CUDA_FUNCTION(cuCtxSynchronize, cuda_cu_ctx_synchronize); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
503 LOOKUP_CUDA_FUNCTION(cuCtxSetCurrent, cuda_cu_ctx_set_current); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
504 LOOKUP_CUDA_FUNCTION(cuDeviceGetCount, cuda_cu_device_get_count); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
505 LOOKUP_CUDA_FUNCTION(cuDeviceGetName, cuda_cu_device_get_name); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
506 LOOKUP_CUDA_FUNCTION(cuDeviceGet, cuda_cu_device_get); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
507 LOOKUP_CUDA_FUNCTION(cuDeviceComputeCapability, cuda_cu_device_compute_capability); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
508 LOOKUP_CUDA_FUNCTION(cuDeviceGetAttribute, cuda_cu_device_get_attribute); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
509 LOOKUP_CUDA_FUNCTION(cuModuleGetFunction, cuda_cu_module_get_function); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
510 LOOKUP_CUDA_FUNCTION(cuModuleLoadDataEx, cuda_cu_module_load_data_ex); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
511 LOOKUP_CUDA_FUNCTION(cuLaunchKernel, cuda_cu_launch_kernel); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
512 LOOKUP_CUDA_FUNCTION(cuMemHostRegister, cuda_cu_mem_host_register); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
513 LOOKUP_CUDA_FUNCTION(cuMemHostUnregister, cuda_cu_mem_host_unregister); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
514 #if defined(__x86_64) || defined(AMD64) || defined(_M_AMD64) |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
515 LOOKUP_CUDA_V2_FUNCTION(cuCtxCreate, cuda_cu_ctx_create); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
516 LOOKUP_CUDA_V2_FUNCTION(cuCtxDestroy, cuda_cu_ctx_destroy); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
517 LOOKUP_CUDA_V2_FUNCTION(cuMemAlloc, cuda_cu_memalloc); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
518 LOOKUP_CUDA_V2_FUNCTION(cuMemFree, cuda_cu_memfree); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
519 LOOKUP_CUDA_V2_FUNCTION(cuMemcpyHtoD, cuda_cu_memcpy_htod); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
520 LOOKUP_CUDA_V2_FUNCTION(cuMemcpyDtoH, cuda_cu_memcpy_dtoh); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
521 LOOKUP_CUDA_V2_FUNCTION(cuMemHostGetDevicePointer, cuda_cu_mem_host_get_device_pointer); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
522 #else |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
523 LOOKUP_CUDA_FUNCTION(cuCtxCreate, cuda_cu_ctx_create); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
524 LOOKUP_CUDA_FUNCTION(cuCtxDestroy, cuda_cu_ctx_destroy); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
525 LOOKUP_CUDA_FUNCTION(cuMemAlloc, cuda_cu_memalloc); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
526 LOOKUP_CUDA_FUNCTION(cuMemFree, cuda_cu_memfree); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
527 LOOKUP_CUDA_FUNCTION(cuMemcpyHtoD, cuda_cu_memcpy_htod); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
528 LOOKUP_CUDA_FUNCTION(cuMemcpyDtoH, cuda_cu_memcpy_dtoh); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
529 LOOKUP_CUDA_FUNCTION(cuMemHostGetDevicePointer, cuda_cu_mem_host_get_device_pointer); |
|
1a7e7011a341
* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
12519
diff
changeset
|
530 #endif |
|
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
|
531 |
| 10879 | 532 if (TraceGPUInteraction) { |
| 533 tty->print_cr("[CUDA] Success: library linkage"); | |
| 534 } | |
| 535 return true; | |
| 536 } else { | |
| 537 // Unable to dlopen libcuda | |
| 538 return false; | |
| 539 } | |
| 540 } else { | |
| 541 tty->print_cr("Unsupported CUDA platform"); | |
| 542 return false; | |
| 543 } | |
| 544 tty->print_cr("Failed to find CUDA linkage"); | |
| 545 return false; | |
| 546 } |