[Mesa-dev] [PATCH] radeonsi: add a debug flag for unsafe math LLVM optimizations

[Marek Olšák]

On Tue, Jun 21, 2016 at 7:21 PM, Tom Stellard <tom at stellard.net> wrote:
> On Mon, Jun 13, 2016 at 06:27:02PM +0200, Marek Olšák wrote:
>> From: Marek Olšák <marek.olsak at amd.com>
>>
>> ---
>>  src/gallium/drivers/radeon/r600_pipe_common.c |  1 +
>>  src/gallium/drivers/radeon/r600_pipe_common.h |  1 +
>>  src/gallium/drivers/radeonsi/si_shader.c      | 16 ++++++++++++++++
>>  3 files changed, 18 insertions(+)
>>
>> diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
>> index fa9f70d..5d4a679 100644
>> --- a/src/gallium/drivers/radeon/r600_pipe_common.c
>> +++ b/src/gallium/drivers/radeon/r600_pipe_common.c
>> @@ -482,6 +482,7 @@ static const struct debug_named_value common_debug_options[] = {
>>       { "sisched", DBG_SI_SCHED, "Enable LLVM SI Machine Instruction Scheduler." },
>>       { "mono", DBG_MONOLITHIC_SHADERS, "Use old-style monolithic shaders compiled on demand" },
>>       { "noce", DBG_NO_CE, "Disable the constant engine"},
>> +     { "unsafemath", DBG_UNSAFE_MATH, "Enable unsafe math shader optimizations" },
>>
>>       DEBUG_NAMED_VALUE_END /* must be last */
>>  };
>> diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
>> index 77dfc4f..263ef5e 100644
>> --- a/src/gallium/drivers/radeon/r600_pipe_common.h
>> +++ b/src/gallium/drivers/radeon/r600_pipe_common.h
>> @@ -99,6 +99,7 @@
>>  #define DBG_SI_SCHED         (1llu << 46)
>>  #define DBG_MONOLITHIC_SHADERS       (1llu << 47)
>>  #define DBG_NO_CE            (1llu << 48)
>> +#define DBG_UNSAFE_MATH              (1llu << 49)
>>
>>  #define R600_MAP_BUFFER_ALIGNMENT 64
>>  #define R600_MAX_VIEWPORTS        16
>> diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
>> index 6dc4985..bba6a55 100644
>> --- a/src/gallium/drivers/radeonsi/si_shader.c
>> +++ b/src/gallium/drivers/radeonsi/si_shader.c
>> @@ -5255,6 +5255,22 @@ static void si_create_function(struct si_shader_context *ctx,
>>               else
>>                       LLVMAddAttribute(P, LLVMInRegAttribute);
>>       }
>> +
>> +     if (ctx->screen->b.debug_flags & DBG_UNSAFE_MATH) {
>> +             /* These were copied from some LLVM test. */
>> +             LLVMAddTargetDependentFunctionAttr(ctx->radeon_bld.main_fn,
>> +                                                "less-precise-fpmad",
>> +                                                "true");
>> +             LLVMAddTargetDependentFunctionAttr(ctx->radeon_bld.main_fn,
>> +                                                "no-infs-fp-math",
>> +                                                "true");
>> +             LLVMAddTargetDependentFunctionAttr(ctx->radeon_bld.main_fn,
>> +                                                "no-nans-fp-math",
>> +                                                "true");
>> +             LLVMAddTargetDependentFunctionAttr(ctx->radeon_bld.main_fn,
>> +                                                "unsafe-fp-math",
>> +                                                "true");
>> +     }
>
> You may get better results by also adding the fast-math flags to the
> individual floating-point instructions, but this would be a more
> invasive change.

Is there sample code showing how to do that?

Thanks,
Marek