From: Andrew Stubbs <ams@codesourcery.com>
To: <gcc-patches@gcc.gnu.org>
Subject: [PATCH 16/17] amdgcn, openmp: Auto-detect USM mode and set HSA_XNACK
Date: Thu, 7 Jul 2022 11:34:47 +0100 [thread overview]
Message-ID: <c73170e6b71179306d40ce6a04a3c02cb4c3cef6.1657188329.git.ams@codesourcery.com> (raw)
In-Reply-To: <cover.1657188329.git.ams@codesourcery.com>
[-- Attachment #1: Type: text/plain, Size: 1446 bytes --]
The AMD GCN runtime must be set to the correct mode for Unified Shared Memory
to work, but this is not always clear at compile and link time due to the split
nature of the offload compilation pipeline.
This patch sets a new attribute on OpenMP offload functions to ensure that the
information is passed all the way to the backend. The backend then places a
marker in the assembler code for mkoffload to find. Finally mkoffload places a
constructor function into the final program to ensure that the HSA_XNACK
environment variable passes the correct mode to the GPU.
The HSA_XNACK variable must be set before the HSA runtime is even loaded, so
it makes more sense to have this set within the constructor than at some point
later within libgomp or the GCN plugin.
gcc/ChangeLog:
* config/gcn/gcn.c (unified_shared_memory_enabled): New variable.
(gcn_init_cumulative_args): Handle attribute "omp unified memory".
(gcn_hsa_declare_function_name): Emit "MKOFFLOAD OPTIONS: USM+".
* config/gcn/mkoffload.c (TEST_XNACK_OFF): New macro.
(process_asm): Detect "MKOFFLOAD OPTIONS: USM+".
Emit configure_xnack constructor, as required.
* omp-low.c (create_omp_child_function): Add attribute "omp unified
memory".
---
gcc/config/gcn/gcn.cc | 28 +++++++++++++++++++++++++++-
gcc/config/gcn/mkoffload.cc | 37 ++++++++++++++++++++++++++++++++++++-
gcc/omp-low.cc | 4 ++++
3 files changed, 67 insertions(+), 2 deletions(-)
[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: 0016-amdgcn-openmp-Auto-detect-USM-mode-and-set-HSA_XNACK.patch --]
[-- Type: text/x-patch; name="0016-amdgcn-openmp-Auto-detect-USM-mode-and-set-HSA_XNACK.patch", Size: 5927 bytes --]
diff --git a/gcc/config/gcn/gcn.cc b/gcc/config/gcn/gcn.cc
index 4df05453604..88cc505597e 100644
--- a/gcc/config/gcn/gcn.cc
+++ b/gcc/config/gcn/gcn.cc
@@ -68,6 +68,11 @@ static bool ext_gcn_constants_init = 0;
enum gcn_isa gcn_isa = ISA_GCN3; /* Default to GCN3. */
+/* Record whether the host compiler added "omp unifed memory" attributes to
+ any functions. We can then pass this on to mkoffload to ensure xnack is
+ compatible there too. */
+static bool unified_shared_memory_enabled = false;
+
/* Reserve this much space for LDS (for propagating variables from
worker-single mode to worker-partitioned mode), per workgroup. Global
analysis could calculate an exact bound, but we don't do that yet.
@@ -2542,6 +2547,25 @@ gcn_init_cumulative_args (CUMULATIVE_ARGS *cum /* Argument info to init */ ,
if (!caller && cfun->machine->normal_function)
gcn_detect_incoming_pointer_arg (fndecl);
+ if (fndecl && lookup_attribute ("omp unified memory",
+ DECL_ATTRIBUTES (fndecl)))
+ {
+ unified_shared_memory_enabled = true;
+
+ switch (gcn_arch)
+ {
+ case PROCESSOR_FIJI:
+ case PROCESSOR_VEGA10:
+ case PROCESSOR_VEGA20:
+ error ("GPU architecture does not support Unified Shared Memory");
+ default:
+ ;
+ }
+
+ if (flag_xnack == HSACO_ATTR_OFF)
+ error ("Unified Shared Memory is enabled, but XNACK is disabled");
+ }
+
reinit_regs ();
}
@@ -5458,12 +5482,14 @@ gcn_hsa_declare_function_name (FILE *file, const char *name, tree)
assemble_name (file, name);
fputs (":\n", file);
- /* This comment is read by mkoffload. */
+ /* These comments are read by mkoffload. */
if (flag_openacc)
fprintf (file, "\t;; OPENACC-DIMS: %d, %d, %d : %s\n",
oacc_get_fn_dim_size (cfun->decl, GOMP_DIM_GANG),
oacc_get_fn_dim_size (cfun->decl, GOMP_DIM_WORKER),
oacc_get_fn_dim_size (cfun->decl, GOMP_DIM_VECTOR), name);
+ if (unified_shared_memory_enabled)
+ fprintf (asm_out_file, "\t;; MKOFFLOAD OPTIONS: USM+\n");
}
/* Implement TARGET_ASM_SELECT_SECTION.
