Quoting Ian Romanick (2018-03-27 12:21:36) > I haven't investigated the root cause yet, but something about this > patch *explodes* memory usage. My shader-db runs on an 8-thread, 16GB > RAM + 16GB swap system get OOM killed. Previous to this patch, RAM > usage would never exceed 3.1% (based on watching top during the run). > So... > 60x memory usage seems like a lot. :) This patch is causing some serious issues in the Mesa CI, tests are being killed by the kernel OOM killer on tester systems. I bisected it to this patch (commit 76dfed8ae2)
>
> On 03/25/2018 11:10 AM, Rob Clark wrote:
> > I threatened to do this a long time ago.. I probably *should* have done
> > it a long time ago when there where many fewer intrinsics. But the
> > system of macro/#include magic for dealing with intrinsics is a bit
> > annoying, and python has the nice property of optional fxn params,
> > making it possible to define new intrinsics while ignoring parameters
> > that are not applicable (and naming optional params). And not having to
> > specify various array lengths explicitly is nice too.
> >
> > I think the end result makes it easier to add new intrinsics.
> >
> > v2: couple small fixes found with a test program to compare the old and
> > new tables
> > v3: misc comments, don't rely on capture=true for meson.build, get rid
> > of system_values table to avoid return value of intrinsic() and
> > *mostly* remove side-effects, add autotools build support
> >
> > Signed-off-by: Rob Clark <robdcl...@gmail.com>
> > ---
> > So, new scheme is, I think, a reasonable compromise between keeping the
> > python "clean" and keeping the intrinsic declarations easy to follow.
> > It still has the side-effect that intrinsic() adds to the table, but
> > drops the separate system_values table so that intrinsic() doesn't
> > return a value. The alternative would require the helper for various
> > specialized intrinsic categories to be declared far from where they are
> > used, which is, I think, suboptimal. And it keeps intrinsic() and
> > various wrappers pretty straightforward, so I don't think this should
> > ever pose a problem for refactoring (and certainly less of a problem
> > than the previous solution using cpp macros, so regardless of what your
> > opinion about the py code, I guess anyone could agree that this is an
> > improvement over the current state ;-))
> >
> > Also added autotools build support. Sorry scons and android. (Are we
> > ready to drop either of these in favor of nir?)
> >
> > src/compiler/Makefile.nir.am | 13 +-
> > src/compiler/Makefile.sources | 4 +-
> > src/compiler/nir/meson.build | 24 +-
> > src/compiler/nir/nir.h | 9 -
> > src/compiler/nir/nir_builder.h | 27 +-
> > src/compiler/nir/nir_builder_opcodes_h.py | 24 +-
> > src/compiler/nir/nir_intrinsics.h | 540
> > --------------------
> > src/compiler/nir/nir_intrinsics.py | 546
> > +++++++++++++++++++++
> > src/compiler/nir/nir_intrinsics_c.py | 68 +++
> > .../nir/{nir_intrinsics.c => nir_intrinsics_h.py} | 75 +--
> > 10 files changed, 711 insertions(+), 619 deletions(-)
> > delete mode 100644 src/compiler/nir/nir_intrinsics.h
> > create mode 100644 src/compiler/nir/nir_intrinsics.py
> > create mode 100644 src/compiler/nir/nir_intrinsics_c.py
> > rename src/compiler/nir/{nir_intrinsics.c => nir_intrinsics_h.py} (55%)
> >
> > diff --git a/src/compiler/Makefile.nir.am b/src/compiler/Makefile.nir.am
> > index 28033256762..d805f573d51 100644
> > --- a/src/compiler/Makefile.nir.am
> > +++ b/src/compiler/Makefile.nir.am
> > @@ -32,7 +32,7 @@ nir_libnir_la_SOURCES =
> > \
> > $(SPIRV_GENERATED_FILES) \
> > $(NIR_GENERATED_FILES)
> >
> > -nir/nir_builder_opcodes.h: nir/nir_opcodes.py nir/nir_builder_opcodes_h.py
> > +nir/nir_builder_opcodes.h: nir/nir_opcodes.py nir/nir_builder_opcodes_h.py
> > nir/nir_intrinsics.py
> > $(MKDIR_GEN)
> > $(PYTHON_GEN) $(srcdir)/nir/nir_builder_opcodes_h.py > $@ || ($(RM)
> > $@; false)
> >
> > @@ -40,6 +40,14 @@ nir/nir_constant_expressions.c: nir/nir_opcodes.py
> > nir/nir_constant_expressions.
> > $(MKDIR_GEN)
> > $(PYTHON_GEN) $(srcdir)/nir/nir_constant_expressions.py > $@ ||
> > ($(RM) $@; false)
> >
> > +nir/nir_intrinsics.h: nir/nir_intrinsics.py nir/nir_intrinsics_h.py
> > + $(MKDIR_GEN)
> > + $(PYTHON_GEN) $(srcdir)/nir/nir_intrinsics_h.py --outdir nir ||
> > ($(RM) $@; false)
> > +
> > +nir/nir_intrinsics.c: nir/nir_intrinsics.py nir/nir_intrinsics_c.py
> > + $(MKDIR_GEN)
> > + $(PYTHON_GEN) $(srcdir)/nir/nir_intrinsics_c.py --outdir nir ||
> > ($(RM) $@; false)
> > +
> > nir/nir_opcodes.h: nir/nir_opcodes.py nir/nir_opcodes_h.py
> > $(MKDIR_GEN)
> > $(PYTHON_GEN) $(srcdir)/nir/nir_opcodes_h.py > $@ || ($(RM) $@; false)
> > @@ -112,6 +120,9 @@ EXTRA_DIST += \
> > nir/nir_algebraic.py \
> > nir/nir_builder_opcodes_h.py \
> > nir/nir_constant_expressions.py \
> > + nir/nir_intrinsics.py \
> > + nir/nir_intrinsics_c.py \
> > + nir/nir_intrinsics_h.py \
> > nir/nir_opcodes.py \
> > nir/nir_opcodes_c.py \
> > nir/nir_opcodes_h.py \
> > diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
> > index 2dc48365507..273ce4cde49 100644
> > --- a/src/compiler/Makefile.sources
> > +++ b/src/compiler/Makefile.sources
> > @@ -182,6 +182,8 @@ LIBGLCPP_GENERATED_FILES = \
> > NIR_GENERATED_FILES = \
> > nir/nir_builder_opcodes.h \
> > nir/nir_constant_expressions.c \
> > + nir/nir_intrinsics.c \
> > + nir/nir_intrinsics.h \
> > nir/nir_opcodes.c \
> > nir/nir_opcodes.h \
> > nir/nir_opt_algebraic.c
> > @@ -202,8 +204,6 @@ NIR_FILES = \
> > nir/nir_inline_functions.c \
> > nir/nir_instr_set.c \
> > nir/nir_instr_set.h \
> > - nir/nir_intrinsics.c \
> > - nir/nir_intrinsics.h \
> > nir/nir_linking_helpers.c \
> > nir/nir_liveness.c \
> > nir/nir_loop_analyze.c \
> > diff --git a/src/compiler/nir/meson.build b/src/compiler/nir/meson.build
> > index 887035d7595..cbee73256ef 100644
> > --- a/src/compiler/nir/meson.build
> > +++ b/src/compiler/nir/meson.build
> > @@ -65,6 +65,24 @@ nir_opt_algebraic_c = custom_target(
> > depend_files : files('nir_algebraic.py'),
> > )
> >
> > +nir_intrinsics_h = custom_target(
> > + 'nir_intrinsics.h',
> > + input : 'nir_intrinsics_h.py',
> > + output : 'nir_intrinsics.h',
> > + command : [prog_python2, '@INPUT@', '--outdir',
> > meson.current_build_dir()],
> > + capture : false,
> > + depend_files : files('nir_intrinsics.py'),
> > +)
> > +
> > +nir_intrinsics_c = custom_target(
> > + 'nir_intrinsic.c',
> > + input : 'nir_intrinsics_c.py',
> > + output : 'nir_intrinsics.c',
> > + command : [prog_python2, '@INPUT@', '--outdir',
> > meson.current_build_dir()],
> > + capture: false,
> > + depend_files : files('nir_intrinsics.py'),
> > +)
> > +
> > spirv_info_c = custom_target(
> > 'spirv_info.c',
> > input : files('../spirv/spirv_info_c.py',
> > '../spirv/spirv.core.grammar.json'),
> > @@ -96,8 +114,6 @@ files_libnir = files(
> > 'nir_inline_functions.c',
> > 'nir_instr_set.c',
> > 'nir_instr_set.h',
> > - 'nir_intrinsics.c',
> > - 'nir_intrinsics.h',
> > 'nir_linking_helpers.c',
> > 'nir_liveness.c',
> > 'nir_loop_analyze.c',
> > @@ -202,7 +218,7 @@ libnir = static_library(
> > 'nir',
> > [files_libnir, spirv_info_c, nir_opt_algebraic_c, nir_opcodes_c,
> > nir_opcodes_h, nir_constant_expressions_c, nir_builder_opcodes_h,
> > - vtn_gather_types_c],
> > + vtn_gather_types_c, nir_intrinsics_c, nir_intrinsics_h],
> > include_directories : [inc_common, inc_compiler,
> > include_directories('../spirv')],
> > c_args : [c_vis_args, c_msvc_compat_args, no_override_init_args],
> > link_with : libcompiler,
> > @@ -211,7 +227,7 @@ libnir = static_library(
> >
> > # Headers-only dependency
> > idep_nir_headers = declare_dependency(
> > - sources : [nir_opcodes_h, nir_builder_opcodes_h],
> > + sources : [nir_opcodes_h, nir_builder_opcodes_h, nir_intrinsics_h],
> > include_directories : include_directories('.'),
> > )
> >
> > diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
> > index 0d207d0ea54..b474567be1c 100644
> > --- a/src/compiler/nir/nir.h
> > +++ b/src/compiler/nir/nir.h
> > @@ -945,16 +945,7 @@ typedef struct {
