[Mesa-dev] [PATCH 1/1] r600: Enable FMA on chips that support it

[Marek Olšák]

On Thu, Jun 16, 2016 at 12:47 AM, Glenn Kennard <glenn.kennard at gmail.com> wrote:
> On Wed, 15 Jun 2016 20:13:13 +0200, Jan Vesely <jan.vesely at rutgers.edu>
> wrote:
>
>> Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
>> ---
>> Untested (I don't have the required hw)
>>
>>  src/gallium/drivers/r600/r600_pipe.c   | 5 ++++-
>>  src/gallium/drivers/r600/r600_shader.c | 2 +-
>>  2 files changed, 5 insertions(+), 2 deletions(-)
>>
>> diff --git a/src/gallium/drivers/r600/r600_pipe.c
>> b/src/gallium/drivers/r600/r600_pipe.c
>> index a49b00f..49c3e1d 100644
>> --- a/src/gallium/drivers/r600/r600_pipe.c
>> +++ b/src/gallium/drivers/r600/r600_pipe.c
>> @@ -548,7 +548,6 @@ static int r600_get_shader_param(struct pipe_screen*
>> pscreen, unsigned shader, e
>>                 return 0;
>>         case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
>>         case PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED:
>> -       case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED:
>>         case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
>>         case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
>>                 return 0;
>> @@ -558,6 +557,10 @@ static int r600_get_shader_param(struct pipe_screen*
>> pscreen, unsigned shader, e
>>                  *    https://bugs.freedesktop.org/show_bug.cgi?id=86720
>>                  */
>>                 return 255;
>> +       case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED:
>> +               // Enable on CYPRESS(EG) and CAYMAN(NI)
>> +               return rscreen->b.family == CHIP_CYPRESS ||
>> +                      rscreen->b.family == CHIP_CAYMAN;
>>         }
>>         return 0;
>>  }
>> diff --git a/src/gallium/drivers/r600/r600_shader.c
>> b/src/gallium/drivers/r600/r600_shader.c
>> index 101f666..35019e3 100644
>> --- a/src/gallium/drivers/r600/r600_shader.c
>> +++ b/src/gallium/drivers/r600/r600_shader.c
>> @@ -8917,7 +8917,7 @@ static const struct r600_shader_tgsi_instruction
>> r600_shader_tgsi_instruction[]
>>         [TGSI_OPCODE_MAD]       = { ALU_OP3_MULADD, tgsi_op3},
>>         [TGSI_OPCODE_SUB]       = { ALU_OP2_ADD, tgsi_op2},
>>         [TGSI_OPCODE_LRP]       = { ALU_OP0_NOP, tgsi_lrp},
>> -       [TGSI_OPCODE_FMA]       = { ALU_OP0_NOP, tgsi_unsupported},
>> +       [TGSI_OPCODE_FMA]       = { ALU_OP3_FMA, tgsi_op3},
>>         [TGSI_OPCODE_SQRT]      = { ALU_OP1_SQRT_IEEE,
>> tgsi_trans_srcx_replicate},
>>         [TGSI_OPCODE_DP2A]      = { ALU_OP0_NOP, tgsi_unsupported},
>>         [22]                    = { ALU_OP0_NOP, tgsi_unsupported},
>
>
> You probably meant to add the opcode to the eg_shader_tgsi_instruction and
> cm_shader_tgsi_instruction opcode tables rather than the R600/R700 one?
>
>
> I'll also note in passing that FMA on CYPRESS/HEMLOCK has an issue rate of
> 4/cycle vs MULADD 5/cycle since FMA cannot be issued in the 't' slot,
> may or may not affect performance depending on if the GLSL front end decides
> to use fma for mul+add operations. On Cayman/Aruba they are the same rate.

It depends on the application, not the GLSL frontend. Alien Isolation
uses fma everywhere, other games don't AFAIK.

Marek