[OG9, amdgcn, committed] Use GFX9 granulated sgprs count correctly

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view
|

[OG9, amdgcn, committed] Use GFX9 granulated sgprs count correctly

Andrew Stubbs-4
This patches adjusts the "granulated sgpr count" kernel settings for
GFX9 devices.

I followed the description I found here:
   http://llvm.org/docs/AMDGPUUsage.html

Basically, GFX9 allocates in blocks of 16, not 8, so there was some
danger of requesting too many registers, which would hurt performance.

Andrew

Use GFX9 granulated sgprs count correctly.

2019-09-10  Andrew Stubbs  <[hidden email]>

        gcc/
        * config/gcn/gcn.c (gcn_hsa_declare_function_name): Calculate
        granulated_sgprs according to architecture.

diff --git a/gcc/config/gcn/gcn.c b/gcc/config/gcn/gcn.c
index 66854b6f9c5..f8434e4a4f1 100644
--- a/gcc/config/gcn/gcn.c
+++ b/gcc/config/gcn/gcn.c
@@ -4884,6 +4884,14 @@ gcn_hsa_declare_function_name (FILE *file, const char *name, tree)
  sgpr = 102 - extra_regs;
     }
 
+  /* GFX8 allocates SGPRs in blocks of 8.
+     GFX9 uses blocks of 16.  */
+  int granulated_sgprs;
+  if (TARGET_GCN3)
+    granulated_sgprs = (sgpr + extra_regs + 7) / 8 - 1;
+  else if (TARGET_GCN5)
+    granulated_sgprs = 2 * ((sgpr + extra_regs + 15) / 16 - 1);
+
   fputs ("\t.align\t256\n", file);
   fputs ("\t.type\t", file);
   assemble_name (file, name);
@@ -4922,7 +4930,7 @@ gcn_hsa_declare_function_name (FILE *file, const char *name, tree)
    "\t\tcompute_pgm_rsrc2_excp_en = 0\n",
    (vgpr - 1) / 4,
    /* Must match wavefront_sgpr_count */
-   (sgpr + extra_regs + 7) / 8 - 1,
+   granulated_sgprs,
    /* The total number of SGPR user data registers requested.  This
       number must match the number of user data registers enabled.  */
    cfun->machine->args.nsgprs);