[Mesa-dev] [PATCH] st/glsl_to_tgsi: drop the merge_registers() pass
Rob Clark
robdclark at gmail.com
Mon Apr 24 21:22:45 UTC 2017
On Mon, Apr 24, 2017 at 5:18 PM, Samuel Pitoiset
<samuel.pitoiset at gmail.com> wrote:
>
>
> On 04/24/2017 11:12 PM, Rob Clark wrote:
>>
>> so I guess this is likely to hurt pipe drivers that don't (yet?) have
>> a real compiler backend. (Ie. etnaviv and freedreno/a2xx.) So maybe
>> it should be optional.
>
>
> I can't say for these drivers, but it seems safer to add an option if this
> can hurt them. How about PIPE_SHADER_CAP_TGSI_MERGE_REGISTERS?
so we have already PIPE_CAP_GLSL_OPTIMIZE_CONSERVATIVELY .. not sure
why that is a cap and not a shader-cap, tbh. I guess I'm ok with
either bikeshed color ;-)
BR,
-R
>
>>
>> Also I wonder about the pre-llvm radeon gen's, since sb uses the
>> actual instruction encoding for IR between tgsi->sb and backend opt
>> passes.. iirc they have had problems when the tgsi code uses too many
>> registers.
>>
>> BR,
>> -R
>>
>> On Mon, Apr 24, 2017 at 5:01 PM, Samuel Pitoiset
>> <samuel.pitoiset at gmail.com> wrote:
>>>
>>> The main goal of this pass to merge temporary registers in order
>>> to reduce the total number of registers and also to produce
>>> optimal TGSI code.
>>>
>>> In fact, compilers seem to be confused when temporary variables
>>> are already merged, maybe because it's done too early in the
>>> process.
>>>
>>> Removing the pass, reduce both the register pressure and the code
>>> size (TGSI is no longer optimized, but who cares?).
>>> shader-db results with RadeonSI and Nouveau are interesting.
>>>
>>> Nouveau:
>>>
>>> total instructions in shared programs : 3931608 -> 3929463 (-0.05%)
>>> total gprs used in shared programs : 481255 -> 479014 (-0.47%)
>>> total local used in shared programs : 27481 -> 27381 (-0.36%)
>>> total bytes used in shared programs : 36031256 -> 36011120 (-0.06%)
>>>
>>> local gpr inst bytes
>>> helped 14 1471 1309 1309
>>> hurt 1 88 384 384
>>>
>>> RadeonSI:
>>>
>>> PERCENTAGE DELTAS Shaders SGPRs VGPRs SpillSGPR SpillVGPR
>>> PrivVGPR Scratch CodeSize MaxWaves Waits
>>>
>>> ----------------------------------------------------------------------------------------------------------------------
>>> All affected 4906 -0.31 % -0.40 % -2.93 % -20.00 %
>>> . -20.00 % -0.18 % 0.19 % .
>>>
>>> ----------------------------------------------------------------------------------------------------------------------
>>> Total 47109 -0.04 % -0.05 % -1.97 % -7.14 %
>>> . -0.30 % -0.03 % 0.02 % .
>>>
>>> Found by luck while fixing an issue in the TGSI dead code elimination
>>> pass which affects tex instructions with bindless samplers.
>>>
>>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>>> ---
>>> src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 62
>>> ------------------------------
>>> 1 file changed, 62 deletions(-)
>>>
>>> diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>>> b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>>> index de7fe7837a..d033bdcc5a 100644
>>> --- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>>> +++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>>> @@ -565,7 +565,6 @@ public:
>>> int eliminate_dead_code(void);
>>>
>>> void merge_two_dsts(void);
>>> - void merge_registers(void);
>>> void renumber_registers(void);
>>>
>>> void emit_block_mov(ir_assignment *ir, const struct glsl_type *type,
>>> @@ -5262,66 +5261,6 @@ glsl_to_tgsi_visitor::merge_two_dsts(void)
>>> }
>>> }
>>>
>>> -/* Merges temporary registers together where possible to reduce the
>>> number of
>>> - * registers needed to run a program.
>>> - *
>>> - * Produces optimal code only after copy propagation and dead code
>>> elimination
>>> - * have been run. */
>>> -void
>>> -glsl_to_tgsi_visitor::merge_registers(void)
>>> -{
>>> - int *last_reads = rzalloc_array(mem_ctx, int, this->next_temp);
>>> - int *first_writes = rzalloc_array(mem_ctx, int, this->next_temp);
>>> - struct rename_reg_pair *renames = rzalloc_array(mem_ctx, struct
>>> rename_reg_pair, this->next_temp);
>>> - int i, j;
>>> - int num_renames = 0;
>>> -
>>> - /* Read the indices of the last read and first write to each temp
>>> register
>>> - * into an array so that we don't have to traverse the instruction
>>> list as
>>> - * much. */
>>> - for (i = 0; i < this->next_temp; i++) {
>>> - last_reads[i] = -1;
>>> - first_writes[i] = -1;
>>> - }
>>> - get_last_temp_read_first_temp_write(last_reads, first_writes);
>>> -
>>> - /* Start looking for registers with non-overlapping usages that can
>>> be
>>> - * merged together. */
>>> - for (i = 0; i < this->next_temp; i++) {
>>> - /* Don't touch unused registers. */
>>> - if (last_reads[i] < 0 || first_writes[i] < 0) continue;
>>> -
>>> - for (j = 0; j < this->next_temp; j++) {
>>> - /* Don't touch unused registers. */
>>> - if (last_reads[j] < 0 || first_writes[j] < 0) continue;
>>> -
>>> - /* We can merge the two registers if the first write to j is
>>> after or
>>> - * in the same instruction as the last read from i. Note that
>>> the
>>> - * register at index i will always be used earlier or at the
>>> same time
>>> - * as the register at index j. */
>>> - if (first_writes[i] <= first_writes[j] &&
>>> - last_reads[i] <= first_writes[j]) {
>>> - renames[num_renames].old_reg = j;
>>> - renames[num_renames].new_reg = i;
>>> - num_renames++;
>>> -
>>> - /* Update the first_writes and last_reads arrays with the
>>> new
>>> - * values for the merged register index, and mark the newly
>>> unused
>>> - * register index as such. */
>>> - assert(last_reads[j] >= last_reads[i]);
>>> - last_reads[i] = last_reads[j];
>>> - first_writes[j] = -1;
>>> - last_reads[j] = -1;
>>> - }
>>> - }
>>> - }
>>> -
>>> - rename_temp_registers(num_renames, renames);
>>> - ralloc_free(renames);
>>> - ralloc_free(last_reads);
>>> - ralloc_free(first_writes);
>>> -}
>>> -
>>> /* Reassign indices to temporary registers by reusing unused indices
>>> created
>>> * by optimization passes. */
>>> void
>>> @@ -6712,7 +6651,6 @@ get_mesa_program_tgsi(struct gl_context *ctx,
>>> while (v->eliminate_dead_code());
>>>
>>> v->merge_two_dsts();
>>> - v->merge_registers();
>>> v->renumber_registers();
>>>
>>> /* Write the END instruction. */
>>> --
>>> 2.12.2
>>>
>>> _______________________________________________
>>> mesa-dev mailing list
>>> mesa-dev at lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
More information about the mesa-dev
mailing list