Added FMI LUT on GPU side + Query requests packed & aligned to 8 base…

…s + I/O bugfixes + Added GCC native flags

Makefile.mk.in

@@ -72,7 +72,7 @@ ifeq ($(CC),gcc)
     AR=gcc-ar
     OPT_AR=-fuse-linker-plugin
   endif
-  FLAGS_OPT=-Ofast -msse4.2 $(OPT_LTO)
+  FLAGS_OPT=-Ofast -march=native $(OPT_LTO)
   FLAGS_LINK=$(OPT_AR)
 endif
 

resources/gpu_modules/Makefile

@@ -44,7 +44,7 @@ CUDA_SRCS=$(addprefix $(FOLDER_SOURCE)/, $(addsuffix .cu, $(CUDA_MODULES)))
 CUDA_OBJS=$(addprefix $(FOLDER_BUILD)/, $(addsuffix .o, $(CUDA_MODULES)))
 
 BASICS=gpu_commons gpu_buffer gpu_errors gpu_io gpu_sample gpu_module gpu_devices gpu_index gpu_reference
-FMI_MODULES=gpu_fmi_index gpu_fmi_primitives gpu_fmi_primitives_decode gpu_fmi_primitives_ssearch gpu_fmi_primitives_asearch
+FMI_MODULES=gpu_fmi_index gpu_fmi_table gpu_fmi_primitives gpu_fmi_primitives_decode gpu_fmi_primitives_ssearch gpu_fmi_primitives_asearch
 SA_MODULES=gpu_sa_index gpu_sa_primitives
 BPM_MODULES=gpu_bpm_primitives
 MODULES= $(FMI_MODULES) $(SA_MODULES) $(BPM_MODULES) $(BASICS)

resources/gpu_modules/gpu_index_interface.h

@@ -100,18 +100,25 @@ typedef struct{
 } gpu_fmi_decode_end_pos_t;
 
 typedef struct {
-  uint64_t*          c;               // Occurrences of each character
-  uint64_t*          C;               // The cumulative occurrences ("ranks") of the symbols of the string
-  uint64_t*          mayor_counters;  // Pointer to the Mayor Counters (Rank)
-  uint64_t*          bwt_mem;         // Pointer to the BWT structure in memory
-  uint64_t           bwt_length;      // Length of the BWT
+  uint64_t*          c;                     // Occurrences of each character
+  uint64_t*          C;                     // The cumulative occurrences ("ranks") of the symbols of the string
+  uint64_t*          mayor_counters;        // Pointer to the Mayor Counters (Rank)
+  uint64_t*          bwt_mem;               // Pointer to the BWT structure in memory
+  uint64_t           bwt_length;            // Length of the BWT
+  uint64_t           num_levels_fmi_table;
+  uint64_t           skip_levels_fmi_table;
   gpu_index_coding_t index_coding;
 } gpu_gem_fmi_dto_t;
 
 typedef struct {
   char                *h_plain;
   gpu_fmi_entry_t     *h_fmi;
   uint64_t            bwtSize;
+  uint32_t            *h_offsetsTable;
+  gpu_sa_entry_t      *h_table;
+  uint32_t            numLevelsTable;
+  uint32_t            skipLevelsTable;
+  uint32_t            numElementsTable;
   gpu_index_coding_t  indexCoding;
 } gpu_fmi_dto_t;
 

resources/gpu_modules/gpu_interface.h

@@ -48,14 +48,16 @@ typedef enum
   /* GPU stages          */
   GPU_SEEDING           = GPU_INDEX,
   GPU_FILTERING         = GPU_REFERENCE,
-
+  /* General setups      */
   GPU_NONE_MODULES      = 0,
   GPU_ALL_MODULES       = GPU_SEEDING | GPU_FILTERING
 } gpu_module_t;
 
