[Mesa-dev] [PATCH 06/10] i965/fs: Get rid of load/store_foo_indirect
Kenneth Graunke
kenneth at whitecape.org
Wed Dec 2 00:33:28 PST 2015
On Wednesday, November 25, 2015 09:00:09 PM Jason Ekstrand wrote:
> ---
> src/mesa/drivers/dri/i965/brw_fs.h | 3 +-
> src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 145 ++++++++++++++-----------------
> src/mesa/drivers/dri/i965/brw_nir.c | 36 ++++++--
> 3 files changed, 92 insertions(+), 92 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs.h b/src/mesa/drivers/dri/i965/brw_fs.h
> index 2d408b2..b8891dd 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs.h
> +++ b/src/mesa/drivers/dri/i965/brw_fs.h
> @@ -296,8 +296,7 @@ public:
> unsigned stream_id);
> void emit_gs_thread_end();
> void emit_gs_input_load(const fs_reg &dst, const nir_src &vertex_src,
> - const fs_reg &indirect_offset, unsigned imm_offset,
> - unsigned num_components);
> + const nir_src &offset_src, unsigned num_components);
> void emit_cs_terminate();
> fs_reg *emit_cs_local_invocation_id_setup();
> fs_reg *emit_cs_work_group_id_setup();
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> index c439da2..b2e6ee8 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> @@ -1601,25 +1601,26 @@ fs_visitor::emit_gs_vertex(const nir_src &vertex_count_nir_src,
> void
> fs_visitor::emit_gs_input_load(const fs_reg &dst,
> const nir_src &vertex_src,
> - const fs_reg &indirect_offset,
> - unsigned imm_offset,
> + const nir_src &offset_src,
> unsigned num_components)
> {
> struct brw_gs_prog_data *gs_prog_data = (struct brw_gs_prog_data *) prog_data;
>
> + nir_const_value *vertex_const = nir_src_as_const_value(vertex_src);
> + nir_const_value *offset_const = nir_src_as_const_value(offset_src);
> + const unsigned push_reg_count = gs_prog_data->base.urb_read_length * 8;
> +
> /* Offset 0 is the VUE header, which contains VARYING_SLOT_LAYER [.y],
> * VARYING_SLOT_VIEWPORT [.z], and VARYING_SLOT_PSIZ [.w]. Only
> * gl_PointSize is available as a GS input, however, so it must be that.
> */
> const bool is_point_size =
> - indirect_offset.file == BAD_FILE && imm_offset == 0;
> -
> - nir_const_value *vertex_const = nir_src_as_const_value(vertex_src);
> - const unsigned push_reg_count = gs_prog_data->base.urb_read_length * 8;
> + offset_const != NULL && offset_const->u[0] == 0;
>
> - if (indirect_offset.file == BAD_FILE && vertex_const != NULL &&
> - 4 * imm_offset < push_reg_count) {
> - imm_offset = 4 * imm_offset + vertex_const->u[0] * push_reg_count;
> + if (offset_const != NULL && vertex_const != NULL &&
> + 4 * offset_const->u[0] < push_reg_count) {
Hmm, I thought you were planning to write a NIR helper that looked if an
expression was (<const> + <non-const>) and return both pieces? As is,
this will generate different code...I used to use the global offset for
the constant part, and the per-slot offsets for only the non-constant
part. I think that can save some ADD instructions.
