[Mesa-dev] [PATCH] glsl: optimize copy_propagation_elements pass
Tapani Pälli
tapani.palli at intel.com
Fri Sep 30 11:55:53 UTC 2016
On 09/30/2016 03:18 AM, Ian Romanick wrote:
> On 09/29/2016 12:17 AM, Tapani Pälli wrote:
>>
>> On 09/28/2016 06:14 PM, Ian Romanick wrote:
>>> On 09/16/2016 06:21 PM, Tapani Pälli wrote:
>>>> Changes make copy_propagation_elements pass faster, reducing link
>>>> time spent in test case of bug 94477. Does not fix the actual issue
>>>> but brings down the total time. No regressions seen in CI.
>>>
>>> How does this affect the time of a full shader-db run?
>>
>> Almost none at all, this is the open-source shaders (100 runs):
>>
>> Difference at 95.0% confidence
>> 0.0312 +/- 0.00502746
>> 1.72566% +/- 0.278068%
>> (Student's t, pooled s = 0.0181375)
>>
>> (testing with DOTA-2 shaders gave very similar result)
>>
>> My assumption is that this really helps only the most pathological cases
>> like in the bug where list size becomes enormous (thousands of entries).
>> With just few entries, list is 'fast enough' to walk through anyway (?)
>>
>> BTW Eric was proposing to just remove this pass. However when testing
>> what happens on removal I noticed there's functional failures
>> (arb_gpu_shader5-interpolateAtSample-dynamically-nonuniform starts to
>> fail), so it seems we are currently dependent on this pass.
>>
>>> There are a bunch of bits of this that are confusing to me. I think
>>> some high-level explanation about which hash tables the acp_ref can be
>>> in, which lists it can be in, and how they relate would help. I've
>>> pointed out a couple of the confusing bits below.
>>>
>>>> Signed-off-by: Tapani Pälli <tapani.palli at intel.com>
>>>> ---
>>>>
>>>> For performance measurements, Martina reported in the bug 8x speedup
>>>> to the test case shader link time when using this patch together with
>>>> commit 2cd02e30d2e1677762d34f1831b8e609970ef0f3
>>>>
>>>> .../glsl/opt_copy_propagation_elements.cpp | 187
>>>> ++++++++++++++++-----
>>>> 1 file changed, 145 insertions(+), 42 deletions(-)
>>>>
>>>> diff --git a/src/compiler/glsl/opt_copy_propagation_elements.cpp
>>>> b/src/compiler/glsl/opt_copy_propagation_elements.cpp
>>>> index e4237cc..1c5060a 100644
>>>> --- a/src/compiler/glsl/opt_copy_propagation_elements.cpp
>>>> +++ b/src/compiler/glsl/opt_copy_propagation_elements.cpp
>>>> @@ -46,6 +46,7 @@
>>>> #include "ir_basic_block.h"
>>>> #include "ir_optimization.h"
>>>> #include "compiler/glsl_types.h"
>>>> +#include "util/hash_table.h"
>>>>
>>>> static bool debug = false;
>>>>
>>>> @@ -76,6 +77,18 @@ public:
>>>> int swizzle[4];
>>>> };
>>>>
>>>> +/* Class that refers to acp_entry in another exec_list. Used
>>>> + * when making removals based on rhs.
>>>> + */
>>>> +class acp_ref : public exec_node
>>>
>>> This pattern is called a box, so maybe acp_box would be a better name.
>>> I'm not too hung up on it.
>>>
>>> With this change, can the acp_entry itself still be in a list?
>>
>> The idea here is a class that only refers to a acp_entry but does not
>> take any ownership .. so it's really just a list of pointers. I'm OK
>> with renaming it.
>
> If only a boxed acp_entry can be in a list, then acp_entry doesn't need
> to derive from exec_node. That's why I was asking. I looked at the
> rest of the code again, and I now see that the acp_entry is in the lhs list.
>
> Correct me if I'm wrong, but an entry will effectively be in two lists
> at all times: the lhs list and the rhs list.
