Mercurial > hg > truffle
annotate src/gpu/ptx/vm/gpu_ptx.cpp @ 13828:5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
author | S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> |
---|---|
date | Thu, 30 Jan 2014 17:49:56 -0500 |
parents | 49db2c1e3bee |
children | 4ca607fc94e0 |
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" |
13624
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
32 #include "runtime/interfaceSupport.hpp" |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
33 #include "graal/graalEnv.hpp" |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
34 #include "graal/graalCompiler.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
|
35 #include "ptxKernelArguments.hpp" |
10879 | 36 |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
37 // Entry to GPU native method implementation that transitions current thread to '_thread_in_vm'. |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
38 #define GPU_VMENTRY(result_type, name, signature) \ |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
39 JNIEXPORT result_type JNICALL name signature { \ |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
40 GRAAL_VM_ENTRY_MARK; \ |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
41 |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
42 // Entry to GPU native method implementation that calls a JNI function |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
43 // and hence cannot transition current thread to '_thread_in_vm'. |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
44 #define GPU_ENTRY(result_type, name, signature) \ |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
45 JNIEXPORT result_type JNICALL name signature { \ |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
46 |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
47 #define GPU_END } |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
48 |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
49 #define CC (char*) /*cast a literal from (const char*)*/ |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
50 #define FN_PTR(f) CAST_FROM_FN_PTR(void*, &(f)) |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
51 |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
52 #define STRING "Ljava/lang/String;" |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
53 |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
54 JNINativeMethod gpu::Ptx::PTX_methods[] = { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
55 {CC"initialize", CC"()Z", FN_PTR(gpu::Ptx::initialize)}, |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
56 {CC"generateKernel", CC"([B" STRING ")J", FN_PTR(gpu::Ptx::generate_kernel)}, |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
57 {CC"getLaunchKernelAddress", CC"()J", FN_PTR(gpu::Ptx::get_execute_kernel_from_vm_address)}, |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
58 {CC"getAvailableProcessors0", CC"()I", FN_PTR(gpu::Ptx::get_total_cores)}, |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
59 }; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
60 |
10879 | 61 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
|
62 int gpu::Ptx::_cu_device = 0; |
10879 | 63 |
64 gpu::Ptx::cuda_cu_init_func_t gpu::Ptx::_cuda_cu_init; | |
65 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
|
66 gpu::Ptx::cuda_cu_ctx_destroy_func_t gpu::Ptx::_cuda_cu_ctx_destroy; |
10879 | 67 gpu::Ptx::cuda_cu_ctx_synchronize_func_t gpu::Ptx::_cuda_cu_ctx_synchronize; |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
68 gpu::Ptx::cuda_cu_ctx_get_current_func_t gpu::Ptx::_cuda_cu_ctx_get_current; |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
69 gpu::Ptx::cuda_cu_ctx_set_current_func_t gpu::Ptx::_cuda_cu_ctx_set_current; |
10879 | 70 gpu::Ptx::cuda_cu_device_get_count_func_t gpu::Ptx::_cuda_cu_device_get_count; |
71 gpu::Ptx::cuda_cu_device_get_name_func_t gpu::Ptx::_cuda_cu_device_get_name; | |
72 gpu::Ptx::cuda_cu_device_get_func_t gpu::Ptx::_cuda_cu_device_get; | |
73 gpu::Ptx::cuda_cu_device_compute_capability_func_t gpu::Ptx::_cuda_cu_device_compute_capability; | |
74 gpu::Ptx::cuda_cu_device_get_attribute_func_t gpu::Ptx::_cuda_cu_device_get_attribute; | |
75 gpu::Ptx::cuda_cu_launch_kernel_func_t gpu::Ptx::_cuda_cu_launch_kernel; | |
76 gpu::Ptx::cuda_cu_module_get_function_func_t gpu::Ptx::_cuda_cu_module_get_function; | |
77 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
|
78 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
|
79 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
|
80 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
|
81 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
|
82 gpu::Ptx::cuda_cu_mem_host_unregister_func_t gpu::Ptx::_cuda_cu_mem_host_unregister; |
10879 | 83 |
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
|
84 #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
|
85 |
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
|
86 #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
|
87 _##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
|
88 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
|
89 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
|
90 tty->print_cr("[CUDA] ***** Error: Failed to lookup %s", STRINGIFY(name)); \ |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
91 return false; \ |
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
|
92 } \ |
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
|
93 |
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
|
94 #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
|
95 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
96 /* |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
97 * 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
|
98 */ |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
99 int gpu::Ptx::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
|
100 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
|
101 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
102 switch (device_type) { |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
103 case 0x10: return 8; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
104 case 0x11: return 8; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
105 case 0x12: return 8; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
106 case 0x13: return 8; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
107 case 0x20: return 32; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
108 case 0x21: return 48; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
