From dbc5076945d08959c0106ff580f42adf3f86d7d9 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Tue, 20 Nov 2018 15:34:03 -0500 Subject: Fixing the compile in gcc 4.4.x --- src/cuda-sim/instructions.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src/cuda-sim/instructions.cc') diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 31a33c6..c85654c 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -195,7 +195,7 @@ void ptx_thread_info::print_reg_thread(char * fname) { reg_map_t reg = m_regs.back(); - typename reg_map_t::const_iterator it; + reg_map_t::const_iterator it; for (it = reg.begin(); it != reg.end(); ++it) { const std::string &name = it->first->name(); -- cgit v1.3 From 73050b730b31e9f58f8c2a0bfc4b3a8cd92be7ff Mon Sep 17 00:00:00 2001 From: Suchita Pati Date: Wed, 10 Apr 2019 23:27:42 -0500 Subject: Adding texture Fix --- libcuda/cuda_runtime_api.cc | 28 ++++++++++++++++++++++++---- src/abstract_hardware_model.cc | 4 +++- src/abstract_hardware_model.h | 22 +++++++++++++++++++++- src/cuda-sim/cuda-sim.cc | 5 +++-- src/cuda-sim/cuda_device_runtime.cc | 3 ++- src/cuda-sim/instructions.cc | 5 +++-- 6 files changed, 56 insertions(+), 11 deletions(-) (limited to 'src/cuda-sim/instructions.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index cf16d9b..7569b20 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1204,8 +1204,24 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetLimit ( size_t* pValue, cudaLimit li if(g_debug_execution >= 3){ announce_call(__my_func__); } - cuda_not_implemented(__my_func__,__LINE__); - return g_last_cudaError = cudaSuccess; + switch(limit) { + case 0: + *pValue=1024; + break; + case 2: + *pValue=8388608; + break; + case 3: + *pValue=2; + break; + case 4: + *pValue=2048; + break; + default: + printf("ERROR:Limit %s unimplemented \n",limit); + abort(); + } + return g_last_cudaError = cudaSuccess; } @@ -1571,7 +1587,7 @@ __host__ cudaError_t CUDARTAPI cudaStreamDestroy(cudaStream_t stream) #if (CUDART_VERSION >= 3000) //synchronization required for application using external libraries without explicit synchronization in the code to //avoid the stream_manager from spinning forever to destroy non-empty streams without making any forward progress. - synchronize(); + stream->synchronize(); g_stream_manager->destroy_stream(stream); #endif return g_last_cudaError = cudaSuccess; @@ -2702,6 +2718,7 @@ cudaError_t CUDARTAPI cudaDeviceSynchronize(void){ if(g_debug_execution >= 3){ announce_call(__my_func__); } + synchronize(); return g_last_cudaError = cudaSuccess; } @@ -3135,6 +3152,8 @@ __host__ cudaError_t CUDARTAPI cudaDeviceSetLimit(enum cudaLimit limit, size_t v } return g_last_cudaError = cudaSuccess; } + + #endif #endif @@ -3313,7 +3332,8 @@ kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *hostFun, announce_call(__my_func__); } function_info *entry = context->get_kernel(hostFun); - kernel_info_t *result = new kernel_info_t(gridDim,blockDim,entry); + gpgpu_t* gpu= context->get_device()->get_gpgpu(); + kernel_info_t *result = new kernel_info_t(gridDim,blockDim,entry,gpu->getNameArrayMapping(),gpu->getNameInfoMapping()); if( entry == NULL ) { printf("GPGPU-Sim PTX: ERROR launching kernel -- no PTX implementation found for %p\n", hostFun); abort(); diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 307df40..3db87c9 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -688,7 +688,7 @@ unsigned g_kernel_launch_latency; unsigned kernel_info_t::m_next_uid = 1; -kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry ) +kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry, std::map NameToCudaArray, std::map NameToTexureInfo) { m_kernel_entry=entry; m_grid_dim=gridDim; @@ -708,6 +708,8 @@ kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info * m_launch_latency = g_kernel_launch_latency; volta_cache_config_set=false; + m_NameToCudaArray = NameToCudaArray; + m_NameToTexureInfo = NameToTexureInfo; } kernel_info_t::~kernel_info_t() diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 28a22e0..bbc31e5 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -212,7 +212,7 @@ public: // m_num_cores_running=0; // m_param_mem=NULL; // } - kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry ); + kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry, std::map m_NameToCudaArray, std::map m_NameToTexureInfo); ~kernel_info_t(); void