truffle: src/gpu/ptx/vm/gpu

annotate src/gpu/ptx/vm/gpu_ptx.cpp @ 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

author	Doug Simon <doug.simon@oracle.com>
date	Mon, 13 Jan 2014 22:32:27 +0100
parents	1a7e7011a341
children	27acedac70b7

rev	line source
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	1 /*
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	2 * Copyright (c) 2013, Oracle and/or its affiliates. All rights reserved.
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	4 *
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	5 * This code is free software; you can redistribute it and/or modify it
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	6 * under the terms of the GNU General Public License version 2 only, as
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	7 * published by the Free Software Foundation.
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	8 *
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	9 * This code is distributed in the hope that it will be useful, but WITHOUT
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	12 * version 2 for more details (a copy is included in the LICENSE file that
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	13 * accompanied this code).
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	14 *
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	15 * You should have received a copy of the GNU General Public License version
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	16 * 2 along with this work; if not, write to the Free Software Foundation,
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	18 *
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	20 * or visit www.oracle.com if you need additional information or have any
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	21 * questions.
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	22 *
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	23 */
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	24
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	25 #include "precompiled.hpp"
11284 d876002b98e6 ptx: fix include for non-precompiledheader build in gpu_ptx.cpp Bernhard Urban <bernhard.urban@jku.at> parents: 11283 diff changeset	26 #include "runtime/javaCalls.hpp"
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	27 #include "runtime/gpu.hpp"
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	28 #include "utilities/globalDefinitions.hpp"
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	29 #include "utilities/ostream.hpp"
10883 5fcb30bcb90a Fix compilation without precompiled headers. Roland Schatz <roland.schatz@oracle.com> parents: 10882 diff changeset	30 #include "memory/allocation.hpp"
5fcb30bcb90a Fix compilation without precompiled headers. Roland Schatz <roland.schatz@oracle.com> parents: 10882 diff changeset	31 #include "memory/allocation.inline.hpp"
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"
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	33 #include "ptxKernelArguments.hpp"
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	34
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	35 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	36 int gpu::Ptx::_cu_device = 0;
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	37
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	38 gpu::Ptx::cuda_cu_init_func_t gpu::Ptx::_cuda_cu_init;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	39 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	40 gpu::Ptx::cuda_cu_ctx_destroy_func_t gpu::Ptx::_cuda_cu_ctx_destroy;
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	41 gpu::Ptx::cuda_cu_ctx_synchronize_func_t gpu::Ptx::_cuda_cu_ctx_synchronize;
11527 c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	42 gpu::Ptx::cuda_cu_ctx_set_current_func_t gpu::Ptx::_cuda_cu_ctx_set_current;
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	43 gpu::Ptx::cuda_cu_device_get_count_func_t gpu::Ptx::_cuda_cu_device_get_count;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	44 gpu::Ptx::cuda_cu_device_get_name_func_t gpu::Ptx::_cuda_cu_device_get_name;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	45 gpu::Ptx::cuda_cu_device_get_func_t gpu::Ptx::_cuda_cu_device_get;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	46 gpu::Ptx::cuda_cu_device_compute_capability_func_t gpu::Ptx::_cuda_cu_device_compute_capability;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	47 gpu::Ptx::cuda_cu_device_get_attribute_func_t gpu::Ptx::_cuda_cu_device_get_attribute;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	48 gpu::Ptx::cuda_cu_launch_kernel_func_t gpu::Ptx::_cuda_cu_launch_kernel;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	49 gpu::Ptx::cuda_cu_module_get_function_func_t gpu::Ptx::_cuda_cu_module_get_function;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	50 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	51 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	52 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	53 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	54 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	55 gpu::Ptx::cuda_cu_mem_host_unregister_func_t gpu::Ptx::_cuda_cu_mem_host_unregister;
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	56
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	57 #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	58
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	59 #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	60 _##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	61 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	62 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	63 tty->print_cr("[CUDA] ***** Error: Failed to lookup %s", STRINGIFY(name)); \
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	64 return 0; \
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	65 } \
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	66
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	67 #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	68
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	69 /*
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	70 * 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	71 */
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	72 int ncores(int major, int minor) {
11843 372bacc13022 gtx_ptx.cpp: fix switch default label and operator precedence in ncores(). Andreas Woess <andreas.woess@jku.at> parents: 11842 diff changeset	73 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	74
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	75 switch (device_type) {
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	76 case 0x10: return 8;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	77 case 0x11: return 8;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	78 case 0x12: return 8;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	79 case 0x13: return 8;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	80 case 0x20: return 32;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	81 case 0x21: return 48;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	82 case 0x30: return 192;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	83 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	84 default:
11842 8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	85 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	86 return 0;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	87 }
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	88 }
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	89
