[Mesa-dev] [PATCH 1/9] i965/fs: Rewrite assign_constant_locations
Jason Ekstrand
jason at jlekstrand.net
Thu Dec 7 18:57:31 UTC 2017
On Thu, Dec 7, 2017 at 9:57 AM, Pohjolainen, Topi <
topi.pohjolainen at gmail.com> wrote:
> On Thu, Dec 07, 2017 at 09:25:08AM -0800, Jason Ekstrand wrote:
> > On Thu, Dec 7, 2017 at 9:10 AM, Pohjolainen, Topi <
> > topi.pohjolainen at gmail.com> wrote:
> >
> > > On Wed, Dec 06, 2017 at 08:34:19PM -0800, Jason Ekstrand wrote:
> > > > This rewires the logic for assigning uniform locations to work in
> terms
> > > > of "complex alignments". The basic idea is that, as we walk the
> list of
> > > > instructions, we keep track of the alignment and continuity
> requirements
> > > > of each slot and assert that the alignments all match up. We then
> use
> > > > those alignments in the compaction stage to ensure that everything
> gets
> > > > placed at a properly aligned register. The old mechanism handled
> > > > alignments by special-casing each of the bit sizes and placing 64-bit
> > > > values first followed by 32-bit values.
> > > >
> > > > The old scheme had the advantage of never leaving a hole since all
> the
> > > > 64-bit values could be tightly packed and so could the 32-bit values.
> > > > However, the new scheme has no type size special cases so it handles
> not
> > > > only 32 and 64-bit types but should gracefully extend to 16 and 8-bit
> > > > types as the need arises.
> > > >
> > > > Tested-by: Jose Maria Casanova Crespo <jmcasanova at igalia.com>
> > > > ---
> > > > src/intel/compiler/brw_fs.cpp | 307 ++++++++++++++++++++++++------
> > > ------------
> > > > 1 file changed, 174 insertions(+), 133 deletions(-)
> > > >
> > > > diff --git a/src/intel/compiler/brw_fs.cpp
> b/src/intel/compiler/brw_fs.
> > > cpp
> > > > index 93bb6b4..41260b4 100644
> > > > --- a/src/intel/compiler/brw_fs.cpp
> > > > +++ b/src/intel/compiler/brw_fs.cpp
> > > > @@ -1906,62 +1906,6 @@ fs_visitor::compact_virtual_grfs()
> > > > return progress;
> > > > }
> > > >
> > > > -static void
> > > > -set_push_pull_constant_loc(unsigned uniform, int *chunk_start,
> > > > - unsigned *max_chunk_bitsize,
> > > > - bool contiguous, unsigned bitsize,
> > > > - const unsigned target_bitsize,
> > > > - int *push_constant_loc, int
> > > *pull_constant_loc,
> > > > - unsigned *num_push_constants,
> > > > - unsigned *num_pull_constants,
> > > > - const unsigned max_push_components,
> > > > - const unsigned max_chunk_size,
> > > > - bool allow_pull_constants,
> > > > - struct brw_stage_prog_data
> *stage_prog_data)
> > > > -{
> > > > - /* This is the first live uniform in the chunk */
> > > > - if (*chunk_start < 0)
> > > > - *chunk_start = uniform;
> > > > -
> > > > - /* Keep track of the maximum bit size access in contiguous
> uniforms
> > > */
> > > > - *max_chunk_bitsize = MAX2(*max_chunk_bitsize, bitsize);
> > > > -
> > > > - /* If this element does not need to be contiguous with the next,
> we
> > > > - * split at this point and everything between chunk_start and u
> > > forms a
> > > > - * single chunk.