> > struct nir_function *callee;
> > } nir_call_instr;
> >
> > -#define INTRINSIC(name, num_srcs, src_components, has_dest,
> > dest_components, \
> > - num_variables, num_indices, idx0, idx1, idx2, flags) \
> > - nir_intrinsic_##name,
> > -
> > -#define LAST_INTRINSIC(name) nir_last_intrinsic = nir_intrinsic_##name,
> > -
> > -typedef enum {
> > #include "nir_intrinsics.h"
> > - nir_num_intrinsics = nir_last_intrinsic + 1
> > -} nir_intrinsic_op;
> >
> > #define NIR_INTRINSIC_MAX_CONST_INDEX 3
> >
> > diff --git a/src/compiler/nir/nir_builder.h b/src/compiler/nir/nir_builder.h
> > index 36e0ae3ac63..8f7ddf1483c 100644
> > --- a/src/compiler/nir/nir_builder.h
> > +++ b/src/compiler/nir/nir_builder.h
> > @@ -610,32 +610,7 @@ nir_copy_var(nir_builder *build, nir_variable *dest,
> > nir_variable *src)
> > nir_builder_instr_insert(build, ©->instr);
> > }
> >
> > -/* Generic builder for system values. */
> > -static inline nir_ssa_def *
> > -nir_load_system_value(nir_builder *build, nir_intrinsic_op op, int index)
> > -{
> > - nir_intrinsic_instr *load = nir_intrinsic_instr_create(build->shader,
> > op);
> > - load->num_components = nir_intrinsic_infos[op].dest_components;
> > - load->const_index[0] = index;
> > - nir_ssa_dest_init(&load->instr, &load->dest,
> > - nir_intrinsic_infos[op].dest_components, 32, NULL);
> > - nir_builder_instr_insert(build, &load->instr);
> > - return &load->dest.ssa;
> > -}
> > -
> > -/* Generate custom builders for system values. */
> > -#define INTRINSIC(name, num_srcs, src_components, has_dest,
> > dest_components, \
> > - num_variables, num_indices, idx0, idx1, idx2, flags)
> > -#define LAST_INTRINSIC(name)
> > -
> > -#define DEFINE_SYSTEM_VALUE(name) \
> > - static inline nir_ssa_def * \
> > - nir_load_##name(nir_builder *build) \
> > - { \
> > - return nir_load_system_value(build, nir_intrinsic_load_##name, 0); \
> > - }
> > -
> > -#include "nir_intrinsics.h"
> > +#include "nir_builder_opcodes.h"
> >
> > static inline nir_ssa_def *
> > nir_load_barycentric(nir_builder *build, nir_intrinsic_op op,
> > diff --git a/src/compiler/nir/nir_builder_opcodes_h.py
> > b/src/compiler/nir/nir_builder_opcodes_h.py
> > index bdabe578ce2..4a41e6079ed 100644
> > --- a/src/compiler/nir/nir_builder_opcodes_h.py
> > +++ b/src/compiler/nir/nir_builder_opcodes_h.py
> > @@ -41,9 +41,31 @@ nir_${name}(nir_builder *build,
> > ${src_decl_list(opcode.num_inputs)})
> > }
> > % endfor
> >
> > +/* Generic builder for system values. */
> > +static inline nir_ssa_def *
> > +nir_load_system_value(nir_builder *build, nir_intrinsic_op op, int index)
> > +{
> > + nir_intrinsic_instr *load = nir_intrinsic_instr_create(build->shader,
> > op);
> > + load->num_components = nir_intrinsic_infos[op].dest_components;
> > + load->const_index[0] = index;
> > + nir_ssa_dest_init(&load->instr, &load->dest,
> > + nir_intrinsic_infos[op].dest_components, 32, NULL);
> > + nir_builder_instr_insert(build, &load->instr);
> > + return &load->dest.ssa;
> > +}
> > +
> > +% for name, opcode in filter(lambda v: v[1].sysval,
> > sorted(INTR_OPCODES.iteritems())):
> > +static inline nir_ssa_def *
> > +nir_${name}(nir_builder *build)
> > +{
> > + return nir_load_system_value(build, nir_intrinsic_${name}, 0);
> > +}
> > +% endfor
> > +
> > #endif /* _NIR_BUILDER_OPCODES_ */"""
> >
> > from nir_opcodes import opcodes
> > +from nir_intrinsics import INTR_OPCODES
> > from mako.template import Template
> >
> > -print Template(template).render(opcodes=opcodes)
> > +print Template(template).render(opcodes=opcodes, INTR_OPCODES=INTR_OPCODES)
> > diff --git a/src/compiler/nir/nir_intrinsics.h
> > b/src/compiler/nir/nir_intrinsics.h
> > deleted file mode 100644
> > index df2790869ba..00000000000
> > --- a/src/compiler/nir/nir_intrinsics.h
> > +++ /dev/null
> > @@ -1,540 +0,0 @@
> > -/*
> > - * Copyright © 2014 Intel Corporation
> > - *
> > - * Permission is hereby granted, free of charge, to any person obtaining a
> > - * copy of this software and associated documentation files (the
> > "Software"),
> > - * to deal in the Software without restriction, including without
> > limitation
> > - * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> > - * and/or sell copies of the Software, and to permit persons to whom the
> > - * Software is furnished to do so, subject to the following conditions:
> > - *
> > - * The above copyright notice and this permission notice (including the
> > next
> > - * paragraph) shall be included in all copies or substantial portions of
> > the
> > - * Software.
> > - *
> > - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
> > OR
> > - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> > - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> > - * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
> > OTHER
> > - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> > - * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
> > DEALINGS
> > - * IN THE SOFTWARE.
> > - *
> > - * Authors:
> > - * Connor Abbott (cwabbo...@gmail.com)
> > - *
> > - */
> > -
> > -/**
> > - * This header file defines all the available intrinsics in one place. It
> > - * expands to a list of macros of the form:
> > - *
> > - * INTRINSIC(name, num_srcs, src_components, has_dest, dest_components,
> > - * num_variables, num_indices, idx0, idx1, idx2, flags)
> > - *
> > - * Which should correspond one-to-one with the nir_intrinsic_info
> > structure. It
> > - * is included in both ir.h to create the nir_intrinsic enum (with members
> > of
> > - * the form nir_intrinsic_(name)) and and in opcodes.c to create
> > - * nir_intrinsic_infos, which is a const array of nir_intrinsic_info
> > structures
> > - * for each intrinsic.
> > - */
> > -
> > -#define ARR(...) { __VA_ARGS__ }
> > -
> > -INTRINSIC(nop, 0, ARR(0), false, 0, 0, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE)
> > -
> > -INTRINSIC(load_var, 0, ARR(0), true, 0, 1, 0, xx, xx, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(store_var, 1, ARR(0), false, 0, 1, 1, WRMASK, xx, xx, 0)
> > -INTRINSIC(copy_var, 0, ARR(0), false, 0, 2, 0, xx, xx, xx, 0)
> > -
> > -/*
> > - * Interpolation of input. The interp_var_at* intrinsics are similar to
> > the
> > - * load_var intrinsic acting on a shader input except that they interpolate
> > - * the input differently. The at_sample and at_offset intrinsics take an
> > - * additional source that is an integer sample id or a vec2 position offset
> > - * respectively.
> > - */
> > -
> > -INTRINSIC(interp_var_at_centroid, 0, ARR(0), true, 0, 1, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -INTRINSIC(interp_var_at_sample, 1, ARR(1), true, 0, 1, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -INTRINSIC(interp_var_at_offset, 1, ARR(2), true, 0, 1, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -
> > -/*
> > - * Ask the driver for the size of a given buffer. It takes the buffer index
> > - * as source.
> > - */
> > -INTRINSIC(get_buffer_size, 1, ARR(1), true, 1, 0, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -
> > -/*
> > - * a barrier is an intrinsic with no inputs/outputs but which can't be
> > moved
> > - * around/optimized in general
> > - */
> > -#define BARRIER(name) INTRINSIC(name, 0, ARR(0), false, 0, 0, 0, xx, xx,
> > xx, 0)
> > -
> > -BARRIER(barrier)
> > -BARRIER(discard)
> > -
> > -/*
> > - * Memory barrier with semantics analogous to the memoryBarrier() GLSL
> > - * intrinsic.
> > - */
> > -BARRIER(memory_barrier)
> > -
> > -/*
> > - * Shader clock intrinsic with semantics analogous to the clock2x32ARB()
> > - * GLSL intrinsic.
> > - * The latter can be used as code motion barrier, which is currently not
> > - * feasible with NIR.
> > - */
> > -INTRINSIC(shader_clock, 0, ARR(0), true, 2, 0, 0, xx, xx, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE)
> > -
> > -/*
> > - * Shader ballot intrinsics with semantics analogous to the
> > - *
> > - * ballotARB()
> > - * readInvocationARB()
> > - * readFirstInvocationARB()
> > - *
> > - * GLSL functions from ARB_shader_ballot.
> > - */
> > -INTRINSIC(ballot, 1, ARR(1), true, 0, 0, 0, xx, xx, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(read_invocation, 2, ARR(0, 1), true, 0, 0, 0, xx, xx, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(read_first_invocation, 1, ARR(0), true, 0, 0, 0, xx, xx, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE)
> > -
> > -/** Additional SPIR-V ballot intrinsics
> > - *
> > - * These correspond to the SPIR-V opcodes
> > - *
> > - * OpGroupUniformElect
> > - * OpSubgroupFirstInvocationKHR
> > - */
> > -INTRINSIC(elect, 0, ARR(0), true, 1, 0, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(first_invocation, 0, ARR(0), true, 1, 0, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE)
> > -
> > -/*
> > - * Memory barrier with semantics analogous to the compute shader
> > - * groupMemoryBarrier(), memoryBarrierAtomicCounter(),
> > memoryBarrierBuffer(),
> > - * memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
> > - */
> > -BARRIER(group_memory_barrier)
> > -BARRIER(memory_barrier_atomic_counter)
> > -BARRIER(memory_barrier_buffer)
> > -BARRIER(memory_barrier_image)
> > -BARRIER(memory_barrier_shared)
> > -
> > -/** A conditional discard, with a single boolean source. */
> > -INTRINSIC(discard_if, 1, ARR(1), false, 0, 0, 0, xx, xx, xx, 0)
> > -
> > -/** ARB_shader_group_vote intrinsics */
> > -INTRINSIC(vote_any, 1, ARR(1), true, 1, 0, 0, xx, xx, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(vote_all, 1, ARR(1), true, 1, 0, 0, xx, xx, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(vote_feq, 1, ARR(0), true, 1, 0, 0, xx, xx, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(vote_ieq, 1, ARR(0), true, 1, 0, 0, xx, xx, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE)
> > -
> > -/** Ballot ALU operations from SPIR-V.
> > - *
> > - * These operations work like their ALU counterparts except that the
> > operate
> > - * on a uvec4 which is treated as a 128bit integer. Also, they are, in
> > - * general, free to ignore any bits which are above the subgroup size.
> > - */
> > -INTRINSIC(ballot_bitfield_extract, 2, ARR(4, 1), true, 1, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(ballot_bit_count_reduce, 1, ARR(4), true, 1, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(ballot_bit_count_inclusive, 1, ARR(4), true, 1, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(ballot_bit_count_exclusive, 1, ARR(4), true, 1, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(ballot_find_lsb, 1, ARR(4), true, 1, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(ballot_find_msb, 1, ARR(4), true, 1, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -
> > -/** Shuffle operations from SPIR-V. */
> > -INTRINSIC(shuffle, 2, ARR(0, 1), true, 0, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(shuffle_xor, 2, ARR(0, 1), true, 0, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(shuffle_up, 2, ARR(0, 1), true, 0, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(shuffle_down, 2, ARR(0, 1), true, 0, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -
> > -/** Quad operations from SPIR-V. */
> > -INTRINSIC(quad_broadcast, 2, ARR(0, 1), true, 0, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(quad_swap_horizontal, 1, ARR(0), true, 0, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(quad_swap_vertical, 1, ARR(0), true, 0, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(quad_swap_diagonal, 1, ARR(0), true, 0, 0,
> > - 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -
> > -INTRINSIC(reduce, 1, ARR(0), true, 0, 0,
> > - 2, REDUCTION_OP, CLUSTER_SIZE, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(inclusive_scan, 1, ARR(0), true, 0, 0,
> > - 1, REDUCTION_OP, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(exclusive_scan, 1, ARR(0), true, 0, 0,
> > - 1, REDUCTION_OP, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -
> > -/**
> > - * Basic Geometry Shader intrinsics.
> > - *
> > - * emit_vertex implements GLSL's EmitStreamVertex() built-in. It takes a
> > single
> > - * index, which is the stream ID to write to.
> > - *
> > - * end_primitive implements GLSL's EndPrimitive() built-in.
> > - */
> > -INTRINSIC(emit_vertex, 0, ARR(0), false, 0, 0, 1, STREAM_ID, xx, xx, 0)
> > -INTRINSIC(end_primitive, 0, ARR(0), false, 0, 0, 1, STREAM_ID, xx, xx, 0)
> > -
> > -/**
> > - * Geometry Shader intrinsics with a vertex count.
> > - *
> > - * Alternatively, drivers may implement these intrinsics, and use
> > - * nir_lower_gs_intrinsics() to convert from the basic intrinsics.
> > - *
> > - * These maintain a count of the number of vertices emitted, as an
> > additional
> > - * unsigned integer source.
> > - */
> > -INTRINSIC(emit_vertex_with_counter, 1, ARR(1), false, 0, 0, 1, STREAM_ID,
> > xx, xx, 0)
> > -INTRINSIC(end_primitive_with_counter, 1, ARR(1), false, 0, 0, 1,
> > STREAM_ID, xx, xx, 0)
> > -INTRINSIC(set_vertex_count, 1, ARR(1), false, 0, 0, 0, xx, xx, xx, 0)
> > -
> > -/*
> > - * Atomic counters
> > - *
> > - * The *_var variants take an atomic_uint nir_variable, while the other,
> > - * lowered, variants take a constant buffer index and register offset.
> > - */
> > -
> > -#define ATOMIC(name, flags) \
> > - INTRINSIC(name##_var, 0, ARR(0), true, 1, 1, 0, xx, xx, xx, flags) \
> > - INTRINSIC(name, 1, ARR(1), true, 1, 0, 1, BASE, xx, xx, flags)
> > -#define ATOMIC2(name) \
> > - INTRINSIC(name##_var, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) \
> > - INTRINSIC(name, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
> > -#define ATOMIC3(name) \
> > - INTRINSIC(name##_var, 2, ARR(1, 1), true, 1, 1, 0, xx, xx, xx, 0) \
> > - INTRINSIC(name, 3, ARR(1, 1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
> > -
> > -ATOMIC(atomic_counter_inc, 0)
> > -ATOMIC(atomic_counter_dec, 0)
> > -ATOMIC(atomic_counter_read, NIR_INTRINSIC_CAN_ELIMINATE)
> > -ATOMIC2(atomic_counter_add)
> > -ATOMIC2(atomic_counter_min)
> > -ATOMIC2(atomic_counter_max)
> > -ATOMIC2(atomic_counter_and)
> > -ATOMIC2(atomic_counter_or)
> > -ATOMIC2(atomic_counter_xor)
> > -ATOMIC2(atomic_counter_exchange)
> > -ATOMIC3(atomic_counter_comp_swap)
> > -
> > -/*
> > - * Image load, store and atomic intrinsics.
> > - *
> > - * All image intrinsics take an image target passed as a nir_variable.
> > Image
> > - * variables contain a number of memory and layout qualifiers that
> > influence
> > - * the semantics of the intrinsic.
> > - *
> > - * All image intrinsics take a four-coordinate vector and a sample index as
> > - * first two sources, determining the location within the image that will
> > be
> > - * accessed by the intrinsic. Components not applicable to the image
> > target
> > - * in use are undefined. Image store takes an additional four-component
> > - * argument with the value to be written, and image atomic operations take
> > - * either one or two additional scalar arguments with the same meaning as
> > in
> > - * the ARB_shader_image_load_store specification.
> > - */
> > -INTRINSIC(image_var_load, 2, ARR(4, 1), true, 4, 1, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE)
> > -INTRINSIC(image_var_store, 3, ARR(4, 1, 4), false, 0, 1, 0, xx, xx, xx, 0)
> > -INTRINSIC(image_var_atomic_add, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx,
> > xx, 0)
> > -INTRINSIC(image_var_atomic_min, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx,
> > xx, 0)
> > -INTRINSIC(image_var_atomic_max, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx,
> > xx, 0)
> > -INTRINSIC(image_var_atomic_and, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx,
> > xx, 0)
> > -INTRINSIC(image_var_atomic_or, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx,
> > 0)
> > -INTRINSIC(image_var_atomic_xor, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx,
> > xx, 0)
> > -INTRINSIC(image_var_atomic_exchange, 3, ARR(4, 1, 1), true, 1, 1, 0, xx,
> > xx, xx, 0)
> > -INTRINSIC(image_var_atomic_comp_swap, 4, ARR(4, 1, 1, 1), true, 1, 1, 0,
> > xx, xx, xx, 0)
> > -INTRINSIC(image_var_size, 0, ARR(0), true, 0, 1, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -INTRINSIC(image_var_samples, 0, ARR(0), true, 1, 1, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -
> > -/*
> > - * Vulkan descriptor set intrinsics
> > - *
> > - * The Vulkan API uses a different binding model from GL. In the Vulkan
> > - * API, all external resources are represented by a tuple:
> > - *
> > - * (descriptor set, binding, array index)
> > - *
> > - * where the array index is the only thing allowed to be indirect. The
> > - * vulkan_surface_index intrinsic takes the descriptor set and binding as
> > - * its first two indices and the array index as its source. The third
> > - * index is a nir_variable_mode in case that's useful to the backend.
> > - *
> > - * The intended usage is that the shader will call vulkan_surface_index to
> > - * get an index and then pass that as the buffer index ubo/ssbo calls.
> > - *
> > - * The vulkan_resource_reindex intrinsic takes a resource index in src0
> > - * (the result of a vulkan_resource_index or vulkan_resource_reindex) which
> > - * corresponds to the tuple (set, binding, index) and computes an index
> > - * corresponding to tuple (set, binding, idx + src1).
> > - */
> > -INTRINSIC(vulkan_resource_index, 1, ARR(1), true, 1, 0, 2,
> > - DESC_SET, BINDING, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -INTRINSIC(vulkan_resource_reindex, 2, ARR(1, 1), true, 1, 0, 0, xx, xx, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -
> > -/*
> > - * variable atomic intrinsics
> > - *
> > - * All of these variable atomic memory operations read a value from memory,
> > - * compute a new value using one of the operations below, write the new
> > value
> > - * to memory, and return the original value read.
> > - *
> > - * All operations take 1 source except CompSwap that takes 2. These sources
> > - * represent:
> > - *
> > - * 0: The data parameter to the atomic function (i.e. the value to add
> > - * in shared_atomic_add, etc).
