[amdgcn] Scale number of threads/workers with VGPR usage

2020-01-31  Kwok Cheung Yeung  <kcy@codesourcery.com>

	gcc/
	* config/gcn/mkoffload.c (process_asm): Add sgpr_count and vgpr_count
	to definition of hsa_kernel_description.  Parse assembly to find SGPR
	and VGPR count of kernel and store in hsa_kernel_description.

	libgomp/
	* plugin/plugin-gcn.c (struct hsa_kernel_description): Add sgpr_count
	and vgpr_count fields.
	(struct kernel_info): Add a field for a hsa_kernel_description.
	(run_kernel): Reduce the number of threads/workers if the requested
	number would require too many VGPRs.
	(init_basic_kernel_info): Initialize description field with
	the hsa_kernel_description entry for the kernel.

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index f44a09d782231d920faac91e9d290060849c98ed..78a831086775c96ef071dd3a24a590722523959a 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,9 @@
+2020-01-31  Kwok Cheung Yeung  <kcy@codesourcery.com>
+
+       * config/gcn/mkoffload.c (process_asm): Add sgpr_count and vgpr_count
+       to definition of hsa_kernel_description.  Parse assembly to find SGPR
+       and VGPR count of kernel and store in hsa_kernel_description.
+
 2020-01-31  Tamar Christina  <tamar.christina@arm.com>
 
        PR rtl-optimization/91838

diff --git a/gcc/config/gcn/mkoffload.c b/gcc/config/gcn/mkoffload.c
index 0062f158d66d4fa6ad3d4486e6ac0d9557e99749..723da108b655b148b5e06b1f9ed20b7de4e9482b 100644 (file)
--- a/gcc/config/gcn/mkoffload.c
+++ b/gcc/config/gcn/mkoffload.c
@@ -211,12 +211,13 @@ access_check (const char *name, int mode)
 static void
 process_asm (FILE *in, FILE *out, FILE *cfile)
 {
-  int fn_count = 0, var_count = 0, dims_count = 0;
-  struct obstack fns_os, vars_os, varsizes_os, dims_os;
+  int fn_count = 0, var_count = 0, dims_count = 0, regcount_count = 0;
+  struct obstack fns_os, vars_os, varsizes_os, dims_os, regcounts_os;
   obstack_init (&fns_os);
   obstack_init (&vars_os);
   obstack_init (&varsizes_os);
   obstack_init (&dims_os);
+  obstack_init (&regcounts_os);
 
   struct oaccdims
   {
@@ -224,13 +225,20 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
     char *name;
   } dim;
 
+  struct regcount
+  {
+    int sgpr_count;
+    int vgpr_count;
+    char *kernel_name;
+  } regcount;
+
   /* Always add _init_array and _fini_array as kernels.  */
   obstack_ptr_grow (&fns_os, xstrdup ("_init_array"));
   obstack_ptr_grow (&fns_os, xstrdup ("_fini_array"));
   fn_count += 2;
 
   char buf[1000];
-  enum { IN_CODE, IN_VARS, IN_FUNCS } state = IN_CODE;
+  enum { IN_CODE, IN_AMD_KERNEL_CODE_T, IN_VARS, IN_FUNCS } state = IN_CODE;
   while (fgets (buf, sizeof (buf), in))
     {
       switch (state)
@@ -243,6 +251,22 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
                obstack_grow (&dims_os, &dim, sizeof (dim));
                dims_count++;
              }
+           else if (sscanf (buf, " .amdgpu_hsa_kernel %ms\n",
+                            &regcount.kernel_name) == 1)
+             break;
+
+           break;
+         }
+       case IN_AMD_KERNEL_CODE_T:
+         {
+           gcc_assert (regcount.kernel_name);
+           if (sscanf (buf, " wavefront_sgpr_count = %d\n",
+                       &regcount.sgpr_count) == 1)
+             break;
+           else if (sscanf (buf, " workitem_vgpr_count = %d\n",
+                            &regcount.vgpr_count) == 1)
+             break;
+
            break;
          }
        case IN_VARS:
@@ -282,19 +306,36 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
        state = IN_VARS;
       else if (sscanf (buf, " .section .gnu.offload_funcs%c", &dummy) > 0)
        state = IN_FUNCS;
+      else if (sscanf (buf, " .amd_kernel_code_%c", &dummy) > 0)
+       {
+         state = IN_AMD_KERNEL_CODE_T;
+         regcount.sgpr_count = regcount.vgpr_count = -1;
+       }
       else if (sscanf (buf, " .section %c", &dummy) > 0
               || sscanf (buf, " .text%c", &dummy) > 0
               || sscanf (buf, " .bss%c", &dummy) > 0
               || sscanf (buf, " .data%c", &dummy) > 0
               || sscanf (buf, " .ident %c", &dummy) > 0)
        state = IN_CODE;
+      else if (sscanf (buf, " .end_amd_kernel_code_%c", &dummy) > 0)
+       {
+         state = IN_CODE;
+         gcc_assert (regcount.kernel_name != NULL
+                     && regcount.sgpr_count >= 0
+                     && regcount.vgpr_count >= 0);
+         obstack_grow (&regcounts_os, &regcount, sizeof (regcount));
+         regcount_count++;
+         regcount.kernel_name = NULL;
+         regcount.sgpr_count = regcount.vgpr_count = -1;
+       }
 