> > > > - */
> > > > - if (!contiguous) {
> > > > - /* If bitsize doesn't match the target one, skip it */
> > > > - if (*max_chunk_bitsize != target_bitsize) {
> > > > - /* FIXME: right now we only support 32 and 64-bit accesses
> */
> > > > - assert(*max_chunk_bitsize == 4 || *max_chunk_bitsize == 8);
> > > > - *max_chunk_bitsize = 0;
> > > > - *chunk_start = -1;
> > > > - return;
> > > > - }
> > > > -
> > > > - unsigned chunk_size = uniform - *chunk_start + 1;
> > > > -
> > > > - /* Decide whether we should push or pull this parameter. In
> the
> > > > - * Vulkan driver, push constants are explicitly exposed via
> the
> > > API
> > > > - * so we push everything. In GL, we only push small arrays.
> > > > - */
> > > > - if (!allow_pull_constants ||
> > > > - (*num_push_constants + chunk_size <= max_push_components
> &&
> > > > - chunk_size <= max_chunk_size)) {
> > > > - assert(*num_push_constants + chunk_size <=
> > > max_push_components);
> > > > - for (unsigned j = *chunk_start; j <= uniform; j++)
> > > > - push_constant_loc[j] = (*num_push_constants)++;
> > > > - } else {
> > > > - for (unsigned j = *chunk_start; j <= uniform; j++)
> > > > - pull_constant_loc[j] = (*num_pull_constants)++;
> > > > - }
> > > > -
> > > > - *max_chunk_bitsize = 0;
> > > > - *chunk_start = -1;
> > > > - }
> > > > -}
> > > > -
> > > > static int
> > > > get_subgroup_id_param_index(const brw_stage_prog_data *prog_data)
> > > > {
> > > > @@ -1977,6 +1921,98 @@ get_subgroup_id_param_index(const
> > > brw_stage_prog_data *prog_data)
> > > > }
> > > >
> > > > /**
> > > > + * Struct for handling complex alignments.
> > > > + *
> > > > + * A complex alignment is stored as multiplier and an offset. A
> value
> > > is
> > > > + * considered to be aligned if it is congruent to the offset modulo
> the
> > > > + * multiplier.
> > > > + */
> > > > +struct cplx_align {
> > > > + unsigned mul:4;
> > > > + unsigned offset:4;
> > > > +};
> > > > +
> > > > +#define CPLX_ALIGN_MAX_MUL 8
> > > > +
> > > > +static void
> > > > +cplx_align_assert_sane(struct cplx_align a)
> > > > +{
> > > > + assert(a.mul > 0 && util_is_power_of_two(a.mul));
> > > > + assert(a.offset < a.mul);
> > > > +}
> > > > +
> > > > +/**
> > > > + * Combines two alignments to produce a least multiple of sorts.
> > > > + *
> > > > + * The returned alignment is the smallest (in terms of multiplier)
> such
> > > that
> > > > + * anything aligned to both a and b will be aligned to the new
> > > alignment.
> > > > + * This function will assert-fail if a and b are not compatible,
> i.e.
> > > if the
> > > > + * offset parameters are such that no common alignment is possible.
> > > > + */
> > > > +static struct cplx_align
> > > > +cplx_align_combine(struct cplx_align a, struct cplx_align b)
> > > > +{
> > > > + cplx_align_assert_sane(a);
> > > > + cplx_align_assert_sane(b);
> > > > +
> > > > + /* Assert that the alignments agree. */
> > > > + assert((a.offset & (b.mul - 1)) == (b.offset & (a.mul - 1)));
> > > > +
> > > > + return a.mul > b.mul ? a : b;
> > > > +}
> > > > +
> > > > +/**
> > > > + * Apply a complex alignment
> > > > + *
> > > > + * This function will return the smallest number greater than or
> equal
> > > to
> > > > + * offset that is aligned to align.
> > > > + */
> > > > +static unsigned
> > > > +cplx_align_apply(struct cplx_align align, unsigned offset)
> > > > +{
> > >
> > > Should we assert that offset >= 4?
> > >
> >
> > Not here, no. I intentionally want this code to be as generic as
> > possible. We may have reason to re-use it (or something similar)
> elsewhere
> > and I don't want to build in lots of push-constant assumptions.