 typedef enum
 {
+  /* Non-supported GPU architectures */
   GPU_ARCH_TESLA      = GPU_UINT32_ONE_MASK << 0,
+  /* Supported GPU architectures     */
   GPU_ARCH_FERMI_1G   = GPU_UINT32_ONE_MASK << 1,
   GPU_ARCH_FERMI_2G   = GPU_UINT32_ONE_MASK << 2,
   GPU_ARCH_KEPLER_1G  = GPU_UINT32_ONE_MASK << 3,
@@ -64,12 +66,12 @@ typedef enum
   GPU_ARCH_MAXWELL_2G = GPU_UINT32_ONE_MASK << 6,
   GPU_ARCH_PASCAL_1G  = GPU_UINT32_ONE_MASK << 7,
   GPU_ARCH_PASCAL_2G  = GPU_UINT32_ONE_MASK << 8,
-
+  /* Main GPU Architectures          */
   GPU_ARCH_FERMI      = GPU_ARCH_FERMI_1G   | GPU_ARCH_FERMI_2G,
   GPU_ARCH_KEPLER     = GPU_ARCH_KEPLER_1G  | GPU_ARCH_KEPLER_2G,
   GPU_ARCH_MAXWELL    = GPU_ARCH_MAXWELL_1G | GPU_ARCH_MAXWELL_2G,
   GPU_ARCH_PASCAL     = GPU_ARCH_PASCAL_1G  | GPU_ARCH_PASCAL_2G,
-
+  /* General setups                  */
   GPU_ARCH_NEWGEN     = GPU_UINT32_ONE_MASK << 31,
   GPU_ARCH_SUPPORTED  = GPU_ARCH_FERMI | GPU_ARCH_KEPLER | GPU_ARCH_MAXWELL | GPU_ARCH_NEWGEN
 } gpu_dev_arch_t;

resources/gpu_modules/include/gpu_devices.h

@@ -34,13 +34,19 @@ typedef enum
 typedef enum
 {
   /* Types of page-lock allocations */
-  GPU_PAGE_LOCKED_PORTABLE,
-  GPU_PAGE_LOCKED_MAPPED,
-  GPU_PAGE_LOCKED_WRITECOMBINED,
+  GPU_PAGE_LOCKED_PORTABLE       = GPU_UINT32_ONE_MASK << 0,
+  GPU_PAGE_LOCKED_MAPPED         = GPU_UINT32_ONE_MASK << 1,
+  GPU_PAGE_LOCKED_WRITECOMBINED  = GPU_UINT32_ONE_MASK << 2,
   /* Types of host allocations */
-  GPU_PAGE_LOCKED = GPU_PAGE_LOCKED_PORTABLE | GPU_PAGE_LOCKED_MAPPED | GPU_PAGE_LOCKED_WRITECOMBINED,
-  GPU_PAGE_UNLOCKED,
-  GPU_NONE_ALLOCATED
+  GPU_PAGE_UNLOCKED              = GPU_UINT32_ONE_MASK << 3,
+  GPU_PAGE_LOCKED                = GPU_PAGE_LOCKED_PORTABLE | GPU_PAGE_LOCKED_MAPPED | GPU_PAGE_LOCKED_WRITECOMBINED,
+  /* Types for non-allocated pages */
+  GPU_PAGE_UNALLOCATED           = 0,
+  /* Types for assigned pages (contains data) */
+  GPU_PAGE_ASSIGNED              = GPU_UINT32_ONE_MASK << 31,
+  /*  */
+  GPU_PAGE_ASSIGNED_AND_UNLOCKED = GPU_PAGE_ASSIGNED | GPU_PAGE_UNLOCKED,
+  GPU_PAGE_ASSIGNED_AND_LOCKED   = GPU_PAGE_ASSIGNED | GPU_PAGE_LOCKED
 } memory_stats_t;
 
 typedef struct {

resources/gpu_modules/include/gpu_errors.h

@@ -29,6 +29,7 @@ typedef enum
   E_DATA_NOT_ALLOCATED,
   E_NOT_SUPPORTED_ALLOC_POLICY,
   E_OVERFLOWING_BUFFER,
+  E_FMI_TABLE_INCOMPATIBLE_SIZE,
   E_NOT_IMPLEMENTED
 } gpu_error_t;
 

resources/gpu_modules/include/gpu_fmi_index.h

@@ -13,12 +13,14 @@
 #include "gpu_commons.h"
 #include "gpu_devices.h"
 #include "gpu_fmi_structure.h"
+#include "gpu_fmi_table.h"
 
 /*****************************
 Internal Objects (General)
 *****************************/
 
 typedef struct {
+  gpu_fmi_table_t table;
   uint64_t        bwtSize;
   uint64_t        numEntries;
   gpu_fmi_entry_t *h_fmi;
@@ -42,6 +44,7 @@ gpu_error_t gpu_fmi_index_transfer_CPU_to_GPUs(gpu_fmi_buffer_t* const fmi, gpu_
 /* Stream index functions */
 gpu_error_t gpu_fmi_index_read_specs(FILE* fp, gpu_fmi_buffer_t* const fmi);
 gpu_error_t gpu_fmi_index_read(FILE* fp, gpu_fmi_buffer_t* const fmi);
+gpu_error_t gpu_fmi_index_write_specs(FILE* fp, const gpu_fmi_buffer_t* const fmi);
 gpu_error_t gpu_fmi_index_write(FILE* fp, const gpu_fmi_buffer_t* const fmi);
 