> + int imm_offset = offset_const->u[0] * 4 +
> + vertex_const->u[0] * push_reg_count;
> /* This input was pushed into registers. */
> if (is_point_size) {
> /* gl_PointSize comes in .w */
> @@ -1681,21 +1682,21 @@ fs_visitor::emit_gs_input_load(const fs_reg &dst,
> }
>
> fs_inst *inst;
> - if (indirect_offset.file == BAD_FILE) {
> + if (offset_const) {
> /* Constant indexing - use global offset. */
> inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, dst, icp_handle);
> - inst->offset = imm_offset;
> + inst->offset = offset_const->u[0];
> inst->base_mrf = -1;
> inst->mlen = 1;
> inst->regs_written = num_components;
> } else {
> /* Indirect indexing - use per-slot offsets as well. */
> - const fs_reg srcs[] = { icp_handle, indirect_offset };
> + const fs_reg srcs[] = { icp_handle, get_nir_src(offset_src) };
> fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, 2);
> bld.LOAD_PAYLOAD(payload, srcs, ARRAY_SIZE(srcs), 0);
>
> inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, dst, payload);
> - inst->offset = imm_offset;
> + inst->offset = 0;
> inst->base_mrf = -1;
> inst->mlen = 2;
> inst->regs_written = num_components;
> @@ -1761,16 +1762,11 @@ fs_visitor::nir_emit_gs_intrinsic(const fs_builder &bld,
> retype(fs_reg(brw_vec8_grf(2, 0)), BRW_REGISTER_TYPE_UD));
> break;
>
> - case nir_intrinsic_load_input_indirect:
> case nir_intrinsic_load_input:
> unreachable("load_input intrinsics are invalid for the GS stage");
>
> - case nir_intrinsic_load_per_vertex_input_indirect:
> - indirect_offset = retype(get_nir_src(instr->src[1]), BRW_REGISTER_TYPE_D);
> - /* fallthrough */
> case nir_intrinsic_load_per_vertex_input:
> - emit_gs_input_load(dest, instr->src[0],
> - indirect_offset, instr->const_index[0],
> + emit_gs_input_load(dest, instr->src[0], instr->src[1],
> instr->num_components);
> break;
>
> @@ -2104,8 +2100,6 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
> if (nir_intrinsic_infos[instr->intrinsic].has_dest)
> dest = get_nir_dest(instr->dest);
>
> - bool has_indirect = false;
> -
> switch (instr->intrinsic) {
> case nir_intrinsic_atomic_counter_inc:
> case nir_intrinsic_atomic_counter_dec:
> @@ -2294,27 +2288,23 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
> bld.MOV(retype(dest, BRW_REGISTER_TYPE_D), brw_imm_d(1));
> break;
>
> - case nir_intrinsic_load_uniform_indirect:
> - has_indirect = true;
> - /* fallthrough */
> case nir_intrinsic_load_uniform: {
> fs_reg uniform_reg(UNIFORM, instr->const_index[0]);
> - uniform_reg.reg_offset = instr->const_index[1];
> + uniform_reg.type = dest.type;
>
> - for (unsigned j = 0; j < instr->num_components; j++) {
> - fs_reg src = offset(retype(uniform_reg, dest.type), bld, j);
> - if (has_indirect)
> - src.reladdr = new(mem_ctx) fs_reg(get_nir_src(instr->src[0]));
> + nir_const_value *const_offset = nir_src_as_const_value(instr->src[0]);
> + if (const_offset) {
> + uniform_reg.reg_offset = const_offset->u[0];
> + } else {
> + uniform_reg.reladdr = new(mem_ctx) fs_reg(get_nir_src(instr->src[0]));
> + }
>
> - bld.MOV(dest, src);
> - dest = offset(dest, bld, 1);
> + for (unsigned j = 0; j < instr->num_components; j++) {
> + bld.MOV(offset(dest, bld, j), offset(uniform_reg, bld, j));
> }
> break;
> }
>
> - case nir_intrinsic_load_ubo_indirect:
> - has_indirect = true;
> - /* fallthrough */
> case nir_intrinsic_load_ubo: {
> nir_const_value *const_index = nir_src_as_const_value(instr->src[0]);
> fs_reg surf_index;
> @@ -2342,27 +2332,28 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
> nir->info.num_ubos - 1);
> }
>
> - if (has_indirect) {
> + nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
> +
> + if (const_offset == NULL) {
> /* Turn the byte offset into a dword offset. */
> fs_reg base_offset = vgrf(glsl_type::int_type);
> bld.SHR(base_offset, retype(get_nir_src(instr->src[1]),
> BRW_REGISTER_TYPE_D),
> brw_imm_d(2));
>
> - unsigned vec4_offset = instr->const_index[0] / 4;
> for (int i = 0; i < instr->num_components; i++)
> VARYING_PULL_CONSTANT_LOAD(bld, offset(dest, bld, i), surf_index,
> - base_offset, vec4_offset + i);
> + base_offset, i);
Shouldn't converting UBOs from vec4 offsets to byte offsets be done
separately from removing the _indirect intrinsics? We're already
squashing a lot together, and that seems like an actual separable
change.