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	90 bool gpu::Ptx::initialize_gpu() {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	91
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	92 /* Initialize CUDA driver API */
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	93 int status = _cuda_cu_init(0);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	94 if (status != GRAAL_CUDA_SUCCESS) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	95 tty->print_cr("Failed to initialize CUDA device");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	96 return false;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	97 }
11527 c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	98
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	99 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	100 tty->print_cr("CUDA driver initialization: Success");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	101 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	102
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	103 /* Get the number of compute-capable device count */
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	104 int device_count = 0;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	105 status = _cuda_cu_device_get_count(&device_count);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	106 if (status != GRAAL_CUDA_SUCCESS) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	107 tty->print_cr("[CUDA] Failed to get compute-capable device count");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	108 return false;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	109 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	110
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	111 if (device_count == 0) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	112 tty->print_cr("[CUDA] Found no device supporting CUDA");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	113 return false;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	114 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	115
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	116 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	117 tty->print_cr("[CUDA] Number of compute-capable devices found: %d", device_count);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	118 }
11527 c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	119
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	120 /* Get the handle to the first compute device */
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	121 int device_id = 0;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	122 /* 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	123 status = _cuda_cu_device_get(&_cu_device, device_id);
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	124
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	125 if (status != GRAAL_CUDA_SUCCESS) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	126 tty->print_cr("[CUDA] Failed to get handle of first compute-capable device i.e., the one at ordinal: %d", device_id);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	127 return false;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	128 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	129
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	130 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	131 tty->print_cr("[CUDA] Got the handle of first compute-device");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	132 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	133
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	134 /* Get device attributes */
11842 8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	135 int unified_addressing;
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	136
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	137 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	138
1cd1f8ff70a1 CR-20: PTX kernel invocation with arguments - from Bharadwaj Morris Meyer <morris.meyer@oracle.com> parents: 10884 diff changeset	139 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	140 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	141 return false;
1cd1f8ff70a1 CR-20: PTX kernel invocation with arguments - from Bharadwaj Morris Meyer <morris.meyer@oracle.com> parents: 10884 diff changeset	142 }
1cd1f8ff70a1 CR-20: PTX kernel invocation with arguments - from Bharadwaj Morris Meyer <morris.meyer@oracle.com> parents: 10884 diff changeset	143
1cd1f8ff70a1 CR-20: PTX kernel invocation with arguments - from Bharadwaj Morris Meyer <morris.meyer@oracle.com> parents: 10884 diff changeset	144 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	145 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	146 }
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
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	149 /* Get device name */
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	150 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	151 status = _cuda_cu_device_get_name(device_name, 256, _cu_device);
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	152
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	153 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	154 tty->print_cr("[CUDA] Failed to get name of device: %d", _cu_device);
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	155 return false;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	156 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	157
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	158 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	159 tty->print_cr("[CUDA] Using %s", device_name);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	160 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	161
11842 8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	162
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	163 return true;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	164 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	165
11842 8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	166 unsigned int gpu::Ptx::total_cores() {
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	167
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	168 int minor, major, nmp;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	169 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	170 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	171 _cu_device);
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	172
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	173 if (status != GRAAL_CUDA_SUCCESS) {
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	174 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	175 return 0;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	176 }
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	177
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	178 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	179 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	180 _cu_device);
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	181
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	182 if (status != GRAAL_CUDA_SUCCESS) {
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	183 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	184 return 0;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	185 }
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	186
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	187 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	188 GRAAL_CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	189 _cu_device);
