annotate src/gpu/ptx/vm/gpu_ptx.cpp @ 11843:372bacc13022

gtx_ptx.cpp: fix switch default label and operator precedence in ncores().
author Andreas Woess <andreas.woess@jku.at>
date Mon, 30 Sep 2013 21:46:41 +0200
parents 8d8f63069f58
children c7abc8411011
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
1 /*
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
2 * Copyright (c) 2013, Oracle and/or its affiliates. All rights reserved.
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
4 *
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
5 * This code is free software; you can redistribute it and/or modify it
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
6 * under the terms of the GNU General Public License version 2 only, as
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
7 * published by the Free Software Foundation.
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
8 *
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
9 * This code is distributed in the hope that it will be useful, but WITHOUT
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
12 * version 2 for more details (a copy is included in the LICENSE file that
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
13 * accompanied this code).
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
14 *
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
15 * You should have received a copy of the GNU General Public License version
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
16 * 2 along with this work; if not, write to the Free Software Foundation,
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
18 *
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
20 * or visit www.oracle.com if you need additional information or have any
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
21 * questions.
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
22 *
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
23 */
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
24
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
27 #include "runtime/gpu.hpp"
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
28 #include "utilities/globalDefinitions.hpp"
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
33
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
36
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
37 gpu::Ptx::cuda_cu_init_func_t gpu::Ptx::_cuda_cu_init;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
42 gpu::Ptx::cuda_cu_device_get_count_func_t gpu::Ptx::_cuda_cu_device_get_count;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
43 gpu::Ptx::cuda_cu_device_get_name_func_t gpu::Ptx::_cuda_cu_device_get_name;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
44 gpu::Ptx::cuda_cu_device_get_func_t gpu::Ptx::_cuda_cu_device_get;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
45 gpu::Ptx::cuda_cu_device_compute_capability_func_t gpu::Ptx::_cuda_cu_device_compute_capability;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
46 gpu::Ptx::cuda_cu_device_get_attribute_func_t gpu::Ptx::_cuda_cu_device_get_attribute;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
47 gpu::Ptx::cuda_cu_launch_kernel_func_t gpu::Ptx::_cuda_cu_launch_kernel;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
48 gpu::Ptx::cuda_cu_module_get_function_func_t gpu::Ptx::_cuda_cu_module_get_function;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
75 bool gpu::Ptx::initialize_gpu() {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
76
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
77 /* Initialize CUDA driver API */
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
78 int status = _cuda_cu_init(0);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
79 if (status != GRAAL_CUDA_SUCCESS) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
80 tty->print_cr("Failed to initialize CUDA device");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
81 return false;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
84 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
85 tty->print_cr("CUDA driver initialization: Success");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
86 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
87
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
88 /* Get the number of compute-capable device count */
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
89 int device_count = 0;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
90 status = _cuda_cu_device_get_count(&device_count);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
91 if (status != GRAAL_CUDA_SUCCESS) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
92 tty->print_cr("[CUDA] Failed to get compute-capable device count");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
93 return false;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
94 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
95
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
96 if (device_count == 0) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
97 tty->print_cr("[CUDA] Found no device supporting CUDA");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
98 return false;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
99 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
100
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
101 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
102 tty->print_cr("[CUDA] Number of compute-capable devices found: %d", device_count);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
105 /* Get the handle to the first compute device */
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
106 int device_id = 0;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
109
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
110 if (status != GRAAL_CUDA_SUCCESS) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
111 tty->print_cr("[CUDA] Failed to get handle of first compute-capable device i.e., the one at ordinal: %d", device_id);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
112 return false;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
113 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
114
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
115 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
116 tty->print_cr("[CUDA] Got the handle of first compute-device");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
117 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
118
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
134 /* Get device name */
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
137
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
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 get name of device: %d", _cu_device);
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
140 return false;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
141 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
142
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
143 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
144 tty->print_cr("[CUDA] Using %s", device_name);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
145 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
146
11842
8d8f63069f58 PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents: 11822
diff changeset
147
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
148 return true;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
149 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
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
8d8f63069f58 PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents: 11822
diff changeset
183 if (TraceGPUInteraction) {
8d8f63069f58 PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents: 11822
diff changeset
184 tty->print_cr("[CUDA] Compatibility version of device %d: %d.%d", _cu_device, major, minor);
8d8f63069f58 PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents: 11822
diff changeset
185 tty->print_cr("[CUDA] Number of cores: %d", total);
8d8f63069f58 PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents: 11822
diff changeset
186 }
8d8f63069f58 PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents: 11822
diff changeset
187 return (total);
8d8f63069f58 PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents: 11822
diff changeset
188
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
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
191 void *gpu::Ptx::generate_kernel(unsigned char *code, int code_len, const char *name) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
192
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
193 struct CUmod_st * cu_module;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
194 // Use three JIT compiler options
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
195 const unsigned int jit_num_options = 3;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
196 int *jit_options = NEW_C_HEAP_ARRAY(int, jit_num_options, mtCompiler);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
197 void **jit_option_values = NEW_C_HEAP_ARRAY(void *, jit_num_options, mtCompiler);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
198
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
199 // Set up PTX JIT compiler options
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
200 // 1. set size of compilation log buffer
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
201 int jit_log_buffer_size = 1024;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
