[Mesa-dev] [PATCH 1/2] radeonsi: add support for indirect samplers

[Dave Airlie]

>
>> +
>>  static void tex_fetch_args(
>>         struct lp_build_tgsi_context * bld_base,
>>         struct lp_build_emit_data * emit_data)
>> @@ -1640,10 +1658,36 @@ static void tex_fetch_args(
>>         unsigned num_coords = tgsi_util_get_texture_coord_dim(target, &ref_pos);
>>         unsigned count = 0;
>>         unsigned chan;
>> -       unsigned sampler_src = emit_data->inst->Instruction.NumSrcRegs - 1;
>> -       unsigned sampler_index = emit_data->inst->Src[sampler_src].Register.Index;
>> +       unsigned sampler_src;
>> +       unsigned sampler_index;
>>         bool has_offset = HAVE_LLVM >= 0x0305 ? inst->Texture.NumOffsets > 0 : false;
>> +       bool has_sampler_indirect = false;
>> +       LLVMValueRef res_ptr, samp_ptr;
>> +
>> +       sampler_src = emit_data->inst->Instruction.NumSrcRegs - 1;
>> +       sampler_index = emit_data->inst->Src[sampler_src].Register.Index;
>> +
>> +       if (emit_data->inst->Src[sampler_src].Register.Indirect) {
>> +               const struct tgsi_full_src_register *reg = &emit_data->inst->Src[sampler_src];
>> +               unsigned first = reg->Register.Index;
>> +               LLVMValueRef ind_index;
>>
>> +               ind_index = get_indirect_index(si_shader_ctx, &reg->Indirect,
>> +                                          reg->Register.Index - first);
>
> It looks like the last parameter is always 0. (first ==
> reg->Register.Index) Therefore, reg->Register.Index is unused. Is that
> the correct behavior?

Good point, I should just be passing reg->Register.Index here and
below, fixed in v2.

>>         } else if (opcode == TGSI_OPCODE_TG4 ||
>>                    opcode == TGSI_OPCODE_LODQ ||
>> -                  has_offset) {
>> +                  has_offset || has_sampler_indirect) {
>
> Why is this here? This codepath doesn't support all sampler types and
> tex instructions yet.

Oh I put that in while I was debugging things to make it clearer what
was going on inside llvm (this intrinsic is much cleaner), turns out I
don't need it, this and the other one dropped in v2.

Thanks,
Dave.