Modify default threads

app/oapv_app_dec.c

@@ -120,7 +120,7 @@ static args_var_t *args_init_vars(args_parser_t *args)
     args_set_variable_by_key_long(opts, "verbose", &op_verbose);
     op_verbose = VERBOSE_SIMPLE; /* default */
     args_set_variable_by_key_long(opts, "threads", &vars->threads);
-    vars->threads = 1; /* default */
+    vars->threads = OAPV_CDESC_THREADS_AUTO; /* default */
     args_set_variable_by_key_long(opts, "output-depth", &vars->output_depth);
     args_set_variable_by_key_long(opts, "output-csp", &vars->output_csp);
     vars->output_csp = 0; /* default: coded CSP */

app/oapv_app_enc.c

@@ -275,7 +275,7 @@ static args_var_t *args_init_vars(args_parser_t *args, oapve_param_t *param)
     args_set_variable_by_key_long(opts, "q-matrix-c3", vars->q_matrix_c3);
 
     args_set_variable_by_key_long(opts, "threads", &vars->threads);
-    vars->threads = OAPVE_CDESC_THREADS_AUTO; /* default */
+    vars->threads = OAPV_CDESC_THREADS_AUTO; /* default */
 
     args_set_variable_by_key_long(opts, "tile-w", vars->tile_w);
     args_set_variable_by_key_long(opts, "tile-h", vars->tile_h);

inc/oapv.h

@@ -469,6 +469,8 @@ static const oapv_dict_str_int_t oapv_param_opts_color_matrix[] = {
 /*****************************************************************************
  * coding parameters
  *****************************************************************************/
+#define OAPV_LEVEL_TO_LEVEL_IDC(level) (int)(((level) * 30.0) + 0.5)
+
 typedef struct oapve_param oapve_param_t;
 struct oapve_param {
     /* profile_idc defined in spec. */
@@ -522,11 +524,10 @@ struct oapve_param {
     int           full_range_flag;
 };
 
+#define OAPV_CDESC_THREADS_AUTO          0
 /*****************************************************************************
  * description for encoder creation
  *****************************************************************************/
-#define OAPVE_CDESC_THREADS_AUTO          0
-
 typedef struct oapve_cdesc oapve_cdesc_t;
 struct oapve_cdesc {
     int           max_bs_buf_size;            // max bitstream buffer size

src/oapv.c

@@ -754,8 +754,9 @@ static int enc_ready(oapve_ctx_t *ctx)
         min_num_tiles = oapv_min(min_num_tiles, num_tiles);
     }
 
-    if(ctx->cdesc.threads == OAPVE_CDESC_THREADS_AUTO) {
-        ctx->threads = oapv_min(OAPV_MAX_THREADS, oapv_min(oapv_get_num_cpu_cores(), min_num_tiles));
+    if(ctx->cdesc.threads == OAPV_CDESC_THREADS_AUTO) {
+        int num_cores = oapv_get_num_cpu_cores();
+        ctx->threads = oapv_min(OAPV_MAX_THREADS, oapv_min(num_cores, min_num_tiles));
     }
     else {
         ctx->threads = ctx->cdesc.threads;
@@ -1820,11 +1821,11 @@ static int dec_thread_tile(void *arg)
 
 static void dec_flush(oapvd_ctx_t *ctx)
 {
-    if(ctx->cdesc.threads >= 2) {
+    if(ctx->threads >= 2) {
         if(ctx->tpool) {
             // thread controller instance is present
             // terminate the created thread
-            for(int i = 0; i < ctx->cdesc.threads - 1; i++) {
+            for(int i = 0; i < ctx->threads - 1; i++) {
                 if(ctx->thread_id[i]) {
                     // valid thread instance
                     ctx->tpool->release(&ctx->thread_id[i]);
@@ -1839,7 +1840,7 @@ static void dec_flush(oapvd_ctx_t *ctx)
 
     oapv_tpool_sync_obj_delete(&(ctx->sync_obj));
 
-    for(int i = 0; i < ctx->cdesc.threads; i++) {
+    for(int i = 0; i < ctx->threads; i++) {
         dec_core_free(ctx->core[i]);
     }
 }
@@ -1848,28 +1849,37 @@ static int dec_ready(oapvd_ctx_t *ctx)
 {
     int i, ret = OAPV_OK;
 
+    if (ctx->cdesc.threads == OAPV_CDESC_THREADS_AUTO) {
+        int num_cores = oapv_get_num_cpu_cores();
+        ctx->threads = oapv_min(OAPV_MAX_THREADS, num_cores);
+    }
+    else {
+        ctx->threads = ctx->cdesc.threads;
+    }
+    oapv_assert_gv(ctx->threads > 0 && ctx->threads <= OAPV_MAX_THREADS, ret, OAPV_ERR_INVALID_ARGUMENT, ERR);
+
     if(ctx->core[0] == NULL) {
         // create cores
-        for(i = 0; i < ctx->cdesc.threads; i++) {
+        for(i = 0; i < ctx->threads; i++) {
             ctx->core[i] = dec_core_alloc();
             oapv_assert_gv(ctx->core[i], ret, OAPV_ERR_OUT_OF_MEMORY, ERR);
             ctx->core[i]->ctx = ctx;
         }
     }
 
     // initialize the threads to NULL
-    for(i = 0; i < OAPV_MAX_THREADS; i++) {
+    for(i = 0; i < ctx->threads; i++) {
         ctx->thread_id[i] = 0;
     }
 
     // get the context synchronization handle
     ctx->sync_obj = oapv_tpool_sync_obj_create();
     oapv_assert_gv(ctx->sync_obj != NULL, ret, OAPV_ERR_UNKNOWN, ERR);
 
-    if(ctx->cdesc.threads >= 2) {
+    if(ctx->threads >= 2) {
         ctx->tpool = oapv_malloc(sizeof(oapv_tpool_t));
-        oapv_tpool_init(ctx->tpool, ctx->cdesc.threads - 1);
-        for(i = 0; i < ctx->cdesc.threads - 1; i++) {
+        oapv_tpool_init(ctx->tpool, ctx->threads - 1);
+        for(i = 0; i < ctx->threads - 1; i++) {
             ctx->thread_id[i] = ctx->tpool->create(ctx->tpool, i);
             oapv_assert_gv(ctx->thread_id[i] != NULL, ret, OAPV_ERR_UNKNOWN, ERR);
         }
@@ -1919,7 +1929,7 @@ oapvd_t oapvd_create(oapvd_cdesc_t *cdesc, int *err)
     ctx = NULL;
 
     /* check if any decoder argument is correctly set */
-    oapv_assert_gv(cdesc->threads > 0 && cdesc->threads <= OAPV_MAX_THREADS, ret, OAPV_ERR_INVALID_ARGUMENT, ERR);
+    oapv_assert_gv((cdesc->threads > 0 && cdesc->threads <= OAPV_MAX_THREADS) || cdesc->threads == OAPV_CDESC_THREADS_AUTO , ret, OAPV_ERR_INVALID_ARGUMENT, ERR);
 
     /* memory allocation for ctx and core structure */
     ctx = (oapvd_ctx_t *)dec_ctx_alloc();
@@ -2008,7 +2018,7 @@ int oapvd_decode(oapvd_t did, oapv_bitb_t *bitb, oapv_frms_t *ofrms, oapvm_t mid
             int           parallel_task = 1;
             int           tidx = 0;
 
-            parallel_task = (ctx->cdesc.threads > ctx->num_tiles) ? ctx->num_tiles : ctx->cdesc.threads;
+            parallel_task = (ctx->threads > ctx->num_tiles) ? ctx->num_tiles : ctx->threads;
 
             /* decode tiles ************************************/
             for(tidx = 0; tidx < (parallel_task - 1); tidx++) {

src/oapv_def.h

@@ -347,7 +347,7 @@ typedef struct oapvd_ctx  oapvd_ctx_t;
 
 struct oapvd_core {
     ALIGNED_16(s16 coef[OAPV_MB_D]);
-    oapvd_ctx_t *ctx;
+    s16          q_mat[N_C][OAPV_BLK_D];
 
     int          prev_dc_ctx[N_C];
     int          prev_1st_ac_ctx[N_C];
@@ -356,10 +356,9 @@ struct oapvd_core {
                           /* and coded as abs_dc_coeff_diff and sign_dc_coeff_diff */
     int          qp[N_C];
     int          dq_shift[N_C];
-    s16          q_mat[N_C][OAPV_BLK_D];
-
     int          tile_idx;
 
