Mercurial > hg > truffle
annotate src/gpu/ptx/vm/gpu_ptx.cpp @ 11485:49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
| author | bharadwaj |
|---|---|
| date | Fri, 30 Aug 2013 16:39:05 -0400 |
| parents | d876002b98e6 |
| children | c99e65785936 |
| 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" |
|
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
|
32 #include "kernelArguments.hpp" |
| 10879 | 33 |
| 34 void * gpu::Ptx::_device_context; | |
|
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
35 int gpu::Ptx::_cu_device = 0; |
| 10879 | 36 |
| 37 gpu::Ptx::cuda_cu_init_func_t gpu::Ptx::_cuda_cu_init; | |
| 38 gpu::Ptx::cuda_cu_ctx_create_func_t gpu::Ptx::_cuda_cu_ctx_create; | |
|
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
39 gpu::Ptx::cuda_cu_ctx_destroy_func_t gpu::Ptx::_cuda_cu_ctx_destroy; |
| 10879 | 40 gpu::Ptx::cuda_cu_ctx_synchronize_func_t gpu::Ptx::_cuda_cu_ctx_synchronize; |
| 41 gpu::Ptx::cuda_cu_device_get_count_func_t gpu::Ptx::_cuda_cu_device_get_count; | |
| 42 gpu::Ptx::cuda_cu_device_get_name_func_t gpu::Ptx::_cuda_cu_device_get_name; | |
| 43 gpu::Ptx::cuda_cu_device_get_func_t gpu::Ptx::_cuda_cu_device_get; | |
| 44 gpu::Ptx::cuda_cu_device_compute_capability_func_t gpu::Ptx::_cuda_cu_device_compute_capability; | |
| 45 gpu::Ptx::cuda_cu_device_get_attribute_func_t gpu::Ptx::_cuda_cu_device_get_attribute; | |
| 46 gpu::Ptx::cuda_cu_launch_kernel_func_t gpu::Ptx::_cuda_cu_launch_kernel; | |
| 47 gpu::Ptx::cuda_cu_module_get_function_func_t gpu::Ptx::_cuda_cu_module_get_function; | |
| 48 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
|
49 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
|
50 gpu::Ptx::cuda_cu_memfree_func_t gpu::Ptx::_cuda_cu_memfree; |
| 10879 | 51 |
| 52 void gpu::probe_linkage() { | |
| 53 #if defined(__APPLE__) || defined(LINUX) | |
| 54 set_gpu_linkage(gpu::Ptx::probe_linkage()); | |
| 55 #else | |
| 56 set_gpu_linkage(false); | |
| 57 #endif | |
| 58 } | |
| 59 | |
| 60 void gpu::initialize_gpu() { | |
| 61 if (gpu::has_gpu_linkage()) { | |
| 62 set_initialized(gpu::Ptx::initialize_gpu()); | |
| 63 } | |
| 64 } | |
| 65 | |
| 66 void * gpu::generate_kernel(unsigned char *code, int code_len, const char *name) { | |
| 67 if (gpu::has_gpu_linkage()) { | |
| 68 return (gpu::Ptx::generate_kernel(code, code_len, name)); | |
| 69 } else { | |
| 70 return NULL; | |
| 71 } | |
| 72 } | |
| 73 | |
|
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
|
74 bool gpu::execute_kernel(address kernel, PTXKernelArguments & ptxka, JavaValue& ret) { |
| 10879 | 75 if (gpu::has_gpu_linkage()) { |
|
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
|
76 return (gpu::Ptx::execute_kernel(kernel, ptxka, ret)); |
| 10879 | 77 } else { |
| 78 return false; | |
| 79 } | |
| 80 } | |
| 81 | |
| 82 bool gpu::Ptx::initialize_gpu() { | |
| 83 | |
| 84 /* Initialize CUDA driver API */ | |
| 85 int status = _cuda_cu_init(0); | |
| 86 if (status != GRAAL_CUDA_SUCCESS) { | |
| 87 tty->print_cr("Failed to initialize CUDA device"); | |
| 88 return false; | |
| 89 } | |
| 90 | |
| 91 if (TraceGPUInteraction) { | |
| 92 tty->print_cr("CUDA driver initialization: Success"); | |
| 93 } | |
| 94 | |
| 95 /* Get the number of compute-capable device count */ | |
| 96 int device_count = 0; | |
| 97 status = _cuda_cu_device_get_count(&device_count); | |
| 98 if (status != GRAAL_CUDA_SUCCESS) { | |
| 99 tty->print_cr("[CUDA] Failed to get compute-capable device count"); | |
| 100 return false; | |
| 101 } | |
| 102 | |
| 103 if (device_count == 0) { | |
| 104 tty->print_cr("[CUDA] Found no device supporting CUDA"); | |
| 105 return false; | |
| 106 } | |
| 107 | |
| 108 if (TraceGPUInteraction) { | |
| 109 tty->print_cr("[CUDA] Number of compute-capable devices found: %d", device_count); | |
| 110 } | |
| 111 | |
| 112 /* Get the handle to the first compute device */ | |
| 113 int device_id = 0; | |
| 114 /* 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
|
115 status = _cuda_cu_device_get(&_cu_device, device_id); |
| 10879 | 116 |
| 117 if (status != GRAAL_CUDA_SUCCESS) { | |
| 118 tty->print_cr("[CUDA] Failed to get handle of first compute-capable device i.e., the one at ordinal: %d", device_id); | |
| 119 return false; | |
| 120 } | |
| 121 | |
| 122 if (TraceGPUInteraction) { | |
| 123 tty->print_cr("[CUDA] Got the handle of first compute-device"); | |
| 124 } | |
| 125 | |
| 126 /* Get device attributes */ | |
|
11283
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
127 int minor, major, unified_addressing; |
