[Mesa-dev] [PATCH 1/9] tgsi/scan: add a new pass that analyzes tess factor writes
Nicolai Hähnle
nhaehnle at gmail.com
Thu Sep 7 12:23:02 UTC 2017
On 07.09.2017 12:55, Marek Olšák wrote:
> We can also say if gl_TessLevel* is written multiple times, then one
> these must be true:
> - there must not be a barrier between the writes
> - gl_TessLevel* writes aren't inside conditional blocks
>
> I see that these games use a barrier in TCS:
> - Hitman
> - Grid Autosport
> - Tomb Raider
>
> All constraints are true for those games. If we do this, it should be
> enough: "there must not be a barrier between the writes"
Right, if there are no barriers between the writes, we can always assume
that the writes from invocation 0 land last (obviously we still need to
know that invocation 0 did write, but your patch already checks that).
Cheers,
Nicolai
>
> Marek
>
>
> On Thu, Sep 7, 2017 at 9:13 AM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
>> On 06.09.2017 19:03, Marek Olšák wrote:
>>>
>>> From: Marek Olšák <marek.olsak at amd.com>
>>>
>>> The pass tries to deduce whether tess factors are always written by
>>> invocation 0 (at least).
>>>
>>> The implication for radeonsi is that it doesn't have to use a barrier
>>> near the end of TCS, and doesn't have to use LDS for passing the tess
>>> factors to the epilog.
>>
>>
>> What about the following shader snippet, which is particularly nasty on
>> purpose:
>>
>> gl_TessLevelInner = ...;
>>
>> barrier();
>>
>> if (gl_InvocationID == 1) {
>> gl_TessLevelInner = ...;
>> }
>>
>> The final gl_TessLevelInner must be the one written by invocation 1.
>>
>> The point is, this would be a nice improvement, but it requires us to check
>> that whenever tess factors are written, then that path includes invocation
>> 0.
>>
>> (The written value itself might depend on the invocation ID, but in that
>> case, we can do whatever we like, which means we can just take the value
>> written by invocation 0.)
>>
>> I don't think we can determine this easily with the tools we have, and the
>> effort is likely not worth it.
>>
>> Although you could do a very simplistic approximation of "tess factors are
>> only written outside of control flow", which is probably reasonably common.
>> (Uniform control flow would also work -- but that would require going to the
>> LLVM level).
>>
>> Cheers,
>> Nicolai
>>
>>
>>
>>> ---
>>> src/gallium/auxiliary/tgsi/tgsi_scan.c | 188
>>> +++++++++++++++++++++++++++++++++
>>> src/gallium/auxiliary/tgsi/tgsi_scan.h | 11 ++
>>> 2 files changed, 199 insertions(+)
>>>
>>> diff --git a/src/gallium/auxiliary/tgsi/tgsi_scan.c
>>> b/src/gallium/auxiliary/tgsi/tgsi_scan.c
>>> index db87ce3..612a8c6 100644
>>> --- a/src/gallium/auxiliary/tgsi/tgsi_scan.c
>>> +++ b/src/gallium/auxiliary/tgsi/tgsi_scan.c
>>> @@ -930,10 +930,198 @@ tgsi_scan_arrays(const struct tgsi_token *tokens,
>>> array->writemask |= dst->Register.WriteMask;
>>> }
>>> }
>>> }
>>> }
>>> tgsi_parse_free(&parse);
>>> return;
>>> }
>>> +
>>> +static void
>>> +check_no_subroutines(const struct tgsi_full_instruction *inst)
>>> +{
>>> + switch (inst->Instruction.Opcode) {
>>> + case TGSI_OPCODE_BGNSUB:
>>> + case TGSI_OPCODE_ENDSUB:
>>> + case TGSI_OPCODE_CAL:
>>> + unreachable("subroutines unhandled");
>>> + }
>>> +}
>>> +
>>> +static unsigned
>>> +get_inst_tessfactor_writemask(const struct tgsi_shader_info *info,
>>> + const struct tgsi_full_instruction *inst)
>>> +{
>>> + unsigned writemask = 0;
>>> +
>>> + for (unsigned i = 0; i < inst->Instruction.NumDstRegs; i++) {
>>> + const struct tgsi_full_dst_register *dst = &inst->Dst[i];
>>> +
>>> + if (dst->Register.File == TGSI_FILE_OUTPUT &&
>>> + !dst->Register.Indirect) {
>>> + unsigned name = info->output_semantic_name[dst->Register.Index];
>>> +
>>> + if (name == TGSI_SEMANTIC_TESSINNER)
>>> + writemask |= dst->Register.WriteMask;