inc_running() { m_num_cores_running++; } @@ -275,6 +275,20 @@ public: std::list &active_threads() { return m_active_threads; } class memory_space *get_param_memory() { return m_param_mem; } + const struct cudaArray* get_texarray( const std::string &texname ) const + { + std::map::const_iterator t=m_NameToCudaArray.find(texname); + assert(t != m_NameToCudaArray.end()); + return t->second; + } + + const struct textureInfo* get_texinfo( const std::string &texname ) const + { + std::map::const_iterator t=m_NameToTexureInfo.find(texname); + assert(t != m_NameToTexureInfo.end()); + return t->second; + } + private: kernel_info_t( const kernel_info_t & ); // disable copy constructor void operator=( const kernel_info_t & ); // disable copy operator @@ -284,6 +298,9 @@ private: unsigned m_uid; static unsigned m_next_uid; + std::map m_NameToCudaArray; + std::map m_NameToTexureInfo; + dim3 m_grid_dim; dim3 m_block_dim; dim3 m_next_cta; @@ -595,6 +612,9 @@ public: const gpgpu_functional_sim_config &get_config() const { return m_function_model_config; } FILE* get_ptx_inst_debug_file() { return ptx_inst_debug_file; } + std::map getNameArrayMapping() {return m_NameToCudaArray;} + std::map getNameInfoMapping() {return m_NameToTexureInfo;} + protected: const gpgpu_functional_sim_config &m_function_model_config; FILE* ptx_inst_debug_file; diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index ec68b5b..a51ea3e 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1507,7 +1507,8 @@ static unsigned get_tex_datasize( const ptx_instruction *pI, ptx_thread_info *th std::string texname = src1.name(); gpgpu_t *gpu = thread->get_gpu(); - const struct textureInfo* texInfo = gpu->get_texinfo(texname); + kernel_info_t& k = thread->get_kernel(); + const struct textureInfo* texInfo = k.get_texinfo(texname); unsigned data_size = texInfo->texel_size; return data_size; @@ -1918,7 +1919,7 @@ kernel_info_t *gpgpu_opencl_ptx_sim_init_grid(class function_info *entry, struct dim3 blockDim, gpgpu_t *gpu ) { - kernel_info_t *result = new kernel_info_t(gridDim,blockDim,entry); + kernel_info_t *result = new kernel_info_t(gridDim,blockDim,entry,gpu->getNameArrayMapping(),gpu->getNameInfoMapping()); unsigned argcount=args.size(); unsigned argn=1; for( gpgpu_ptx_sim_arg_list_t::iterator a = args.begin(); a != args.end(); a++ ) { diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc index b399133..917e7a8 100644 --- a/src/cuda-sim/cuda_device_runtime.cc +++ b/src/cuda-sim/cuda_device_runtime.cc @@ -198,7 +198,8 @@ void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * memory_space *device_kernel_param_mem; //create child kernel_info_t and index it with parameter_buffer address - device_grid = new kernel_info_t(config.grid_dim, config.block_dim, device_kernel_entry); + gpgpu_t* gpu=thread->get_gpu(); + device_grid = new kernel_info_t(config.grid_dim, config.block_dim, device_kernel_entry, gpu->getNameArrayMapping(), gpu->getNameInfoMapping()); device_grid->launch_cycle = gpu_sim_cycle + gpu_tot_sim_cycle; kernel_info_t & parent_grid = thread->get_kernel(); DEV_RUNTIME_REPORT("child kernel launched by " << parent_grid.name() << ", cta (" << diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index c85654c..d29fc8b 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -5260,9 +5260,10 @@ void tex_impl( const ptx_instruction *pI, ptx_thread_info *thread ) thread->get_vector_operand_values(src2, ptx_tex_regs, nelem); //ptx_reg should be 4 entry vector type...coordinates into texture gpgpu_t *gpu = thread->get_gpu(); + kernel_info_t &k = thread->get_kernel(); const struct textureReference* texref = gpu->get_texref(texname); - const struct cudaArray* cuArray = gpu->get_texarray(texname); - const struct textureInfo* texInfo = gpu->get_texinfo(texname); + const struct cudaArray* cuArray = k.get_texarray(texname); + const struct textureInfo* texInfo = k.get_texinfo(texname); const struct textureReferenceAttr* texAttr = gpu->get_texattr(texname); //assume always 2D f32 input -- cgit v1.3 From 69af057daf1999f17b81d761c7a616f67f56adbb Mon Sep 17 00:00:00 2001 From: Suchita Pati Date: Tue, 16 Apr 2019 23:36:02 -0500 Subject: Adding additional fixes for the texture bugs --- libcuda/cuda_runtime_api.cc | 22 ++++++++++++---------- src/abstract_hardware_model.cc | 13 ++++++++++--- src/abstract_hardware_model.h | 36 +++++++++++++++++++++--------------- src/cuda-sim/cuda-sim.cc | 7 +++++++ src/cuda-sim/instructions.cc | 8 +++++++- 5 files changed, 57 insertions(+), 29 deletions(-) (limited to 'src/cuda-sim/instructions.