Re: [Mesa3d-dev] Mesa (instanced-arrays): translate: Implement instancing for linear SSE run .
Michal,
Did you update the 'C' version of translate for this new functionality?
You can't just extend the fast path - the fallback/default mode needs to
be updated as well.
Also, I'm sure it's not necessary to do a divide per vertex-element to
achieve instancing. It can't be that hard to throw some more counters
at the problem and do this with a couple of adds instead of a divide...
Keith
On Wed, 2009-12-30 at 05:23 -0800, Micha?? Kr??l wrote:
Module: Mesa
Branch: instanced-arrays
Commit: 09c0287b84725098c0b365668231ddf00487c84c
URL:
http://cgit.freedesktop.org/mesa/mesa/commit/?id=09c0287b84725098c0b365668231ddf00487c84c
Author: Michal Krol mic...@vmware.com
Date: Wed Dec 30 14:23:12 2009 +0100
translate: Implement instancing for linear SSE run.
---
src/gallium/auxiliary/translate/translate_sse.c | 154
++-
1 files changed, 120 insertions(+), 34 deletions(-)
diff --git a/src/gallium/auxiliary/translate/translate_sse.c
b/src/gallium/auxiliary/translate/translate_sse.c
index edd0be1..ddfa4c6 100644
--- a/src/gallium/auxiliary/translate/translate_sse.c
+++ b/src/gallium/auxiliary/translate/translate_sse.c
@@ -49,6 +49,7 @@
typedef void (PIPE_CDECL *run_func)( struct translate *translate,
unsigned start,
unsigned count,
+ unsigned instance_id,
void *output_buffer );
typedef void (PIPE_CDECL *run_elts_func)( struct translate *translate,
@@ -59,7 +60,12 @@ typedef void (PIPE_CDECL *run_elts_func)( struct translate
*translate,
struct translate_buffer {
const void *base_ptr;
unsigned stride;
- void *ptr; /* updated per vertex */
+};
+
+struct translate_buffer_varient {
+ unsigned buffer_index;
+ unsigned instance_divisor;
+ void *ptr;/* updated either per vertex or per
instance */
};
@@ -81,6 +87,16 @@ struct translate_sse {
struct translate_buffer buffer[PIPE_MAX_ATTRIBS];
unsigned nr_buffers;
+ /* Multiple buffer varients can map to a single buffer. */
+ struct translate_buffer_varient buffer_varient[PIPE_MAX_ATTRIBS];
+ unsigned nr_buffer_varients;
+
+ /* Multiple elements can map to a single buffer varient. */
+ unsigned element_to_buffer_varient[PIPE_MAX_ATTRIBS];
+
+ boolean use_instancing;
+ unsigned instance_id;
+
run_func gen_run;
run_elts_func gen_run_elts;
@@ -360,31 +376,59 @@ static boolean init_inputs( struct translate_sse *p,
{
unsigned i;
if (linear) {
- for (i = 0; i p-nr_buffers; i++) {
+ struct x86_reg instance_id = x86_make_disp(p-machine_EDX,
+ get_offset(p,
p-instance_id));
+
+ for (i = 0; i p-nr_buffer_varients; i++) {
+ struct translate_buffer_varient *varient = p-buffer_varient[i];
+ struct translate_buffer *buffer = p-buffer[varient-buffer_index];
struct x86_reg buf_stride = x86_make_disp(p-machine_EDX,
- get_offset(p,
p-buffer[i].stride));
+ get_offset(p,
buffer-stride));
struct x86_reg buf_ptr = x86_make_disp(p-machine_EDX,
- get_offset(p,
p-buffer[i].ptr));
+ get_offset(p,
varient-ptr));
struct x86_reg buf_base_ptr = x86_make_disp(p-machine_EDX,
- get_offset(p,
p-buffer[i].base_ptr));
+ get_offset(p,
buffer-base_ptr));
struct x86_reg elt = p-idx_EBX;
- struct x86_reg tmp = p-tmp_EAX;
-
+ struct x86_reg tmp_EAX = p-tmp_EAX;
/* Calculate pointer to first attrib:
+ * base_ptr + stride * index, where index depends on instance
divisor
*/
- x86_mov(p-func, tmp, buf_stride);
- x86_imul(p-func, tmp, elt);
- x86_add(p-func, tmp, buf_base_ptr);
+ if (varient-instance_divisor) {
+/* Our index is instance ID divided by instance divisor.
+ */
+x86_mov(p-func, tmp_EAX, instance_id);
+
+if (varient-instance_divisor != 1) {
+ struct x86_reg tmp_EDX = p-machine_EDX;
+ struct x86_reg tmp_ECX = p-outbuf_ECX;
+
+ /* TODO: Add x86_shr() to rtasm and use it whenever
+* instance divisor is power of two.
