Merge pull request #190 from accel-sim/dev

Merging the final bits of the 4.0 release from the ISCA paper.

5 files changed, 34 insertions, 20 deletions

diff --git a/src/cuda-sim/cuda_device_runtime.h b/src/cuda-sim/cuda_device_runtime.h
index 1d661b2..b06dd24 100644
--- a/src/cuda-sim/cuda_device_runtime.h
+++ b/src/cuda-sim/cuda_device_runtime.h
@@ -43,6 +43,7 @@ class cuda_device_runtime {
   std::map<void*, device_launch_config_t> g_cuda_device_launch_param_map;
   std::list<device_launch_operation_t> g_cuda_device_launch_op;
   unsigned g_kernel_launch_latency;
+  unsigned g_TB_launch_latency;
   unsigned long long g_max_total_param_size;
   bool g_cdp_enabled;
 
diff --git a/src/cuda-sim/ptx-stats.cc b/src/cuda-sim/ptx-stats.cc
index 9f7e760..3e96984 100644
--- a/src/cuda-sim/ptx-stats.cc
+++ b/src/cuda-sim/ptx-stats.cc
@@ -168,9 +168,10 @@ void ptx_file_line_stats_add_exec_count(const ptx_instruction *pInsn) {
 void ptx_stats::ptx_file_line_stats_add_latency(unsigned pc, unsigned latency) {
   const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc);
 
-  ptx_file_line_stats_tracker[ptx_file_line(pInsn->source_file(),
-                                            pInsn->source_line())]
-      .latency += latency;
+  if (pInsn != NULL)
+    ptx_file_line_stats_tracker[ptx_file_line(pInsn->source_file(),
+                                              pInsn->source_line())]
+        .latency += latency;
 }
 
 // attribute dram traffic to this ptx instruction (specified by the pc)
@@ -179,9 +180,10 @@ void ptx_stats::ptx_file_line_stats_add_dram_traffic(unsigned pc,
                                                      unsigned dram_traffic) {
   const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc);
 
-  ptx_file_line_stats_tracker[ptx_file_line(pInsn->source_file(),
-                                            pInsn->source_line())]
-      .dram_traffic += dram_traffic;
+  if (pInsn != NULL)
+    ptx_file_line_stats_tracker[ptx_file_line(pInsn->source_file(),
+                                              pInsn->source_line())]
+        .dram_traffic += dram_traffic;
 }
 
 // attribute the number of shared memory access cycles to a ptx instruction
@@ -191,10 +193,12 @@ void ptx_stats::ptx_file_line_stats_add_smem_bank_conflict(
     unsigned pc, unsigned n_way_bkconflict) {
   const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc);
 
-  ptx_file_line_stats &line_stats = ptx_file_line_stats_tracker[ptx_file_line(
-      pInsn->source_file(), pInsn->source_line())];
-  line_stats.smem_n_way_bank_conflict_total += n_way_bkconflict;
-  line_stats.smem_warp_count += 1;
+  if (pInsn != NULL) {
+    ptx_file_line_stats &line_stats = ptx_file_line_stats_tracker[ptx_file_line(
+        pInsn->source_file(), pInsn->source_line())];
+    line_stats.smem_n_way_bank_conflict_total += n_way_bkconflict;
+    line_stats.smem_warp_count += 1;
+  }
 }
 
 // attribute a non-coalesced mem access to a ptx instruction
@@ -204,10 +208,12 @@ void ptx_stats::ptx_file_line_stats_add_uncoalesced_gmem(unsigned pc,
                                                          unsigned n_access) {
   const ptx_instruction *pInsn = gpgpu_ctx->pc_to_instruction(pc);
 
-  ptx_file_line_stats &line_stats = ptx_file_line_stats_tracker[ptx_file_line(
-      pInsn->source_file(), pInsn->source_line())];
-  line_stats.gmem_n_access_total += n_access;
-  line_stats.gmem_warp_count += 1;
+  if (pInsn != NULL) {
+    ptx_file_line_stats &line_stats = ptx_file_line_stats_tracker[ptx_file_line(
+        pInsn->source_file(), pInsn->source_line())];
+    line_stats.gmem_n_access_total += n_access;
+    line_stats.gmem_warp_count += 1;
+  }
 }
 