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	190
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	191 if (status != GRAAL_CUDA_SUCCESS) {
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	192 tty->print_cr("[CUDA] Failed to get numberof MPs on device: %d", _cu_device);
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	193 return 0;
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	194 }
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	195
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	196 int total = nmp * ncores(major, minor);
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	197
11894 c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	198 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	199
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	200 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	201 GRAAL_CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK,
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	202 _cu_device);
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	203
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	204 if (status != GRAAL_CUDA_SUCCESS) {
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	205 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	206 return 0;
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	207 }
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	208
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	209 status = _cuda_cu_device_get_attribute(&warp_size,
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	210 GRAAL_CU_DEVICE_ATTRIBUTE_WARP_SIZE,
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	211 _cu_device);
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	212
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	213 if (status != GRAAL_CUDA_SUCCESS) {
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	214 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	215 return 0;
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	216 }
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	217
11894 c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	218 status = _cuda_cu_device_get_attribute(&async_engines,
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	219 GRAAL_CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT,
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	220 _cu_device);
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	221
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	222 if (status != GRAAL_CUDA_SUCCESS) {
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	223 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	224 return 0;
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	225 }
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	226
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	227 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	228 GRAAL_CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY,
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	229 _cu_device);
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	230
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	231 if (status != GRAAL_CUDA_SUCCESS) {
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	232 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	233 return 0;
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	234 }
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	235
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	236 status = _cuda_cu_device_get_attribute(&concurrent_kernels,
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	237 GRAAL_CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS,
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	238 _cu_device);
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	239
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	240 if (status != GRAAL_CUDA_SUCCESS) {
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	241 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	242 return 0;
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	243 }
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	244
11842 8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	245 if (TraceGPUInteraction) {
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	246 tty->print_cr("[CUDA] Compatibility version of device %d: %d.%d", _cu_device, major, minor);
11894 c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	247 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	248 total, async_engines, can_map_host_memory, concurrent_kernels);
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	249 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	250 }
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	251 return (total);
12653 1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	252
11842 8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	253 }
8d8f63069f58 PTX warp limiter to available GPU processors Morris Meyer <morris.meyer@oracle.com> parents: 11822 diff changeset	254
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	255 void gpu::Ptx::generate_kernel(unsigned char code, int code_len, const char *name) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	256
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	257 struct CUmod_st * cu_module;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	258 // Use three JIT compiler options
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	259 const unsigned int jit_num_options = 3;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	260 int *jit_options = NEW_C_HEAP_ARRAY(int, jit_num_options, mtCompiler);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	261 void *jit_option_values = NEW_C_HEAP_ARRAY(void , jit_num_options, mtCompiler);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	262
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	263 // Set up PTX JIT compiler options
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	264 // 1. set size of compilation log buffer
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	265 int jit_log_buffer_size = 1024;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	266 jit_options[0] = GRAAL_CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	267 jit_option_values[0] = (void *)(size_t)jit_log_buffer_size;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	268
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	269 // 2. set pointer to compilation log buffer
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	270 char *jit_log_buffer = NEW_C_HEAP_ARRAY(char, jit_log_buffer_size, mtCompiler);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	271 jit_options[1] = GRAAL_CU_JIT_INFO_LOG_BUFFER;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	272 jit_option_values[1] = jit_log_buffer;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	273
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	274 // 3. set pointer to set the Maximum # of registers (32) for the kernel
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	275 int jit_register_count = 32;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	276 jit_options[2] = GRAAL_CU_JIT_MAX_REGISTERS;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	277 jit_option_values[2] = (void *)(size_t)jit_register_count;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	278
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	279 /* 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	280 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	281
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	282 if (status != GRAAL_CUDA_SUCCESS) {