> > - * 1: For CompSwap only: the second data parameter.
> > - *
> > - * All operations take 1 variable deref.
> > - */
> > -INTRINSIC(var_atomic_add, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
> > -INTRINSIC(var_atomic_imin, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
> > -INTRINSIC(var_atomic_umin, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
> > -INTRINSIC(var_atomic_imax, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
> > -INTRINSIC(var_atomic_umax, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
> > -INTRINSIC(var_atomic_and, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
> > -INTRINSIC(var_atomic_or, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
> > -INTRINSIC(var_atomic_xor, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
> > -INTRINSIC(var_atomic_exchange, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
> > -INTRINSIC(var_atomic_comp_swap, 2, ARR(1, 1), true, 1, 1, 0, xx, xx, xx, 0)
> > -
> > -/*
> > - * SSBO atomic intrinsics
> > - *
> > - * All of the SSBO atomic memory operations read a value from memory,
> > - * compute a new value using one of the operations below, write the new
> > - * value to memory, and return the original value read.
> > - *
> > - * All operations take 3 sources except CompSwap that takes 4. These
> > - * sources represent:
> > - *
> > - * 0: The SSBO buffer index.
> > - * 1: The offset into the SSBO buffer of the variable that the atomic
> > - * operation will operate on.
> > - * 2: The data parameter to the atomic function (i.e. the value to add
> > - * in ssbo_atomic_add, etc).
> > - * 3: For CompSwap only: the second data parameter.
> > - */
> > -INTRINSIC(ssbo_atomic_add, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
> > -INTRINSIC(ssbo_atomic_imin, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
> > -INTRINSIC(ssbo_atomic_umin, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
> > -INTRINSIC(ssbo_atomic_imax, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
> > -INTRINSIC(ssbo_atomic_umax, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
> > -INTRINSIC(ssbo_atomic_and, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
> > -INTRINSIC(ssbo_atomic_or, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
> > -INTRINSIC(ssbo_atomic_xor, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
> > -INTRINSIC(ssbo_atomic_exchange, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx,
> > xx, 0)
> > -INTRINSIC(ssbo_atomic_comp_swap, 4, ARR(1, 1, 1, 1), true, 1, 0, 0, xx,
> > xx, xx, 0)
> > -
> > -/*
> > - * CS shared variable atomic intrinsics
> > - *
> > - * All of the shared variable atomic memory operations read a value from
> > - * memory, compute a new value using one of the operations below, write the
> > - * new value to memory, and return the original value read.
> > - *
> > - * All operations take 2 sources except CompSwap that takes 3. These
> > - * sources represent:
> > - *
> > - * 0: The offset into the shared variable storage region that the atomic
> > - * operation will operate on.
> > - * 1: The data parameter to the atomic function (i.e. the value to add
> > - * in shared_atomic_add, etc).
> > - * 2: For CompSwap only: the second data parameter.
> > - */
> > -INTRINSIC(shared_atomic_add, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
> > -INTRINSIC(shared_atomic_imin, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
> > -INTRINSIC(shared_atomic_umin, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
> > -INTRINSIC(shared_atomic_imax, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
> > -INTRINSIC(shared_atomic_umax, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
> > -INTRINSIC(shared_atomic_and, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
> > -INTRINSIC(shared_atomic_or, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
> > -INTRINSIC(shared_atomic_xor, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
> > -INTRINSIC(shared_atomic_exchange, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx,
> > xx, 0)
> > -INTRINSIC(shared_atomic_comp_swap, 3, ARR(1, 1, 1), true, 1, 0, 1, BASE,
> > xx, xx, 0)
> > -
> > -/* Used by nir_builder.h to generate loader helpers for the system values.
> > */
> > -#ifndef DEFINE_SYSTEM_VALUE
> > -#define DEFINE_SYSTEM_VALUE(name)
> > -#endif
> > -
> > -#define SYSTEM_VALUE(name, components, num_indices, idx0, idx1, idx2) \
> > - DEFINE_SYSTEM_VALUE(name) \
> > - INTRINSIC(load_##name, 0, ARR(0), true, components, 0, num_indices, \
> > - idx0, idx1, idx2, \
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -
> > -SYSTEM_VALUE(frag_coord, 4, 0, xx, xx, xx)
> > -SYSTEM_VALUE(front_face, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(vertex_id, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(vertex_id_zero_base, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(base_vertex, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(instance_id, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(base_instance, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(draw_id, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(sample_id, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(sample_pos, 2, 0, xx, xx, xx)
> > -SYSTEM_VALUE(sample_mask_in, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(primitive_id, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(invocation_id, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(tess_coord, 3, 0, xx, xx, xx)
> > -SYSTEM_VALUE(tess_level_outer, 4, 0, xx, xx, xx)
> > -SYSTEM_VALUE(tess_level_inner, 2, 0, xx, xx, xx)
> > -SYSTEM_VALUE(patch_vertices_in, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(local_invocation_id, 3, 0, xx, xx, xx)
> > -SYSTEM_VALUE(local_invocation_index, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(work_group_id, 3, 0, xx, xx, xx)
> > -SYSTEM_VALUE(user_clip_plane, 4, 1, UCP_ID, xx, xx)
> > -SYSTEM_VALUE(num_work_groups, 3, 0, xx, xx, xx)
> > -SYSTEM_VALUE(helper_invocation, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(alpha_ref_float, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(layer_id, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(view_index, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(subgroup_size, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(subgroup_invocation, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(subgroup_eq_mask, 0, 0, xx, xx, xx)
> > -SYSTEM_VALUE(subgroup_ge_mask, 0, 0, xx, xx, xx)
> > -SYSTEM_VALUE(subgroup_gt_mask, 0, 0, xx, xx, xx)
> > -SYSTEM_VALUE(subgroup_le_mask, 0, 0, xx, xx, xx)
> > -SYSTEM_VALUE(subgroup_lt_mask, 0, 0, xx, xx, xx)
> > -SYSTEM_VALUE(num_subgroups, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(subgroup_id, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(local_group_size, 3, 0, xx, xx, xx)
> > -
> > -/* Blend constant color values. Float values are clamped. */
> > -SYSTEM_VALUE(blend_const_color_r_float, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(blend_const_color_g_float, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(blend_const_color_b_float, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(blend_const_color_a_float, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(blend_const_color_rgba8888_unorm, 1, 0, xx, xx, xx)
> > -SYSTEM_VALUE(blend_const_color_aaaa8888_unorm, 1, 0, xx, xx, xx)
> > -
> > -/**
> > - * Barycentric coordinate intrinsics.
> > - *
> > - * These set up the barycentric coordinates for a particular interpolation.
> > - * The first three are for the simple cases: pixel, centroid, or per-sample
> > - * (at gl_SampleID). The next two handle interpolating at a specified
> > - * sample location, or interpolating with a vec2 offset,
> > - *
> > - * The interp_mode index should be either the INTERP_MODE_SMOOTH or
> > - * INTERP_MODE_NOPERSPECTIVE enum values.
> > - *
> > - * The vec2 value produced by these intrinsics is intended for use as the
> > - * barycoord source of a load_interpolated_input intrinsic.
> > - */
> > -
> > -#define BARYCENTRIC(name, sources, source_components) \
> > - INTRINSIC(load_barycentric_##name, sources, ARR(source_components), \
> > - true, 2, 0, 1, INTERP_MODE, xx, xx, \
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -
> > -/* no sources. const_index[] = { interp_mode } */
> > -BARYCENTRIC(pixel, 0, 0)
> > -BARYCENTRIC(centroid, 0, 0)
> > -BARYCENTRIC(sample, 0, 0)
> > -/* src[] = { sample_id }. const_index[] = { interp_mode } */
> > -BARYCENTRIC(at_sample, 1, 1)
> > -/* src[] = { offset.xy }. const_index[] = { interp_mode } */
> > -BARYCENTRIC(at_offset, 1, 2)
> > -
> > -/*
> > - * Load operations pull data from some piece of GPU memory. All load
> > - * operations operate in terms of offsets into some piece of theoretical
> > - * memory. Loads from externally visible memory (UBO and SSBO) simply
> > take a
> > - * byte offset as a source. Loads from opaque memory (uniforms, inputs,
> > etc.)
> > - * take a base+offset pair where the base (const_index[0]) gives the
> > location
> > - * of the start of the variable being loaded and and the offset source is a
> > - * offset into that variable.
> > - *
> > - * Uniform load operations have a second "range" index that specifies the
> > - * range (starting at base) of the data from which we are loading. If
> > - * const_index[1] == 0, then the range is unknown.
> > - *
> > - * Some load operations such as UBO/SSBO load and per_vertex loads take an
> > - * additional source to specify which UBO/SSBO/vertex to load from.
> > - *
> > - * The exact address type depends on the lowering pass that generates the
> > - * load/store intrinsics. Typically, this is vec4 units for things such as
> > - * varying slots and float units for fragment shader inputs. UBO and SSBO
> > - * offsets are always in bytes.