+*/
+
+ x86_push(p-func, tmp_EDX);
+ x86_push(p-func, tmp_ECX);
+ x86_xor(p-func, tmp_EDX, tmp_EDX);
+ x86_mov_reg_imm(p-func, tmp_ECX, varient-instance_divisor);
+ x86_div(p-func,
Re: [Mesa3d-dev] Mesa (instanced-arrays): translate: Implement instancing for linear SSE run .
Keith Whitwell wrote on 2009-12-30 16:22:
Michal,
Did you update the 'C' version of translate for this new functionality?
You can't just extend the fast path - the fallback/default mode needs to
be updated as well.
Yes, I did that in the previous commit.
Also, I'm sure it's not necessary to do a divide per vertex-element to
achieve instancing. It can't be that hard to throw some more counters
at the problem and do this with a couple of adds instead of a divide...
The division is done once per instance, not per every vertex attrib. Are
you serious about optimising such low-profile things?
Keith
On Wed, 2009-12-30 at 05:23 -0800, Micha?? Kr??l wrote:
Module: Mesa
Branch: instanced-arrays
Commit: 09c0287b84725098c0b365668231ddf00487c84c
URL:
http://cgit.freedesktop.org/mesa/mesa/commit/?id=09c0287b84725098c0b365668231ddf00487c84c
Author: Michal Krol mic...@vmware.com
Date: Wed Dec 30 14:23:12 2009 +0100
translate: Implement instancing for linear SSE run.
---
src/gallium/auxiliary/translate/translate_sse.c | 154
++-
1 files changed, 120 insertions(+), 34 deletions(-)
diff --git a/src/gallium/auxiliary/translate/translate_sse.c
b/src/gallium/auxiliary/translate/translate_sse.c
index edd0be1..ddfa4c6 100644
--- a/src/gallium/auxiliary/translate/translate_sse.c
+++ b/src/gallium/auxiliary/translate/translate_sse.c
@@ -49,6 +49,7 @@
typedef void (PIPE_CDECL *run_func)( struct translate *translate,
unsigned start,
unsigned count,
+ unsigned instance_id,
void *output_buffer );
typedef void (PIPE_CDECL *run_elts_func)( struct translate *translate,
@@ -59,7 +60,12 @@ typedef void (PIPE_CDECL *run_elts_func)( struct
translate *translate,
struct translate_buffer {
const void *base_ptr;
unsigned stride;
- void *ptr; /* updated per vertex */
+};
+
+struct translate_buffer_varient {
+ unsigned buffer_index;
+ unsigned instance_divisor;
+ void *ptr;/* updated either per vertex or per
instance */
};
@@ -81,6 +87,16 @@ struct translate_sse {
struct translate_buffer buffer[PIPE_MAX_ATTRIBS];
unsigned nr_buffers;
+ /* Multiple buffer varients can map to a single buffer. */
+ struct translate_buffer_varient buffer_varient[PIPE_MAX_ATTRIBS];
+ unsigned nr_buffer_varients;
+
+ /* Multiple elements can map to a single buffer varient. */
+ unsigned element_to_buffer_varient[PIPE_MAX_ATTRIBS];
+
+ boolean use_instancing;
+ unsigned instance_id;
+
run_func gen_run;
run_elts_func gen_run_elts;
@@ -360,31 +376,59 @@ static boolean init_inputs( struct translate_sse *p,
{
unsigned i;
if (linear) {
- for (i = 0; i p-nr_buffers; i++) {
+ struct x86_reg instance_id = x86_make_disp(p-machine_EDX,
+ get_offset(p,
p-instance_id));
+
+ for (i = 0; i p-nr_buffer_varients; i++) {
+ struct translate_buffer_varient *varient = p-buffer_varient[i];
+ struct translate_buffer *buffer =
p-buffer[varient-buffer_index];
struct x86_reg buf_stride = x86_make_disp(p-machine_EDX,
- get_offset(p,
p-buffer[i].stride));
+ get_offset(p,
buffer-stride));
struct x86_reg buf_ptr = x86_make_disp(p-machine_EDX,
- get_offset(p,
p-buffer[i].ptr));
+ get_offset(p,
varient-ptr));
struct x86_reg buf_base_ptr = x86_make_disp(p-machine_EDX,
- get_offset(p,
p-buffer[i].base_ptr));
+ get_offset(p,
buffer-base_ptr));
struct x86_reg elt = p-idx_EBX;
- struct x86_reg tmp = p-tmp_EAX;
-
+ struct x86_reg tmp_EAX = p-tmp_EAX;
/* Calculate pointer to first attrib:
+ * base_ptr + stride * index, where index depends on instance
divisor
*/
- x86_mov(p-func, tmp, buf_stride);
- x86_imul(p-func, tmp, elt);
- x86_add(p-func, tmp, buf_base_ptr);
+ if (varient-instance_divisor) {
+/* Our index is instance ID divided by instance divisor.
+ */
+x86_mov(p-func, tmp_EAX, instance_id);
+
+if (varient-instance_divisor != 1) {
+ struct x86_reg tmp_EDX = p-machine_EDX;
+ struct x86_reg tmp_ECX = p-outbuf_ECX;
+
+ /* TODO: Add x86_shr() to rtasm and use it whenever
+* instance divisor is power of two.