11634 bff2b88444f5 Start of PTX array passing Morris Meyer <morris.meyer@oracle.com> parents: 11606 diff changeset	283 tty->print_cr("[CUDA] Failed to create CUDA context for device(%d): %d", _cu_device, status);
11485 49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	284 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	285 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	286
49bb1bc983c6 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 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	288 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	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
11527 c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	291 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	292
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	293 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	294 tty->print_cr("[CUDA] Failed to set current context for device: %d", _cu_device);
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	295 return NULL;
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	296 }
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	297
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	298 if (TraceGPUInteraction) {
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	299 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	300 }
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	301
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	302 if (TraceGPUInteraction) {
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	303 tty->print_cr("[CUDA] 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	304 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	305
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	306 }
c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	307
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	308 /* 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	309 status = _cuda_cu_module_load_data_ex(&cu_module, (void*) code, jit_num_options,
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	310 jit_options, (void **)jit_option_values);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	311 if (status != GRAAL_CUDA_SUCCESS) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	312 if (status == GRAAL_CUDA_ERROR_NO_BINARY_FOR_GPU) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	313 tty->print_cr("[CUDA] Check for malformed PTX kernel or incorrect PTX compilation options");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	314 }
11527 c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	315 tty->print_cr("[CUDA] *** Error (%d) Failed to load module data with online compiler options for method %s",
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	316 status, name);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	317 return NULL;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	318 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	319
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	320 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	321 tty->print_cr("[CUDA] Loaded data for PTX Kernel");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	322 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	323
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	324 struct CUfunc_st * cu_function;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	325
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	326 status = _cuda_cu_module_get_function(&cu_function, cu_module, name);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	327
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	328 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	329 tty->print_cr("[CUDA] *** Error: Failed to get function %s", name);
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	330 return NULL;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	331 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	332
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	333 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	334 tty->print_cr("[CUDA] Got function handle for %s", name);
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	335 }
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	336
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	337 return cu_function;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	338 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	339
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	340 JRT_ENTRY(jlong, gpu::Ptx::execute_kernel_from_vm(JavaThread* thread, jlong kernel, jlong parametersAndReturnValueBuffer, jint parametersAndReturnValueBufferSize, int encodedReturnTypeSize))
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	341 tty->print_cr("*** gpu::Ptx::execute_kernel_from_vm(kernel=%p, parametersAndReturnValueBuffer=%p, parametersAndReturnValueBufferSize=%d, encodedReturnTypeSize=%d)",
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	342 kernel, parametersAndReturnValueBuffer, parametersAndReturnValueBufferSize, encodedReturnTypeSize);
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	343 tty->print(" buffer as bytes: ");
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	344 for (int i = 0; i < parametersAndReturnValueBufferSize; i++) {
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	345 tty->print(" 0x%02x", ((jbyte*) (address) parametersAndReturnValueBuffer)[i] & 0xFF);
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	346 }
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	347 tty->cr();
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	348 tty->print(" buffer as ints: ");
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	349 for (int i = 0; i < (parametersAndReturnValueBufferSize / 4); i++) {
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	350 tty->print(" %d", ((jint*) (address) parametersAndReturnValueBuffer)[i]);
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	351 }
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	352 tty->cr();
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	353 tty->print(" buffer as words: ");
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	354 for (unsigned i = 0; i < (parametersAndReturnValueBufferSize / sizeof(void*)); i++) {
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	355 tty->print(" "INTPTR_FORMAT, ((void**) (address) parametersAndReturnValueBuffer)[i]);
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	356 }
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	357 tty->cr();
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	358 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	359 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	360 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	361 }
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	362
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	363
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	364 // grid dimensionality
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	365 unsigned int gridX = 1;
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	366 unsigned int gridY = 1;
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	367 unsigned int gridZ = 1;
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	368
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	369 // thread dimensionality
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	370 unsigned int blockX = 1;
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	371 unsigned int blockY = 1;
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	372 unsigned int blockZ = 1;
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	373