> > - */
> > -
> > -#define LOAD(name, srcs, num_indices, idx0, idx1, idx2, flags) \
> > - INTRINSIC(load_##name, srcs, ARR(1, 1, 1, 1), true, 0, 0, num_indices,
> > idx0, idx1, idx2, flags)
> > -
> > -/* src[] = { offset }. const_index[] = { base, range } */
> > -LOAD(uniform, 1, 2, BASE, RANGE, xx, NIR_INTRINSIC_CAN_ELIMINATE |
> > NIR_INTRINSIC_CAN_REORDER)
> > -/* src[] = { buffer_index, offset }. No const_index */
> > -LOAD(ubo, 2, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE |
> > NIR_INTRINSIC_CAN_REORDER)
> > -/* src[] = { offset }. const_index[] = { base, component } */
> > -LOAD(input, 1, 2, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE |
> > NIR_INTRINSIC_CAN_REORDER)
> > -/* src[] = { vertex, offset }. const_index[] = { base, component } */
> > -LOAD(per_vertex_input, 2, 2, BASE, COMPONENT, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -/* src[] = { barycoord, offset }. const_index[] = { base, component } */
> > -INTRINSIC(load_interpolated_input, 2, ARR(2, 1), true, 0, 0,
> > - 2, BASE, COMPONENT, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -
> > -/* src[] = { buffer_index, offset }. No const_index */
> > -LOAD(ssbo, 2, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -/* src[] = { offset }. const_index[] = { base, component } */
> > -LOAD(output, 1, 2, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -/* src[] = { vertex, offset }. const_index[] = { base, component } */
> > -LOAD(per_vertex_output, 2, 2, BASE, COMPONENT, xx,
> > NIR_INTRINSIC_CAN_ELIMINATE)
> > -/* src[] = { offset }. const_index[] = { base } */
> > -LOAD(shared, 1, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
> > -/* src[] = { offset }. const_index[] = { base, range } */
> > -LOAD(push_constant, 1, 2, BASE, RANGE, xx,
> > - NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> > -
> > -/*
> > - * Stores work the same way as loads, except now the first source is the
> > value
> > - * to store and the second (and possibly third) source specify where to
> > store
> > - * the value. SSBO and shared memory stores also have a write mask as
> > - * const_index[0].
> > - */
> > -
> > -#define STORE(name, srcs, num_indices, idx0, idx1, idx2, flags) \
> > - INTRINSIC(store_##name, srcs, ARR(0, 1, 1, 1), false, 0, 0,
> > num_indices, idx0, idx1, idx2, flags)
> > -
> > -/* src[] = { value, offset }. const_index[] = { base, write_mask,
> > component } */
> > -STORE(output, 2, 3, BASE, WRMASK, COMPONENT, 0)
> > -/* src[] = { value, vertex, offset }.
> > - * const_index[] = { base, write_mask, component }
> > - */
> > -STORE(per_vertex_output, 3, 3, BASE, WRMASK, COMPONENT, 0)
> > -/* src[] = { value, block_index, offset }. const_index[] = { write_mask }
> > */
> > -STORE(ssbo, 3, 1, WRMASK, xx, xx, 0)
> > -/* src[] = { value, offset }. const_index[] = { base, write_mask } */
> > -STORE(shared, 2, 2, BASE, WRMASK, xx, 0)
> > -
> > -LAST_INTRINSIC(store_shared)
> > -
> > -#undef DEFINE_SYSTEM_VALUE
> > -#undef INTRINSIC
> > -#undef LAST_INTRINSIC
> > diff --git a/src/compiler/nir/nir_intrinsics.py
> > b/src/compiler/nir/nir_intrinsics.py
> > new file mode 100644
> > index 00000000000..1bc99552cd7
> > --- /dev/null
> > +++ b/src/compiler/nir/nir_intrinsics.py
> > @@ -0,0 +1,546 @@
> > +#
> > +# Copyright (C) 2018 Red Hat
> > +# Copyright (C) 2014 Intel Corporation
> > +#
> > +# Permission is hereby granted, free of charge, to any person obtaining a
> > +# copy of this software and associated documentation files (the
> > "Software"),
> > +# to deal in the Software without restriction, including without limitation
> > +# the rights to use, copy, modify, merge, publish, distribute, sublicense,
> > +# and/or sell copies of the Software, and to permit persons to whom the
> > +# Software is furnished to do so, subject to the following conditions:
> > +#
> > +# The above copyright notice and this permission notice (including the next
> > +# paragraph) shall be included in all copies or substantial portions of the
> > +# Software.
> > +#
> > +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
> > OR
> > +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> > +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> > +# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
> > OTHER
> > +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> > +# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
> > DEALINGS
> > +# IN THE SOFTWARE.
> > +#
> > +
> > +# This file defines all the available intrinsics in one place.
> > +#
> > +# The Intrinsic class corresponds one-to-one with nir_intrinsic_info
> > +# structure.
> > +
> > +class Intrinsic(object):
> > + """Class that represents all the information about an intrinsic opcode.
> > + NOTE: this must be kept in sync with nir_intrinsic_info.
> > + """
> > + def __init__(self, name, src_components, dest_components, num_variables,
> > + indices, flags, sysval):
> > + """Parameters:
> > +
> > + - name: the intrinsic name
> > + - src_components: list of the number of components per src, 0 means
> > + vectorized instruction with number of components given in the
> > + num_components field in nir_intrinsic_instr.
> > + - dest_components: number of destination components, -1 means no
> > + dest, 0 means number of components given in num_components field
> > + in nir_intrinsic_instr.
> > + - num_variables: the number of variables
> > + - indices: list of constant indicies
> > + - flags: list of semantic flags
> > + - sysval: is this a system-value intrinsic
> > + """
> > + assert isinstance(name, str)
> > + assert isinstance(src_components, list)
> > + if src_components:
> > + assert isinstance(src_components[0], int)
> > + assert isinstance(dest_components, int)
> > + assert isinstance(num_variables, int)
> > + assert isinstance(indices, list)
> > + if indices:
> > + assert isinstance(indices[0], str)
> > + assert isinstance(flags, list)
> > + if flags:
> > + assert isinstance(flags[0], str)
> > + assert isinstance(sysval, bool)
> > +
> > + self.name = name
> > + self.num_srcs = len(src_components)
> > + self.src_components = src_components
> > + self.has_dest = (dest_components >= 0)
> > + self.dest_components = dest_components
> > + self.num_variables = num_variables
> > + self.num_indices = len(indices)
> > + self.indices = indices
> > + self.flags = flags
> > + self.sysval = sysval
> > +
> > +#
> > +# Possible indices:
> > +#
> > +
> > +# A constant 'base' value that is added to an offset src:
> > +BASE = "NIR_INTRINSIC_BASE"
> > +# For store instructions, a writemask:
> > +WRMASK = "NIR_INTRINSIC_WRMASK"
> > +# The stream-id for GS emit_vertex/end_primitive intrinsics:
> > +STREAM_ID = "NIR_INTRINSIC_STREAM_ID"
> > +# The clip-plane id for load_user_clip_plane intrinsics:
> > +UCP_ID = "NIR_INTRINSIC_UCP_ID"
> > +# The amount of data, starting from BASE, that this instruction
> > +# may access. This is used to provide bounds if the offset is
> > +# not constant.
> > +RANGE = "NIR_INTRINSIC_RANGE"
> > +# The vulkan descriptor set binding for vulkan_resource_index
> > +# intrinsic
> > +DESC_SET = "NIR_INTRINSIC_DESC_SET"
> > +# The vulkan descriptor set binding for vulkan_resource_index
> > +# intrinsic
> > +BINDING = "NIR_INTRINSIC_BINDING"
> > +# Component offset
> > +COMPONENT = "NIR_INTRINSIC_COMPONENT"
> > +# Interpolation mode (only meaningful for FS inputs)
> > +INTERP_MODE = "NIR_INTRINSIC_INTERP_MODE"
> > +# A binary nir_op to use when performing a reduction or scan operation
> > +REDUCTION_OP = "NIR_INTRINSIC_REDUCTION_OP"
> > +# Cluster size for reduction operations
> > +CLUSTER_SIZE = "NIR_INTRINSIC_CLUSTER_SIZE"
> > +
> > +#
> > +# Possible flags:
> > +#
> > +
> > +CAN_ELIMINATE = "NIR_INTRINSIC_CAN_ELIMINATE"
> > +CAN_REORDER = "NIR_INTRINSIC_CAN_REORDER"
> > +
> > +INTR_OPCODES = {}
> > +
> > +def intrinsic(name, src_comp=[], dest_comp=-1, num_vars=0, indices=[],
> > + flags=[], sysval=False):
> > + assert name not in INTR_OPCODES
> > + INTR_OPCODES[name] = Intrinsic(name, src_comp, dest_comp, num_vars,
> > + indices, flags, sysval)
> > +
> > +intrinsic("nop", flags=[CAN_ELIMINATE])
> > +
> > +intrinsic("load_var", dest_comp=0, num_vars=1, flags=[CAN_ELIMINATE])
> > +intrinsic("store_var", src_comp=[0], num_vars=1, indices=[WRMASK])
> > +intrinsic("copy_var", num_vars=2)
> > +
> > +# Interpolation of input. The interp_var_at* intrinsics are similar to the
> > +# load_var intrinsic acting on a shader input except that they interpolate
> > +# the input differently. The at_sample and at_offset intrinsics take an
> > +# additional source that is an integer sample id or a vec2 position offset
> > +# respectively.
> > +
> > +intrinsic("interp_var_at_centroid", dest_comp=0, num_vars=1,
> > + flags=[ CAN_ELIMINATE, CAN_REORDER])
> > +intrinsic("interp_var_at_sample", src_comp=[1], dest_comp=0, num_vars=1,
> > + flags=[CAN_ELIMINATE, CAN_REORDER])
> > +intrinsic("interp_var_at_offset", src_comp=[2], dest_comp=0, num_vars=1,
> > + flags=[CAN_ELIMINATE, CAN_REORDER])
> > +
> > +# Ask the driver for the size of a given buffer. It takes the buffer index
> > +# as source.
> > +intrinsic("get_buffer_size", src_comp=[1], dest_comp=1,
> > + flags=[CAN_ELIMINATE, CAN_REORDER])
> > +
> > +# a barrier is an intrinsic with no inputs/outputs but which can't be moved
> > +# around/optimized in general
> > +def barrier(name):
> > + intrinsic(name)
> > +
> > +barrier("barrier")
> > +barrier("discard")
> > +
> > +# Memory barrier with semantics analogous to the memoryBarrier() GLSL
> > +# intrinsic.