109 case 0x30: return 192; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
110 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
|
111 default: |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
112 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
|
113 return 0; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
114 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
115 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
116 |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
117 bool gpu::Ptx::register_natives(JNIEnv* env) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
118 jclass klass = env->FindClass("com/oracle/graal/hotspot/ptx/PTXHotSpotBackend"); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
119 if (klass == NULL) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
120 if (TraceGPUInteraction) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
121 tty->print_cr("PTXHotSpotBackend class not found"); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
122 } |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
123 return false; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
124 } |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
125 jint status = env->RegisterNatives(klass, PTX_methods, sizeof(PTX_methods) / sizeof(JNINativeMethod)); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
126 if (status != JNI_OK) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
127 if (TraceGPUInteraction) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
128 tty->print_cr("Error registering natives for PTXHotSpotBackend: %d", status); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
129 } |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
130 return false; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
131 } |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
132 return true; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
133 } |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
134 |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
135 GPU_ENTRY(jboolean, gpu::Ptx::initialize, (JNIEnv *env, jclass)) |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
136 |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
137 if (!link()) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
138 return false; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
139 } |
10879 | 140 |
141 /* Initialize CUDA driver API */ | |
142 int status = _cuda_cu_init(0); | |
143 if (status != GRAAL_CUDA_SUCCESS) { | |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
144 if (TraceGPUInteraction) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
145 tty->print_cr("Failed to initialize CUDA device: %d", status); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
146 } |
10879 | 147 return false; |
148 } | |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
149 |
10879 | 150 if (TraceGPUInteraction) { |
151 tty->print_cr("CUDA driver initialization: Success"); | |
152 } | |
153 | |
154 /* Get the number of compute-capable device count */ | |
155 int device_count = 0; | |
156 status = _cuda_cu_device_get_count(&device_count); | |
157 if (status != GRAAL_CUDA_SUCCESS) { | |
158 tty->print_cr("[CUDA] Failed to get compute-capable device count"); | |
159 return false; | |
160 } | |
161 | |
162 if (device_count == 0) { | |
163 tty->print_cr("[CUDA] Found no device supporting CUDA"); | |
164 return false; | |
165 } | |
166 | |
167 if (TraceGPUInteraction) { | |
168 tty->print_cr("[CUDA] Number of compute-capable devices found: %d", device_count); | |
169 } | |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
170 |
10879 | 171 /* Get the handle to the first compute device */ |
172 int device_id = 0; | |
173 /* 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
|
174 status = _cuda_cu_device_get(&_cu_device, device_id); |
10879 | 175 |
176 if (status != GRAAL_CUDA_SUCCESS) { | |
177 tty->print_cr("[CUDA] Failed to get handle of first compute-capable device i.e., the one at ordinal: %d", device_id); | |
178 return false; | |
179 } | |
180 | |
181 if (TraceGPUInteraction) { | |
182 tty->print_cr("[CUDA] Got the handle of first compute-device"); | |
183 } | |
184 | |
185 /* Get device attributes */ | |
13828
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
186 int minor, major; |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
187 int unified_addressing; |
13828
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
188 float version = 0.0; |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
189 |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
190 /* Get the compute capability of the device found */ |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
191 status = _cuda_cu_device_get_attribute(&minor, GRAAL_CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, _cu_device); |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
192 if (status != GRAAL_CUDA_SUCCESS) { |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
193 tty->print_cr("[CUDA] Failed to get minor attribute of device: %d", _cu_device); |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
194 return false; |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
195 } |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
196 status = _cuda_cu_device_get_attribute(&major, GRAAL_CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, _cu_device); |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
197 if (status != GRAAL_CUDA_SUCCESS) { |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
198 tty->print_cr("[CUDA] Failed to get major attribute of device: %d", _cu_device); |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
199 return false; |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
200 } |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
201 |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
202 /* Check if the device supports atleast GRAAL_SUPPORTED_COMPUTE_CAPABILITY_VERSION */ |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
203 version = (float) major + ((float) minor)/10; |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
204 |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
205 if (version < GRAAL_SUPPORTED_COMPUTE_CAPABILITY_VERSION) { |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
206 tty->print_cr("[CUDA] Only cuda compute capability 3.0 and later supported. Device %d supports %.1f", |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
207 _cu_device, version); |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
208 return false; |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