diff --git a/gcc/config/gcn/mkoffload.cc b/gcc/config/gcn/mkoffload.cc
index cb8903c27cb..5741d0a917b 100644
--- a/gcc/config/gcn/mkoffload.cc
+++ b/gcc/config/gcn/mkoffload.cc
@@ -80,6 +80,8 @@
== EF_AMDGPU_FEATURE_XNACK_ANY_V4)
#define TEST_XNACK_ON(VAR) ((VAR & EF_AMDGPU_FEATURE_XNACK_V4) \
== EF_AMDGPU_FEATURE_XNACK_ON_V4)
+#define TEST_XNACK_OFF(VAR) ((VAR & EF_AMDGPU_FEATURE_XNACK_V4) \
+ == EF_AMDGPU_FEATURE_XNACK_OFF_V4)
#define SET_SRAM_ECC_ON(VAR) VAR = ((VAR & ~EF_AMDGPU_FEATURE_SRAMECC_V4) \
| EF_AMDGPU_FEATURE_SRAMECC_ON_V4)
@@ -474,6 +476,7 @@ static void
process_asm (FILE *in, FILE *out, FILE *cfile)
{
int fn_count = 0, var_count = 0, dims_count = 0, regcount_count = 0;
+ bool unified_shared_memory_enabled = false;
struct obstack fns_os, dims_os, regcounts_os;
obstack_init (&fns_os);
obstack_init (&dims_os);
@@ -498,6 +501,7 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
fn_count += 2;
char buf[1000];
+ char dummy;
enum
{ IN_CODE,
IN_METADATA,
@@ -517,6 +521,9 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
dims_count++;
}
+ if (sscanf (buf, " ;; MKOFFLOAD OPTIONS: USM+%c", &dummy) > 0)
+ unified_shared_memory_enabled = true;
+
break;
}
case IN_METADATA:
@@ -565,7 +572,6 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
}
}
- char dummy;
if (sscanf (buf, " .section .gnu.offload_vars%c", &dummy) > 0)
{
state = IN_VARS;
@@ -617,6 +623,7 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
fprintf (cfile, "#include <stdlib.h>\n");
fprintf (cfile, "#include <stdint.h>\n");
fprintf (cfile, "#include <stdbool.h>\n\n");
+ fprintf (cfile, "#include <stdio.h>\n\n");
fprintf (cfile, "static const int gcn_num_vars = %d;\n\n", var_count);
@@ -657,6 +664,34 @@ process_asm (FILE *in, FILE *out, FILE *cfile)
}
fprintf (cfile, "\n};\n\n");
+ /* Emit a constructor function to set the HSA_XNACK environment variable.
+ This must be done before the ROCr runtime library is loaded.