|
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
|
128 status = _cuda_cu_device_get_attribute(&minor, GRAAL_CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, _cu_device); |
| 10879 | 129 |
| 130 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
|
131 tty->print_cr("[CUDA] Failed to get minor attribute of device: %d", _cu_device); |
| 10879 | 132 return false; |
| 133 } | |
| 134 | |
|
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
|
135 status = _cuda_cu_device_get_attribute(&major, GRAAL_CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, _cu_device); |
| 10879 | 136 |
| 137 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
|
138 tty->print_cr("[CUDA] Failed to get major attribute of device: %d", _cu_device); |
| 10879 | 139 return false; |
| 140 } | |
| 141 | |
| 142 if (TraceGPUInteraction) { | |
|
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
143 tty->print_cr("[CUDA] Compatibility version of device %d: %d.%d", _cu_device, major, minor); |
| 10879 | 144 } |
| 145 | |
|
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
|
146 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
|
147 |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
148 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
|
149 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
|
150 return false; |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
151 } |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
152 |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
153 if (TraceGPUInteraction) { |
|
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
154 tty->print_cr("[CUDA] Unified addressing support on device %d: %d", _cu_device, unified_addressing); |
|
11283
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
155 } |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
156 |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
157 |
| 10879 | 158 /* Get device name */ |
| 159 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
|
160 status = _cuda_cu_device_get_name(device_name, 256, _cu_device); |
| 10879 | 161 |
| 162 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
|
163 tty->print_cr("[CUDA] Failed to get name of device: %d", _cu_device); |
| 10879 | 164 return false; |
| 165 } | |
| 166 | |
| 167 if (TraceGPUInteraction) { | |
| 168 tty->print_cr("[CUDA] Using %s", device_name); | |
| 169 } | |
| 170 | |
| 171 return true; | |
| 172 } | |
| 173 | |
| 174 void *gpu::Ptx::generate_kernel(unsigned char *code, int code_len, const char *name) { | |
| 175 | |
| 176 struct CUmod_st * cu_module; | |
| 177 // Use three JIT compiler options | |
| 178 const unsigned int jit_num_options = 3; | |
| 179 int *jit_options = NEW_C_HEAP_ARRAY(int, jit_num_options, mtCompiler); | |
| 180 void **jit_option_values = NEW_C_HEAP_ARRAY(void *, jit_num_options, mtCompiler); | |
| 181 | |
| 182 // Set up PTX JIT compiler options | |
| 183 // 1. set size of compilation log buffer | |
| 184 int jit_log_buffer_size = 1024; | |
| 185 jit_options[0] = GRAAL_CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES; | |
| 186 jit_option_values[0] = (void *)(size_t)jit_log_buffer_size; | |
| 187 | |
| 188 // 2. set pointer to compilation log buffer | |
| 189 char *jit_log_buffer = NEW_C_HEAP_ARRAY(char, jit_log_buffer_size, mtCompiler); | |
| 190 jit_options[1] = GRAAL_CU_JIT_INFO_LOG_BUFFER; | |
| 191 jit_option_values[1] = jit_log_buffer; | |
| 192 | |
| 193 // 3. set pointer to set the Maximum # of registers (32) for the kernel | |
| 194 int jit_register_count = 32; | |
| 195 jit_options[2] = GRAAL_CU_JIT_MAX_REGISTERS; | |
| 196 jit_option_values[2] = (void *)(size_t)jit_register_count; | |
| 197 | |
| 198 if (TraceGPUInteraction) { | |
| 199 tty->print_cr("[CUDA] PTX Kernel\n%s", code); | |
| 200 tty->print_cr("[CUDA] Function name : %s", name); | |
| 201 | |
| 202 } | |
| 203 | |
|
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
|
204 /* Create CUDA context to compile and execute the kernel */ |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
205 int status = _cuda_cu_ctx_create(&_device_context, 0, _cu_device); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
206 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
207 if (status != GRAAL_CUDA_SUCCESS) { |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
208 tty->print_cr("[CUDA] Failed to create CUDA 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
|
209 return NULL; |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
210 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