209 } |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
210 |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
211 if (TraceGPUInteraction) { |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
212 tty->print_cr("[CUDA] Device %d supports cuda compute capability %.1f", _cu_device, version); |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
213 } |
10879 | 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 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
|
216 |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
217 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
|
218 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
|
219 return false; |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
220 } |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
221 |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
222 if (TraceGPUInteraction) { |
13828
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
223 tty->print_cr("[CUDA] Device %d %s unified addressing support", _cu_device, |
5c8a3c09397b
Add additional device checks and compute capability checks for CUDA devices found on Linux.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13819
diff
changeset
|
224 ((unified_addressing == 0) ? "does not have" : "has")); |
11283
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
225 } |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
226 |
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
227 |
10879 | 228 /* Get device name */ |
229 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
|
230 status = _cuda_cu_device_get_name(device_name, 256, _cu_device); |
10879 | 231 |
232 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
|
233 tty->print_cr("[CUDA] Failed to get name of device: %d", _cu_device); |
10879 | 234 return false; |
235 } | |
236 | |
237 if (TraceGPUInteraction) { | |
238 tty->print_cr("[CUDA] Using %s", device_name); | |
239 } | |
240 | |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
241 gpu::initialized_gpu(device_name); |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
242 |
10879 | 243 return true; |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
244 GPU_END |
10879 | 245 |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
246 GPU_ENTRY(jint, gpu::Ptx::get_total_cores, (JNIEnv *env, jobject)) |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
247 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
248 int minor, major, nmp; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
249 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
|
250 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
|
251 _cu_device); |
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 if (status != GRAAL_CUDA_SUCCESS) { |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
254 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
|
255 return 0; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
256 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
257 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
258 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
|
259 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
|
260 _cu_device); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
261 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
262 if (status != GRAAL_CUDA_SUCCESS) { |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
263 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
|
264 return 0; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
265 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
266 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
267 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
|
268 GRAAL_CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
269 _cu_device); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
270 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
271 if (status != GRAAL_CUDA_SUCCESS) { |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
272 tty->print_cr("[CUDA] Failed to get number of MPs on device: %d", _cu_device); |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
273 return 0; |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
274 } |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
275 |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
276 int total = nmp * ncores(major, minor); |
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
277 |
11894
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
278 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
|
279 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
280 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
|
281 GRAAL_CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
282 _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
283 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
284 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
285 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
|
286 return 0; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
287 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
288 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
289 status = _cuda_cu_device_get_attribute(&warp_size, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
290 GRAAL_CU_DEVICE_ATTRIBUTE_WARP_SIZE, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
291 _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
292 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
293 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
294 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
|
295 return 0; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
296 } |
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
|
297 |
11894
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
298 status = _cuda_cu_device_get_attribute(&async_engines, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
299 GRAAL_CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
300 _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
301 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
302 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
303 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
|
304 return 0; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
305 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
306 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
307 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
|
308 GRAAL_CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
309 _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
310 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