202 jit_options[0] = GRAAL_CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
203 jit_option_values[0] = (void *)(size_t)jit_log_buffer_size;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
204
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
205 // 2. set pointer to compilation log buffer
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
206 char *jit_log_buffer = NEW_C_HEAP_ARRAY(char, jit_log_buffer_size, mtCompiler);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
207 jit_options[1] = GRAAL_CU_JIT_INFO_LOG_BUFFER;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
208 jit_option_values[1] = jit_log_buffer;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
209
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
210 // 3. set pointer to set the Maximum # of registers (32) for the kernel
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
211 int jit_register_count = 32;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
212 jit_options[2] = GRAAL_CU_JIT_MAX_REGISTERS;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
213 jit_option_values[2] = (void *)(size_t)jit_register_count;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
214
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
215 /* 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
216 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
217
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
218 if (status != GRAAL_CUDA_SUCCESS) {
11634
bff2b88444f5 Start of PTX array passing
Morris Meyer <morris.meyer@oracle.com>
parents: 11606
diff changeset
219 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
220 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
221 }
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
222
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
223 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
224 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
225 }
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
226
11527
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
227 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
228
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
229 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
230 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
231 return NULL;
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
232 }
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
233
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
234 if (TraceGPUInteraction) {
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
235 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
236 }
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
237
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
238 if (TraceGPUInteraction) {
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
239 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
240 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
241
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
242 }
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
243
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
244 /* 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
245 status = _cuda_cu_module_load_data_ex(&cu_module, (void*) code, jit_num_options,
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
246 jit_options, (void **)jit_option_values);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
247 if (status != GRAAL_CUDA_SUCCESS) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
248 if (status == GRAAL_CUDA_ERROR_NO_BINARY_FOR_GPU) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
249 tty->print_cr("[CUDA] Check for malformed PTX kernel or incorrect PTX compilation options");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
250 }
11527
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents: 11485
diff changeset
251 tty->print_cr("[CUDA] *** Error (%d) Failed to load module data with online compiler options for method %s",
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
252 status, name);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
253 return NULL;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
254 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
255
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
256 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
257 tty->print_cr("[CUDA] Loaded data for PTX Kernel");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
258 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
259
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
260 struct CUfunc_st * cu_function;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
261
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
262 status = _cuda_cu_module_get_function(&cu_function, cu_module, name);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
263
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
264 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
265 tty->print_cr("[CUDA] *** Error: Failed to get function %s", name);
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
266 return NULL;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
267 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
268
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
269 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
270 tty->print_cr("[CUDA] Got function handle for %s", name);
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
271 }
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
272
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
273 return cu_function;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
274 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
275
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
276 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
277 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
278 }
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11821
diff changeset
279
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11821
diff changeset
280 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
281 address kernel, PTXKernelArguments &ptxka, JavaValue &ret) {
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
282 // grid dimensionality
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
283 unsigned int gridX = 1;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
284 unsigned int gridY = 1;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
285 unsigned int gridZ = 1;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
286
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
287 // thread dimensionality
11822
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11821
diff changeset
288 unsigned int blockX = dimX;
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11821
diff changeset
289 unsigned int blockY = dimY;
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11821
diff changeset
290 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
291
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
292 struct CUfunc_st* cu_function = (struct CUfunc_st*) kernel;
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
293
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 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
295 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
296 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
297 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
298 };
1cd1f8ff70a1 CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents: 10884
diff changeset
299
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
300 if (kernel == NULL) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
301 return false;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
302 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
303
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
304 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
305 tty->print_cr("[CUDA] launching kernel");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
306 }
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
307
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
308 int status = _cuda_cu_launch_kernel(cu_function,
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
309 gridX, gridY, gridZ,
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
310 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
311 0, NULL, NULL, (void **) &config);
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
312 if (status != GRAAL_CUDA_SUCCESS) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
313 tty->print_cr("[CUDA] Failed to launch kernel");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
314 return false;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
315 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
316
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
317 if (TraceGPUInteraction) {
11822
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11821
diff changeset
318 tty->print_cr("[CUDA] Success: Kernel Launch: X: %d Y: %d Z: %d", blockX, blockY, blockZ);
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
319 }
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
320
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 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
322
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
323 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
324 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
325 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
326 }
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
