Mesa (master): nir: add pass to lower disjoint wrmask's
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Thu May 14 03:52:05 UTC 2020
Module: Mesa
Branch: master
Commit: 42d38ad02816d72e8c3ef1d15dabb28ee9eb2f88
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=42d38ad02816d72e8c3ef1d15dabb28ee9eb2f88
Author: Rob Clark <robdclark at chromium.org>
Date: Wed May 6 14:20:09 2020 -0700
nir: add pass to lower disjoint wrmask's
Signed-off-by: Rob Clark <robdclark at chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg at google.com>
Reviewed-by: Eric Anholt <eric at anholt.net>
---
src/compiler/Makefile.sources | 1 +
src/compiler/nir/meson.build | 1 +
src/compiler/nir/nir.h | 2 +
src/compiler/nir/nir_lower_wrmasks.c | 232 +++++++++++++++++++++++++++++++++++
4 files changed, 236 insertions(+)
diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index 6a400677529..15aa8b596d1 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -296,6 +296,7 @@ NIR_FILES = \
nir/nir_lower_viewport_transform.c \
nir/nir_lower_wpos_center.c \
nir/nir_lower_wpos_ytransform.c \
+ nir/nir_lower_wrmasks.c \
nir/nir_metadata.c \
nir/nir_move_vec_src_uses_to_dest.c \
nir/nir_normalize_cubemap_coords.c \
diff --git a/src/compiler/nir/meson.build b/src/compiler/nir/meson.build
index fbe403d42a7..15d73b5aede 100644
--- a/src/compiler/nir/meson.build
+++ b/src/compiler/nir/meson.build
@@ -174,6 +174,7 @@ files_libnir = files(
'nir_lower_viewport_transform.c',
'nir_lower_wpos_center.c',
'nir_lower_wpos_ytransform.c',
+ 'nir_lower_wrmasks.c',
'nir_lower_bit_size.c',
'nir_lower_uniforms_to_ubo.c',
'nir_metadata.c',
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index b12deb2cd78..e393fb49100 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -4338,6 +4338,8 @@ bool nir_lower_wpos_ytransform(nir_shader *shader,
const nir_lower_wpos_ytransform_options *options);
bool nir_lower_wpos_center(nir_shader *shader, const bool for_sample_shading);
+bool nir_lower_wrmasks(nir_shader *shader, nir_instr_filter_cb cb, const void *data);
+
bool nir_lower_fb_read(nir_shader *shader);
typedef struct nir_lower_drawpixels_options {
diff --git a/src/compiler/nir/nir_lower_wrmasks.c b/src/compiler/nir/nir_lower_wrmasks.c
new file mode 100644
index 00000000000..d8b4c2423ac
--- /dev/null
+++ b/src/compiler/nir/nir_lower_wrmasks.c
@@ -0,0 +1,232 @@
+/*
+ * Copyright © 2020 Google, Inc.
+ *
+ * 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"
+#include "nir_builder.h"
+
+/* A pass to split intrinsics with discontinuous writemasks into ones
+ * with contiguous writemasks starting with .x, ie:
+ *
+ * vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35
+ * intrinsic store_ssbo (ssa_76, ssa_105, ssa_106) (2, 0, 4, 0) // wrmask=y
+ *
+ * is turned into:
+ *
+ * vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35
+ * vec1 32 ssa_107 = load_const (0x00000001)
+ * vec1 32 ssa_108 = iadd ssa_106, ssa_107
+ * vec1 32 ssa_109 = mov ssa_76.y
+ * intrinsic store_ssbo (ssa_109, ssa_105, ssa_108) (1, 0, 4, 0) // wrmask=x
+ *
+ * and likewise:
+ *
+ * vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35
+ * intrinsic store_ssbo (ssa_76, ssa_105, ssa_106) (15, 0, 4, 0) // wrmask=xzw
+ *
+ * is split into:
+ *
+ * // .x component:
+ * vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35
+ * vec1 32 ssa_107 = load_const (0x00000000)
+ * vec1 32 ssa_108 = iadd ssa_106, ssa_107
+ * vec1 32 ssa_109 = mov ssa_76.x
+ * intrinsic store_ssbo (ssa_109, ssa_105, ssa_108) (1, 0, 4, 0) // wrmask=x
+ * // .zw components:
+ * vec1 32 ssa_110 = load_const (0x00000002)
+ * vec1 32 ssa_111 = iadd ssa_106, ssa_110
+ * vec2 32 ssa_112 = mov ssa_76.zw
+ * intrinsic store_ssbo (ssa_112, ssa_105, ssa_111) (3, 0, 4, 0) // wrmask=xy
+ */
+
+static int
+value_src(nir_intrinsic_op intrinsic)
+{
+ switch (intrinsic) {
+ case nir_intrinsic_store_output:
+ case nir_intrinsic_store_per_vertex_output:
+ case nir_intrinsic_store_ssbo:
+ case nir_intrinsic_store_shared:
+ case nir_intrinsic_store_global:
+ case nir_intrinsic_store_scratch:
+ return 0;
+ default:
+ return -1;
+ }
+}
+
+static int
+offset_src(nir_intrinsic_op intrinsic)
+{
+ switch (intrinsic) {