311 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
312 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
|
313 return 0; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
314 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
315 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
316 status = _cuda_cu_device_get_attribute(&concurrent_kernels, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
317 GRAAL_CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS, |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
318 _cu_device); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
319 |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
320 if (status != GRAAL_CUDA_SUCCESS) { |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
321 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
|
322 return 0; |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
323 } |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
324 |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
325 if (TraceGPUInteraction) { |
11894
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
326 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
|
327 total, async_engines, can_map_host_memory, concurrent_kernels); |
c7abc8411011
Fixed BasicPTXTest and IntegerPTXTest
Morris Meyer <morris.meyer@oracle.com>
parents:
11843
diff
changeset
|
328 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
|
329 } |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
330 return total; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
331 GPU_END |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
332 |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
333 GPU_ENTRY(jlong, gpu::Ptx::generate_kernel, (JNIEnv *env, jclass, jbyteArray code_handle, jstring name_handle)) |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
334 ResourceMark rm; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
335 jsize name_len = env->GetStringLength(name_handle); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
336 jsize code_len = env->GetArrayLength(code_handle); |
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
|
337 |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
338 char* name = NEW_RESOURCE_ARRAY(char, name_len + 1); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
339 unsigned char *code = NEW_RESOURCE_ARRAY(unsigned char, code_len + 1); |
11842
8d8f63069f58
PTX warp limiter to available GPU processors
Morris Meyer <morris.meyer@oracle.com>
parents:
11822
diff
changeset
|
340 |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
341 code[code_len] = 0; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
342 name[name_len] = 0; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
343 |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
344 env->GetByteArrayRegion(code_handle, 0, code_len, (jbyte*) code); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
345 env->GetStringUTFRegion(name_handle, 0, name_len, name); |
10879 | 346 |
347 struct CUmod_st * cu_module; | |
348 // Use three JIT compiler options | |
349 const unsigned int jit_num_options = 3; | |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
350 int *jit_options = NEW_RESOURCE_ARRAY(int, jit_num_options); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
351 void **jit_option_values = NEW_RESOURCE_ARRAY(void *, jit_num_options); |
10879 | 352 |
353 // Set up PTX JIT compiler options | |
354 // 1. set size of compilation log buffer | |
355 int jit_log_buffer_size = 1024; | |
356 jit_options[0] = GRAAL_CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES; | |
357 jit_option_values[0] = (void *)(size_t)jit_log_buffer_size; | |
358 | |
359 // 2. set pointer to compilation log buffer | |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
360 char *jit_log_buffer = NEW_RESOURCE_ARRAY(char, jit_log_buffer_size); |
10879 | 361 jit_options[1] = GRAAL_CU_JIT_INFO_LOG_BUFFER; |
362 jit_option_values[1] = jit_log_buffer; | |
363 | |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
364 // 3. set pointer to set the maximum number of registers (32) for the kernel |
10879 | 365 int jit_register_count = 32; |
366 jit_options[2] = GRAAL_CU_JIT_MAX_REGISTERS; | |
367 jit_option_values[2] = (void *)(size_t)jit_register_count; | |
368 | |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
369 // 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
|
370 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
|
371 |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
372 if (status != GRAAL_CUDA_SUCCESS) { |
11634
bff2b88444f5
Start of PTX array passing
Morris Meyer <morris.meyer@oracle.com>
parents:
11606
diff
changeset
|
373 tty->print_cr("[CUDA] Failed to create CUDA context for device(%d): %d", _cu_device, status); |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
374 return 0L; |
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
|
375 } |
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
376 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
|
377 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
|
378 } |
49bb1bc983c6
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 |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
380 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
|
381 |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
382 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
|
383 tty->print_cr("[CUDA] Failed to set current context for device: %d", _cu_device); |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
384 return 0L; |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
385 } |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
386 if (TraceGPUInteraction) { |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
387 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
|
388 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
|
389 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
|
390 } |
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
391 |
10879 | 392 /* 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
|
393 status = _cuda_cu_module_load_data_ex(&cu_module, (void*) code, jit_num_options, |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
394 jit_options, (void **)jit_option_values); |
10879 | 395 if (status != GRAAL_CUDA_SUCCESS) { |
396 if (status == GRAAL_CUDA_ERROR_NO_BINARY_FOR_GPU) { | |
397 tty->print_cr("[CUDA] Check for malformed PTX kernel or incorrect PTX compilation options"); | |
398 } | |
11527
c99e65785936
Improvements to PTX codegen; allows more PTX tests that run on the device to pass.
