libdrm: Move intel_atomic.h to libdrm core for sharing. (V2)
PATCH 1/2 libdrm: Move intel_atomic.h to libdrm core for sharing. (V2)
intel_atomic.h includes very usefull atomic operations for lock free parrallel access of variables. Moving these to core libdrm for code sharing with radeon.
V2: Fix remaining references to intel_atomic.h and libdrm-intel.
--- Makefile.am | 2 +- configure.ac | 2 +- intel/intel_atomic.h | 56 +----------------------------- xf86atomic.h | 93 ++++++++++++++++++++++++++++++++++++++++++++++++++ 4 files changed, 96 insertions(+), 57 deletions(-) create mode 100644 xf86atomic.h
Makefile.am +1 -1
@@ -59,7 +59,7 @@ libdrm_la_SOURCES = \ libdrm_lists.h libdrmincludedir = ${includedir} -libdrminclude_HEADERS = xf86drm.h xf86drmMode.h +libdrminclude_HEADERS = xf86drm.h xf86drmMode.h xf86atomic.h EXTRA_DIST = libdrm.pc.in include/drm/*
configure.ac +1 -1
@@ -198,7 +198,7 @@ if test "x$INTEL" != "xno"; then ]) if test "x$drm_cv_atomic_primitives" = xIntel; then - AC_DEFINE(HAVE_INTEL_ATOMIC_PRIMITIVES, 1, + AC_DEFINE(HAVE_LIBDRM_ATOMIC_PRIMITIVES, 1, [Enable if your compiler supports the Intel __sync_* atomic primitives]) fi if test "x$drm_cv_atomic_primitives" = "xlibatomic-ops"; then
intel/intel_atomic.h +1 -55
61 lines not on this page · raw message
xf86atomic.h new file +93 -0
93 lines not on this page · raw message
Signature · 1 line
1.6.3.3
------------------------------------------------------------------------------
Download Intel® Parallel Studio Eval
Try the new software tools for yourself. Speed compiling, find bugs proactively, and fine-tune applications for parallel performance. See why Intel Parallel Studio got high marks during beta.
http://p.sf.net/sfu/intel-sw-dev
--
_______________________________________________
Dri-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/dri-devel
PATCH 2/2 libdrm_radeon: Optimize cs_gem_reloc to do less looping.
bo->referenced_in_cs is checked if bo is already in cs. Adding and removing reference in bo is done with atomic operations to allow parallel access to a bo from multiple contexts.
cs->id generation code quarentees there is not duplicated ids which limits number of cs->ids to 32. If there is more cs objects rest will get id 0.
This optimization decreases cs_write_reloc share of torcs profiling from 4.3% to 2.6%.
--- radeon/radeon_bo_gem.c | 1 + radeon/radeon_bo_int.h | 4 +- radeon/radeon_cs.c | 6 ++ radeon/radeon_cs.h | 2 +- radeon/radeon_cs_gem.c | 123 ++++++++++++++++++++++++++++++++++++------------ radeon/radeon_cs_int.h | 1 + xf86atomic.h | 6 ++ 7 files changed, 110 insertions(+), 33 deletions(-)
radeon/radeon_bo_gem.c +1 -0
@@ -80,6 +80,7 @@ static struct radeon_bo *bo_open(struct radeon_bo_manager *bom, bo->base.domains = domains; bo->base.flags = flags; bo->base.ptr =NULL ; +atomic_set (&bo->base.referenced_in_cs,0 ); bo->map_count =0 ;if (handle) {struct drm_gem_open open_arg;
radeon/radeon_bo_int.h +3 -1
@@ -1,6 +1,8 @@ #ifndef RADEON_BO_INT #define RADEON_BO_INT +#include <xf86atomic.h> + struct radeon_bo_manager { struct radeon_bo_funcs *funcs; int fd;
@@ -17,7 +19,7 @@ struct radeon_bo_int {unsigned cref;struct radeon_bo_manager *bom;uint32_t space_accounted; -uint32_t referenced_in_cs; +atomic_t referenced_in_cs; };/* bo functions */
radeon/radeon_cs.c +6 -0
@@ -88,3 +88,9 @@ void radeon_cs_space_set_flush(struct radeon_cs *cs, void (*fn)(void *), void *d csi->space_flush_fn = fn; csi->space_flush_data = data; } + +uint32_t radeon_cs_get_id (struct radeon_cs *cs) +{ +struct radeon_cs_int *csi = (struct radeon_cs_int *)cs; +return csi->id; +}
radeon/radeon_cs.h +1 -1
@@ -85,7 +85,7 @@ extern int radeon_cs_write_reloc(struct radeon_cs *cs,uint32_t read_domain, uint32_t write_domain, uint32_t flags); - +extern uint32_t radeon_cs_get_id (struct radeon_cs *cs);/* * add a persistent BO to the list * a persistent BO is one that will be referenced across flushes,
radeon/radeon_cs_gem.c +92 -31
178 lines not on this page · raw message
radeon/radeon_cs_int.h +1 -0
@@ -28,6 +28,7 @@ struct radeon_cs_int {int bo_count;void (*space_flush_fn)(void *);void *space_flush_data; +uint32_t id; };/* cs functions */
xf86atomic.h +6 -0