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	374 struct CUfunc_st* cu_function = (struct CUfunc_st*) (address) kernel;
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	375
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	376 void * config[5] = {
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	377 GRAAL_CU_LAUNCH_PARAM_BUFFER_POINTER, (char*) (address) parametersAndReturnValueBuffer,
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	378 GRAAL_CU_LAUNCH_PARAM_BUFFER_SIZE, &parametersAndReturnValueBufferSize,
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	379 GRAAL_CU_LAUNCH_PARAM_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	380 };
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	381
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	382 if (TraceGPUInteraction) {
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	383 tty->print_cr("[CUDA] launching kernel");
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	384 }
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	385
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	386 bool isObjectReturn = encodedReturnTypeSize < 0;
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	387 int returnTypeSize = encodedReturnTypeSize < 0 ? -encodedReturnTypeSize : encodedReturnTypeSize;
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	388 gpu::Ptx::CUdeviceptr device_return_value;
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	389 int status;
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	390 if (returnTypeSize != 0) {
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	391 status = _cuda_cu_memalloc(&device_return_value, returnTypeSize);
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	392 if (status != GRAAL_CUDA_SUCCESS) {
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	393 tty->print_cr("[CUDA] *** Error (%d) Failed to allocate memory for return value pointer on device", status);
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	394 SharedRuntime::throw_and_post_jvmti_exception(thread, vmSymbols::java_lang_Exception(), "[CUDA] Failed to allocate memory for return value pointer on device");
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	395 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	396 }
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	397 // Push device_return_value to kernelParams
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	398 gpu::Ptx::CUdeviceptr* returnValuePtr = (gpu::Ptx::CUdeviceptr*) (address) parametersAndReturnValueBuffer + parametersAndReturnValueBufferSize - sizeof(device_return_value);
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	399 *returnValuePtr = device_return_value;
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	400 }
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	401
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	402 status = _cuda_cu_launch_kernel(cu_function,
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	403 gridX, gridY, gridZ,
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	404 blockX, blockY, blockZ,
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	405 0, NULL, NULL, (void **) &config);
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	406
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	407 if (status != GRAAL_CUDA_SUCCESS) {
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	408 tty->print_cr("[CUDA] Failed to launch kernel");
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	409 SharedRuntime::throw_and_post_jvmti_exception(thread, vmSymbols::java_lang_Exception(), "[CUDA] Failed to launch kernel");
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	410 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	411 }
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	412
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	413 if (TraceGPUInteraction) {
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	414 tty->print_cr("[CUDA] Success: Kernel Launch: X: %d Y: %d Z: %d", blockX, blockY, blockZ);
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	415 }
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	416
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	417 status = _cuda_cu_ctx_synchronize();
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	418
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	419 if (status != GRAAL_CUDA_SUCCESS) {
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	420 tty->print_cr("[CUDA] Failed to synchronize launched kernel (%d)", status);
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	421 SharedRuntime::throw_and_post_jvmti_exception(thread, vmSymbols::java_lang_Exception(), "[CUDA] Failed to synchronize launched kernel");
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	422 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	423 }
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	424
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	425 if (TraceGPUInteraction) {
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	426 tty->print_cr("[CUDA] Success: Synchronized launch kernel");
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	427 }
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	428
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	429 jlong primitiveReturnValue = 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	430 if (isObjectReturn) {
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	431 oop return_val;
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	432 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, device_return_value, T_OBJECT_BYTE_SIZE);
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	433 if (status != GRAAL_CUDA_SUCCESS) {
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	434 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value from device argument", status);
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	435 SharedRuntime::throw_and_post_jvmti_exception(thread, vmSymbols::java_lang_Exception(), "[CUDA] Failed to copy value from device argument");
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	436 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	437 }
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	438 thread->set_vm_result(return_val);
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	439 } else if (returnTypeSize > 0) {
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	440 jlong result;
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	441 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&primitiveReturnValue, device_return_value, T_LONG_BYTE_SIZE);
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	442 if (status != GRAAL_CUDA_SUCCESS) {
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	443 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value from device argument", status);
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	444 SharedRuntime::throw_and_post_jvmti_exception(thread, vmSymbols::java_lang_Exception(), "[CUDA] Failed to copy value from device argument");
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	445 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	446 }
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	447 }
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	448
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	449 // Free device memory allocated for result
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	450 if (returnTypeSize != 0) {
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	451 status = gpu::Ptx::_cuda_cu_memfree(device_return_value);