bharadwaj
parents:
11485
diff
changeset
|
399 tty->print_cr("[CUDA] *** Error (%d) Failed to load module data with online compiler options for method %s", |
10879 | 400 status, name); |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
401 return 0L; |
10879 | 402 } |
403 | |
404 if (TraceGPUInteraction) { | |
405 tty->print_cr("[CUDA] Loaded data for PTX Kernel"); | |
406 } | |
407 | |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
408 struct CUfunc_st* cu_function; |
10879 | 409 status = _cuda_cu_module_get_function(&cu_function, cu_module, name); |
410 | |
411 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
|
412 tty->print_cr("[CUDA] *** Error: Failed to get function %s", name); |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
413 return 0L; |
10879 | 414 } |
415 | |
416 if (TraceGPUInteraction) { | |
13683
de839ec35cc7
schedule lambda method compilation and execution on GPU (PTX) when possible; fix a couple of bugs.
S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>
parents:
13655
diff
changeset
|
417 tty->print_cr("[CUDA] Got function handle for %s kernel address %p", name, cu_function); |
10879 | 418 } |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
419 return (jlong) cu_function; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
420 GPU_END |
10879 | 421 |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
422 // A PtxCall is used to manage executing a GPU kernel. In addition to launching |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
423 // the kernel, this class releases resources allocated for the execution. |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
424 class PtxCall: StackObj { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
425 private: |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
426 JavaThread* _thread; // the thread on which this call is made |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
427 address _buffer; // buffer containing parameters and _return_value |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
428 int _buffer_size; // size (in bytes) of _buffer |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
429 oop* _pinned; // objects that have been pinned with cuMemHostRegister |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
430 int _pinned_length; // length of _pinned |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
431 gpu::Ptx::CUdeviceptr _ret_value; // pointer to slot in GPU memory holding the return value |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
432 int _ret_type_size; // size of the return type value |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
433 bool _ret_is_object; // specifies if the return type is Object |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
434 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
435 bool check(int status, const char *action) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
436 if (status != GRAAL_CUDA_SUCCESS) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
437 Thread* THREAD = _thread; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
438 char* message = NEW_RESOURCE_ARRAY_IN_THREAD(THREAD, char, O_BUFLEN + 1); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
439 jio_snprintf(message, O_BUFLEN, "[CUDA] *** Error (status=%d): %s", status, action); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
440 if (TraceGPUInteraction || HAS_PENDING_EXCEPTION) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
441 tty->print_cr(message); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
442 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
443 if (!HAS_PENDING_EXCEPTION) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
444 SharedRuntime::throw_and_post_jvmti_exception(_thread, vmSymbols::java_lang_RuntimeException(), message); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
445 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
446 return false; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
447 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
448 if (TraceGPUInteraction) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
449 tty->print_cr("[CUDA] Success: %s", action); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
450 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
451 return true; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
452 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
453 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
454 public: |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
455 PtxCall(JavaThread* thread, address buffer, int buffer_size, oop* pinned, int encodedReturnTypeSize) : _thread(thread), |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
456 _buffer(buffer), _buffer_size(buffer_size), _pinned(pinned), _pinned_length(0), _ret_value(0), _ret_is_object(encodedReturnTypeSize < 0) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
457 _ret_type_size = _ret_is_object ? -encodedReturnTypeSize : encodedReturnTypeSize; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
458 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
459 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
460 bool is_object_return() { return _ret_is_object; } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
461 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
462 void alloc_return_value() { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
463 if (_ret_type_size != 0) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
464 if (check(gpu::Ptx::_cuda_cu_memalloc(&_ret_value, _ret_type_size), "Allocate device memory for return value")) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
465 gpu::Ptx::CUdeviceptr* retValuePtr = (gpu::Ptx::CUdeviceptr*) ((_buffer + _buffer_size) - sizeof(_ret_value)); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
466 *retValuePtr = _ret_value; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
467 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
468 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
469 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