@@ -50,6 +50,8 @@ typedef struct {# define atomic_set (x, val) ((x)->atomic = (val))# define atomic_inc (x) ((void) __sync_fetch_and_add (&(x)->atomic, 1))# define atomic_dec_and_test (x) (__sync_fetch_and_add (&(x)->atomic, -1) == 1) +# define atomic_add (x, v) ((void) __sync_add_and_fetch(&(x)->atomic, (v))) +# define atomic_dec (x, v) ((void) __sync_sub_and_fetch(&(x)->atomic, (v)))# define atomic_cmpxchg (x, oldv, newv) __sync_val_compare_and_swap (&(x)->atomic, oldv, newv)#endif
@@ -66,6 +68,8 @@ typedef struct {# define atomic_read (x) AO_load_full(&(x)->atomic)# define atomic_set (x, val) AO_store_full(&(x)->atomic, (val))# define atomic_inc (x) ((void) AO_fetch_and_add1_full(&(x)->atomic)) +# define atomic_add (x, v) ((void) AO_fetch_and_add_full(&(x)->atomic, (v))) +# define atomic_dec (x, v) ((void) AO_fetch_and_add_full(&(x)->atomic, -(v)))# define atomic_dec_and_test (x) (AO_fetch_and_sub1_full(&(x)->atomic) == 1)# define atomic_cmpxchg (x, oldv, newv) AO_compare_and_swap_full(&(x)->atomic, oldv, newv)
@@ -82,6 +86,8 @@ typedef struct { uint_t atomic; } atomic_t;# define atomic_set (x, val) ((x)->atomic = (uint_t)(val))# define atomic_inc (x) (atomic_inc_uint (&(x)->atomic))# define atomic_dec_and_test (x) (atomic_dec_uint_nv(&(x)->atomic) == 1) +# define atomic_add (x, v) (atomic_add_uint(&(x)->atomic, (v))) +# define atomic_dec (x, v) (atomic_dec_uint(&(x)->atomic, (v)))# define atomic_cmpxchg (x, oldv, newv) atomic_cas_uint (&(x)->atomic, oldv, newv)#endif
Signature · 10 lines
1.6.3.3 ------------------------------------------------------------------------------ Download Intel® Parallel Studio Eval Try the new software tools for yourself. Speed compiling, find bugs proactively, and fine-tune applications for parallel performance. See why Intel Parallel Studio got high marks during beta. http://p.sf.net/sfu/intel-sw-dev --
Re: [PATCH 2/2] libdrm_radeon: Optimize cs_gem_reloc to do less looping.
Quoted 11 lines · Pauli Nieminen
bo->referenced_in_csis checked if bo is already in cs. Adding and removing reference in bo is done with atomic operations to allow parallel access to a bo from multiple contexts.
cs->idgeneration code quarentees there is not duplicated ids which limits number ofcs->idsto 32. If there is more cs objects rest will get id 0.This optimization decreases
cs_write_relocshare of torcs profiling from 4.3% to 2.6%.Signed-off-by Pauli Nieminen <[email protected]>
This also needs something like the below for configure to verify that atomic ops are available.
configure.ac +13 -3
@@ -173,7 +173,7 @@ if test "x$HAVE_LIBUDEV" = xyes; then fi AM_CONDITIONAL(HAVE_LIBUDEV, [test "x$HAVE_LIBUDEV" = xyes]) -if test "x$INTEL" != "xno"; then +if test "x$INTEL" != "xno" || test "x$RADEON" != "xno"; then # Check for atomic intrinsics AC_CACHE_CHECK([for native atomic primitives], drm_cv_atomic_primitives, [
@@ -206,13 +206,23 @@ if test "x$INTEL" != "xno"; then fi if test "x$drm_cv_atomic_primitives" = "xnone"; then - if test "x$INTEL" != "xauto"; then + if test "x$INTEL" = "xyes"; then AC_MSG_ERROR([libdrm_intel depends upon atomic operations, which were not found for your compiler/cpu. Try compiling with -march=native, or install the libatomics-op-dev package, or, failing both of those, disable support for Intel GPUs by passing --disable-intel to ./configure]) else INTEL=no fi + if test "x$RADEON" = "xyes"; then + AC_MSG_ERROR([libdrm_radeon depends upon atomic operations, which were not found for your compiler/cpu. Try compiling with -march=native, or install the libatomics-op-dev package, or, failing both of those, disable support for Radeon GPUs by passing --disable-radeon to ./configure]) + else + RADEON=no + fi else - INTEL=yes + if test "x$INTEL" != "xno"; then + INTEL=yes + fi + if test "x$RADEON" != "xno"; then + RADEON=yes + fi fi fi
Signature · 14 lines
Earthling Michel Dänzer | http://www.vmware.com Libre software enthusiast | Debian, X and DRI developer ------------------------------------------------------------------------------ Download Intel® Parallel Studio Eval Try the new software tools for yourself. Speed compiling, find bugs proactively, and fine-tune applications for parallel performance. See why Intel Parallel Studio got high marks during beta. http://p.sf.net/sfu/intel-sw-dev -- _______________________________________________ Dri-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/dri-devel