>
> Assuming that previous assumption is correct, I might suggest a
> different structure that makes it all less confusing. Don't make
> acp_entry drive from exec_node. Instead, embed two acp_ref (or whatever
> it ends up being called) nodes in the acp_entry:
>
> acp_ref lhs_node;
> acp_ref rhs_node;
>
> When adding an entry to the lhs list, use
>
> lhs_list->push_tail(&entry->lhs_node);
>
> Similar for rhs list:
>
> rhs_list->push_tail(&entry->rhs_node);
>
> Walk the lists like:
>
> foreach_in_list_safe(acp_ref, ref, rhs_list) {
> acp_entry *entry = ref->entry;
>
> ...
> }
>
> I think this would be a lot more clear because both lists are handled in
> the same way. It also avoids the overhead of allocating the boxes.
TBH I'm not sure at this point if this will make end implementation
simpler or more complex but I'll give this change a try :)
>>>> +{
>>>> +public:
>>>> + acp_ref(acp_entry *e)
>>>> + {
>>>> + entry = e;
>>>> + }
>>>> + acp_entry *entry;
>>>> +};
>>>>
>>>> class kill_entry : public exec_node
>>>> {
>>>> @@ -98,14 +111,42 @@ public:
>>>> this->killed_all = false;
>>>> this->mem_ctx = ralloc_context(NULL);
>>>> this->shader_mem_ctx = NULL;
>>>> - this->acp = new(mem_ctx) exec_list;
>>>> this->kills = new(mem_ctx) exec_list;
>>>> +
>>>> + create_acp();
>>>> }
>>>> ~ir_copy_propagation_elements_visitor()
>>>> {
>>>> ralloc_free(mem_ctx);
>>>> }
>>>>
>>>> + void create_acp()
>>>> + {
>>>> + lhs_ht = _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
>>>> + _mesa_key_pointer_equal);
>>>> + rhs_ht = _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
>>>> + _mesa_key_pointer_equal);
>>>> + }
>>>> +
>>>> + void destroy_acp()
>>>> + {
>>>> + _mesa_hash_table_destroy(lhs_ht, NULL);
>>>> + _mesa_hash_table_destroy(rhs_ht, NULL);
>>>> + }
>>>> +
>>>> + void populate_acp(hash_table *lhs, hash_table *rhs)
>>>> + {
>>>> + struct hash_entry *entry;
>>>> + hash_table_foreach(lhs, entry)
>>>> + {
>>>
>>> Opening { on the hash_table_foreach line.
>>>
>>>> + _mesa_hash_table_insert(lhs_ht, entry->key, entry->data);
>>>> + }
>>>
>>> Blank line here.
>>>
>>>> + hash_table_foreach(rhs, entry)
>>>> + {
>>>
>>> Opening { on the hash_table_foreach line.
>>>
>>>> + _mesa_hash_table_insert(rhs_ht, entry->key, entry->data);
>>>> + }
>>>> + }
>>
>> thanks, will fix these
>>
>>>> +
>>>> void handle_loop(ir_loop *, bool keep_acp);
>>>> virtual ir_visitor_status visit_enter(class ir_loop *);
>>>> virtual ir_visitor_status visit_enter(class ir_function_signature
>>>> *);
>>>> @@ -120,8 +161,10 @@ public:
>>>> void kill(kill_entry *k);
>>>> void handle_if_block(exec_list *instructions);
>>>>
>>>> - /** List of acp_entry: The available copies to propagate */
>>>> - exec_list *acp;
>>>> + /** Hash of acp_entry: The available copies to propagate */
>>>> + hash_table *lhs_ht;
>>>> + hash_table *rhs_ht;
>>>> +
>>>> /**
>>>> * List of kill_entry: The variables whose values were killed in
>>>> this
>>>> * block.
>>>> @@ -147,23 +190,29 @@
>>>> ir_copy_propagation_elements_visitor::visit_enter(ir_function_signature
>>>> *ir)
>>>> * block. Any instructions at global scope will be shuffled into
>>>> * main() at link time, so they're irrelevant to us.