> } else {
> fs_reg packed_consts = vgrf(glsl_type::float_type);
> packed_consts.type = dest.type;
>
> - struct brw_reg const_offset_reg = brw_imm_ud(instr->const_index[0] & ~15);
> + struct brw_reg const_offset_reg = brw_imm_ud(const_offset->u[0] & ~15);
> bld.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD, packed_consts,
> surf_index, const_offset_reg);
>
> for (unsigned i = 0; i < instr->num_components; i++) {
> - packed_consts.set_smear(instr->const_index[0] % 16 / 4 + i);
> + packed_consts.set_smear(const_offset->u[0] % 16 / 4 + i);
>
> /* The std140 packing rules don't allow vectors to cross 16-byte
> * boundaries, and a reg is 32 bytes.
> @@ -2376,9 +2367,6 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
> break;
> }
>
> - case nir_intrinsic_load_ssbo_indirect:
> - has_indirect = true;
> - /* fallthrough */
> case nir_intrinsic_load_ssbo: {
> assert(devinfo->gen >= 7);
>
> @@ -2404,12 +2392,12 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
> nir->info.num_ssbos - 1);
> }
>
> - /* Get the offset to read from */
> fs_reg offset_reg;
> - if (has_indirect) {
> - offset_reg = get_nir_src(instr->src[1]);
> + nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
> + if (const_offset) {
> + offset_reg = brw_imm_ud(const_offset->u[0]);
> } else {
> - offset_reg = brw_imm_ud(instr->const_index[0]);
> + offset_reg = get_nir_src(instr->src[1]);
> }
This pattern comes up quite a bit - I actually have a patch in my tess
branch that adds a get_nir_src_imm() helpers so you can do this in one
line. I just mailed it out, in case you're interested.
>
> /* Read the vector */
> @@ -2424,32 +2412,27 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
> break;
> }
>
> - case nir_intrinsic_load_input_indirect:
> - has_indirect = true;
> - /* fallthrough */
> case nir_intrinsic_load_input: {
> - unsigned index = 0;
> - for (unsigned j = 0; j < instr->num_components; j++) {
> - fs_reg src;
> - if (stage == MESA_SHADER_VERTEX) {
> - src = offset(fs_reg(ATTR, instr->const_index[0], dest.type), bld, index);
> - } else {
> - src = offset(retype(nir_inputs, dest.type), bld,
> - instr->const_index[0] + index);
> - }
> - if (has_indirect)
> - src.reladdr = new(mem_ctx) fs_reg(get_nir_src(instr->src[0]));
> - index++;
> + fs_reg src;
> + if (stage == MESA_SHADER_VERTEX) {
> + src = fs_reg(ATTR, 0, dest.type);
> + } else {
> + src = retype(nir_inputs, dest.type);
> + }
>
> - bld.MOV(dest, src);
> - dest = offset(dest, bld, 1);
> + nir_const_value *const_offset = nir_src_as_const_value(instr->src[0]);
> + if (const_offset) {
> + src = offset(src, bld, const_offset->u[0]);
> + } else {
> + src.reladdr = new(mem_ctx) fs_reg(get_nir_src(instr->src[0]));
> + }
> +
> + for (unsigned j = 0; j < instr->num_components; j++) {
> + bld.MOV(offset(dest, bld, j), offset(src, bld, j));
> }
> break;
> }
>
> - case nir_intrinsic_store_ssbo_indirect:
> - has_indirect = true;
> - /* fallthrough */
> case nir_intrinsic_store_ssbo: {
> assert(devinfo->gen >= 7);
>
> @@ -2476,7 +2459,7 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
> fs_reg val_reg = get_nir_src(instr->src[0]);
>
> /* Writemask */
> - unsigned writemask = instr->const_index[1];
> + unsigned writemask = instr->const_index[0];
>
> /* Combine groups of consecutive enabled channels in one write
> * message. We use ffs to find the first enabled channel and then ffs on
> @@ -2486,10 +2469,11 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
> while (writemask) {
> unsigned first_component = ffs(writemask) - 1;
> unsigned length = ffs(~(writemask >> first_component)) - 1;
> - fs_reg offset_reg;
>
> - if (!has_indirect) {
> - offset_reg = brw_imm_ud(instr->const_index[0] + 4 * first_component);
> + fs_reg offset_reg;
> + nir_const_value *const_offset = nir_src_as_const_value(instr->src[2]);
> + if (const_offset) {
> + offset_reg = brw_imm_ud(const_offset->u[0] + 4 * first_component);