470 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
471 void pin_objects(int count, int* objectOffsets) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
472 if (count == 0) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
473 return; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
474 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
475 for (int i = 0; i < count; i++) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
476 int offset = objectOffsets[i]; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
477 oop* argPtr = (oop*) (_buffer + offset); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
478 oop obj = *argPtr; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
479 if (obj != NULL) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
480 // Size (in bytes) of object |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
481 int objSize = obj->size() * HeapWordSize; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
482 //tty->print_cr("Pinning object %d at offset %d: %p", i, offset, obj); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
483 if (!check(gpu::Ptx::_cuda_cu_mem_host_register(obj, objSize, GRAAL_CU_MEMHOSTREGISTER_DEVICEMAP), "Pin object")) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
484 return; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
485 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
486 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
487 // Record original oop so that its memory can be unpinned |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
488 _pinned[_pinned_length++] = obj; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
489 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
490 // Replace host pointer to object with device pointer |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
491 // to object in kernel parameters buffer |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
492 if (!check(gpu::Ptx::_cuda_cu_mem_host_get_device_pointer((gpu::Ptx::CUdeviceptr*) argPtr, obj, 0), "Get device pointer for pinned object")) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
493 return; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
494 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
495 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
496 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
497 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
498 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
499 void launch(address kernel, jint dimX, jint dimY, jint dimZ) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
500 // grid dimensionality |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
501 unsigned int gridX = 1; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
502 unsigned int gridY = 1; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
503 unsigned int gridZ = 1; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
504 void * config[] = { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
505 GRAAL_CU_LAUNCH_PARAM_BUFFER_POINTER, (char*) (address) _buffer, |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
506 GRAAL_CU_LAUNCH_PARAM_BUFFER_SIZE, &_buffer_size, |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
507 GRAAL_CU_LAUNCH_PARAM_END |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
508 }; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
509 if (check(gpu::Ptx::_cuda_cu_launch_kernel((struct CUfunc_st*) (address) kernel, |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
510 gridX, gridY, gridZ, |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
511 dimX, dimY, dimZ, |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
512 0, NULL, NULL, (void**) &config), "Launch kernel")) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
513 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
514 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
515 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
516 void synchronize() { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
517 check(gpu::Ptx::_cuda_cu_ctx_synchronize(), "Synchronize kernel"); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
518 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
519 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
520 void unpin_objects() { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
521 while (_pinned_length > 0) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
522 oop obj = _pinned[--_pinned_length]; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
523 assert(obj != NULL, "npe"); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
524 //tty->print_cr("Unpinning object %d: %p", _pinned_length, obj); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
525 if (!check(gpu::Ptx::_cuda_cu_mem_host_unregister(obj), "Unpin object")) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
526 return; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
527 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
528 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
529 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
530 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
531 oop get_object_return_value() { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
532 oop return_val; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
533 check(gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, _ret_value, T_OBJECT_BYTE_SIZE), "Copy return value from device"); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
534 return return_val; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
535 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
536 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
537 jlong get_primitive_return_value() { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
538 jlong return_val; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
539 check(gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, _ret_value, _ret_type_size), "Copy return value from device"); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
540 return return_val; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