> >
> >
> > > > + return ALIGN(offset - align.offset, align.mul) + align.offset;
> > >
> > > Soon I'm going to continue with glsl mediump support and thought
> better to
> > > read these patches. I stumbled on something (probably missing something
> > > again):
> > >
> >
> > :-)
> >
> >
> > > To me it looks there can ever be only three different instances of
> > > cplx_align:
> > >
> >
> > Today, yes, but there may come a day when we want to re-arrange with a
> > finer granularity than dwords.
> >
> >
> > > {8, 0}, {8, 4} and {4, 0} and that "offset" argument here is always
> > > multiple
> > > of 4.
> > >
> > > In case of align == {8, 0} "offset" gets aligned to 8, and in case of
> > > align == {4, 0} cplx_align_apply() is nop.
> > >
> > > But in case of {8, 4} and "offset" not multiple of 8, "offset" - 4 is
> > > multiple of 8 and returned value becomes ALIGN(offset - 4, 8) + 4 which
> > > in turn isn't aligned by 8.
> > >
> >
> > Correct, and it shouldn't be. Let's walk through the calculation with a
> > more complex alignment: {8, 6} (no, that isn't possible today but the
> > calculations should work with it) and some concrete numbers: 6, 8, 10,
> 12,
> > 14, 16
> >
> > 6 -> ALIGN(6 - 6, 8) + 6 = 6
> > 8 -> ALIGN(8 - 6, 8) + 6 = 14
> > 10 -> ALIGN(10 - 6, 8) + 6 = 14
> > 12 -> ALIGN(12 - 6, 8) + 6 = 14
> > 14 -> ALIGN(14 - 6, 8) + 6 = 14
> > 16 -> ALIGN(16 - 6, 8) + 6 = 22
> >
> > As you can see, the result of each is the next number that is 6 more
> than a
> > multiple of 8. There is also a question about negative numbers. Let's
> > look at one more calculation, this time with more detail about the guts
> of
> > ALIGN():
> >
> > 4 -> ALIGN(4 - 6, 8) + 6 = (((4 - 6) + (8 - 1)) & ~(8-1)) + 6 = (5 & ~7)
> +
> > 6 = 6
> >
> > That one's a bit more complicated but I think you see what's going on.
> If
> > it makes things more clear, we could do something such as
> >
> > return (offset + (align.mul - align.offset - 1)) & ~(align.mul - 1)) +
> > align.offset;
>
> Big thanks for the good explanation. I was also wondering why I didn't
> fully
> get the documentation of cplx_align. I kept asking myself "but offset is
> always modulo of mul". Now the comment there makes sense :) Some of your
> reply
> here would probably go nicely there.
>
Any preference whether it goes in the documentation of struct cplx_align or
cplx_align_apply? I'm leaning towards cplx_align
> I'll read the whole thing again to see if there was anything else I wanted
> to
> ask.
>
Cool
--Jason
>
> >
> > > +}
> > > > +
> > > > +#define UNIFORM_SLOT_SIZE 4
> > > > +
> > > > +struct uniform_slot_info {
> > > > + /** True if the given uniform slot is live */
> > > > + unsigned is_live:1;
> > > > +
> > > > + /** True if this slot and the slot must remain contiguous */
> > > > + unsigned contiguous:1;
> > > > +
> > > > + struct cplx_align align;
> > > > +};
> > > > +
> > > > +static void
> > > > +mark_uniform_slots_read(struct uniform_slot_info *slots,
> > > > + unsigned num_slots, unsigned alignment)
> > > > +{
> > > > + assert(alignment > 0 && util_is_power_of_two(alignment));
> > > > + assert(alignment <= CPLX_ALIGN_MAX_MUL);
> > > > +
> > > > + /* We can't align a slot to anything less than the slot size */
> > > > + alignment = MAX2(alignment, UNIFORM_SLOT_SIZE);
> > > > +
> > > > + struct cplx_align align = {alignment, 0};
> > > > + cplx_align_assert_sane(align);
> > > > +
> > > > + for (unsigned i = 0; i < num_slots; i++) {
> > > > + slots[i].is_live = true;
> > > > + if (i < num_slots - 1)
> > > > + slots[i].contiguous = true;
> > > > +
> > > > + align.offset = (i * UNIFORM_SLOT_SIZE) & (align.mul - 1);
> > > > + if (slots[i].align.mul == 0) {
> > > > + slots[i].align = align;
> > > > + } else {
> > > > + slots[i].align = cplx_align_combine(slots[i].align,
> align);
> > > > + }
> > > > + }
> > > > +}
> > > > +
> > > > +/**
> > > > * Assign UNIFORM file registers to either push constants or pull
> > > constants.