>>> + else if (name == TGSI_SEMANTIC_TESSOUTER)
>>> + writemask |= dst->Register.WriteMask << 4;
>>> + }
>>> + }
>>> + return writemask;
>>> +}
>>> +
>>> +static unsigned
>>> +get_block_tessfactor_writemask(const struct tgsi_shader_info *info,
>>> + struct tgsi_parse_context *parse,
>>> + unsigned end_opcode)
>>> +{
>>> + struct tgsi_full_instruction *inst;
>>> + unsigned writemask = 0;
>>> +
>>> + do {
>>> + tgsi_parse_token(parse);
>>> + assert(parse->FullToken.Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION);
>>> + inst = &parse->FullToken.FullInstruction;
>>> + check_no_subroutines(inst);
>>> +
>>> + /* Recursively process nested blocks. */
>>> + switch (inst->Instruction.Opcode) {
>>> + case TGSI_OPCODE_IF:
>>> + case TGSI_OPCODE_UIF:
>>> + writemask |=
>>> + get_block_tessfactor_writemask(info, parse,
>>> TGSI_OPCODE_ENDIF);
>>> + continue;
>>> +
>>> + case TGSI_OPCODE_BGNLOOP:
>>> + writemask |=
>>> + get_block_tessfactor_writemask(info, parse,
>>> TGSI_OPCODE_ENDLOOP);
>>> + continue;
>>> + }
>>> +
>>> + writemask |= get_inst_tessfactor_writemask(info, inst);
>>> + } while (inst->Instruction.Opcode != end_opcode);
>>> +
>>> + return writemask;
>>> +}
>>> +
>>> +static void
>>> +get_if_block_tessfactor_writemask(const struct tgsi_shader_info *info,
>>> + struct tgsi_parse_context *parse,
>>> + unsigned *upper_block_tf_writemask,
>>> + unsigned *cond_block_tf_writemask)
>>> +{
>>> + struct tgsi_full_instruction *inst;
>>> + unsigned then_tessfactor_writemask = 0;
>>> + unsigned else_tessfactor_writemask = 0;
>>> + bool is_then = true;
>>> +
>>> + do {
>>> + tgsi_parse_token(parse);
>>> + assert(parse->FullToken.Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION);
>>> + inst = &parse->FullToken.FullInstruction;
>>> + check_no_subroutines(inst);
>>> +
>>> + switch (inst->Instruction.Opcode) {
>>> + case TGSI_OPCODE_ELSE:
>>> + is_then = false;
>>> + continue;
>>> +
>>> + /* Recursively process nested blocks. */
>>> + case TGSI_OPCODE_IF:
>>> + case TGSI_OPCODE_UIF:
>>> + get_if_block_tessfactor_writemask(info, parse,
>>> + is_then ?
>>> &then_tessfactor_writemask :
>>> +
>>> &else_tessfactor_writemask,
>>> + cond_block_tf_writemask);
>>> + continue;
>>> +
>>> + case TGSI_OPCODE_BGNLOOP:
>>> + *cond_block_tf_writemask |=
>>> + get_block_tessfactor_writemask(info, parse,
>>> TGSI_OPCODE_ENDLOOP);
>>> + continue;
>>> + }
>>> +
>>> + /* Process an instruction in the current block. */
>>> + unsigned writemask = get_inst_tessfactor_writemask(info, inst);
>>> +
>>> + if (writemask) {
>>> + if (is_then)
>>> + then_tessfactor_writemask |= writemask;
>>> + else
>>> + else_tessfactor_writemask |= writemask;
>>> + }
>>> + } while (inst->Instruction.Opcode != TGSI_OPCODE_ENDIF);
>>> +
>>> + if (then_tessfactor_writemask || else_tessfactor_writemask) {
>>> + /* If both statements write the same tess factor channels,
>>> + * we can say that the upper block writes them too. */
>>> + *upper_block_tf_writemask |= then_tessfactor_writemask &
>>> + else_tessfactor_writemask;
>>> + *cond_block_tf_writemask |= then_tessfactor_writemask |
>>> + else_tessfactor_writemask;
>>> + }
>>> +}
>>> +
>>> +void
>>> +tgsi_scan_tess_ctrl(const struct tgsi_token *tokens,
>>> + const struct tgsi_shader_info *info,
>>> + struct tgsi_tessctrl_info *out)
>>> +{
>>> + memset(out, 0, sizeof(*out));
>>> +
>>> + if (info->processor != PIPE_SHADER_TESS_CTRL)
>>> + return;
>>> +
>>> + struct tgsi_parse_context parse;
>>> + if (tgsi_parse_init(&parse, tokens) != TGSI_PARSE_OK) {
>>> + debug_printf("tgsi_parse_init() failed in tgsi_scan_arrays()!\n");
>>> + return;
>>> + }
>>> +
>>> + /* The pass works as follows:
>>> + * If all codepaths write tess factors, we can say that all
>>> invocations
>>> + * define tess factors, therefore invocation 0 defines tess factors.