+    oapvd_ctx_t *ctx;
     /* platform specific data, if needed */
     void        *pf;
 };
@@ -385,6 +384,7 @@ struct oapvd_ctx {
     int                     num_tile_rows;
     int                     w;
     int                     h;
+    int                     threads;
     oapv_tpool_t           *tpool;
     oapv_thread_t           thread_id[OAPV_MAX_THREADS];
     oapv_sync_obj_t         sync_obj;

src/oapv_param.c

@@ -31,8 +31,6 @@
 
 #include "oapv_def.h"
 
-#define LEVEL_TO_LEVEL_IDC(level) (int)(((level) * 30.0) + 0.5)
-
 int oapve_param_default(oapve_param_t *param)
 {
     oapv_mset(param, 0, sizeof(oapve_param_t));
@@ -47,7 +45,7 @@ int oapve_param_default(oapve_param_t *param)
     param->tile_h = 16 * OAPV_MB_H; // default: 256
 
     param->profile_idc = OAPV_PROFILE_422_10;
-    param->level_idc = LEVEL_TO_LEVEL_IDC(4.1);
+    param->level_idc = OAPV_LEVEL_TO_LEVEL_IDC(4.1);
     param->band_idc = 2;
 
     param->use_q_matrix = 0;
@@ -103,6 +101,10 @@ static int kbps_str_to_int(const char *str)
         char *tmp = strtok(s, "Mm ");
         kbps = (int)(atof(tmp) * 1000);
     }
+    else if(strchr(s, 'G') || strchr(s, 'g')) {
+        char *tmp = strtok(s, "Gg ");
+        kbps = (int)(atof(tmp) * 1000000);
+    }
     else {
         kbps = atoi(s);
     }
@@ -127,7 +129,6 @@ static int get_q_matrix(const char *str, u8 q_matrix[OAPV_BLK_D])
     return 0;
 }
 
-
 #define NAME_CMP(VAL)      else if(strcmp(name, VAL)== 0)
 #define GET_INTEGER_OR_ERR(STR, F) { \
     char * left; (F) = strtol(STR, &left, 10); \
@@ -181,7 +182,7 @@ int oapve_param_parse(oapve_param_t *param, const char *name,  const char *value
             tf0 == 3.0f || tf0 == 3.1f || tf0 == 4.0f || tf0 == 4.1f ||\
             tf0 == 5.0f || tf0 == 5.1f || tf0 == 6.0f || tf0 == 6.1f ||\
             tf0 == 7.0f || tf0 == 7.1f) {
-            param->level_idc = LEVEL_TO_LEVEL_IDC(tf0);
+            param->level_idc = OAPV_LEVEL_TO_LEVEL_IDC(tf0);
         }
         else {
             return OAPV_ERR_INVALID_ARGUMENT;

src/oapv_port.c

@@ -29,12 +29,6 @@
  * POSSIBILITY OF SUCH DAMAGE.
  */
 
-#if defined(LINUX)
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE // for sched_getaffinity()
-#endif
-#endif
-
 #include <stdarg.h>
 #include "oapv_port.h"
 
@@ -95,36 +89,34 @@ void oapv_trace_line(char *pre)
     printf("%s\n", str);
 }
 
-#if defined(WIN32) || defined(WIN64)
+#if defined(WIN32) || defined(WIN64) || defined(_WIN32)
 #include <windows.h>
 #include <sysinfoapi.h>
-#elif defined(LINUX)
-#include <sched.h>
-#elif defined(MACOS)
+#else /* LINUX, MACOS, Android */
 #include <unistd.h>
 #endif
 
 int oapv_get_num_cpu_cores(void)
 {
     int num_cores = 1; // default
-#if defined(WIN32) || defined(WIN64)
+#if defined(WIN32) || defined(WIN64) || defined(_WIN32)
     {
         SYSTEM_INFO si;
         GetNativeSystemInfo(&si);
         num_cores = si.dwNumberOfProcessors;
     }
-#elif defined(LINUX)
+#elif defined(_SC_NPROCESSORS_ONLN)
+    {
+        num_cores = (int)sysconf(_SC_NPROCESSORS_ONLN);
+    }
+#elif defined(CPU_COUNT)
     {
         cpu_set_t cset;
         memset(&cset, 0, sizeof(cset));
         if(!sched_getaffinity(0, sizeof(cset), &cset)) {
             num_cores = CPU_COUNT(&cset);
         }
     }
-#elif defined(MACOS)
-    {
-        num_cores = sysconf(_SC_NPROCESSORS_ONLN);
-    }
 #endif
     return num_cores;
 }