-      if (state == IN_CODE)
+      if (state == IN_CODE || state == IN_AMD_KERNEL_CODE_T)
        fputs (buf, out);
     }
 
   char **fns = XOBFINISH (&fns_os, char **);
   struct oaccdims *dims = XOBFINISH (&dims_os, struct oaccdims *);
+  struct regcount *regcounts = XOBFINISH (&regcounts_os, struct regcount *);
 
   fprintf (cfile, "#include <stdlib.h>\n");
   fprintf (cfile, "#include <stdbool.h>\n\n");
@@ -322,6 +363,8 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
   fprintf (cfile, "static const struct hsa_kernel_description {\n"
           "  const char *name;\n"
           "  int oacc_dims[3];\n"
+          "  int sgpr_count;\n"
+          "  int vgpr_count;\n"
           "} gcn_kernels[] = {\n  ");
   dim.d[0] = dim.d[1] = dim.d[2] = 0;
   const char *comma;
@@ -329,15 +372,24 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
     {
       /* Find if we recorded dimensions for this function.  */
       int *d = dim.d;          /* Previously zeroed.  */
+      int sgpr_count = 0;
+      int vgpr_count = 0;
       for (int j = 0; j < dims_count; j++)
        if (strcmp (fns[i], dims[j].name) == 0)
          {
            d = dims[j].d;
            break;
          }
+      for (int j = 0; j < regcount_count; j++)
+       if (strcmp (fns[i], regcounts[j].kernel_name) == 0)
+         {
+           sgpr_count = regcounts[j].sgpr_count;
+           vgpr_count = regcounts[j].vgpr_count;
+           break;
+         }
 
-      fprintf (cfile, "%s{\"%s\", {%d, %d, %d}}", comma,
-              fns[i], d[0], d[1], d[2]);
+      fprintf (cfile, "%s{\"%s\", {%d, %d, %d}, %d, %d}", comma,
+              fns[i], d[0], d[1], d[2], sgpr_count, vgpr_count);
 
       free (fns[i]);
     }
@@ -346,7 +398,10 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
   obstack_free (&fns_os, NULL);
   for (i = 0; i < dims_count; i++)
     free (dims[i].name);
+  for (i = 0; i < regcount_count; i++)
+    free (regcounts[i].kernel_name);
   obstack_free (&dims_os, NULL);
+  obstack_free (&regcounts_os, NULL);
 }
 
 /* Embed an object file into a C source file.  */

diff --git a/libgomp/ChangeLog b/libgomp/ChangeLog
index 09707247068ad3549cddbef91dd4a6eba39bb05d..557bec34f2f067f8ff451ed53b1f2cf1cfb3b860 100644 (file)
--- a/libgomp/ChangeLog
+++ b/libgomp/ChangeLog
@@ -1,3 +1,13 @@
+2020-01-31  Kwok Cheung Yeung  <kcy@codesourcery.com>
+
+       * plugin/plugin-gcn.c (struct hsa_kernel_description): Add sgpr_count
+       and vgpr_count fields.
+       (struct kernel_info): Add a field for a hsa_kernel_description.
+       (run_kernel): Reduce the number of threads/workers if the requested
+       number would require too many VGPRs.
+       (init_basic_kernel_info): Initialize description field with
+       the hsa_kernel_description entry for the kernel.
+
 2020-01-29  Tobias Burnus  <tobias@codesourcery.com>
 
        PR bootstrap/93409

diff --git a/libgomp/plugin/plugin-gcn.c b/libgomp/plugin/plugin-gcn.c
index 22676b45efb73ff72efdff9a26ec53e38c0f4ee9..25547ef12e81b70990d7be246f39be23f97d23f3 100644 (file)
--- a/libgomp/plugin/plugin-gcn.c
+++ b/libgomp/plugin/plugin-gcn.c
@@ -371,6 +371,8 @@ struct hsa_kernel_description
 {
   const char *name;
   int oacc_dims[3];  /* Only present for GCN kernels.  */
+  int sgpr_count;
+  int vpgr_count;
 };
 
 /* Mkoffload uses this structure to describe an offload variable.  */
@@ -478,6 +480,8 @@ struct kernel_info
   struct agent_info *agent;
   /* The specific module where the kernel takes place.  */
   struct module_info *module;
+  /* Information provided by mkoffload associated with the kernel.  */
+  struct hsa_kernel_description *description;
   /* Mutex enforcing that at most once thread ever initializes a kernel for
      use.  A thread should have locked agent->module_rwlock for reading before
      acquiring it.  */
@@ -2102,6 +2106,24 @@ run_kernel (struct kernel_info *kernel, void *vars,
            struct GOMP_kernel_launch_attributes *kla,
            struct goacc_asyncqueue *aq, bool module_locked)
 {
+  GCN_DEBUG ("SGPRs: %d, VGPRs: %d\n", kernel->description->sgpr_count,
+            kernel->description->vpgr_count);
+
+  /* Reduce the number of threads/workers if there are insufficient
+     VGPRs available to run the kernels together.  */
+  if (kla->ndim == 3 && kernel->description->vpgr_count > 0)
+    {
+      int granulated_vgprs = (kernel->description->vpgr_count + 3) & ~3;
+      int max_threads = (256 / granulated_vgprs) * 4;
+      if (kla->gdims[2] > max_threads)
+       {
+         GCN_WARNING ("Too many VGPRs required to support %d threads/workers"
+                      " per team/gang - reducing to %d threads/workers.\n",
+                      kla->gdims[2], max_threads);
+         kla->gdims[2] = max_threads;
+       }
+    }
+
   GCN_DEBUG ("GCN launch on queue: %d:%d\n", kernel->agent->device_id,
             (aq ? aq->id : 0));
   GCN_DEBUG ("GCN launch attribs: gdims:[");
@@ -2303,6 +2325,7 @@ init_basic_kernel_info (struct kernel_info *kernel,
   kernel->agent = agent;
   kernel->module = module;
   kernel->name = d->name;
+  kernel->description = d;
   if (pthread_mutex_init (&kernel->init_mutex, NULL))
     {
       GOMP_PLUGIN_error ("Failed to initialize a GCN kernel mutex");