 /* Data transform functions */

resources/gpu_modules/include/gpu_fmi_primitives_asearch.h

@@ -20,7 +20,9 @@ Common constants for Device & Host
 *********************************/
 
 /* Defines related to FMI adaptative BACKWARD-SEARCH primitives */
-#define GPU_FMI_ENTRIES_PER_QUERY 2
+#define GPU_FMI_ENTRIES_PER_QUERY      2
+#define GPU_FMI_BASE_QUERY_LENGTH      8 //Bits per base query
+#define GPU_FMI_BASES_PER_QUERY_ENTRY  (GPU_UINT64_LENGTH / GPU_FMI_BASE_QUERY_LENGTH) //Bases per internal entry
 
 #define GPU_FMI_THREADS_PER_QUERY (GPU_FMI_THREADS_PER_ENTRY * GPU_FMI_ENTRIES_PER_SEED)
 #define GPU_FMI_D                 2

resources/gpu_modules/include/gpu_fmi_table.h

@@ -0,0 +1,64 @@
+/*
+ *  GEM-Cutter "Highly optimized genomic resources for GPUs"
+ *  Copyright (c) 2013-2016 by Alejandro Chacon    <alejandro.chacond@gmail.com>
+ *
+ *  Licensed under GNU General Public License 3.0 or later.
+ *  Some rights reserved. See LICENSE, AUTHORS.
+ *  @license GPL-3.0+ <http://www.gnu.org/licenses/gpl-3.0.en.html>
+ */
+
+#ifndef GPU_FMI_TABLE_H_
+#define GPU_FMI_TABLE_H_
+
+#include "gpu_commons.h"
+#include "gpu_devices.h"
+
+#define GPU_FMI_TABLE_ALPHABET_SIZE       4
+#define GPU_FMI_TABLE_MIN_ELEMENTS        2
+#define GPU_FMI_TABLE_MIN_LEVELS          1
+#define GPU_FMI_TABLE_DEFAULT_LEVELS      11
+#define GPU_FMI_TABLE_DEFAULT_SKIP_LEVELS 0
+
+typedef struct{
+  uint32_t init;
+  uint32_t top;
+} offset_table_t;
+
+typedef struct{
+  uint32_t         maxLevelsTableLUT;
+  uint32_t         skipLevelsTableLUT;
+  uint32_t         totalElemTableLUT;
+  offset_table_t*  h_offsetsTableLUT;
+  offset_table_t** d_offsetsTableLUT;
+  gpu_sa_entry_t*  h_fmiTableLUT;
+  gpu_sa_entry_t** d_fmiTableLUT;
+  memory_stats_t   hostAllocStats;
+  memory_alloc_t*  memorySpace;
+} gpu_fmi_table_t;
+
+/* High level functions */
+gpu_error_t gpu_fmi_table_get_num_elements(const uint32_t numLevels, uint32_t* const totalElemTableLUT);
+gpu_error_t gpu_fmi_table_init_dto(gpu_fmi_table_t* const fmiTable);
+gpu_error_t gpu_fmi_table_init(gpu_fmi_table_t* const fmiTable, const uint32_t numLevels, const uint32_t numSupportedDevices);
+gpu_error_t gpu_fmi_table_allocate(gpu_fmi_table_t* const fmiTable);
+gpu_error_t gpu_fmi_table_read_specs(FILE* fp, gpu_fmi_table_t* const fmiTable);
+gpu_error_t gpu_fmi_table_load_default_specs(gpu_fmi_table_t* const fmiTable);
+gpu_error_t gpu_fmi_table_read(FILE* fp, gpu_fmi_table_t* const fmiTable);
+gpu_error_t gpu_fmi_table_write_specs(FILE* fp, const gpu_fmi_table_t* const fmiTable);
+gpu_error_t gpu_fmi_table_write(FILE* fp, const gpu_fmi_table_t* const fmiTable);
+gpu_error_t gpu_fmi_table_transfer_CPU_to_GPUs(gpu_fmi_table_t* const fmiTable, gpu_device_info_t** const devices);
+gpu_error_t gpu_fmi_table_build(gpu_fmi_table_t* const fmiTable, const gpu_fmi_entry_t* const h_fmi, const uint64_t bwtSize);
+gpu_error_t gpu_fmi_table_get_size(const gpu_fmi_table_t* const fmiTable, size_t* const bytesPerFmiTable);
+gpu_error_t gpu_fmi_table_free_host(gpu_fmi_table_t* const fmiTable);
+gpu_error_t gpu_fmi_table_free_unused_host(gpu_fmi_table_t* const fmiTable, gpu_device_info_t** const devices);
+gpu_error_t gpu_fmi_table_free_device(gpu_fmi_table_t* const fmiTable, gpu_device_info_t** const devices);
+gpu_error_t gpu_fmi_table_free_metainfo(gpu_fmi_table_t* const fmiTable);
+
+/* Low level functions */
+gpu_error_t gpu_fmi_table_init_offsets(offset_table_t* const offsetsTableLUT, const uint32_t numLevels);
+void        gpu_fmi_table_process_entry(const gpu_fmi_entry_t* const h_fmi, gpu_sa_entry_t* const currentIntervals, const gpu_sa_entry_t L, const gpu_sa_entry_t R);
+gpu_error_t gpu_fmi_table_process_forward_level(const gpu_fmi_entry_t* const h_fmi, const uint32_t idLevel, offset_table_t* const offsetsTableLUT, gpu_sa_entry_t* const fmiTableLUT);
+gpu_error_t gpu_fmi_table_process_backward_level(const gpu_fmi_entry_t* const h_fmi, const uint32_t idLevel, offset_table_t* const offsetsTableLUT, gpu_sa_entry_t* const fmiTableLUT);
+
+
+#endif /* GPU_FMI_TABLE_H_ */