541 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
542 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
543 void free_return_value() { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
544 if (_ret_value != 0) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
545 check(gpu::Ptx::_cuda_cu_memfree(_ret_value), "Free device memory"); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
546 _ret_value = 0; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
547 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
548 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
549 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
550 void destroy_context() { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
551 if (gpu::Ptx::_device_context != NULL) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
552 check(gpu::Ptx::_cuda_cu_ctx_destroy(gpu::Ptx::_device_context), "Destroy context"); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
553 gpu::Ptx::_device_context = NULL; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
554 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
555 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
556 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
557 ~PtxCall() { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
558 unpin_objects(); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
559 free_return_value(); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
560 destroy_context(); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
561 } |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
562 }; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
563 |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
564 GPU_VMENTRY(jlong, gpu::Ptx::get_execute_kernel_from_vm_address, (JNIEnv *env, jclass)) |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
565 return (jlong) gpu::Ptx::execute_kernel_from_vm; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
566 GPU_END |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
567 |
13648
27acedac70b7
added support for @ParallelOver annotation in PTX kernel call wrapper
Doug Simon <doug.simon@oracle.com>
parents:
13624
diff
changeset
|
568 JRT_ENTRY(jlong, gpu::Ptx::execute_kernel_from_vm(JavaThread* thread, jlong kernel, jint dimX, jint dimY, jint dimZ, |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
569 jlong buffer, |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
570 jint bufferSize, |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
571 jint objectParametersCount, |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
572 jlong objectParametersOffsets, |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
573 jlong pinnedObjects, |
13648
27acedac70b7
added support for @ParallelOver annotation in PTX kernel call wrapper
Doug Simon <doug.simon@oracle.com>
parents:
13624
diff
changeset
|
574 int encodedReturnTypeSize)) |
13624
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
575 if (kernel == 0L) { |
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
576 SharedRuntime::throw_and_post_jvmti_exception(thread, vmSymbols::java_lang_NullPointerException(), NULL); |
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
577 return 0L; |
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
578 } |
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
579 |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
580 #if 0 |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
581 Universe::heap()->collect(GCCause::_jvmti_force_gc); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
582 #endif |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
583 |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
584 PtxCall call(thread, (address) buffer, bufferSize, (oop*) (address) pinnedObjects, encodedReturnTypeSize); |
13624
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
585 |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
586 #define TRY(action) do { \ |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
587 action; \ |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
588 if (HAS_PENDING_EXCEPTION) return 0L; \ |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
589 } while (0) |
13624
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
590 |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
591 TRY(call.alloc_return_value()); |
13624
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
592 |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
593 TRY(call.pin_objects(objectParametersCount, (int*) (address) objectParametersOffsets)); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
594 |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
595 TRY(call.launch((address) kernel, dimX, dimY, dimZ)); |
13624
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
596 |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
597 TRY(call.synchronize()); |
13624
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
598 |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
599 if (call.is_object_return()) { |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
600 oop return_val; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
601 TRY(return_val = call.get_object_return_value()); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
602 thread->set_vm_result(return_val); |
13624
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
603 return 0L; |
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
604 } |
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
605 |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
606 jlong return_val; |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
607 TRY(return_val = call.get_primitive_return_value()); |
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
608 return return_val; |
13624
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
609 |
13753
80cd5c3b8827
partially fixed passing of object parameters to PTX kernels; use a C++ object for managing resource allocation and cleanup around a PTX kernel execution
Doug Simon <doug.simon@oracle.com>