>>>> */
>>>> - exec_list *orig_acp = this->acp;
>>>> exec_list *orig_kills = this->kills;
>>>> bool orig_killed_all = this->killed_all;
>>>>
>>>> - this->acp = new(mem_ctx) exec_list;
>>>> + hash_table *orig_lhs_ht = lhs_ht;
>>>> + hash_table *orig_rhs_ht = rhs_ht;
>>>> +
>>>> this->kills = new(mem_ctx) exec_list;
>>>
>>> Orthogonal to this patch, there is some efficiency to be gained by not
>>> doing this extra memory allocation. Doing
>>>
>>> this->kills.move_nodes_to(&orig_kills);
>>>
>>> here, and
>>>
>>> assert(this->kills.is_empty());
>>> orig_kills.move_nodes_to(&this->kills);
>>>
>>> below is sufficient. You'll also have to convert kills from exec_list*
>>> to just exec_list. One nice thing about that is it will just delete
>>> code. :)
>>
>> Sounds good to me if we can squeeze a bit more off, I can try adding this.
>
> I'd suggest doing this as a separate patch that comes first. It should
> be a small change, so we ought to be able to land that without too much
> debate. :)
agreed, will make this a separate patch
>>>> this->killed_all = false;
>>>>
>>>> + create_acp();
>>>> +
>>>> visit_list_elements(this, &ir->body);
>>>>
>>>> - ralloc_free(this->acp);
>>>> ralloc_free(this->kills);
>>>>
>>>> + destroy_acp();
>>>> +
>>>> this->kills = orig_kills;
>>>> - this->acp = orig_acp;
>>>> this->killed_all = orig_killed_all;
>>>>
>>>> + lhs_ht = orig_lhs_ht;
>>>> + rhs_ht = orig_rhs_ht;
>>>> +
>>>> return visit_continue_with_parent;
>>>> }
>>>>
>>>> @@ -249,17 +298,19 @@
>>>> ir_copy_propagation_elements_visitor::handle_rvalue(ir_rvalue **ir)
>>>> /* Try to find ACP entries covering swizzle_chan[], hoping they're
>>>> * the same source variable.
>>>> */
>>>> - foreach_in_list(acp_entry, entry, this->acp) {
>>>> - if (var == entry->lhs) {
>>>> - for (int c = 0; c < chans; c++) {
>>>> - if (entry->write_mask & (1 << swizzle_chan[c])) {
>>>> - source[c] = entry->rhs;
>>>> - source_chan[c] = entry->swizzle[swizzle_chan[c]];
>>>> + hash_entry *ht_entry = _mesa_hash_table_search(lhs_ht, var);
>>>> + if (ht_entry) {
>>>> + exec_list *ht_list = (exec_list *) ht_entry->data;
>>>> + foreach_in_list(acp_entry, entry, ht_list) {
>>>> + for (int c = 0; c < chans; c++) {
>>>> + if (entry->write_mask & (1 << swizzle_chan[c])) {
>>>> + source[c] = entry->rhs;
>>>> + source_chan[c] = entry->swizzle[swizzle_chan[c]];
>>>>
>>>> if (source_chan[c] != swizzle_chan[c])
>>>> noop_swizzle = false;
>>>> - }
>>>> - }
>>>> + }
>>>
>>> I think this indentation is off.
>>
>> will fix
>>
>>>> + }
>>>> }
>>>> }
>>>>
>>>> @@ -319,7 +370,9 @@
>>>> ir_copy_propagation_elements_visitor::visit_enter(ir_call *ir)
>>>> /* Since we're unlinked, we don't (necessarily) know the side
>>>> effects of
>>>> * this call. So kill all copies.
