[Mesa-dev] [PATCH 2/2] i965/fs: Lower arithmetic instructions with register regions of unsupported width.

[Jason Ekstrand]

You can have my R-B on both patches too.

On Wed, Aug 5, 2015 at 11:14 AM, Connor Abbott <cwabbott0 at gmail.com> wrote:
> FWIW, both patches are:
>
> Reviewed-by: Connor Abbott <connor.w.abbott at intel.com>
>
> I'm working on FP64 support (I've been using no16 up till now) so this
> is obviously very useful to me.
>
> On Wed, Aug 5, 2015 at 10:38 AM, Francisco Jerez <currojerez at riseup.net> wrote:
>> This extends the SIMD lowering pass to enforce the hardware limitation
>> that no directly-addressed source may read more than 2 physical GRFs.
>> One can easily go over this limit when doing 64-bit arithmetic
>> (e.g. FP64 or extended-precision integer MULs) or SIMD32, so it's nice
>> to be able to just emit an instruction of the intended execution size
>> from the visitor and let the lowering pass deal with this restriction
>> transparently.
>>
>> Some hardware arithmetic instructions are not handled here, including
>> all instructions that use the accumulator implicitly (which the SIMD
>> lowering pass deliberately doesn't handle), instructions with
>> non-per-channel sources (e.g. LINE or PLANE) and SEND-like
>> instructions, which need special handling most likely as virtual
>> opcodes.
>> ---
>>  src/mesa/drivers/dri/i965/brw_fs.cpp | 62 ++++++++++++++++++++++++++++++++++++
>>  1 file changed, 62 insertions(+)
>>
>> diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
>> index f9773bd..fa5ed4f 100644
>> --- a/src/mesa/drivers/dri/i965/brw_fs.cpp
>> +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
>> @@ -4130,6 +4130,68 @@ get_lowered_simd_width(const struct brw_device_info *devinfo,
>>                         const fs_inst *inst)
>>  {
>>     switch (inst->opcode) {
>> +   case BRW_OPCODE_MOV:
>> +   case BRW_OPCODE_SEL:
>> +   case BRW_OPCODE_NOT:
>> +   case BRW_OPCODE_AND:
>> +   case BRW_OPCODE_OR:
>> +   case BRW_OPCODE_XOR:
>> +   case BRW_OPCODE_SHR:
>> +   case BRW_OPCODE_SHL:
>> +   case BRW_OPCODE_ASR:
>> +   case BRW_OPCODE_CMP:
>> +   case BRW_OPCODE_CMPN:
>> +   case BRW_OPCODE_CSEL:
>> +   case BRW_OPCODE_F32TO16:
>> +   case BRW_OPCODE_F16TO32:
>> +   case BRW_OPCODE_BFREV:
>> +   case BRW_OPCODE_BFE:
>> +   case BRW_OPCODE_BFI1:
>> +   case BRW_OPCODE_BFI2:
>> +   case BRW_OPCODE_ADD:
>> +   case BRW_OPCODE_MUL:
>> +   case BRW_OPCODE_AVG:
>> +   case BRW_OPCODE_FRC:
>> +   case BRW_OPCODE_RNDU:
>> +   case BRW_OPCODE_RNDD:
>> +   case BRW_OPCODE_RNDE:
>> +   case BRW_OPCODE_RNDZ:
>> +   case BRW_OPCODE_LZD:
>> +   case BRW_OPCODE_FBH:
>> +   case BRW_OPCODE_FBL:
>> +   case BRW_OPCODE_CBIT:
>> +   case BRW_OPCODE_SAD2:
>> +   case BRW_OPCODE_MAD:
>> +   case BRW_OPCODE_LRP:
>> +   case SHADER_OPCODE_RCP:
>> +   case SHADER_OPCODE_RSQ:
>> +   case SHADER_OPCODE_SQRT:
>> +   case SHADER_OPCODE_EXP2:
>> +   case SHADER_OPCODE_LOG2:
>> +   case SHADER_OPCODE_POW:
>> +   case SHADER_OPCODE_INT_QUOTIENT:
>> +   case SHADER_OPCODE_INT_REMAINDER:
>> +   case SHADER_OPCODE_SIN:
>> +   case SHADER_OPCODE_COS: {
>> +      /* According to the PRMs:
>> +       *  "A. In Direct Addressing mode, a source cannot span more than 2
>> +       *      adjacent GRF registers.
>> +       *   B. A destination cannot span more than 2 adjacent GRF registers."
>> +       *
>> +       * Look for the source or destination with the largest register region
>> +       * which is the one that is going to limit the overal execution size of
>> +       * the instruction due to this rule.
>> +       */
>> +      unsigned reg_count = inst->regs_written;
>> +
>> +      for (unsigned i = 0; i < inst->sources; i++)
>> +         reg_count = MAX2(reg_count, (unsigned)inst->regs_read(i));
>> +
>> +      /* Calculate the maximum execution size of the instruction based on the
>> +       * factor by which it goes over the hardware limit of 2 GRFs.
>> +       */
>> +      return inst->exec_size / DIV_ROUND_UP(reg_count, 2);
>> +   }
>>     case SHADER_OPCODE_MULH:
>>        /* MULH is lowered to the MUL/MACH sequence using the accumulator, which
>>         * is 8-wide on Gen7+.
>> --
>> 2.4.6
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev