On Mon, Jan 21, 2019 at 2:09 AM Iago Toral <ito...@igalia.com> wrote:
> On Fri, 2019-01-18 at 06:46 -0600, Jason Ekstrand wrote:
>
> On January 18, 2019 01:48:25 Iago Toral <ito...@igalia.com> wrote:
>
> On Thu, 2019-01-17 at 13:42 -0600, Jason Ekstrand wrote:
>
> On Tue, Jan 15, 2019 at 7:54 AM Iago Toral Quiroga <ito...@igalia.com>
> wrote
>
>
> (...)
>
>
> + nir_ssa_def *tmp = nir_build_alu(b, op1, src, NULL, NULL, NULL);
> + nir_ssa_def *res = nir_build_alu(b, op2, tmp, NULL, NULL, NULL);
> + nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(res));
> + nir_instr_remove(&alu->instr);
> +}
> +
> +static bool
> +lower_instr(nir_builder *b, nir_alu_instr *alu)
> +{
> + unsigned src_bit_size = nir_src_bit_size(alu->src[0].src);
> + nir_alu_type src_type = nir_op_infos[alu->op].input_types[0];
> + nir_alu_type src_full_type = (nir_alu_type) (src_type | src_bit_size);
> +
> + unsigned dst_bit_size = nir_dest_bit_size(alu->dest.dest);
> + nir_alu_type dst_type = nir_op_infos[alu->op].output_type;
> + nir_alu_type dst_full_type = (nir_alu_type) (dst_type | dst_bit_size);
>
>
> The nir_op_infos[*].output_type will already be a full type. Maybe just
> delete the dst_type temporary so that no one gets confused and thinks it's
> a base type.
>
>
> Right, will do that.
>
>
> Actually, we want the base type too, since that is what we pass to the
> get_conversion_op() helper, I'll just pick it from the dst_full_type with
> the nir_alu_type_get_base_type() helper.
>
Sounds like the right thing to do.
> +
> + /* BDW PRM, vol02, Command Reference Instructions, mov - MOVE:
> + *
> + * "There is no direct conversion from HF to DF or DF to HF.
> + * Use two instructions and F (Float) as an intermediate type.
> + *
> + * There is no direct conversion from HF to Q/UQ or Q/UQ to HF.
> + * Use two instructions and F (Float) or a word integer type
> + * or a DWord integer type as an intermediate type."
> + */
> + if ((src_full_type == nir_type_float16 && dst_bit_size == 64) ||
> + (src_bit_size == 64 && dst_full_type == nir_type_float16)) {
> + nir_op op1 = get_conversion_op(src_type, src_bit_size, src_type, 32,
> + nir_rounding_mode_undef);
> + nir_op op2 = get_conversion_op(src_type, 32, dst_type, dst_bit_size,
> + get_opcode_rounding_mode(alu->op));
> + split_conversion(b, alu, op1, op2);
> + return true;
> + }
> +
> + /* SKL PRM, vol 02a, Command Reference: Instructions, Move:
> + *
> + * "There is no direct conversion from B/UB to DF or DF to B/UB. Use
> + * two instructions and a word or DWord intermediate type."
> + *
> + * "There is no direct conversion from B/UB to Q/UQ or Q/UQ to B/UB.
> + * Use two instructions and a word or DWord intermediate integer
> + * type."
> + */
> + if ((src_bit_size == 8 && dst_bit_size == 64) ||
> + (src_bit_size == 64 && dst_bit_size == 8)) {
> + nir_op op1 = get_conversion_op(src_type, src_bit_size, src_type, 32,
> + nir_rounding_mode_undef);
> + nir_op op2 = get_conversion_op(src_type, 32, dst_type, dst_bit_size,
> + nir_rounding_mode_undef);
> + split_conversion(b, alu, op1, op2);
> + return true;
> + }
> +
> + return false;
> +}
> +
> +static bool
> +lower_impl(nir_function_impl *impl)
> +{
> + nir_builder b;
> + nir_builder_init(&b, impl);
> + bool progress = false;
> +
> + nir_foreach_block(block, impl) {
> + nir_foreach_instr_safe(instr, block) {
> + if (instr->type != nir_instr_type_alu)
> + continue;
> +
> + nir_alu_instr *alu = nir_instr_as_alu(instr);
> + assert(alu->dest.dest.is_ssa);
> +
> + if (nir_op_infos[alu->op].num_inputs > 1)
>
>
> I don't think this is a reliable way to detect conversion opcodes. There
> are many unary operations that aren't conversions. You're only getting
> lucky because there are non which do float16 <-> 64=bit or 8-bit <->
> 64-bit. I've thought many times about throwing a boolean in nir_op_infos
> for is_conversion. Maybe now is a good time to do it?
>
>
> Yes, I knew that this was not going to only let conversions through and
> was relying on what you say above. Adding an is_conversion field souns good
> to me, alternatively I guess we could add a is_conversion() helper, but I
> guess having the field in nir_op_infos maye be a bit easier to maintain if
> we want to add more opcodes in the future.
>
>
> Thinking about this, I guess we only want to set this to True for the
> opcodes that are explicit numeric conversions, that is, the ones we
> generate right after "# Generate all of the numeric conversion opcodes". I
> am saying this because even if we have unop_convert(), that is also used
> for things like unpack opcodes, but_count, etc that are not real
> conversions. I was thinking about creating a new unop_numeric_convert()
> helper and have all other helpers set this new field to False. Sounds good?
>
Seems reasonable.
> + continue;
> +
> + progress = lower_instr(&b, alu) || progress;
> + }
> + }
> +
> + if (progress) {
> + nir_metadata_preserve(impl, nir_metadata_block_index |
> + nir_metadata_dominance);
> + }
> +
> + return progress;
> +}
> +
> +bool
> +brw_nir_lower_conversions(nir_shader *shader)
> +{
> + bool progress = false;
> +
> + nir_foreach_function(function, shader) {
> + if (function->impl)
> + progress |= lower_impl(function->impl);
> + }
> +
> + return progress;
> +}
> diff --git a/src/intel/compiler/meson.build
> b/src/intel/compiler/meson.build
> index f8e5e2518fe..9daf2a94260 100644
> --- a/src/intel/compiler/meson.build
> +++ b/src/intel/compiler/meson.build
> @@ -76,6 +76,7 @@ libintel_compiler_files = files(
> 'brw_nir_analyze_boolean_resolves.c',
> 'brw_nir_analyze_ubo_ranges.c',
> 'brw_nir_attribute_workarounds.c',
> + 'brw_nir_lower_conversions.c',
> 'brw_nir_lower_cs_intrinsics.c',
> 'brw_nir_lower_image_load_store.c',
> 'brw_nir_lower_mem_access_bit_sizes.c',
>
>
>
>
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