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	452 if (status != GRAAL_CUDA_SUCCESS) {
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	453 tty->print_cr("[CUDA] *** Error (%d) Failed to free device memory of return value", status);
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	454 SharedRuntime::throw_and_post_jvmti_exception(thread, vmSymbols::java_lang_Exception(), "[CUDA] Failed to free device memory of return value");
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	455 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	456 }
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	457 }
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	458
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	459 if (TraceGPUInteraction) {
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	460 tty->print_cr("[CUDA] Success: Freed device memory of return value");
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	461 }
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	462
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	463 // Destroy context
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	464 status = gpu::Ptx::_cuda_cu_ctx_destroy(_device_context);
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	465 if (status != GRAAL_CUDA_SUCCESS) {
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	466 tty->print_cr("[CUDA] *** Error (%d) Failed to destroy context", status);
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	467 SharedRuntime::throw_and_post_jvmti_exception(thread, vmSymbols::java_lang_Exception(), "[CUDA] Failed to destroy context");
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	468 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	469 }
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	470
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	471 if (TraceGPUInteraction) {
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	472 tty->print_cr("[CUDA] Success: Destroy context");
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	473 }
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	474
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	475 return primitiveReturnValue;
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	476 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	477
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	478 bool gpu::Ptx::execute_kernel(address kernel, PTXKernelArguments &ptxka, JavaValue &ret) {
11822 365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation Morris Meyer <morris.meyer@oracle.com> parents: 11821 diff changeset	479 return gpu::Ptx::execute_warp(1, 1, 1, kernel, ptxka, ret);
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation Morris Meyer <morris.meyer@oracle.com> parents: 11821 diff changeset	480 }
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation Morris Meyer <morris.meyer@oracle.com> parents: 11821 diff changeset	481
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation Morris Meyer <morris.meyer@oracle.com> parents: 11821 diff changeset	482 bool gpu::Ptx::execute_warp(int dimX, int dimY, int dimZ,
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation Morris Meyer <morris.meyer@oracle.com> parents: 11821 diff changeset	483 address kernel, PTXKernelArguments &ptxka, JavaValue &ret) {
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	484 // grid dimensionality
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	485 unsigned int gridX = 1;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	486 unsigned int gridY = 1;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	487 unsigned int gridZ = 1;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	488
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	489 // thread dimensionality
11822 365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation Morris Meyer <morris.meyer@oracle.com> parents: 11821 diff changeset	490 unsigned int blockX = dimX;
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation Morris Meyer <morris.meyer@oracle.com> parents: 11821 diff changeset	491 unsigned int blockY = dimY;
365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation Morris Meyer <morris.meyer@oracle.com> parents: 11821 diff changeset	492 unsigned int blockZ = dimZ;
11527 c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	493
11485 49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	494 struct CUfunc_st* cu_function = (struct CUfunc_st*) kernel;
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	495
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	496 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	497 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	498 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	499 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	500 };
1cd1f8ff70a1 CR-20: PTX kernel invocation with arguments - from Bharadwaj Morris Meyer <morris.meyer@oracle.com> parents: 10884 diff changeset	501
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	502 if (kernel == NULL) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	503 return false;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	504 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	505
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	506 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	507 tty->print_cr("[CUDA] launching kernel");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	508 }
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	509
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	510 int status = _cuda_cu_launch_kernel(cu_function,
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	511 gridX, gridY, gridZ,
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	512 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	513 0, NULL, NULL, (void **) &config);
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	514 if (status != GRAAL_CUDA_SUCCESS) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	515 tty->print_cr("[CUDA] Failed to launch kernel");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	516 return false;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	517 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	518
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	519 if (TraceGPUInteraction) {
11822 365d8f385fb5 PTX: one-dimensional parallel warp invocation, ParallelOver annotation Morris Meyer <morris.meyer@oracle.com> parents: 11821 diff changeset	520 tty->print_cr("[CUDA] Success: Kernel Launch: X: %d Y: %d Z: %d", blockX, blockY, blockZ);
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	521 }
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	522
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	523 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	524
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	525 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	526 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	527 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	528 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	529
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	530 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	531 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	532 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	533
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	534
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	535 // 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	536 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	537 switch (return_type) {