327
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
328 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
329 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
330 }
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
331
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
332
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
333 // 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
334 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
335 switch (return_type) {
11821
d8659ad83fcc PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11634
diff changeset
336 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
337 {
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
338 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
339 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
340 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
341 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
342 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
343 }
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 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
345 }
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
346 break;
11821
d8659ad83fcc PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11634
diff changeset
347 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
348 {
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
349 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
350 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
351 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
352 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
353 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
354 }
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 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
356 }
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 break;
11821
d8659ad83fcc PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11634
diff changeset
358 case T_VOID:
d8659ad83fcc PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11634
diff changeset
359 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
360 default:
11821
d8659ad83fcc PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11634
diff changeset
361 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
362 }
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
363
11821
d8659ad83fcc PTX single-threaded array store, Warp annotation
Morris Meyer <morris.meyer@oracle.com>
parents: 11634
diff changeset
364 // 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
365 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
366
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
367 // 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
368 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
369 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
370 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
371 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
372 }
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
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 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
375 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
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
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 // 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
379 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
380 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
381 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
382 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
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
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 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
386 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
387 }
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
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 return (status == GRAAL_CUDA_SUCCESS);
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
390 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
391
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
392 #if defined(LINUX)
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
393 static const char cuda_library_name[] = "libcuda.so";
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
394 #elif defined(__APPLE__)
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
395 static char const cuda_library_name[] = "/usr/local/cuda/lib/libcuda.dylib";
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
396 #else
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
397 static char const cuda_library_name[] = "";
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
398 #endif
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
399
10882
6a2d65cb5d7d fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents: 10879
diff changeset
400 #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
401
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
402 bool gpu::Ptx::probe_linkage() {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
403 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
404 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
405 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
406 free(buffer);
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
407 if (handle != NULL) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
408 _cuda_cu_init =
10882
6a2d65cb5d7d fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents: 10879
diff changeset
409 CAST_TO_FN_PTR(cuda_cu_init_func_t, os::dll_lookup(handle, "cuInit"));
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
410 _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
411 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
412 _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
413 CAST_TO_FN_PTR(cuda_cu_ctx_destroy_func_t, os::dll_lookup(handle, "cuCtxDestroy"));
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
414 _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
415 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
416 _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
417 CAST_TO_FN_PTR(cuda_cu_ctx_set_current_func_t, os::dll_lookup(handle, "cuCtxSetCurrent"));
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
418 _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
419 CAST_TO_FN_PTR(cuda_cu_device_get_count_func_t, os::dll_lookup(handle, "cuDeviceGetCount"));
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
420 _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
421 CAST_TO_FN_PTR(cuda_cu_device_get_name_func_t, os::dll_lookup(handle, "cuDeviceGetName"));
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
422 _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
423 CAST_TO_FN_PTR(cuda_cu_device_get_func_t, os::dll_lookup(handle, "cuDeviceGet"));
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
424 _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
425 CAST_TO_FN_PTR(cuda_cu_device_compute_capability_func_t, os::dll_lookup(handle, "cuDeviceComputeCapability"));
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
426 _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
427 CAST_TO_FN_PTR(cuda_cu_device_get_attribute_func_t, os::dll_lookup(handle, "cuDeviceGetAttribute"));
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
428 _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
429 CAST_TO_FN_PTR(cuda_cu_module_get_function_func_t, os::dll_lookup(handle, "cuModuleGetFunction"));
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
430 _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
431 CAST_TO_FN_PTR(cuda_cu_module_load_data_ex_func_t, os::dll_lookup(handle, "cuModuleLoadDataEx"));
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
432 _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
433 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
434 _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
435 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
436 _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
437 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
438 _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
439 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
440 _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
441 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
442
10879
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
443 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
444 tty->print_cr("[CUDA] Success: library linkage");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
445 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
446 return true;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
447 } else {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
448 // Unable to dlopen libcuda
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
449 return false;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
450 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
451 } else {
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
452 tty->print_cr("Unsupported CUDA platform");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
453 return false;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
454 }
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
455 tty->print_cr("Failed to find CUDA linkage");
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
456 return false;
d55f24eac4b1 PTX support for Linux
Morris Meyer <morris.meyer@oracle.com>
parents:
diff changeset
457 }