+ case nir_intrinsic_store_output:
+ case nir_intrinsic_store_shared:
+ case nir_intrinsic_store_global:
+ case nir_intrinsic_store_scratch:
+ return 1;
+ case nir_intrinsic_store_per_vertex_output:
+ case nir_intrinsic_store_ssbo:
+ return 2;
+ default:
+ return -1;
+ }
+}
+
+static void
+split_wrmask(nir_builder *b, nir_intrinsic_instr *intr)
+{
+ const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
+
+ b->cursor = nir_before_instr(&intr->instr);
+
+ assert(!info->has_dest); /* expecting only store intrinsics */
+
+ unsigned num_srcs = info->num_srcs;
+ unsigned value_idx = value_src(intr->intrinsic);
+ unsigned offset_idx = offset_src(intr->intrinsic);
+ unsigned num_comp = nir_intrinsic_src_components(intr, value_idx);
+
+ unsigned wrmask = nir_intrinsic_write_mask(intr);
+ while (wrmask) {
+ unsigned first_component = ffs(wrmask) - 1;
+ unsigned length = ffs(~(wrmask >> first_component)) - 1;
+
+ nir_ssa_def *value = nir_ssa_for_src(b, intr->src[value_idx], num_comp);
+ nir_ssa_def *offset = nir_ssa_for_src(b, intr->src[offset_idx], 1);
+
+ /* swizzle out the consecutive components that we'll store
+ * in this iteration:
+ */
+ unsigned cur_mask = (BITFIELD_MASK(length) << first_component);
+ value = nir_channels(b, value, cur_mask);
+
+ /* and create the replacement intrinsic: */
+ nir_intrinsic_instr *new_intr =
+ nir_intrinsic_instr_create(b->shader, intr->intrinsic);
+
+ nir_intrinsic_copy_const_indices(new_intr, intr);
+ nir_intrinsic_set_write_mask(new_intr, BITFIELD_MASK(length));
+
+ const int offset_units = value->bit_size / 8;
+
+ if (info->index_map[NIR_INTRINSIC_ALIGN_MUL]) {
+ assert(info->index_map[NIR_INTRINSIC_ALIGN_OFFSET]);
+ unsigned align_mul = nir_intrinsic_align_mul(intr);
+ unsigned align_off = nir_intrinsic_align_offset(intr);
+
+ align_off += offset_units * first_component;
+ align_off = align_off % align_mul;
+
+ nir_intrinsic_set_align(new_intr, align_mul, align_off);
+ }
+
+ /* if the instruction has a BASE, fold the offset adjustment
+ * into that instead of adding alu instructions, otherwise add
+ * instructions
+ */
+ unsigned offset_adj = offset_units * first_component;
+ if (info->index_map[NIR_INTRINSIC_BASE]) {
+ nir_intrinsic_set_base(new_intr,
+ nir_intrinsic_base(intr) + offset_adj);
+ } else {
+ offset = nir_iadd(b, offset,
+ nir_imm_intN_t(b, offset_adj, offset->bit_size));
+ }
+
+ new_intr->num_components = length;
+
+ /* Copy the sources, replacing value/offset, and passing everything
+ * else through to the new instrution:
+ */
+ for (unsigned i = 0; i < num_srcs; i++) {
+ if (i == value_idx) {
+ new_intr->src[i] = nir_src_for_ssa(value);
+ } else if (i == offset_idx) {
+ new_intr->src[i] = nir_src_for_ssa(offset);
+ } else {
+ new_intr->src[i] = intr->src[i];
+ }
+ }
+
+ nir_builder_instr_insert(b, &new_intr->instr);
+
+ /* Clear the bits in the writemask that we just wrote, then try
+ * again to see if more channels are left.
+ */
+ wrmask &= ~cur_mask;
+ }
+
+ /* Finally remove the original intrinsic. */
+ nir_instr_remove(&intr->instr);
+}
+
+bool
+nir_lower_wrmasks(nir_shader *shader, nir_instr_filter_cb cb, const void *data)
+{
+ bool progress = false;
+
+ nir_foreach_function(function, shader) {
+ nir_function_impl *impl = function->impl;
+
+ if (!impl)
+ continue;
+
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+ const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
+
+ /* if no wrmask, then skip it: */
+ if (!info->index_map[NIR_INTRINSIC_WRMASK])
+ continue;
+
+ /* if wrmask is already contiguous, then nothing to do: */
+ if (nir_intrinsic_write_mask(intr) == BITFIELD_MASK(intr->num_components))
+ continue;
+
+ /* do we know how to lower this instruction? */
+ if (value_src(intr->intrinsic) < 0)
+ continue;
+
+ assert(offset_src(intr->intrinsic) >= 0);
+
+ /* does backend need us to lower this intrinsic? */
+ if (cb && !cb(instr, data))
+ continue;
+
+ nir_builder b;
+ nir_builder_init(&b, impl);
+ split_wrmask(&b, intr);
+ progress = true;
+ }
+ }
+
+ nir_metadata_preserve(impl, nir_metadata_block_index |
+ nir_metadata_dominance);
+
+ }
+
+ return progress;
+}
More information about the mesa-commit
mailing list