11821 d8659ad83fcc PTX single-threaded array store, Warp annotation Morris Meyer <morris.meyer@oracle.com> parents: 11634 diff changeset	538 case T_INT:
11485 49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	539 {
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	540 int return_val;
12519 f020e149c1b6 PTX codegen enhancements; fixes to PTX test regressions. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 11902 diff changeset	541 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_INT_BYTE_SIZE);
11485 49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	542 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	543 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	544 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	545 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	546 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	547 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	548 break;
11901 61767ccd4600 PTX boolean return value, emitIntegerTestMove, warnings Morris Meyer <morris.meyer@oracle.com> parents: 11894 diff changeset	549 case T_BOOLEAN:
61767ccd4600 PTX boolean return value, emitIntegerTestMove, warnings Morris Meyer <morris.meyer@oracle.com> parents: 11894 diff changeset	550 {
61767ccd4600 PTX boolean return value, emitIntegerTestMove, warnings Morris Meyer <morris.meyer@oracle.com> parents: 11894 diff changeset	551 int return_val;
12519 f020e149c1b6 PTX codegen enhancements; fixes to PTX test regressions. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 11902 diff changeset	552 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_INT_BYTE_SIZE);
11901 61767ccd4600 PTX boolean return value, emitIntegerTestMove, warnings Morris Meyer <morris.meyer@oracle.com> parents: 11894 diff changeset	553 if (status != GRAAL_CUDA_SUCCESS) {
61767ccd4600 PTX boolean return value, emitIntegerTestMove, warnings Morris Meyer <morris.meyer@oracle.com> parents: 11894 diff changeset	554 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status);
61767ccd4600 PTX boolean return value, emitIntegerTestMove, warnings Morris Meyer <morris.meyer@oracle.com> parents: 11894 diff changeset	555 return false;
61767ccd4600 PTX boolean return value, emitIntegerTestMove, warnings Morris Meyer <morris.meyer@oracle.com> parents: 11894 diff changeset	556 }
61767ccd4600 PTX boolean return value, emitIntegerTestMove, warnings Morris Meyer <morris.meyer@oracle.com> parents: 11894 diff changeset	557 ret.set_jint(return_val);
61767ccd4600 PTX boolean return value, emitIntegerTestMove, warnings Morris Meyer <morris.meyer@oracle.com> parents: 11894 diff changeset	558 }
61767ccd4600 PTX boolean return value, emitIntegerTestMove, warnings Morris Meyer <morris.meyer@oracle.com> parents: 11894 diff changeset	559 break;
11894 c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	560 case T_FLOAT:
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	561 {
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	562 float return_val;
12519 f020e149c1b6 PTX codegen enhancements; fixes to PTX test regressions. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 11902 diff changeset	563 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_FLOAT_BYTE_SIZE);
11894 c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	564 if (status != GRAAL_CUDA_SUCCESS) {
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	565 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status);
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	566 return false;
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	567 }
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	568 ret.set_jfloat(return_val);
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	569 }
c7abc8411011 Fixed BasicPTXTest and IntegerPTXTest Morris Meyer <morris.meyer@oracle.com> parents: 11843 diff changeset	570 break;
11902 67a1e27a8dbb PTX initial float and double Morris Meyer <morris.meyer@oracle.com> parents: 11901 diff changeset	571 case T_DOUBLE:
67a1e27a8dbb PTX initial float and double Morris Meyer <morris.meyer@oracle.com> parents: 11901 diff changeset	572 {
67a1e27a8dbb PTX initial float and double Morris Meyer <morris.meyer@oracle.com> parents: 11901 diff changeset	573 double return_val;
12519 f020e149c1b6 PTX codegen enhancements; fixes to PTX test regressions. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 11902 diff changeset	574 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_DOUBLE_BYTE_SIZE);
11902 67a1e27a8dbb PTX initial float and double Morris Meyer <morris.meyer@oracle.com> parents: 11901 diff changeset	575 if (status != GRAAL_CUDA_SUCCESS) {
67a1e27a8dbb PTX initial float and double Morris Meyer <morris.meyer@oracle.com> parents: 11901 diff changeset	576 tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status);
67a1e27a8dbb PTX initial float and double Morris Meyer <morris.meyer@oracle.com> parents: 11901 diff changeset	577 return false;
67a1e27a8dbb PTX initial float and double Morris Meyer <morris.meyer@oracle.com> parents: 11901 diff changeset	578 }
67a1e27a8dbb PTX initial float and double Morris Meyer <morris.meyer@oracle.com> parents: 11901 diff changeset	579 ret.set_jdouble(return_val);
67a1e27a8dbb PTX initial float and double Morris Meyer <morris.meyer@oracle.com> parents: 11901 diff changeset	580 }
67a1e27a8dbb PTX initial float and double Morris Meyer <morris.meyer@oracle.com> parents: 11901 diff changeset	581 break;
11821 d8659ad83fcc PTX single-threaded array store, Warp annotation Morris Meyer <morris.meyer@oracle.com> parents: 11634 diff changeset	582 case T_LONG:
11485 49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	583 {
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	584 long return_val;
12519 f020e149c1b6 PTX codegen enhancements; fixes to PTX test regressions. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 11902 diff changeset	585 status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_LONG_BYTE_SIZE);
11485 49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	586 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	587 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	588 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	589 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	590 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	591 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	592 break;
11821 d8659ad83fcc PTX single-threaded array store, Warp annotation Morris Meyer <morris.meyer@oracle.com> parents: 11634 diff changeset	593 case T_VOID:
d8659ad83fcc PTX single-threaded array store, Warp annotation Morris Meyer <morris.meyer@oracle.com> parents: 11634 diff changeset	594 break;
11485 49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	595 default:
11821 d8659ad83fcc PTX single-threaded array store, Warp annotation Morris Meyer <morris.meyer@oracle.com> parents: 11634 diff changeset	596 tty->print_cr("[CUDA] TODO *** Unhandled return type: %d", return_type);
11485 49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	597 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	598
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	599 // Free device memory allocated for result
12519 f020e149c1b6 PTX codegen enhancements; fixes to PTX test regressions. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 11902 diff changeset	600 status = gpu::Ptx::_cuda_cu_memfree(ptxka._dev_return_value);
11485 49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	601 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	602 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	603 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	604 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	605