resources/gpu_modules/include/gpu_sa_index.h

@@ -38,6 +38,7 @@ gpu_error_t gpu_sa_index_transfer_CPU_to_GPUs(gpu_sa_buffer_t* const sa, gpu_dev
 /* Stream index functions  */
 gpu_error_t gpu_sa_index_read_specs(FILE* fp, gpu_sa_buffer_t* const sa);
 gpu_error_t gpu_sa_index_read(FILE* fp, gpu_sa_buffer_t* const sa);
+gpu_error_t gpu_sa_index_write_specs(FILE* fp, const gpu_sa_buffer_t* const sa);
 gpu_error_t gpu_sa_index_write(FILE* fp, const gpu_sa_buffer_t* const sa);
 
 /* Data load functions */

resources/gpu_modules/src/gpu_buffer.c

@@ -48,12 +48,15 @@ gpu_error_t gpu_buffer_get_min_memory_size(size_t *bytesPerBuffer)
 {
   const uint32_t averarageNumPEQEntries = 1;
   const uint32_t candidatesPerQuery     = 1;
+  const uint32_t averageQuerySize       = 100;
+  const uint32_t averageRegionsPerQuery = 10;
 
-  const size_t bytesPerBPMBuffer    = GPU_BPM_MIN_ELEMENTS        * gpu_bpm_size_per_candidate(averarageNumPEQEntries,candidatesPerQuery);
-  const size_t bytesPerSearchBuffer = GPU_FMI_SEARCH_MIN_ELEMENTS * gpu_fmi_ssearch_input_size();
-  const size_t bytesPerDecodeBuffer = GPU_FMI_DECODE_MIN_ELEMENTS * gpu_fmi_decode_input_size();
+  const size_t bytesPerBPMBuffer     = GPU_BPM_MIN_ELEMENTS        * gpu_bpm_size_per_candidate(averarageNumPEQEntries,candidatesPerQuery);
+  const size_t bytesPerSSearchBuffer = GPU_FMI_SEARCH_MIN_ELEMENTS * gpu_fmi_asearch_size_per_query(averageQuerySize, averageRegionsPerQuery);
+  const size_t bytesPerASearchBuffer = GPU_FMI_SEARCH_MIN_ELEMENTS * gpu_fmi_ssearch_input_size();
+  const size_t bytesPerDecodeBuffer  = GPU_FMI_DECODE_MIN_ELEMENTS * gpu_fmi_decode_input_size();
 
-  (* bytesPerBuffer) = GPU_MAX(bytesPerBPMBuffer,GPU_MAX(bytesPerSearchBuffer,bytesPerDecodeBuffer));
+  (* bytesPerBuffer) = GPU_MAX(bytesPerBPMBuffer,GPU_MAX(bytesPerDecodeBuffer,GPU_MAX(bytesPerSSearchBuffer,bytesPerASearchBuffer)));
   return (SUCCESS);
 }
 

resources/gpu_modules/src/gpu_errors.c

@@ -38,6 +38,7 @@ const char* gpuGetErrorString(gpu_error_t error)
     case E_NOT_IMPLEMENTED:             return "GEM GPU - Error: functionality not implemented";
     case E_DATA_NOT_ALLOCATED:          return "GEM GPU - Error: structure not allocated to any memory";
     case E_OVERFLOWING_BUFFER:          return "GEM GPU - Error: overflowing elements per buffer";
+    case E_FMI_TABLE_INCOMPATIBLE_SIZE: return "GEM GPU - Error: fmi table num levels incompatible";
     default:                            return "GEM GPU - Unknown error";
   }
 }