> } else {
> offset_reg = vgrf(glsl_type::uint_type);
> bld.ADD(offset_reg,
> @@ -2510,20 +2494,19 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
> break;
> }
>
> - case nir_intrinsic_store_output_indirect:
> - has_indirect = true;
> - /* fallthrough */
> case nir_intrinsic_store_output: {
> fs_reg src = get_nir_src(instr->src[0]);
> - unsigned index = 0;
> + fs_reg new_dest = retype(nir_outputs, src.type);
> +
> + nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
> + if (const_offset) {
> + new_dest = offset(new_dest, bld, const_offset->u[0]);
> + } else {
> + src.reladdr = new(mem_ctx) fs_reg(get_nir_src(instr->src[1]));
> + }
> +
> for (unsigned j = 0; j < instr->num_components; j++) {
> - fs_reg new_dest = offset(retype(nir_outputs, src.type), bld,
> - instr->const_index[0] + index);
> - if (has_indirect)
> - src.reladdr = new(mem_ctx) fs_reg(get_nir_src(instr->src[1]));
> - index++;
> - bld.MOV(new_dest, src);
> - src = offset(src, bld, 1);
> + bld.MOV(offset(new_dest, bld, j), offset(src, bld, j));
> }
> break;
> }
> diff --git a/src/mesa/drivers/dri/i965/brw_nir.c b/src/mesa/drivers/dri/i965/brw_nir.c
> index bf3bc2d..85b4acc 100644
> --- a/src/mesa/drivers/dri/i965/brw_nir.c
> +++ b/src/mesa/drivers/dri/i965/brw_nir.c
> @@ -25,21 +25,26 @@
> #include "brw_shader.h"
> #include "glsl/glsl_parser_extras.h"
> #include "glsl/nir/glsl_to_nir.h"
> +#include "glsl/nir/nir_builder.h"
> #include "program/prog_to_nir.h"
>
> +struct remap_vs_attrs_state {
> + nir_builder b;
> + GLbitfield64 inputs_read;
> +};
> +
> static bool
> remap_vs_attrs(nir_block *block, void *closure)
> {
> - GLbitfield64 inputs_read = *((GLbitfield64 *) closure);
> + struct remap_vs_attrs_state *state = closure;
>
> - nir_foreach_instr(block, instr) {
> + nir_foreach_instr_safe(block, instr) {
> if (instr->type != nir_instr_type_intrinsic)
> continue;
>
> nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
>
> - /* We set EmitNoIndirect for VS inputs, so there are no indirects. */
> - assert(intrin->intrinsic != nir_intrinsic_load_input_indirect);
> + state->b.cursor = nir_before_instr(&intrin->instr);
>
> if (intrin->intrinsic == nir_intrinsic_load_input) {
> /* Attributes come in a contiguous block, ordered by their
> @@ -47,9 +52,15 @@ remap_vs_attrs(nir_block *block, void *closure)
> * number for an attribute by masking out the enabled attributes
> * before it and counting the bits.
> */
> - int attr = intrin->const_index[0];
> - int slot = _mesa_bitcount_64(inputs_read & BITFIELD64_MASK(attr));
> - intrin->const_index[0] = 4 * slot;
> + nir_const_value *const_attr = nir_src_as_const_value(intrin->src[0]);
> +
> + /* We set EmitNoIndirect for VS inputs, so there are no indirects. */
> + assert(const_attr != NULL);
> +
> + int attr = const_attr->u[0];
> + int slot = _mesa_bitcount_64(state->inputs_read & BITFIELD64_MASK(attr));
> + nir_instr_rewrite_src(&intrin->instr, &intrin->src[0],
> + nir_src_for_ssa(nir_imm_int(&state->b, 4 * slot)));
> }
> }
> return true;
> @@ -80,10 +91,17 @@ brw_nir_lower_inputs(nir_shader *nir,
> * key->inputs_read since the two are identical aside from Gen4-5
> * edge flag differences.
> */
> - GLbitfield64 inputs_read = nir->info.inputs_read;
> +
> + /* This pass needs actual constants */
> + nir_opt_constant_folding(nir);
Is this sufficient? We don't need copy prop to eliminate imovs/fmovs
as well?
> +
> + struct remap_vs_attrs_state state = {
> + .inputs_read = nir->info.inputs_read
> + };
> nir_foreach_overload(nir, overload) {
> if (overload->impl) {
> - nir_foreach_block(overload->impl, remap_vs_attrs, &inputs_read);
> + nir_builder_init(&state.b, overload->impl);
> + nir_foreach_block(overload->impl, remap_vs_attrs, &state);
> }
> }
> }
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20151202/27ee2647/attachment-0001.sig>
More information about the mesa-dev
mailing list