 // a class that tracks the inflight memory instructions of a shader core
diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h
index 8c4ad4d..4243941 100644
--- a/src/cuda-sim/ptx_ir.h
+++ b/src/cuda-sim/ptx_ir.h
@@ -1339,12 +1339,12 @@ class function_info {
                       memory_space *param_mem, gpgpu_t *gpu, dim3 gridDim,
                       dim3 blockDim);
 
-  const struct gpgpu_ptx_sim_info *get_kernel_info() const {
+  virtual const struct gpgpu_ptx_sim_info *get_kernel_info() const {
     assert(m_kernel_info.maxthreads == maxnt_id);
     return &m_kernel_info;
   }
 
-  const void set_kernel_info(const struct gpgpu_ptx_sim_info &info) {
+  virtual const void set_kernel_info(const struct gpgpu_ptx_sim_info &info) {
     m_kernel_info = info;
     m_kernel_info.ptx_version = 10 * get_ptx_version().ver();
     m_kernel_info.sm_target = get_ptx_version().target();
@@ -1380,6 +1380,11 @@ class function_info {
   // backward pointer
   class gpgpu_context *gpgpu_ctx;
 
+ protected:
+  // Registers/shmem/etc. used (from ptxas -v), loaded from ___.ptxinfo along
+  // with ___.ptx
+  struct gpgpu_ptx_sim_info m_kernel_info;
+
  private:
   unsigned maxnt_id;
   unsigned m_uid;
@@ -1405,7 +1410,7 @@ class function_info {
 
   // Registers/shmem/etc. used (from ptxas -v), loaded from ___.ptxinfo along
   // with ___.ptx
-  struct gpgpu_ptx_sim_info m_kernel_info;
+  // with ___.ptx
 
   symbol_table *m_symtab;
 
diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc
index 372bda4..4e91763 100644
--- a/src/cuda-sim/ptx_loader.cc
+++ b/src/cuda-sim/ptx_loader.cc
@@ -38,7 +38,8 @@
 
 /// extern prototypes
 
-extern int ptx_error(yyscan_t yyscanner, const char *s);
+extern int ptx_error(yyscan_t yyscanner, ptx_recognizer *recognizer,
+                     const char *s);
 extern int ptx_lex_init(yyscan_t *scanner);
 extern void ptx_set_in(FILE *_in_str, yyscan_t yyscanner);
 extern int ptx_parse(yyscan_t scanner, ptx_recognizer *recognizer);
diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc
index 549c08c..afdb41b 100644
--- a/src/cuda-sim/ptx_parser.cc
+++ b/src/cuda-sim/ptx_parser.cc
@@ -36,7 +36,8 @@ typedef void *yyscan_t;
 
 extern int ptx_get_lineno(yyscan_t yyscanner);
 extern YYSTYPE *ptx_get_lval(yyscan_t yyscanner);
-extern int ptx_error(yyscan_t yyscanner, const char *s);
+extern int ptx_error(yyscan_t yyscanner, ptx_recognizer *recognizer,
+                     const char *s);
 extern int ptx_lex_init(yyscan_t *scanner);
 extern void ptx_set_in(FILE *_in_str, yyscan_t yyscanner);
 extern FILE *ptx_get_in(yyscan_t yyscanner);
@@ -225,7 +226,7 @@ void ptx_recognizer::parse_error_impl(const char *file, unsigned line,
   g_error_detected = 1;
   printf("%s:%u: Parse error: %s (%s:%u)\n\n", gpgpu_ctx->g_filename,
          ptx_get_lineno(scanner), buf, file, line);
-  ptx_error(scanner, NULL);
+  ptx_error(scanner, this, NULL);
   abort();
   exit(1);
 }