On Thu, Oct 12, 2017 at 11:38 AM, Jose Maria Casanova Crespo <
jmcasan...@igalia.com> wrote:
> From: Eduardo Lima Mitev <el...@igalia.com>
>
> Currently, we use byte-scattered write messages for storing 16-bit
> into an SSBO. This is because untyped surface messages have a fixed
> 32-bit size.
>
> This patch optimizes these 16-bit writes by combining 2 values (e.g,
> two consecutive components) into a 32-bit register, packing the two
> 16-bit words.
>
> 16-bit single component values will continue to use byte-scattered
> write messages.
>
> This optimization reduces the number of SEND messages used for storing
> 16-bit values potentially by 2 or 4, which cuts down execution time
> significantly because byte-scattered writes are an expensive
> operation.
>
> v2: Removed use of stride = 2 on sources (Jason Ekstrand)
> Rework optimization using shuffle 16 write and enable writes
> of 16bit vec4 with only one message of 32-bits. (Chema Casanova)
>
> Signed-off-by: Jose Maria Casanova Crespo <jmcasan...@igalia.com>
> Signed-off-by: Eduardo Lima <el...@igalia.com>
> ---
> src/intel/compiler/brw_fs_nir.cpp | 64 ++++++++++++++++++++++++++++++
> +--------
> 1 file changed, 52 insertions(+), 12 deletions(-)
>
> diff --git a/src/intel/compiler/brw_fs_nir.cpp
> b/src/intel/compiler/brw_fs_nir.cpp
> index 2d0b3e139e..c07b3e4d8d 100644
> --- a/src/intel/compiler/brw_fs_nir.cpp
> +++ b/src/intel/compiler/brw_fs_nir.cpp
> @@ -4218,6 +4218,9 @@ fs_visitor::nir_emit_intrinsic(const fs_builder
> &bld, nir_intrinsic_instr *instr
> instr->num_components);
> val_reg = tmp;
> }
> + if (bit_size == 16) {
> + val_reg=retype(val_reg, BRW_REGISTER_TYPE_HF);
> + }
>
> /* 16-bit types would use a minimum of 1 slot */
> unsigned type_slots = MAX2(type_size / 4, 1);
> @@ -4231,6 +4234,9 @@ fs_visitor::nir_emit_intrinsic(const fs_builder
> &bld, nir_intrinsic_instr *instr
> unsigned first_component = ffs(writemask) - 1;
> unsigned length = ffs(~(writemask >> first_component)) - 1;
>
> + fs_reg current_val_reg =
> + offset(val_reg, bld, first_component * type_slots);
> +
> /* We can't write more than 2 64-bit components at once. Limit
> the
> * length of the write to what we can do and let the next
> iteration
> * handle the rest
> @@ -4238,11 +4244,40 @@ fs_visitor::nir_emit_intrinsic(const fs_builder
> &bld, nir_intrinsic_instr *instr
> if (type_size > 4) {
> length = MIN2(2, length);
> } else if (type_size == 2) {
> - /* For 16-bit types we are using byte scattered writes, that
> can
> - * only write one component per call. So we limit the length,
> and
> - * let the write happening in several iterations.
> + /* For 16-bit types we pack two consecutive values into a
> + * 32-bit word and use an untyped write message. For single
> values
> + * we need to use byte-scattered writes because untyped
> writes work
> + * on multiples of 32 bits.
> + *
> + * For example, if there is a 3-component vector we submit one
> + * untyped-write message of 32-bit (first two components),
> and one
> + * byte-scattered write message (the last component).
> */
> - length = 1;
> + if (length >= 2) {
> + /* pack two consecutive 16-bit words into a 32-bit
> register,
> + * using the same original source register.
> + */
>
This doesn't work if you have a writemask of .xz
> + length -= length % 2;
>
I'm very confused by this bit of math.
> + fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_F, length / 2);
> + shuffle_16bit_data_for_32bit_write(bld,
> + tmp,
> + current_val_reg,
> + length);
> + current_val_reg = tmp;
> +
> + } else {
> + /* For single 16-bit values, we just limit the length to 1
> and
> + * use a byte-scattered write message below.
> + */
> + length = 1;
>
I think this can be an assert. Also, why do we need the shuffle? I
thought this case would work if we just set length == 1.
I answered my own question about the shuffle. It lets us delete some code
below.
> + fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_F);
> + shuffle_16bit_data_for_32bit_write(bld,
> + tmp,
> + current_val_reg,
> + length);
>
+ current_val_reg = tmp;
> +
> + }
> }
>
> fs_reg offset_reg;
> @@ -4257,24 +4292,29 @@ fs_visitor::nir_emit_intrinsic(const fs_builder
> &bld, nir_intrinsic_instr *instr
> brw_imm_ud(type_size * first_component));
> }
>
> - if (type_size == 2) {
> + if (type_size == 2 && length == 1) {
>
I'd rather see us do a type_size correction above and just make this one
"if (type_size == 2) {" or, even better yet "if (type_size < 4) {
assert(type_size == 2);"
> /* Untyped Surface messages have a fixed 32-bit size, so we
> need
> * to rely on byte scattered in order to write 16-bit
> elements.
> * The byte_scattered_write message needs that every written
> 16-bit
> * type to be aligned 32-bits (stride=2).
> */
> - fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_F);
> - val_reg.type = BRW_REGISTER_TYPE_HF;
> - bld.MOV (subscript(tmp, BRW_REGISTER_TYPE_HF, 0),
> - offset(val_reg, bld, first_component));
> emit_byte_scattered_write(bld, surf_index, offset_reg,
> - tmp,
> + current_val_reg,
> 1 /* dims */, length * type_slots,
> BRW_PREDICATE_NONE);
> } else {
> + unsigned write_size = length * type_slots;
> +
> + /* We need to half the write size of the untyped write
> message when
> + * submitting two packed 16-bit values, because length is 2
> but the
> + * type size is 16-bit. So we are effectively writing just 32
> bits
> + * (size = 1).
> + */
> + if (type_size == 2)
> + write_size /= 2;
>
Can we not do this correction higher up where we decide to pack two 16-bit
things into a single 32-bit thing?
> emit_untyped_write(bld, surf_index, offset_reg,
> - offset(val_reg, bld, first_component *
> type_slots),
> - 1 /* dims */, length * type_slots,
> + current_val_reg,
> + 1 /* dims */, write_size,
> BRW_PREDICATE_NONE);
> }
>
> --
> 2.13.6
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