+ We never override a user value (exit empty string), but we do emit a
+ useful diagnostic in the wrong mode (the ROCr message is not good. */
+ if (TEST_XNACK_OFF (elf_flags) && unified_shared_memory_enabled)
+ fatal_error (input_location,
+ "conflicting settings; XNACK is forced off but Unified "
+ "Shared Memory is on");
+ if (!TEST_XNACK_ANY (elf_flags) || unified_shared_memory_enabled)
+ fprintf (cfile,
+ "static __attribute__((constructor))\n"
+ "void configure_xnack (void)\n"
+ "{\n"
+ " const char *val = getenv (\"HSA_XNACK\");\n"
+ " if (!val || val[0] == '\\0')\n"
+ " setenv (\"HSA_XNACK\", \"%d\", true);\n"
+ " else if (%s)\n"
+ " {\n"
+ " fprintf (stderr, \"error: HSA_XNACK=%%s is incompatible; "
+ "please unset\\n\", val);\n"
+ " exit (1);\n"
+ " }\n"
+ "}\n\n",
+ unified_shared_memory_enabled || TEST_XNACK_ON (elf_flags),
+ (unified_shared_memory_enabled || TEST_XNACK_ON (elf_flags)
+ ? "val[0] != '1' || val[1] != '\\0'"
+ : "val[0] == '1' && val[1] == '\\0'"));
+
obstack_free (&fns_os, NULL);
for (i = 0; i < dims_count; i++)
free (dims[i].name);
diff --git a/gcc/omp-low.cc b/gcc/omp-low.cc
index 7d1a2a0d795..239446beb52 100644
--- a/gcc/omp-low.cc
+++ b/gcc/omp-low.cc
@@ -2107,6 +2107,10 @@ create_omp_child_function (omp_context *ctx, bool task_copy)
DECL_ATTRIBUTES (decl)
= tree_cons (get_identifier (target_attr),
NULL_TREE, DECL_ATTRIBUTES (decl));
+ if (flag_offload_memory == OFFLOAD_MEMORY_UNIFIED)
+ DECL_ATTRIBUTES (decl)
+ = tree_cons (get_identifier ("omp unified memory"),
+ NULL_TREE, DECL_ATTRIBUTES (decl));
}
t = build_decl (DECL_SOURCE_LOCATION (decl),
next prev parent reply other threads:[~2022-07-07 10:38 UTC|newest]
Thread overview: 30+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-07-07 10:34 [PATCH 00/17] openmp, nvptx, amdgcn: 5.0 Memory Allocators Andrew Stubbs
2022-07-07 10:34 ` [PATCH 01/17] libgomp, nvptx: low-latency memory allocator Andrew Stubbs
2022-12-08 11:40 ` Jakub Jelinek
2022-07-07 10:34 ` [PATCH 02/17] libgomp: pinned memory Andrew Stubbs
2022-12-08 12:11 ` Jakub Jelinek
2022-12-08 12:51 ` Andrew Stubbs
2022-12-08 14:02 ` Tobias Burnus
2022-12-08 14:35 ` Andrew Stubbs
2022-12-08 15:02 ` Tobias Burnus
2022-07-07 10:34 ` [PATCH 03/17] libgomp, openmp: Add ompx_pinned_mem_alloc Andrew Stubbs
2022-07-07 10:34 ` [PATCH 04/17] openmp, nvptx: low-lat memory access traits Andrew Stubbs
2022-07-07 10:34 ` [PATCH 05/17] openmp, nvptx: ompx_unified_shared_mem_alloc Andrew Stubbs
2022-07-07 10:34 ` [PATCH 06/17] openmp: Add -foffload-memory Andrew Stubbs
2022-07-07 10:34 ` [PATCH 07/17] openmp: allow requires unified_shared_memory Andrew Stubbs
2022-07-07 10:34 ` [PATCH 08/17] openmp: -foffload-memory=pinned Andrew Stubbs
2022-07-07 11:54 ` Tobias Burnus
2022-07-07 22:18 ` Andrew Stubbs
2022-07-08 9:00 ` Tobias Burnus
2022-07-08 9:55 ` Andrew Stubbs
2022-07-08 9:57 ` Tobias Burnus
2023-02-20 14:59 ` Prototype 'GOMP_enable_pinned_mode' (was: [PATCH 08/17] openmp: -foffload-memory=pinned) Thomas Schwinge
2022-07-07 10:34 ` [PATCH 09/17] openmp: Use libgomp memory allocation functions with unified shared memory Andrew Stubbs
2022-07-07 10:34 ` [PATCH 10/17] Add parsing support for allocate directive (OpenMP 5.0) Andrew Stubbs
2022-07-07 10:34 ` [PATCH 11/17] Translate " Andrew Stubbs
2022-07-07 10:34 ` [PATCH 12/17] Handle cleanup of omp allocated variables " Andrew Stubbs
2022-07-07 10:34 ` [PATCH 13/17] Gimplify allocate directive " Andrew Stubbs
2022-07-07 10:34 ` [PATCH 14/17] Lower " Andrew Stubbs
2022-07-07 10:34 ` [PATCH 15/17] amdgcn: Support XNACK mode Andrew Stubbs
2022-07-07 10:34 ` Andrew Stubbs [this message]
2022-07-07 10:34 ` [PATCH 17/17] amdgcn: libgomp plugin USM implementation Andrew Stubbs
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=c73170e6b71179306d40ce6a04a3c02cb4c3cef6.1657188329.git.ams@codesourcery.com \
--to=ams@codesourcery.com \
--cc=gcc-patches@gcc.gnu.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).