211 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
212 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
|
213 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
|
214 } |
|
49bb1bc983c6
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 |
| 10879 | 216 /* 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
|
217 status = _cuda_cu_module_load_data_ex(&cu_module, (void*) code, jit_num_options, |
| 10879 | 218 jit_options, (void **)jit_option_values); |
| 219 if (status != GRAAL_CUDA_SUCCESS) { | |
| 220 if (status == GRAAL_CUDA_ERROR_NO_BINARY_FOR_GPU) { | |
| 221 tty->print_cr("[CUDA] Check for malformed PTX kernel or incorrect PTX compilation options"); | |
| 222 } | |
| 223 tty->print_cr("[CUDA] *** Error (%d) Failed to load module data with online compiler options for method %s", | |
| 224 status, name); | |
| 225 return NULL; | |
| 226 } | |
| 227 | |
| 228 if (TraceGPUInteraction) { | |
| 229 tty->print_cr("[CUDA] Loaded data for PTX Kernel"); | |
| 230 } | |
| 231 | |
| 232 struct CUfunc_st * cu_function; | |
| 233 | |
| 234 status = _cuda_cu_module_get_function(&cu_function, cu_module, name); | |
| 235 | |
| 236 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
|
237 tty->print_cr("[CUDA] *** Error: Failed to get function %s", name); |
| 10879 | 238 return NULL; |
| 239 } | |
| 240 | |
| 241 if (TraceGPUInteraction) { | |
| 242 tty->print_cr("[CUDA] Got function handle for %s", name); | |
| 243 } | |
|
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
|
244 |
| 10879 | 245 return cu_function; |
| 246 } | |
| 247 | |
|
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
|
248 bool gpu::Ptx::execute_kernel(address kernel, PTXKernelArguments &ptxka, JavaValue &ret) { |
| 10879 | 249 // grid dimensionality |
| 250 unsigned int gridX = 1; | |
| 251 unsigned int gridY = 1; | |
| 252 unsigned int gridZ = 1; | |
| 253 | |
| 254 // thread dimensionality | |
| 255 unsigned int blockX = 1; | |
| 256 unsigned int blockY = 1; | |
| 257 unsigned int blockZ = 1; | |
| 258 | |
|
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
|
259 struct CUfunc_st* cu_function = (struct CUfunc_st*) kernel; |
| 10879 | 260 |
|
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
|
261 void * config[5] = { |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
262 GRAAL_CU_LAUNCH_PARAM_BUFFER_POINTER, ptxka._kernelArgBuffer, |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
263 GRAAL_CU_LAUNCH_PARAM_BUFFER_SIZE, &(ptxka._bufferOffset), |
|
11283
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
264 GRAAL_CU_LAUNCH_PARAM_END |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
265 }; |
|
1cd1f8ff70a1
CR-20: PTX kernel invocation with arguments - from Bharadwaj
Morris Meyer <morris.meyer@oracle.com>
parents:
10884
diff
changeset
|
266 |
| 10879 | 267 if (kernel == NULL) { |
| 268 return false; | |
| 269 } | |
| 270 | |
| 271 if (TraceGPUInteraction) { | |
| 272 tty->print_cr("[CUDA] launching kernel"); | |
| 273 } | |
|
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
|
274 |
| 10879 | 275 int status = _cuda_cu_launch_kernel(cu_function, |
| 276 gridX, gridY, gridZ, | |
| 277 blockX, blockY, blockZ, | |
|
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
278 0, NULL, NULL, (void **) &config); |
| 10879 | 279 if (status != GRAAL_CUDA_SUCCESS) { |
| 280 tty->print_cr("[CUDA] Failed to launch kernel"); | |
| 281 return false; | |
| 282 } | |
| 283 | |
| 284 if (TraceGPUInteraction) { | |
| 285 tty->print_cr("[CUDA] Success: Kernel Launch"); | |
| 286 } | |
|
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
|
287 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
288 status = _cuda_cu_ctx_synchronize(); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
289 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
290 if (status != GRAAL_CUDA_SUCCESS) { |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
291 tty->print_cr("[CUDA] Failed to synchronize launched kernel (%d)", status); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
292 return false; |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
293 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
294 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
295 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
|
296 tty->print_cr("[CUDA] Success: Synchronized launch kernel"); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
297 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
298 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