parents:
13683
diff
changeset
|
610 #undef TRY |
13624
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
611 |
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
612 JRT_END |
220ed109bf77
initial code for calling PTX kernel code from Java with parameter marshaling and return value unmarshaling performed by a wrapper specified via manual graph construction
Doug Simon <doug.simon@oracle.com>
parents:
12653
diff
changeset
|
613 |
10879 | 614 #if defined(LINUX) |
615 static const char cuda_library_name[] = "libcuda.so"; | |
616 #elif defined(__APPLE__) | |
617 static char const cuda_library_name[] = "/usr/local/cuda/lib/libcuda.dylib"; | |
618 #else | |
619 static char const cuda_library_name[] = ""; | |
620 #endif | |
621 | |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
622 bool gpu::Ptx::link() { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
623 if (cuda_library_name == NULL) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
624 if (TraceGPUInteraction) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
625 tty->print_cr("Failed to find CUDA linkage"); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
626 } |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
627 return false; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
628 } |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
629 char ebuf[O_BUFLEN]; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
630 void *handle = os::dll_load(cuda_library_name, ebuf, O_BUFLEN); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
631 if (handle == NULL) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
632 if (TraceGPUInteraction) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
633 tty->print_cr("Unsupported CUDA platform: %s", ebuf); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
634 } |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
635 return false; |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
636 } |
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
637 |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
638 LOOKUP_CUDA_FUNCTION(cuInit, cuda_cu_init); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
639 LOOKUP_CUDA_FUNCTION(cuCtxSynchronize, cuda_cu_ctx_synchronize); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
640 LOOKUP_CUDA_FUNCTION(cuCtxGetCurrent, cuda_cu_ctx_get_current); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
641 LOOKUP_CUDA_FUNCTION(cuCtxSetCurrent, cuda_cu_ctx_set_current); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
642 LOOKUP_CUDA_FUNCTION(cuDeviceGetCount, cuda_cu_device_get_count); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
643 LOOKUP_CUDA_FUNCTION(cuDeviceGetName, cuda_cu_device_get_name); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
644 LOOKUP_CUDA_FUNCTION(cuDeviceGet, cuda_cu_device_get); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
645 LOOKUP_CUDA_FUNCTION(cuDeviceComputeCapability, cuda_cu_device_compute_capability); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
646 LOOKUP_CUDA_FUNCTION(cuDeviceGetAttribute, cuda_cu_device_get_attribute); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
647 LOOKUP_CUDA_FUNCTION(cuModuleGetFunction, cuda_cu_module_get_function); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
648 LOOKUP_CUDA_FUNCTION(cuModuleLoadDataEx, cuda_cu_module_load_data_ex); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
649 LOOKUP_CUDA_FUNCTION(cuLaunchKernel, cuda_cu_launch_kernel); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
650 LOOKUP_CUDA_FUNCTION(cuMemHostRegister, cuda_cu_mem_host_register); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
651 LOOKUP_CUDA_FUNCTION(cuMemHostUnregister, cuda_cu_mem_host_unregister); |
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
|
652 #if defined(__x86_64) || defined(AMD64) || defined(_M_AMD64) |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
653 LOOKUP_CUDA_V2_FUNCTION(cuCtxCreate, cuda_cu_ctx_create); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
654 LOOKUP_CUDA_V2_FUNCTION(cuCtxDestroy, cuda_cu_ctx_destroy); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
655 LOOKUP_CUDA_V2_FUNCTION(cuMemAlloc, cuda_cu_memalloc); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
656 LOOKUP_CUDA_V2_FUNCTION(cuMemFree, cuda_cu_memfree); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
657 LOOKUP_CUDA_V2_FUNCTION(cuMemcpyHtoD, cuda_cu_memcpy_htod); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
658 LOOKUP_CUDA_V2_FUNCTION(cuMemcpyDtoH, cuda_cu_memcpy_dtoh); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
659 LOOKUP_CUDA_V2_FUNCTION(cuMemHostGetDevicePointer, cuda_cu_mem_host_get_device_pointer); |
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
|
660 #else |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
661 LOOKUP_CUDA_FUNCTION(cuCtxCreate, cuda_cu_ctx_create); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
662 LOOKUP_CUDA_FUNCTION(cuCtxDestroy, cuda_cu_ctx_destroy); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
663 LOOKUP_CUDA_FUNCTION(cuMemAlloc, cuda_cu_memalloc); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
664 LOOKUP_CUDA_FUNCTION(cuMemFree, cuda_cu_memfree); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
665 LOOKUP_CUDA_FUNCTION(cuMemcpyHtoD, cuda_cu_memcpy_htod); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
666 LOOKUP_CUDA_FUNCTION(cuMemcpyDtoH, cuda_cu_memcpy_dtoh); |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
667 LOOKUP_CUDA_FUNCTION(cuMemHostGetDevicePointer, cuda_cu_mem_host_get_device_pointer); |
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
|
668 #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
|
669 |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
670 if (TraceGPUInteraction) { |
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
671 tty->print_cr("[CUDA] Success: library linkage"); |
10879 | 672 } |
13819
49db2c1e3bee
added support for co-existing GPU backends (JBS:GRAAL-1)
Doug Simon <doug.simon@oracle.com>
parents:
13753
diff
changeset
|
673 return true; |
10879 | 674 } |