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 <a...@codesourcery.com>
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);