299 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
300 // Get the result. TODO: Move this code to get_return_oop() |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
301 BasicType return_type = ptxka.get_ret_type(); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
302 switch (return_type) { |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
303 case T_INT : |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
304 { |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
305 int return_val; |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
306 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._return_value_ptr, T_INT_BYTE_SIZE); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
307 if (status != GRAAL_CUDA_SUCCESS) { |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
308 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
309 return false; |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
310 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
311 ret.set_jint(return_val); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
312 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
313 break; |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
314 case T_LONG : |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
315 { |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
316 long return_val; |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
317 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._return_value_ptr, T_LONG_BYTE_SIZE); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
318 if (status != GRAAL_CUDA_SUCCESS) { |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
319 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
320 return false; |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
321 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
322 ret.set_jlong(return_val); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
323 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
324 break; |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
325 default: |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
326 tty->print_cr("[CUDA] TODO *** Unhandled return type"); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
327 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
328 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
329 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
330 // Free device memory allocated for result |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
331 status = gpu::Ptx::_cuda_cu_memfree(ptxka._return_value_ptr); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
332 if (status != GRAAL_CUDA_SUCCESS) { |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
333 tty->print_cr("[CUDA] *** Error (%d) Failed to free device memory of return value", status); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
334 return false; |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
335 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
336 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
337 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
|
338 tty->print_cr("[CUDA] Success: Freed device memory of return value"); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
339 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
340 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
341 // Destroy context |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
342 status = gpu::Ptx::_cuda_cu_ctx_destroy(_device_context); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
343 if (status != GRAAL_CUDA_SUCCESS) { |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
344 tty->print_cr("[CUDA] *** Error (%d) Failed to destroy context", status); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
345 return false; |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
346 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
347 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
348 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
|
349 tty->print_cr("[CUDA] Success: Destroy context"); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
350 } |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
351 |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
352 return (status == GRAAL_CUDA_SUCCESS); |
| 10879 | 353 } |
| 354 | |
| 355 #if defined(LINUX) | |
| 356 static const char cuda_library_name[] = "libcuda.so"; | |
| 357 #elif defined(__APPLE__) | |
| 358 static char const cuda_library_name[] = "/usr/local/cuda/lib/libcuda.dylib"; | |