>>>> */
>>>> - acp->make_empty();
>>>> + _mesa_hash_table_clear(lhs_ht, NULL);
>>>> + _mesa_hash_table_clear(rhs_ht, NULL);
>>>> +
>>>> this->killed_all = true;
>>>>
>>>> return visit_continue_with_parent;
>>>> @@ -328,31 +381,36 @@
>>>> ir_copy_propagation_elements_visitor::visit_enter(ir_call *ir)
>>>> void
>>>> ir_copy_propagation_elements_visitor::handle_if_block(exec_list
>>>> *instructions)
>>>> {
>>>> - exec_list *orig_acp = this->acp;
>>>> exec_list *orig_kills = this->kills;
>>>> bool orig_killed_all = this->killed_all;
>>>>
>>>> - this->acp = new(mem_ctx) exec_list;
>>>> + hash_table *orig_lhs_ht = lhs_ht;
>>>> + hash_table *orig_rhs_ht = rhs_ht;
>>>> +
>>>> this->kills = new(mem_ctx) exec_list;
>>>> this->killed_all = false;
>>>>
>>>> + create_acp();
>>>> +
>>>> /* Populate the initial acp with a copy of the original */
>>>> - foreach_in_list(acp_entry, a, orig_acp) {
>>>> - this->acp->push_tail(new(this->acp) acp_entry(a));
>>>> - }
>>>> + populate_acp(orig_lhs_ht, orig_rhs_ht);
>>>>
>>>> visit_list_elements(this, instructions);
>>>>
>>>> if (this->killed_all) {
>>>> - orig_acp->make_empty();
>>>> + _mesa_hash_table_clear(orig_lhs_ht, NULL);
>>>> + _mesa_hash_table_clear(orig_rhs_ht, NULL);
>>>> }
>>>>
>>>> exec_list *new_kills = this->kills;
>>>> this->kills = orig_kills;
>>>> - ralloc_free(this->acp);
>>>> - this->acp = orig_acp;
>>>> this->killed_all = this->killed_all || orig_killed_all;
>>>>
>>>> + destroy_acp();
>>>> +
>>>> + lhs_ht = orig_lhs_ht;
>>>> + rhs_ht = orig_rhs_ht;
>>>> +
>>>> /* Move the new kills into the parent block's list, removing them
>>>> * from the parent's ACP list in the process.
>>>> */
>>>> @@ -378,37 +436,42 @@
>>>> ir_copy_propagation_elements_visitor::visit_enter(ir_if *ir)
>>>> void
>>>> ir_copy_propagation_elements_visitor::handle_loop(ir_loop *ir, bool
>>>> keep_acp)
>>>> {
>>>> - exec_list *orig_acp = this->acp;
>>>> exec_list *orig_kills = this->kills;
>>>> bool orig_killed_all = this->killed_all;
>>>>
>>>> + hash_table *orig_lhs_ht = lhs_ht;
>>>> + hash_table *orig_rhs_ht = rhs_ht;
>>>> +
>>>> /* FINISHME: For now, the initial acp for loops is totally empty.
>>>> * We could go through once, then go through again with the acp
>>>> * cloned minus the killed entries after the first run through.
>>>> */
>>>> - this->acp = new(mem_ctx) exec_list;
>>>> this->kills = new(mem_ctx) exec_list;
>>>> this->killed_all = false;
>>>>
>>>> + create_acp();
>>>> +
>>>> if (keep_acp) {
>>>> /* Populate the initial acp with a copy of the original */
>>>> - foreach_in_list(acp_entry, a, orig_acp) {
>>>> - this->acp->push_tail(new(this->acp) acp_entry(a));
>>>> - }
>>>> + populate_acp(orig_lhs_ht, orig_rhs_ht);
>>>> }
>>>>
>>>> visit_list_elements(this, &ir->body_instructions);
>>>>
>>>> if (this->killed_all) {
>>>> - orig_acp->make_empty();
>>>> + _mesa_hash_table_clear(orig_lhs_ht, NULL);
>>>> + _mesa_hash_table_clear(orig_rhs_ht, NULL);
>>>> }
>>>>
>>>> exec_list *new_kills = this->kills;
>>>> this->kills = orig_kills;
>>>> - ralloc_free(this->acp);
>>>> - this->acp = orig_acp;
>>>> this->killed_all = this->killed_all || orig_killed_all;
>>>>
>>>> + destroy_acp();
>>>> +
>>>> + lhs_ht = orig_lhs_ht;
>>>> + rhs_ht = orig_rhs_ht;
>>>> +
>>>> foreach_in_list_safe(kill_entry, k, new_kills) {
>>>> kill(k);
>>>> }
>>>> @@ -430,16 +493,33 @@
>>>> ir_copy_propagation_elements_visitor::visit_enter(ir_loop *ir)
>>>> void
>>>> ir_copy_propagation_elements_visitor::kill(kill_entry *k)
>>>> {
>>>> - foreach_in_list_safe(acp_entry, entry, acp) {
>>>> - if (entry->lhs == k->var) {
>>>> - entry->write_mask = entry->write_mask & ~k->write_mask;
>>>> - if (entry->write_mask == 0) {
>>>> - entry->remove();
>>>> - continue;
>>>> - }
>>>> + /* removal of lhs entries */
>>>> + hash_entry *ht_entry = _mesa_hash_table_search(lhs_ht, k->var);
>>>> + if (ht_entry) {
>>>> + exec_list *lhs_list = (exec_list *) ht_entry->data;
>>>> + foreach_in_list_safe(acp_entry, entry, lhs_list) {
>>>> + entry->write_mask = entry->write_mask & ~k->write_mask;
>>>> + if (entry->write_mask == 0) {
>>>> + entry->remove();
>>>> + continue;
>>>> + }
>>>> }
>>>> - if (entry->rhs == k->var) {
>>>> - entry->remove();
>>>> + }
>>>> +
>>>> + /* removal of rhs entries */
>>>> + ht_entry = _mesa_hash_table_search(rhs_ht, k->var);
>>>> + if (ht_entry) {
>>>> + exec_list *rhs_list = (exec_list *) ht_entry->data;
>>>> + foreach_in_list_safe(acp_ref, ref, rhs_list) {
>>>> + acp_entry *entry = ref->entry;
>>>> + /* If entry is still in a list (not already removed by lhs
>>>> entry
>>>> + * removal above), remove it.