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 7569b20..f6e5e6f 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1121,10 +1121,10 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDevic *value= 0; break; case 75: - *value= 9 ; + *value= 7 ; break; case 76: - *value= 3 ; + *value= 0 ; break; case 78: *value= 0 ; //TODO: as of now, we dont support stream priorities. @@ -1199,22 +1199,23 @@ __host__ cudaError_t CUDARTAPI cudaGetDevice(int *device) *device = g_active_device; return g_last_cudaError = cudaSuccess; } + __host__ cudaError_t CUDARTAPI cudaDeviceGetLimit ( size_t* pValue, cudaLimit limit ) { if(g_debug_execution >= 3){ announce_call(__my_func__); } switch(limit) { - case 0: + case 0: // cudaLimitStackSize *pValue=1024; break; - case 2: + case 2: // cudaLimitMallocHeapSize *pValue=8388608; break; - case 3: + case 3: // cudaLimitDevRuntimeSyncDepth *pValue=2; break; - case 4: + case 4: // cudaLimitDevRuntimePendingLaunchCount *pValue=2048; break; default: @@ -1225,7 +1226,6 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetLimit ( size_t* pValue, cudaLimit li } - __host__ cudaError_t CUDARTAPI cudaStreamGetPriority ( cudaStream_t hStream, int* priority ) { if(g_debug_execution >= 3){ @@ -1585,7 +1585,7 @@ __host__ cudaError_t CUDARTAPI cudaStreamDestroy(cudaStream_t stream) announce_call(__my_func__); } #if (CUDART_VERSION >= 3000) - //synchronization required for application using external libraries without explicit synchronization in the code to + //per-stream synchronization required for application using external libraries without explicit synchronization in the code to //avoid the stream_manager from spinning forever to destroy non-empty streams without making any forward progress. stream->synchronize(); g_stream_manager->destroy_stream(stream); @@ -2714,7 +2714,6 @@ cudaError_t cudaDeviceReset ( void ) { return g_last_cudaError = cudaSuccess; } cudaError_t CUDARTAPI cudaDeviceSynchronize(void){ - // I don't know what this should do if(g_debug_execution >= 3){ announce_call(__my_func__); } @@ -2722,7 +2721,6 @@ cudaError_t CUDARTAPI cudaDeviceSynchronize(void){ return g_last_cudaError = cudaSuccess; } - void CUDARTAPI __cudaRegisterFunction( void **fatCubinHandle, const char *hostFun, @@ -3333,6 +3331,10 @@ kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *hostFun, } function_info *entry = context->get_kernel(hostFun); gpgpu_t* gpu= context->get_device()->get_gpgpu(); + /* + Passing a snapshot of the GPU's current texture mapping to the kernel's info + as kernels should use texture bindings present at the time of their launch. + */ kernel_info_t *result = new kernel_info_t(gridDim,blockDim,entry,gpu->getNameArrayMapping(),gpu->getNameInfoMapping()); if( entry == NULL ) { printf("GPGPU-Sim PTX: ERROR launching kernel -- no PTX implementation found for %p\n", hostFun); diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 3db87c9..820f5fd 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -192,6 +192,9 @@ gpgpu_t::gpgpu_t( const gpgpu_functional_sim_config &config ) checkpoint_CTA_t = m_function_model_config.get_checkpoint_CTA_t(); checkpoint_insn_Y = m_function_model_config.get_checkpoint_insn_Y(); + // initialize texture mappings to empty + m_NameToTextureInfo.clear(); + m_NameToCudaArray.clear(); if(m_function_model_config.get_ptx_inst_debug_to_file() != 0) ptx_inst_debug_file = fopen(m_function_model_config.get_ptx_inst_debug_file(), "w"); @@ -688,7 +691,11 @@ unsigned g_kernel_launch_latency; unsigned kernel_info_t::m_next_uid = 1; -kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry, std::map NameToCudaArray, std::map NameToTexureInfo) +/*A snapshot of the texture mappings needs to be stored in the kernel's info as +kernels should use the texture bindings seen at the time of launch and textures + can be bound/unbound asynchronously with respect to streams. */ + +kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry, std::map nameToCudaArray, std::map nameToTexureInfo) { m_kernel_entry=entry; m_grid_dim=gridDim; @@ -708,8 +715,8 @@ kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info * m_launch_latency = g_kernel_launch_latency; volta_cache_config_set=false; - m_NameToCudaArray = NameToCudaArray; - m_NameToTexureInfo = NameToTexureInfo; + m_NameToCudaArray = nameToCudaArray; + m_NameToTexureInfo = nameToTexureInfo; } kernel_info_t::~kernel_info_t() diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index bbc31e5..201b21f 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -212,7 +212,7 @@ public: // m_num_cores_running=0; // m_param_mem=NULL; // } - kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry, std::map m_NameToCudaArray, std::map m_NameToTexureInfo); + kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry, std::map nameToCudaArray, std::map nameToTexureInfo); ~kernel_info_t(); void inc_running() { m_num_cores_running++; } @@ -275,19 +275,22 @@ public: std::list &active_threads() { return m_active_threads; } class memory_space *get_param_memory() { return m_param_mem; } + + //The following functions access texture bindings present at the kernel's launch + const struct cudaArray* get_texarray( const std::string &texname ) const - { - std::map::const_iterator t=m_NameToCudaArray.find(texname); - assert(t != m_NameToCudaArray.end()); - return t->second; - } + { + std::map::const_iterator t=m_NameToCudaArray.find(texname); + assert(t != m_NameToCudaArray.end()); + return t->second; + } - const struct textureInfo* get_texinfo( const std::string &texname ) const - { - std::map::const_iterator t=m_NameToTexureInfo.find(texname); - assert(t != m_NameToTexureInfo.end()); - return t->second; - } + const struct textureInfo* get_texinfo( const std::string &texname ) const + { + std::map::const_iterator t=m_NameToTexureInfo.find(texname); + assert(t != m_NameToTexureInfo.end()); + return t->second; + } private: kernel_info_t( const kernel_info_t & ); // disable copy constructor @@ -297,7 +300,8 @@ private: unsigned m_uid; static unsigned m_next_uid; - + + //These maps contain the snapshot of the texture mappings at kernel launch std::map m_NameToCudaArray; std::map m_NameToTexureInfo; @@ -611,7 +615,9 @@ public: const gpgpu_functional_sim_config &get_config() const { return m_function_model_config; } FILE* get_ptx_inst_debug_file() { return ptx_inst_debug_file; } - + + // These maps return the current texture mappings for the GPU at any given time. + std::map getNameArrayMapping() {return m_NameToCudaArray;} std::map getNameInfoMapping() {return m_NameToTexureInfo;} @@ -624,7 +630,7 @@ protected: class memory_space *m_surf_mem; unsigned long long m_dev_malloc; - + // These maps contain the current texture mappings for the GPU at any given time. std::map > m_NameToTextureRef; std::map m_TextureRefToName; std::map m_NameToCudaArray; diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index a51ea3e..9d59411 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1507,6 +1507,13 @@ static unsigned get_tex_datasize( const ptx_instruction *pI, ptx_thread_info *th std::string texname = src1.name(); gpgpu_t *gpu = thread->get_gpu(); + /* + For programs with many streams, textures can be bound and unbound + asynchronously. This means we need to use the kernel's "snapshot" of + the state of the texture mappings when it was launched (so that we + don't try to access the incorrect texture mapping if it's been updated, + or that we don't access a mapping that has been unbound). + */ kernel_info_t& k = thread->get_kernel(); const struct textureInfo* texInfo = k.get_texinfo(texname); diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index d29fc8b..11001d7 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -5258,7 +5258,13 @@ void tex_impl( const ptx_instruction *pI, ptx_thread_info *thread ) if (!ptx_tex_regs) ptx_tex_regs = new ptx_reg_t[4]; unsigned nelem = src2.get_vect_nelem(); thread->get_vector_operand_values(src2, ptx_tex_regs, nelem); //ptx_reg should be 4 entry vector type...coordinates into texture - + /* + For programs with many streams, textures can be bound and unbound + asynchronously. This means we need to use the kernel's "snapshot" of + the state of the texture mappings when it was launched (so that we + don't try to access the incorrect texture mapping if it's been updated, + or that we don't access a mapping that has been unbound). + */ gpgpu_t *gpu = thread->get_gpu(); kernel_info_t &k = thread->get_kernel(); const struct textureReference* texref = gpu->get_texref(texname); -- cgit v1.3