| 359 #else | |
| 360 static char const cuda_library_name[] = ""; | |
| 361 #endif | |
| 362 | |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
363 #define STD_BUFFER_SIZE 1024 |
|
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
364 |
| 10879 | 365 bool gpu::Ptx::probe_linkage() { |
| 366 if (cuda_library_name != NULL) { | |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
367 char *buffer = (char*)malloc(STD_BUFFER_SIZE); |
|
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
368 void *handle = os::dll_load(cuda_library_name, buffer, STD_BUFFER_SIZE); |
|
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
369 free(buffer); |
| 10879 | 370 if (handle != NULL) { |
| 371 _cuda_cu_init = | |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
372 CAST_TO_FN_PTR(cuda_cu_init_func_t, os::dll_lookup(handle, "cuInit")); |
| 10879 | 373 _cuda_cu_ctx_create = |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
374 CAST_TO_FN_PTR(cuda_cu_ctx_create_func_t, os::dll_lookup(handle, "cuCtxCreate")); |
|
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
375 _cuda_cu_ctx_destroy = |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
376 CAST_TO_FN_PTR(cuda_cu_ctx_destroy_func_t, os::dll_lookup(handle, "cuCtxDestroy")); |
| 10879 | 377 _cuda_cu_ctx_synchronize = |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
378 CAST_TO_FN_PTR(cuda_cu_ctx_synchronize_func_t, os::dll_lookup(handle, "cuCtxSynchronize")); |
| 10879 | 379 _cuda_cu_device_get_count = |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
380 CAST_TO_FN_PTR(cuda_cu_device_get_count_func_t, os::dll_lookup(handle, "cuDeviceGetCount")); |
| 10879 | 381 _cuda_cu_device_get_name = |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
382 CAST_TO_FN_PTR(cuda_cu_device_get_name_func_t, os::dll_lookup(handle, "cuDeviceGetName")); |
| 10879 | 383 _cuda_cu_device_get = |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
384 CAST_TO_FN_PTR(cuda_cu_device_get_func_t, os::dll_lookup(handle, "cuDeviceGet")); |
| 10879 | 385 _cuda_cu_device_compute_capability = |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
386 CAST_TO_FN_PTR(cuda_cu_device_compute_capability_func_t, os::dll_lookup(handle, "cuDeviceComputeCapability")); |
| 10879 | 387 _cuda_cu_device_get_attribute = |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
388 CAST_TO_FN_PTR(cuda_cu_device_get_attribute_func_t, os::dll_lookup(handle, "cuDeviceGetAttribute")); |
| 10879 | 389 _cuda_cu_module_get_function = |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
390 CAST_TO_FN_PTR(cuda_cu_module_get_function_func_t, os::dll_lookup(handle, "cuModuleGetFunction")); |
| 10879 | 391 _cuda_cu_module_load_data_ex = |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
392 CAST_TO_FN_PTR(cuda_cu_module_load_data_ex_func_t, os::dll_lookup(handle, "cuModuleLoadDataEx")); |
| 10879 | 393 _cuda_cu_launch_kernel = |
|
10882
6a2d65cb5d7d
fix: gpu_ptx.cpp did not compile on windows
Christian Wirth <christian.wirth@oracle.com>
parents:
10879
diff
changeset
|
394 CAST_TO_FN_PTR(cuda_cu_launch_kernel_func_t, os::dll_lookup(handle, "cuLaunchKernel")); |
|
11485
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
395 _cuda_cu_memalloc = |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
396 CAST_TO_FN_PTR(cuda_cu_memalloc_func_t, os::dll_lookup(handle, "cuMemAlloc")); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
397 _cuda_cu_memfree = |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
398 CAST_TO_FN_PTR(cuda_cu_memfree_func_t, os::dll_lookup(handle, "cuMemFree")); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
399 _cuda_cu_memcpy_htod = |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
400 CAST_TO_FN_PTR(cuda_cu_memcpy_htod_func_t, os::dll_lookup(handle, "cuMemcpyHtoD")); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
401 _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
|
402 CAST_TO_FN_PTR(cuda_cu_memcpy_dtoh_func_t, os::dll_lookup(handle, "cuMemcpyDtoH")); |
|
49bb1bc983c6
Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU.
bharadwaj
parents:
11284
diff
changeset
|
403 |
| 10879 | 404 if (TraceGPUInteraction) { |
| 405 tty->print_cr("[CUDA] Success: library linkage"); | |
| 406 } | |
| 407 return true; | |
| 408 } else { | |
| 409 // Unable to dlopen libcuda | |
| 410 return false; | |
| 411 } | |
| 412 } else { | |
| 413 tty->print_cr("Unsupported CUDA platform"); | |
| 414 return false; | |
| 415 } | |
| 416 tty->print_cr("Failed to find CUDA linkage"); | |
| 417 return false; | |
| 418 } | |
| 419 |