> > +barrier("memory_barrier")
> > +
> > +# Shader clock intrinsic with semantics analogous to the clock2x32ARB()
> > +# GLSL intrinsic.
> > +# The latter can be used as code motion barrier, which is currently not
> > +# feasible with NIR.
> > +intrinsic("shader_clock", dest_comp=2, flags=[CAN_ELIMINATE])
> > +
> > +# Shader ballot intrinsics with semantics analogous to the
> > +#
> > +# ballotARB()
> > +# readInvocationARB()
> > +# readFirstInvocationARB()
> > +#
> > +# GLSL functions from ARB_shader_ballot.
> > +intrinsic("ballot", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE])
> > +intrinsic("read_invocation", src_comp=[0, 1], dest_comp=0,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("read_first_invocation", src_comp=[0], dest_comp=0,
> > flags=[CAN_ELIMINATE])
> > +
> > +# Additional SPIR-V ballot intrinsics
> > +#
> > +# These correspond to the SPIR-V opcodes
> > +#
> > +# OpGroupUniformElect
> > +# OpSubgroupFirstInvocationKHR
> > +intrinsic("elect", dest_comp=1, flags=[CAN_ELIMINATE])
> > +intrinsic("first_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
> > +
> > +# Memory barrier with semantics analogous to the compute shader
> > +# groupMemoryBarrier(), memoryBarrierAtomicCounter(),
> > memoryBarrierBuffer(),
> > +# memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
> > +barrier("group_memory_barrier")
> > +barrier("memory_barrier_atomic_counter")
> > +barrier("memory_barrier_buffer")
> > +barrier("memory_barrier_image")
> > +barrier("memory_barrier_shared")
> > +
> > +# A conditional discard, with a single boolean source.
> > +intrinsic("discard_if", src_comp=[1])
> > +
> > +# ARB_shader_group_vote intrinsics
> > +intrinsic("vote_any", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
> > +intrinsic("vote_all", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
> > +intrinsic("vote_feq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
> > +intrinsic("vote_ieq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
> > +
> > +# Ballot ALU operations from SPIR-V.
> > +#
> > +# These operations work like their ALU counterparts except that the operate
> > +# on a uvec4 which is treated as a 128bit integer. Also, they are, in
> > +# general, free to ignore any bits which are above the subgroup size.
> > +intrinsic("ballot_bitfield_extract", src_comp=[4, 1], dest_comp=1,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("ballot_bit_count_reduce", src_comp=[4], dest_comp=1,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("ballot_bit_count_inclusive", src_comp=[4], dest_comp=1,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("ballot_bit_count_exclusive", src_comp=[4], dest_comp=1,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("ballot_find_lsb", src_comp=[4], dest_comp=1,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("ballot_find_msb", src_comp=[4], dest_comp=1,
> > flags=[CAN_ELIMINATE])
> > +
> > +# Shuffle operations from SPIR-V.
> > +intrinsic("shuffle", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
> > +intrinsic("shuffle_xor", src_comp=[0, 1], dest_comp=0,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("shuffle_up", src_comp=[0, 1], dest_comp=0,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("shuffle_down", src_comp=[0, 1], dest_comp=0,
> > flags=[CAN_ELIMINATE])
> > +
> > +# Quad operations from SPIR-V.
> > +intrinsic("quad_broadcast", src_comp=[0, 1], dest_comp=0,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("quad_swap_horizontal", src_comp=[0], dest_comp=0,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("quad_swap_vertical", src_comp=[0], dest_comp=0,
> > flags=[CAN_ELIMINATE])
> > +intrinsic("quad_swap_diagonal", src_comp=[0], dest_comp=0,
> > flags=[CAN_ELIMINATE])
> > +
> > +intrinsic("reduce", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP,
> > CLUSTER_SIZE],
> > + flags=[CAN_ELIMINATE])
> > +intrinsic("inclusive_scan", src_comp=[0], dest_comp=0,
> > indices=[REDUCTION_OP],
> > + flags=[CAN_ELIMINATE])
> > +intrinsic("exclusive_scan", src_comp=[0], dest_comp=0,
> > indices=[REDUCTION_OP],
> > + flags=[CAN_ELIMINATE])
> > +
> > +# Basic Geometry Shader intrinsics.
> > +#
> > +# emit_vertex implements GLSL's EmitStreamVertex() built-in. It takes a
> > single
> > +# index, which is the stream ID to write to.
> > +#
> > +# end_primitive implements GLSL's EndPrimitive() built-in.
> > +intrinsic("emit_vertex", indices=[STREAM_ID])
> > +intrinsic("end_primitive", indices=[STREAM_ID])
> > +
> > +# Geometry Shader intrinsics with a vertex count.
> > +#
> > +# Alternatively, drivers may implement these intrinsics, and use
> > +# nir_lower_gs_intrinsics() to convert from the basic intrinsics.
> > +#
> > +# These maintain a count of the number of vertices emitted, as an
> > additional
> > +# unsigned integer source.
> > +intrinsic("emit_vertex_with_counter", src_comp=[1], indices=[STREAM_ID])
> > +intrinsic("end_primitive_with_counter", src_comp=[1], indices=[STREAM_ID])
> > +intrinsic("set_vertex_count", src_comp=[1])
> > +
> > +# Atomic counters
> > +#
> > +# The *_var variants take an atomic_uint nir_variable, while the other,
> > +# lowered, variants take a constant buffer index and register offset.
> > +
> > +def atomic(name, flags=[]):
> > + intrinsic(name + "_var", dest_comp=1, num_vars=1, flags=flags)
> > + intrinsic(name, src_comp=[1], dest_comp=1, indices=[BASE], flags=flags)
> > +
> > +def atomic2(name):
> > + intrinsic(name + "_var", src_comp=[1], dest_comp=1, num_vars=1)
> > + intrinsic(name, src_comp=[1, 1], dest_comp=1, indices=[BASE])
> > +
> > +def atomic3(name):
> > + intrinsic(name + "_var", src_comp=[1, 1], dest_comp=1, num_vars=1)
> > + intrinsic(name, src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
> > +
> > +atomic("atomic_counter_inc")
> > +atomic("atomic_counter_dec")
> > +atomic("atomic_counter_read", flags=[CAN_ELIMINATE])
> > +atomic2("atomic_counter_add")
> > +atomic2("atomic_counter_min")
> > +atomic2("atomic_counter_max")
> > +atomic2("atomic_counter_and")
> > +atomic2("atomic_counter_or")
> > +atomic2("atomic_counter_xor")
> > +atomic2("atomic_counter_exchange")
> > +atomic3("atomic_counter_comp_swap")
> > +
> > +# Image load, store and atomic intrinsics.
> > +#
> > +# All image intrinsics take an image target passed as a nir_variable.
> > Image
> > +# variables contain a number of memory and layout qualifiers that influence
> > +# the semantics of the intrinsic.
> > +#
> > +# All image intrinsics take a four-coordinate vector and a sample index as
> > +# first two sources, determining the location within the image that will be
> > +# accessed by the intrinsic. Components not applicable to the image target
> > +# in use are undefined. Image store takes an additional four-component
> > +# argument with the value to be written, and image atomic operations take
> > +# either one or two additional scalar arguments with the same meaning as in
> > +# the ARB_shader_image_load_store specification.
> > +intrinsic("image_var_load", src_comp=[4, 1], dest_comp=4, num_vars=1,
> > + flags=[CAN_ELIMINATE])
> > +intrinsic("image_var_store", src_comp=[4, 1, 4], num_vars=1)
> > +intrinsic("image_var_atomic_add", src_comp=[4, 1, 1], dest_comp=1,
> > num_vars=1)
> > +intrinsic("image_var_atomic_min", src_comp=[4, 1, 1], dest_comp=1,
> > num_vars=1)
> > +intrinsic("image_var_atomic_max", src_comp=[4, 1, 1], dest_comp=1,
> > num_vars=1)
> > +intrinsic("image_var_atomic_and", src_comp=[4, 1, 1], dest_comp=1,
> > num_vars=1)
> > +intrinsic("image_var_atomic_or", src_comp=[4, 1, 1], dest_comp=1,
> > num_vars=1)
> > +intrinsic("image_var_atomic_xor", src_comp=[4, 1, 1], dest_comp=1,
> > num_vars=1)
> > +intrinsic("image_var_atomic_exchange", src_comp=[4, 1, 1], dest_comp=1,
> > num_vars=1)
> > +intrinsic("image_var_atomic_comp_swap", src_comp=[4, 1, 1, 1],
> > dest_comp=1, num_vars=1)
> > +intrinsic("image_var_size", dest_comp=0, num_vars=1,
> > flags=[CAN_ELIMINATE, CAN_REORDER])
> > +intrinsic("image_var_samples", dest_comp=1, num_vars=1,
> > flags=[CAN_ELIMINATE, CAN_REORDER])
> > +
> > +# Vulkan descriptor set intrinsics
> > +#
> > +# The Vulkan API uses a different binding model from GL. In the Vulkan
> > +# API, all external resources are represented by a tuple:
> > +#
> > +# (descriptor set, binding, array index)
> > +#
> > +# where the array index is the only thing allowed to be indirect. The
> > +# vulkan_surface_index intrinsic takes the descriptor set and binding as
> > +# its first two indices and the array index as its source. The third
> > +# index is a nir_variable_mode in case that's useful to the backend.
> > +#
> > +# The intended usage is that the shader will call vulkan_surface_index to
> > +# get an index and then pass that as the buffer index ubo/ssbo calls.
> > +#
> > +# The vulkan_resource_reindex intrinsic takes a resource index in src0
> > +# (the result of a vulkan_resource_index or vulkan_resource_reindex) which
> > +# corresponds to the tuple (set, binding, index) and computes an index
> > +# corresponding to tuple (set, binding, idx + src1).
> > +intrinsic("vulkan_resource_index", src_comp=[1], dest_comp=1,
> > + indices=[DESC_SET, BINDING], flags=[CAN_ELIMINATE, CAN_REORDER])
> > +intrinsic("vulkan_resource_reindex", src_comp=[1, 1], dest_comp=1,
> > + flags=[CAN_ELIMINATE, CAN_REORDER])
> > +
> > +# variable atomic intrinsics
> > +#
> > +# All of these variable atomic memory operations read a value from memory,
> > +# compute a new value using one of the operations below, write the new
> > value
> > +# to memory, and return the original value read.
> > +#
> > +# All operations take 1 source except CompSwap that takes 2. These sources
> > +# represent:
> > +#
> > +# 0: The data parameter to the atomic function (i.e. the value to add
> > +# in shared_atomic_add, etc).
> > +# 1: For CompSwap only: the second data parameter.
> > +#
> > +# All operations take 1 variable deref.
> > +intrinsic("var_atomic_add", src_comp=[1], dest_comp=1, num_vars=1)
> > +intrinsic("var_atomic_imin", src_comp=[1], dest_comp=1, num_vars=1)
> > +intrinsic("var_atomic_umin", src_comp=[1], dest_comp=1, num_vars=1)
> > +intrinsic("var_atomic_imax", src_comp=[1], dest_comp=1, num_vars=1)
> > +intrinsic("var_atomic_umax", src_comp=[1], dest_comp=1, num_vars=1)
> > +intrinsic("var_atomic_and", src_comp=[1], dest_comp=1, num_vars=1)
> > +intrinsic("var_atomic_or", src_comp=[1], dest_comp=1, num_vars=1)
> > +intrinsic("var_atomic_xor", src_comp=[1], dest_comp=1, num_vars=1)
> > +intrinsic("var_atomic_exchange", src_comp=[1], dest_comp=1, num_vars=1)
> > +intrinsic("var_atomic_comp_swap", src_comp=[1, 1], dest_comp=1, num_vars=1)
> > +
> > +# SSBO atomic intrinsics
> > +#
> > +# All of the SSBO atomic memory operations read a value from memory,
> > +# compute a new value using one of the operations below, write the new
> > +# value to memory, and return the original value read.
> > +#
> > +# All operations take 3 sources except CompSwap that takes 4. These
> > +# sources represent:
> > +#
> > +# 0: The SSBO buffer index.
> > +# 1: The offset into the SSBO buffer of the variable that the atomic
> > +# operation will operate on.
> > +# 2: The data parameter to the atomic function (i.e. the value to add
> > +# in ssbo_atomic_add, etc).
> > +# 3: For CompSwap only: the second data parameter.
> > +intrinsic("ssbo_atomic_add", src_comp=[1, 1, 1], dest_comp=1)
> > +intrinsic("ssbo_atomic_imin", src_comp=[1, 1, 1], dest_comp=1)
> > +intrinsic("ssbo_atomic_umin", src_comp=[1, 1, 1], dest_comp=1)
> > +intrinsic("ssbo_atomic_imax", src_comp=[1, 1, 1], dest_comp=1)
> > +intrinsic("ssbo_atomic_umax", src_comp=[1, 1, 1], dest_comp=1)
> > +intrinsic("ssbo_atomic_and", src_comp=[1, 1, 1], dest_comp=1)
> > +intrinsic("ssbo_atomic_or", src_comp=[1, 1, 1], dest_comp=1)
> > +intrinsic("ssbo_atomic_xor", src_comp=[1, 1, 1], dest_comp=1)
> > +intrinsic("ssbo_atomic_exchange", src_comp=[1, 1, 1], dest_comp=1)
> > +intrinsic("ssbo_atomic_comp_swap", src_comp=[1, 1, 1, 1], dest_comp=1)
> > +
> > +# CS shared variable atomic intrinsics
> > +#
> > +# All of the shared variable atomic memory operations read a value from
> > +# memory, compute a new value using one of the operations below, write the
> > +# new value to memory, and return the original value read.
> > +#
> > +# All operations take 2 sources except CompSwap that takes 3. These
> > +# sources represent:
> > +#
> > +# 0: The offset into the shared variable storage region that the atomic
> > +# operation will operate on.
> > +# 1: The data parameter to the atomic function (i.e. the value to add
> > +# in shared_atomic_add, etc).
> > +# 2: For CompSwap only: the second data parameter.
> > +intrinsic("shared_atomic_add", src_comp=[1, 1], dest_comp=1,
> > indices=[BASE])
> > +intrinsic("shared_atomic_imin", src_comp=[1, 1], dest_comp=1,
> > indices=[BASE])
> > +intrinsic("shared_atomic_umin", src_comp=[1, 1], dest_comp=1,
> > indices=[BASE])
> > +intrinsic("shared_atomic_imax", src_comp=[1, 1], dest_comp=1,
> > indices=[BASE])
> > +intrinsic("shared_atomic_umax", src_comp=[1, 1], dest_comp=1,
> > indices=[BASE])
> > +intrinsic("shared_atomic_and", src_comp=[1, 1], dest_comp=1,
> > indices=[BASE])
> > +intrinsic("shared_atomic_or", src_comp=[1, 1], dest_comp=1,
> > indices=[BASE])
> > +intrinsic("shared_atomic_xor", src_comp=[1, 1], dest_comp=1,
> > indices=[BASE])
> > +intrinsic("shared_atomic_exchange", src_comp=[1, 1], dest_comp=1,
> > indices=[BASE])
> > +intrinsic("shared_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1,
> > indices=[BASE])
> > +
> > +def system_value(name, dest_comp, indices=[]):
> > + intrinsic("load_" + name, [], dest_comp, 0, indices,
> > + flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True)
> > +
> > +system_value("frag_coord", 4)
> > +system_value("front_face", 1)
> > +system_value("vertex_id", 1)
> > +system_value("vertex_id_zero_base", 1)
> > +system_value("base_vertex", 1)
> > +system_value("instance_id", 1)
> > +system_value("base_instance", 1)
> > +system_value("draw_id", 1)
> > +system_value("sample_id", 1)
> > +system_value("sample_pos", 2)
> > +system_value("sample_mask_in", 1)
> > +system_value("primitive_id", 1)
> > +system_value("invocation_id", 1)
> > +system_value("tess_coord", 3)
> > +system_value("tess_level_outer", 4)
> > +system_value("tess_level_inner", 2)
> > +system_value("patch_vertices_in", 1)
> > +system_value("local_invocation_id", 3)
> > +system_value("local_invocation_index", 1)
> > +system_value("work_group_id", 3)
> > +system_value("user_clip_plane", 4, indices=[UCP_ID])
> > +system_value("num_work_groups", 3)
> > +system_value("helper_invocation", 1)
> > +system_value("alpha_ref_float", 1)
> > +system_value("layer_id", 1)
> > +system_value("view_index", 1)
> > +system_value("subgroup_size", 1)
> > +system_value("subgroup_invocation", 1)
> > +system_value("subgroup_eq_mask", 0)
> > +system_value("subgroup_ge_mask", 0)
> > +system_value("subgroup_gt_mask", 0)
> > +system_value("subgroup_le_mask", 0)
> > +system_value("subgroup_lt_mask", 0)
> > +system_value("num_subgroups", 1)
> > +system_value("subgroup_id", 1)
> > +system_value("local_group_size", 3)
> > +
> > +# Blend constant color values. Float values are clamped.#
> > +system_value("blend_const_color_r_float", 1)
> > +system_value("blend_const_color_g_float", 1)
> > +system_value("blend_const_color_b_float", 1)
> > +system_value("blend_const_color_a_float", 1)
> > +system_value("blend_const_color_rgba8888_unorm", 1)
> > +system_value("blend_const_color_aaaa8888_unorm", 1)
> > +
> > +# Barycentric coordinate intrinsics.
> > +#
> > +# These set up the barycentric coordinates for a particular interpolation.
> > +# The first three are for the simple cases: pixel, centroid, or per-sample
> > +# (at gl_SampleID). The next two handle interpolating at a specified
> > +# sample location, or interpolating with a vec2 offset,
> > +#
> > +# The interp_mode index should be either the INTERP_MODE_SMOOTH or
> > +# INTERP_MODE_NOPERSPECTIVE enum values.
> > +#
> > +# The vec2 value produced by these intrinsics is intended for use as the
> > +# barycoord source of a load_interpolated_input intrinsic.
> > +
> > +def barycentric(name, src_comp=[]):
> > + intrinsic("load_barycentric_" + name, src_comp=src_comp, dest_comp=2,
> > + indices=[INTERP_MODE], flags=[CAN_ELIMINATE, CAN_REORDER])
> > +
> > +# no sources. const_index[] = { interp_mode }
> > +barycentric("pixel")
> > +barycentric("centroid")
> > +barycentric("sample")
> > +# src[] = { sample_id }. const_index[] = { interp_mode }
> > +barycentric("at_sample", [1])
> > +# src[] = { offset.xy }. const_index[] = { interp_mode }
> > +barycentric("at_offset", [2])
> > +
> > +# Load operations pull data from some piece of GPU memory. All load
> > +# operations operate in terms of offsets into some piece of theoretical
> > +# memory. Loads from externally visible memory (UBO and SSBO) simply take
> > a
> > +# byte offset as a source. Loads from opaque memory (uniforms, inputs,
> > etc.)