>>>> + */
>>>> + if (entry->prev || entry->next)
>>>> + entry->remove();
>>>
>>> This seems weird.
>>
>> Why it's done this way is to separate rhs entry removal from lhs entry
>> removal so that we would not end up seeking rhs hash for each lhs entry
>> removed. Existing implementation removes from exec_list if lhs matches
>> or rhs matches, I did not find way to do that efficiently in one pass.
>>
>> So what's done instead is that in the first loop (lhs entries) we remove
>> based on lhs. Now when iterating 'rhs_ht' some of the entries have been
>> potentially been removed from list already. This checks if that has
>> happened, if not we removal here instead.
>>
>>>> +
>>>> + /* remove from acp_ref list */
>>>> + ref->remove();
>>>
>>> What list is ref in?
>>
>> This is list of 'acp_ref' in rhs_ht. So first we removed this from
>> exec_list that is in lhs_ht and now we remove from rhs_ht lists. I agree
>> that this overall structure is a bit confusing but I did not find other
>> solution that would be as fast. Comments below (when adding new items to
>> hashes) tries to document the overall structure, maybe some more
>> comments required there?
>>
>>
>>>> }
>>>> }
>>>>
>>>> @@ -513,7 +593,30 @@
>>>> ir_copy_propagation_elements_visitor::add_copy(ir_assignment *ir)
>>>>
>>>> entry = new(this->mem_ctx) acp_entry(lhs->var, rhs->var, write_mask,
>>>> swizzle);
>>>> - this->acp->push_tail(entry);
>>>> +
>>>> + /* lhs hash, hash of lhs -> acp_entry lists */
>>>> + hash_entry *ht_entry = _mesa_hash_table_search(lhs_ht, lhs->var);
>>>> + if (ht_entry) {
>>>> + exec_list *lhs_list = (exec_list *) ht_entry->data;
>>>> + lhs_list->push_tail(entry);
>>>> + } else {
>>>> + exec_list *lhs_list = new(mem_ctx) exec_list;
>>>> + lhs_list->push_tail(entry);
>>>> + _mesa_hash_table_insert(lhs_ht, lhs->var, lhs_list);
>>>> + }
>>>> +
>>>> + acp_ref *ref = new(mem_ctx) acp_ref(entry);
>>>> +
>>>> + /* rhs hash, hash of rhs -> acp_entry pointers to lhs lists */
>>>> + ht_entry = _mesa_hash_table_search(rhs_ht, rhs->var);
>>>> + if (ht_entry) {
>>>> + exec_list *rhs_list = (exec_list *) ht_entry->data;
>>>> + rhs_list->push_tail(ref);
>>>> + } else {
>>>> + exec_list *rhs_list = new(mem_ctx) exec_list;
>>>> + rhs_list->push_tail(ref);
>>>> + _mesa_hash_table_insert(rhs_ht, rhs->var, rhs_list);
>>>> + }
>>>> }
>>>>
>>>> bool
>
More information about the mesa-dev
mailing list