> > > > *
> > > > * We allow a fragment shader to have more than the specified
> minimum
> > > > @@ -1994,60 +2030,44 @@ fs_visitor::assign_constant_locations()
> > > > return;
> > > > }
> > > >
> > > > - bool is_live[uniforms];
> > > > - memset(is_live, 0, sizeof(is_live));
> > > > - unsigned bitsize_access[uniforms];
> > > > - memset(bitsize_access, 0, sizeof(bitsize_access));
> > > > -
> > > > - /* For each uniform slot, a value of true indicates that the
> given
> > > slot and
> > > > - * the next slot must remain contiguous. This is used to keep us
> > > from
> > > > - * splitting arrays and 64-bit values apart.
> > > > - */
> > > > - bool contiguous[uniforms];
> > > > - memset(contiguous, 0, sizeof(contiguous));
> > > > + struct uniform_slot_info slots[uniforms];
> > > > + memset(slots, 0, sizeof(slots));
> > > >
> > > > - /* First, we walk through the instructions and do two things:
> > > > - *
> > > > - * 1) Figure out which uniforms are live.
> > > > - *
> > > > - * 2) Mark any indirectly used ranges of registers as
> contiguous.
> > > > - *
> > > > - * Note that we don't move constant-indexed accesses to arrays.
> No
> > > > - * testing has been done of the performance impact of this
> choice.
> > > > - */
> > > > foreach_block_and_inst_safe(block, fs_inst, inst, cfg) {
> > > > for (int i = 0 ; i < inst->sources; i++) {
> > > > if (inst->src[i].file != UNIFORM)
> > > > continue;
> > > >
> > > > - int constant_nr = inst->src[i].nr + inst->src[i].offset /
> 4;
> > > > + /* NIR tightly packs things so the uniform number might
> not be
> > > > + * aligned (if we have a double right after a float, for
> > > instance).
> > > > + * This is fine because the process of re-arranging them
> will
> > > ensure
> > > > + * that things are properly aligned. The offset into that
> > > uniform,
> > > > + * however, must be aligned.
> > > > + *
> > > > + * In Vulkan, we have explicit offsets but everything is
> > > crammed
> > > > + * into a single "variable" so inst->src[i].nr will always
> be
> > > 0.
> > > > + * Everything will be properly aligned relative to that one
> > > base.