+*/
+
+
Re: [Mesa3d-dev] Mesa (instanced-arrays): translate: Implement instancing for linear SSE run .
On Wed, 2009-12-30 at 07:37 -0800, michal wrote:
Keith Whitwell wrote on 2009-12-30 16:22:
Michal,
Did you update the 'C' version of translate for this new functionality?
You can't just extend the fast path - the fallback/default mode needs to
be updated as well.
Yes, I did that in the previous commit.
OK. I was confused as your other changes to translate are labelled that
way in the commit message, while this was buried inside what looked like
a softpipe change.
Also, I'm sure it's not necessary to do a divide per vertex-element to
achieve instancing. It can't be that hard to throw some more counters
at the problem and do this with a couple of adds instead of a divide...
The division is done once per instance, not per every vertex attrib.
OK, I misread that also...
Are you serious about optimising such low-profile things?
No, once per invocation is fine. Sorry for the confusion.
Keith
Keith
On Wed, 2009-12-30 at 05:23 -0800, Micha?? Kr??l wrote:
Module: Mesa
Branch: instanced-arrays
Commit: 09c0287b84725098c0b365668231ddf00487c84c
URL:
http://cgit.freedesktop.org/mesa/mesa/commit/?id=09c0287b84725098c0b365668231ddf00487c84c
Author: Michal Krol mic...@vmware.com
Date: Wed Dec 30 14:23:12 2009 +0100
translate: Implement instancing for linear SSE run.
---
src/gallium/auxiliary/translate/translate_sse.c | 154
++-
1 files changed, 120 insertions(+), 34 deletions(-)
diff --git a/src/gallium/auxiliary/translate/translate_sse.c
b/src/gallium/auxiliary/translate/translate_sse.c
index edd0be1..ddfa4c6 100644
--- a/src/gallium/auxiliary/translate/translate_sse.c
+++ b/src/gallium/auxiliary/translate/translate_sse.c
@@ -49,6 +49,7 @@
typedef void (PIPE_CDECL *run_func)( struct translate *translate,
unsigned start,
unsigned count,
+ unsigned instance_id,
void *output_buffer );
typedef void (PIPE_CDECL *run_elts_func)( struct translate *translate,
@@ -59,7 +60,12 @@ typedef void (PIPE_CDECL *run_elts_func)( struct
translate *translate,
struct translate_buffer {
const void *base_ptr;
unsigned stride;
- void *ptr; /* updated per vertex */
+};
+
+struct translate_buffer_varient {
+ unsigned buffer_index;
+ unsigned instance_divisor;
+ void *ptr;/* updated either per vertex or per
instance */
};
@@ -81,6 +87,16 @@ struct translate_sse {
struct translate_buffer buffer[PIPE_MAX_ATTRIBS];
unsigned nr_buffers;
+ /* Multiple buffer varients can map to a single buffer. */
+ struct translate_buffer_varient buffer_varient[PIPE_MAX_ATTRIBS];
+ unsigned nr_buffer_varients;
+
+ /* Multiple elements can map to a single buffer varient. */
+ unsigned element_to_buffer_varient[PIPE_MAX_ATTRIBS];
+
+ boolean use_instancing;
+ unsigned instance_id;
+
run_func gen_run;
run_elts_func gen_run_elts;
@@ -360,31 +376,59 @@ static boolean init_inputs( struct translate_sse *p,
{
unsigned i;
if (linear) {
- for (i = 0; i p-nr_buffers; i++) {
+ struct x86_reg instance_id = x86_make_disp(p-machine_EDX,
+ get_offset(p,
p-instance_id));
+
+ for (i = 0; i p-nr_buffer_varients; i++) {
+ struct translate_buffer_varient *varient = p-buffer_varient[i];
+ struct translate_buffer *buffer =
p-buffer[varient-buffer_index];
struct x86_reg buf_stride = x86_make_disp(p-machine_EDX,
- get_offset(p,
p-buffer[i].stride));
+ get_offset(p,
buffer-stride));
struct x86_reg buf_ptr = x86_make_disp(p-machine_EDX,
- get_offset(p,
p-buffer[i].ptr));
+ get_offset(p,
varient-ptr));
struct x86_reg buf_base_ptr = x86_make_disp(p-machine_EDX,
- get_offset(p,
p-buffer[i].base_ptr));
+ get_offset(p,
buffer-base_ptr));
struct x86_reg elt = p-idx_EBX;
- struct x86_reg tmp = p-tmp_EAX;
-
+ struct x86_reg tmp_EAX = p-tmp_EAX;
/* Calculate pointer to first attrib:
+ * base_ptr + stride * index, where index depends on instance
divisor
*/
- x86_mov(p-func, tmp, buf_stride);
- x86_imul(p-func, tmp, elt);
- x86_add(p-func, tmp, buf_base_ptr);
+ if (varient-instance_divisor) {
+/* Our index is instance ID divided by instance