>>> + *
>>> + * Each tess factor channel is tracked separately.
>>> + */
>>> + unsigned main_block_tf_writemask = 0;
>>> + unsigned cond_block_tf_writemask = 0;
>>> +
>>> + while (!tgsi_parse_end_of_tokens(&parse)) {
>>> + tgsi_parse_token(&parse);
>>> +
>>> + if (parse.FullToken.Token.Type != TGSI_TOKEN_TYPE_INSTRUCTION)
>>> + continue;
>>> +
>>> + struct tgsi_full_instruction *inst =
>>> &parse.FullToken.FullInstruction;
>>> + check_no_subroutines(inst);
>>> +
>>> + /* Process nested blocks. */
>>> + switch (inst->Instruction.Opcode) {
>>> + case TGSI_OPCODE_IF:
>>> + case TGSI_OPCODE_UIF:
>>> + get_if_block_tessfactor_writemask(info, &parse,
>>> + &main_block_tf_writemask,
>>> + &cond_block_tf_writemask);
>>> + continue;
>>> +
>>> + case TGSI_OPCODE_BGNLOOP:
>>> + cond_block_tf_writemask |=
>>> + get_block_tessfactor_writemask(info, &parse,
>>> TGSI_OPCODE_ENDIF);
>>> + continue;
>>> + }
>>> +
>>> + main_block_tf_writemask |= get_inst_tessfactor_writemask(info,
>>> inst);
>>> + }
>>> +
>>> + /* If there is a conditional block that writes a tess factor channel
>>> that
>>> + * the main block doesn't write, we can't say (we're not sure) that
>>> + * invocation 0 writes it.
>>> + */
>>> + out->invoc0_tessfactors_are_def =
>>> + main_block_tf_writemask &&
>>> + !(cond_block_tf_writemask & ~main_block_tf_writemask);
>>> +
>>> + tgsi_parse_free(&parse);
>>> +}
>>> diff --git a/src/gallium/auxiliary/tgsi/tgsi_scan.h
>>> b/src/gallium/auxiliary/tgsi/tgsi_scan.h
>>> index 857434f..ff64789 100644
>>> --- a/src/gallium/auxiliary/tgsi/tgsi_scan.h
>>> +++ b/src/gallium/auxiliary/tgsi/tgsi_scan.h
>>> @@ -181,30 +181,41 @@ struct tgsi_array_info
>>> /** Whether an array with this ID was declared. */
>>> bool declared;
>>> /** The OR of all writemasks used to write to this array. */
>>> ubyte writemask;
>>> /** The range with which the array was declared. */
>>> struct tgsi_declaration_range range;
>>> };
>>> +struct tgsi_tessctrl_info
>>> +{
>>> + /** Whether all codepaths of invocation 0 write tess factors. */
>>> + bool invoc0_tessfactors_are_def;
>>> +};
>>> +
>>> extern void
>>> tgsi_scan_shader(const struct tgsi_token *tokens,
>>> struct tgsi_shader_info *info);
>>> void
>>> tgsi_scan_arrays(const struct tgsi_token *tokens,
>>> unsigned file,
>>> unsigned max_array_id,
>>> struct tgsi_array_info *arrays);
>>> +void
>>> +tgsi_scan_tess_ctrl(const struct tgsi_token *tokens,
>>> + const struct tgsi_shader_info *info,
>>> + struct tgsi_tessctrl_info *out);
>>> +
>>> static inline bool
>>> tgsi_is_bindless_image_file(unsigned file)
>>> {
>>> return file != TGSI_FILE_IMAGE &&
>>> file != TGSI_FILE_MEMORY &&
>>> file != TGSI_FILE_BUFFER;
>>> }
>>> #ifdef __cplusplus
>>> } // extern "C"
>>>
>>
>>
>> --
>> Lerne, wie die Welt wirklich ist,
>> Aber vergiss niemals, wie sie sein sollte.
--
Lerne, wie die Welt wirklich ist,
Aber vergiss niemals, wie sie sein sollte.
More information about the mesa-dev
mailing list