> > > > + */
> > > > + assert(inst->src[i].offset % type_sz(inst->src[i].type) ==
> 0);
> > > > +
> > > > + unsigned u = inst->src[i].nr +
> > > > + inst->src[i].offset / UNIFORM_SLOT_SIZE;
> > > >
> > > > + if (u >= uniforms)
> > > > + continue;
> > > > +
> > > > + unsigned slots_read;
> > > > if (inst->opcode == SHADER_OPCODE_MOV_INDIRECT && i == 0) {
> > > > - assert(inst->src[2].ud % 4 == 0);
> > > > - unsigned last = constant_nr + (inst->src[2].ud / 4) - 1;
> > > > - assert(last < uniforms);
> > > > -
> > > > - for (unsigned j = constant_nr; j < last; j++) {
> > > > - is_live[j] = true;
> > > > - contiguous[j] = true;
> > > > - bitsize_access[j] = MAX2(bitsize_access[j],
> > > type_sz(inst->src[i].type));
> > > > - }
> > > > - is_live[last] = true;
> > > > - bitsize_access[last] = MAX2(bitsize_access[last],
> > > type_sz(inst->src[i].type));
> > > > + slots_read = DIV_ROUND_UP(inst->src[2].ud,
> > > UNIFORM_SLOT_SIZE);
> > > > } else {
> > > > - if (constant_nr >= 0 && constant_nr < (int) uniforms) {
> > > > - int regs_read = inst->components_read(i) *
> > > > - type_sz(inst->src[i].type) / 4;
> > > > - assert(regs_read <= 2);
> > > > - if (regs_read == 2)
> > > > - contiguous[constant_nr] = true;
> > > > - for (int j = 0; j < regs_read; j++) {
> > > > - is_live[constant_nr + j] = true;
> > > > - bitsize_access[constant_nr + j] =
> > > > - MAX2(bitsize_access[constant_nr + j],
> > > type_sz(inst->src[i].type));
> > > > - }
> > > > - }
> > > > + unsigned bytes_read = inst->components_read(i) *
> > > > + type_sz(inst->src[i].type);
> > > > + slots_read = DIV_ROUND_UP(bytes_read,
> UNIFORM_SLOT_SIZE);
> > > > }
> > > > +
> > > > + assert(u + slots_read <= uniforms);
> > > > + mark_uniform_slots_read(&slots[u], slots_read,
> > > > + type_sz(inst->src[i].type));
> > > > }
> > > > }
> > > >
> > > > @@ -2082,43 +2102,64 @@ fs_visitor::assign_constant_locations()
> > > > memset(pull_constant_loc, -1, uniforms *
> sizeof(*pull_constant_loc));
> > > >
> > > > int chunk_start = -1;
> > > > - unsigned max_chunk_bitsize = 0;
> > > > -
> > > > - /* First push 64-bit uniforms to ensure they are properly
> aligned */
> > > > - const unsigned uniform_64_bit_size =
> type_sz(BRW_REGISTER_TYPE_DF);
> > > > + struct cplx_align align;
> > > > for (unsigned u = 0; u < uniforms; u++) {
> > > > - if (!is_live[u])
> > > > + if (!slots[u].is_live) {
> > > > + assert(chunk_start == -1);
> > > > continue;
> > > > + }
> > > >
> > > > - set_push_pull_constant_loc(u, &chunk_start,
> &max_chunk_bitsize,
> > > > - contiguous[u], bitsize_access[u],
> > > > - uniform_64_bit_size,
> > > > - push_constant_loc,
> pull_constant_loc,
> > > > - &num_push_constants,
> > > &num_pull_constants,
> > > > - max_push_components,
> max_chunk_size,
> > > > - compiler->supports_pull_constants,
> > > > - stage_prog_data);
> > > > + /* Skip subgroup_id_index to put it in the last push
> register. */
> > > > + if (subgroup_id_index == (int)u)
> > > > + continue;
> > > >
> > > > - }
> > > > + if (chunk_start == -1) {
> > > > + chunk_start = u;
> > > > + align = slots[u].align;
> > > > + } else {
> > > > + /* Offset into the chunk */
> > > > + unsigned chunk_offset = (u - chunk_start) *
> UNIFORM_SLOT_SIZE;
> > > >
> > > > - /* Then push the rest of uniforms */
> > > > - const unsigned uniform_32_bit_size =
> type_sz(BRW_REGISTER_TYPE_F);
> > > > - for (unsigned u = 0; u < uniforms; u++) {
> > > > - if (!is_live[u])
> > > > - continue;
> > > > + /* Shift the slot alignment down by the chunk offset so it
> is
> > > > + * comparable with the base chunk alignment.