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	606 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	607 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	608 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	609
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	610 // 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	611 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	612 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	613 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	614 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	615 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	616
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	617 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	618 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	619 }
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	620
49bb1bc983c6 Implement several missing PTX codegen features; return value capture and method args passing of java method executed on GPU. bharadwaj parents: 11284 diff changeset	621 return (status == GRAAL_CUDA_SUCCESS);
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	622 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	623
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	624 #if defined(LINUX)
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	625 static const char cuda_library_name[] = "libcuda.so";
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	626 #elif defined(__APPLE__)
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	627 static char const cuda_library_name[] = "/usr/local/cuda/lib/libcuda.dylib";
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	628 #else
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	629 static char const cuda_library_name[] = "";
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	630 #endif
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	631
10882 6a2d65cb5d7d fix: gpu_ptx.cpp did not compile on windows Christian Wirth <christian.wirth@oracle.com> parents: 10879 diff changeset	632 #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	633
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	634 bool gpu::Ptx::probe_linkage() {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	635 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	636 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	637 void *handle = os::dll_load(cuda_library_name, buffer, STD_BUFFER_SIZE);
11527 c99e65785936 Improvements to PTX codegen; allows more PTX tests that run on the device to pass. bharadwaj parents: 11485 diff changeset	638 free(buffer);
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	639 if (handle != NULL) {
12653 1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	640 LOOKUP_CUDA_FUNCTION(cuInit, cuda_cu_init);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	641 LOOKUP_CUDA_FUNCTION(cuCtxSynchronize, cuda_cu_ctx_synchronize);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	642 LOOKUP_CUDA_FUNCTION(cuCtxSetCurrent, cuda_cu_ctx_set_current);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	643 LOOKUP_CUDA_FUNCTION(cuDeviceGetCount, cuda_cu_device_get_count);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	644 LOOKUP_CUDA_FUNCTION(cuDeviceGetName, cuda_cu_device_get_name);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	645 LOOKUP_CUDA_FUNCTION(cuDeviceGet, cuda_cu_device_get);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	646 LOOKUP_CUDA_FUNCTION(cuDeviceComputeCapability, cuda_cu_device_compute_capability);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	647 LOOKUP_CUDA_FUNCTION(cuDeviceGetAttribute, cuda_cu_device_get_attribute);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	648 LOOKUP_CUDA_FUNCTION(cuModuleGetFunction, cuda_cu_module_get_function);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	649 LOOKUP_CUDA_FUNCTION(cuModuleLoadDataEx, cuda_cu_module_load_data_ex);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	650 LOOKUP_CUDA_FUNCTION(cuLaunchKernel, cuda_cu_launch_kernel);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	651 LOOKUP_CUDA_FUNCTION(cuMemHostRegister, cuda_cu_mem_host_register);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	652 LOOKUP_CUDA_FUNCTION(cuMemHostUnregister, cuda_cu_mem_host_unregister);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	653 #if defined(__x86_64) \|\| defined(AMD64) \|\| defined(_M_AMD64)
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	654 LOOKUP_CUDA_V2_FUNCTION(cuCtxCreate, cuda_cu_ctx_create);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	655 LOOKUP_CUDA_V2_FUNCTION(cuCtxDestroy, cuda_cu_ctx_destroy);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	656 LOOKUP_CUDA_V2_FUNCTION(cuMemAlloc, cuda_cu_memalloc);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	657 LOOKUP_CUDA_V2_FUNCTION(cuMemFree, cuda_cu_memfree);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	658 LOOKUP_CUDA_V2_FUNCTION(cuMemcpyHtoD, cuda_cu_memcpy_htod);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	659 LOOKUP_CUDA_V2_FUNCTION(cuMemcpyDtoH, cuda_cu_memcpy_dtoh);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	660 LOOKUP_CUDA_V2_FUNCTION(cuMemHostGetDevicePointer, cuda_cu_mem_host_get_device_pointer);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	661 #else
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	662 LOOKUP_CUDA_FUNCTION(cuCtxCreate, cuda_cu_ctx_create);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	663 LOOKUP_CUDA_FUNCTION(cuCtxDestroy, cuda_cu_ctx_destroy);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	664 LOOKUP_CUDA_FUNCTION(cuMemAlloc, cuda_cu_memalloc);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	665 LOOKUP_CUDA_FUNCTION(cuMemFree, cuda_cu_memfree);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	666 LOOKUP_CUDA_FUNCTION(cuMemcpyHtoD, cuda_cu_memcpy_htod);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	667 LOOKUP_CUDA_FUNCTION(cuMemcpyDtoH, cuda_cu_memcpy_dtoh);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	668 LOOKUP_CUDA_FUNCTION(cuMemHostGetDevicePointer, cuda_cu_mem_host_get_device_pointer);
1a7e7011a341 * PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. S.Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com> parents: 12519 diff changeset	669 #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	670
10879 d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	671 if (TraceGPUInteraction) {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	672 tty->print_cr("[CUDA] Success: library linkage");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	673 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	674 return true;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	675 } else {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	676 // Unable to dlopen libcuda
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	677 return false;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	678 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	679 } else {
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	680 tty->print_cr("Unsupported CUDA platform");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	681 return false;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	682 }
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	683 tty->print_cr("Failed to find CUDA linkage");
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	684 return false;
d55f24eac4b1 PTX support for Linux Morris Meyer <morris.meyer@oracle.com> parents: diff changeset	685 }

Mercurial > hg > truffle

annotate src/gpu/ptx/vm/gpu_ptx.cpp @ 13624:220ed109bf77