> > +# take a base+offset pair where the base (const_index[0]) gives the
> > location
> > +# of the start of the variable being loaded and and the offset source is a
> > +# offset into that variable.
> > +#
> > +# Uniform load operations have a second "range" index that specifies the
> > +# range (starting at base) of the data from which we are loading. If
> > +# const_index[1] == 0, then the range is unknown.
> > +#
> > +# Some load operations such as UBO/SSBO load and per_vertex loads take an
> > +# additional source to specify which UBO/SSBO/vertex to load from.
> > +#
> > +# The exact address type depends on the lowering pass that generates the
> > +# load/store intrinsics. Typically, this is vec4 units for things such as
> > +# varying slots and float units for fragment shader inputs. UBO and SSBO
> > +# offsets are always in bytes.
> > +
> > +def load(name, num_srcs, indices=[], flags=[]):
> > + intrinsic("load_" + name, [1] * num_srcs, dest_comp=0, indices=indices,
> > + flags=flags)
> > +
> > +# src[] = { offset }. const_index[] = { base, range }
> > +load("uniform", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
> > +# src[] = { buffer_index, offset }. No const_index
> > +load("ubo", 2, flags=[CAN_ELIMINATE, CAN_REORDER])
> > +# src[] = { offset }. const_index[] = { base, component }
> > +load("input", 1, [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
> > +# src[] = { vertex, offset }. const_index[] = { base, component }
> > +load("per_vertex_input", 2, [BASE, COMPONENT], [CAN_ELIMINATE,
> > CAN_REORDER])
> > +# src[] = { barycoord, offset }. const_index[] = { base, component }
> > +intrinsic("load_interpolated_input", src_comp=[2, 1], dest_comp=0,
> > + indices=[BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
> > +
> > +# src[] = { buffer_index, offset }. No const_index
> > +load("ssbo", 2, flags=[CAN_ELIMINATE])
> > +# src[] = { offset }. const_index[] = { base, component }
> > +load("output", 1, [BASE, COMPONENT], flags=[CAN_ELIMINATE])
> > +# src[] = { vertex, offset }. const_index[] = { base }
> > +load("per_vertex_output", 2, [BASE, COMPONENT], [CAN_ELIMINATE])
> > +# src[] = { offset }. const_index[] = { base }
> > +load("shared", 1, [BASE], [CAN_ELIMINATE])
> > +# src[] = { offset }. const_index[] = { base, range }
> > +load("push_constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
> > +
> > +# Stores work the same way as loads, except now the first source is the
> > value
> > +# to store and the second (and possibly third) source specify where to
> > store
> > +# the value. SSBO and shared memory stores also have a write mask as
> > +# const_index[0].
> > +
> > +def store(name, num_srcs, indices=[], flags=[]):
> > + intrinsic("store_" + name, [0] + ([1] * (num_srcs - 1)),
> > indices=indices, flags=flags)
> > +
> > +# src[] = { value, offset }. const_index[] = { base, write_mask, component
> > }
> > +store("output", 2, [BASE, WRMASK, COMPONENT])
> > +# src[] = { value, vertex, offset }.
> > +# const_index[] = { base, write_mask, component }
> > +store("per_vertex_output", 3, [BASE, WRMASK, COMPONENT])
> > +# src[] = { value, block_index, offset }. const_index[] = { write_mask }
> > +store("ssbo", 3, [WRMASK])
> > +# src[] = { value, offset }. const_index[] = { base, write_mask }
> > +store("shared", 2, [BASE, WRMASK])
> > diff --git a/src/compiler/nir/nir_intrinsics_c.py
> > b/src/compiler/nir/nir_intrinsics_c.py
> > new file mode 100644
> > index 00000000000..556db785c2b
> > --- /dev/null
> > +++ b/src/compiler/nir/nir_intrinsics_c.py
> > @@ -0,0 +1,68 @@
> > +
> > +template = """\
> > +/* Copyright (C) 2018 Red Hat
> > + *
> > + * Permission is hereby granted, free of charge, to any person obtaining a
> > + * copy of this software and associated documentation files (the
> > "Software"),
> > + * to deal in the Software without restriction, including without
> > limitation
> > + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> > + * and/or sell copies of the Software, and to permit persons to whom the
> > + * Software is furnished to do so, subject to the following conditions:
> > + *
> > + * The above copyright notice and this permission notice (including the
> > next
> > + * paragraph) shall be included in all copies or substantial portions of
> > the
> > + * Software.
> > + *
> > + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
> > OR
> > + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> > + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> > + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
> > OTHER
> > + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> > + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
> > DEALINGS
> > + * IN THE SOFTWARE.
> > + */
> > +
> > +#include "nir.h"
> > +
> > +const nir_intrinsic_info nir_intrinsic_infos[nir_num_intrinsics] = {
> > +% for name, opcode in sorted(INTR_OPCODES.iteritems()):
> > +{
> > + .name = "${name}",
> > + .num_srcs = ${opcode.num_srcs},
> > + .src_components = {
> > + ${", ".join(str(comp) for comp in opcode.src_components)}
> > + },
> > + .has_dest = ${"true" if opcode.has_dest else "false"},
> > + .dest_components = ${opcode.dest_components},
> > + .num_variables = ${opcode.num_variables},
> > + .num_indices = ${opcode.num_indices},
> > + .index_map = {
> > +% for i in range(len(opcode.indices)):
> > + [${opcode.indices[i]}] = ${i + 1},
> > +% endfor
> > + },
> > + .flags = ${"0" if len(opcode.flags) == 0 else " | ".join(opcode.flags)},
> > +},
> > +% endfor
> > +};
> > +"""
> > +
> > +from nir_intrinsics import INTR_OPCODES
> > +from mako.template import Template
> > +import argparse
> > +import os
> > +
> > +def main():
> > + parser = argparse.ArgumentParser()
> > + parser.add_argument('--outdir', required=True,
> > + help='Directory to put the generated files in')
> > +
> > + args = parser.parse_args()
> > +
> > + path = os.path.join(args.outdir, 'nir_intrinsics.c')
> > + with open(path, 'wb') as f:
> > + f.write(Template(template).render(INTR_OPCODES=INTR_OPCODES))
> > +
> > +if __name__ == '__main__':
> > + main()
> > +
> > diff --git a/src/compiler/nir/nir_intrinsics.c
> > b/src/compiler/nir/nir_intrinsics_h.py
> > similarity index 55%
> > rename from src/compiler/nir/nir_intrinsics.c
> > rename to src/compiler/nir/nir_intrinsics_h.py
> > index 0257b19b348..8a4f0d501e6 100644
> > --- a/src/compiler/nir/nir_intrinsics.c
> > +++ b/src/compiler/nir/nir_intrinsics_h.py
> > @@ -1,5 +1,6 @@
> > -/*
> > - * Copyright © 2014 Intel Corporation
> > +
> > +template = """\
> > +/* Copyright (C) 2018 Red Hat
> > *
> > * Permission is hereby granted, free of charge, to any person obtaining a
> > * copy of this software and associated documentation files (the
> > "Software"),
> > @@ -19,39 +20,41 @@
> > * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> > * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
> > DEALINGS
> > * IN THE SOFTWARE.
> > - *
> > - * Authors:
> > - * Connor Abbott (cwabbo...@gmail.com)
> > - *
> > */
> >
> > -#include "nir.h"
> > -
> > -#define OPCODE(name) nir_intrinsic_##name
> > -
> > -#define INTRINSIC(_name, _num_srcs, _src_components, _has_dest, \
> > - _dest_components, _num_variables, _num_indices, \
> > - idx0, idx1, idx2, _flags) \
> > -{ \
> > - .name = #_name, \
> > - .num_srcs = _num_srcs, \
> > - .src_components = _src_components, \
> > - .has_dest = _has_dest, \
> > - .dest_components = _dest_components, \
> > - .num_variables = _num_variables, \
> > - .num_indices = _num_indices, \
> > - .index_map = { \
> > - [NIR_INTRINSIC_ ## idx0] = 1, \
> > - [NIR_INTRINSIC_ ## idx1] = 2, \
> > - [NIR_INTRINSIC_ ## idx2] = 3, \
> > - }, \
> > - .flags = _flags \
> > -},
> > -
> > -#define NIR_INTRINSIC_xx 0
> > -
> > -#define LAST_INTRINSIC(name)
> > -
> > -const nir_intrinsic_info nir_intrinsic_infos[nir_num_intrinsics] = {
> > -#include "nir_intrinsics.h"
> > -};
> > \ No newline at end of file
> > +#ifndef _NIR_INTRINSICS_
> > +#define _NIR_INTRINSICS_
> > +
> > +<% opcode_names = sorted(INTR_OPCODES) %>
> > +
> > +typedef enum {
> > +% for name in opcode_names:
> > + nir_intrinsic_${name},
> > +% endfor
> > +
> > + nir_last_intrinsic = nir_intrinsic_${opcode_names[-1]},
> > + nir_num_intrinsics = nir_last_intrinsic + 1
> > +} nir_intrinsic_op;
> > +
> > +#endif /* _NIR_INTRINSICS_ */"""
> > +
> > +from nir_intrinsics import INTR_OPCODES
> > +from mako.template import Template
> > +import argparse
> > +import os
> > +
> > +
> > +def main():
> > + parser = argparse.ArgumentParser()
> > + parser.add_argument('--outdir', required=True,
> > + help='Directory to put the generated files in')
> > +
> > + args = parser.parse_args()
> > +
> > + path = os.path.join(args.outdir, 'nir_intrinsics.h')
> > + with open(path, 'wb') as f:
> > + f.write(Template(template).render(INTR_OPCODES=INTR_OPCODES))
> > +
> > +if __name__ == '__main__':
> > + main()
> > +
> >
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
signature.asc
Description: signature
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