> > > > + */
> > > > + struct cplx_align slot_align = slots[u].align;
> > > > + slot_align.offset =
> > > > + (slot_align.offset - chunk_offset) & (align.mul - 1);
> > > >
> > > > - /* Skip subgroup_id_index to put it in the last push
> register. */
> > > > - if (subgroup_id_index == (int)u)
> > > > + align = cplx_align_combine(align, slot_align);
> > > > + }
> > > > +
> > > > + /* Sanity check the alignment */
> > > > + cplx_align_assert_sane(align);
> > > > +
> > > > + if (slots[u].contiguous)
> > > > continue;
> > > >
> > > > - set_push_pull_constant_loc(u, &chunk_start,
> &max_chunk_bitsize,
> > > > - contiguous[u], bitsize_access[u],
> > > > - uniform_32_bit_size,
> > > > - push_constant_loc,
> pull_constant_loc,
> > > > - &num_push_constants,
> > > &num_pull_constants,
> > > > - max_push_components,
> max_chunk_size,
> > > > - compiler->supports_pull_constants,
> > > > - stage_prog_data);
> > > > + /* Adjust the alignment to be in terms of slots, not bytes */
> > > > + assert((align.mul & (UNIFORM_SLOT_SIZE - 1)) == 0);
> > > > + assert((align.offset & (UNIFORM_SLOT_SIZE - 1)) == 0);
> > > > + align.mul /= UNIFORM_SLOT_SIZE;
> > > > + align.offset /= UNIFORM_SLOT_SIZE;
> > > > +
> > > > + unsigned push_start_align = cplx_align_apply(align,
> > > num_push_constants);
> > > > + unsigned chunk_size = u - chunk_start + 1;
> > > > + if (!compiler->supports_pull_constants ||
> > > > + (chunk_size < max_chunk_size &&
> > > > + push_start_align + chunk_size <= max_push_components)) {
> > > > + /* Align up the number of push constants */
> > > > + num_push_constants = push_start_align;
> > > > + for (unsigned i = 0; i < chunk_size; i++)
> > > > + push_constant_loc[chunk_start + i] =
> num_push_constants++;
> > > > + } else {
> > > > + /* We need to pull this one */
> > > > + num_pull_constants = cplx_align_apply(align,
> > > num_pull_constants);
> > > > + for (unsigned i = 0; i < chunk_size; i++)
> > > > + pull_constant_loc[chunk_start + i] =
> num_pull_constants++;
> > > > + }
> > > > +
> > > > + /* Reset the chunk and start again */
> > > > + chunk_start = -1;
> > > > }
> > > >
> > > > /* Add the CS local thread ID uniform at the end of the push
> > > constants */
> > > > @@ -2131,8 +2172,8 @@ fs_visitor::assign_constant_locations()
> > > > uint32_t *param = stage_prog_data->param;
> > > > stage_prog_data->nr_params = num_push_constants;
> > > > if (num_push_constants) {
> > > > - stage_prog_data->param = ralloc_array(mem_ctx, uint32_t,
> > > > - num_push_constants);
> > > > + stage_prog_data->param = rzalloc_array(mem_ctx, uint32_t,
> > > > + num_push_constants);
> > > > } else {
> > > > stage_prog_data->param = NULL;
> > > > }
> > > > @@ -2140,8 +2181,8 @@ fs_visitor::assign_constant_locations()
> > > > assert(stage_prog_data->pull_param == NULL);
> > > > if (num_pull_constants > 0) {
> > > > stage_prog_data->nr_pull_params = num_pull_constants;
> > > > - stage_prog_data->pull_param = ralloc_array(mem_ctx, uint32_t,
> > > > -
> num_pull_constants);
> > > > + stage_prog_data->pull_param = rzalloc_array(mem_ctx, uint32_t,
> > > > +
> num_pull_constants);
> > > > }
> > > >
> > > > /* Now that we know how many regular uniforms we'll push, reduce
> the
> > > > --
> > > > 2.5.0.400.gff86faf
> > > >
> > > > _______________________________________________
> > > > mesa-dev mailing list
> > > > mesa-dev at lists.freedesktop.org
> > > > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
> > >
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20171207/4f7ef356/attachment-0001.html>
More information about the mesa-dev
mailing list