Mesa (main): d3d12: Handle cubemap gather on int cubemaps
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Thu Dec 30 18:51:46 UTC 2021
Module: Mesa
Branch: main
Commit: bd7fc35d6a03265803263e0ee371d0080157af85
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=bd7fc35d6a03265803263e0ee371d0080157af85
Author: Jesse Natalie <jenatali at microsoft.com>
Date: Thu Dec 9 13:38:58 2021 -0800
d3d12: Handle cubemap gather on int cubemaps
There were 2 options: This, or double-bind the cubemap so we can reference
it as both a cube and as a 2D array, and only run gathers on the cubemap.
As ugly as this is, I think I like it better.
Reviewed-by: Sil Vilerino <sivileri at microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14161>
---
.../d3d12/d3d12_lower_int_cubemap_to_array.c | 169 ++++++++++++++++++++-
1 file changed, 166 insertions(+), 3 deletions(-)
diff --git a/src/gallium/drivers/d3d12/d3d12_lower_int_cubemap_to_array.c b/src/gallium/drivers/d3d12/d3d12_lower_int_cubemap_to_array.c
index 96319a2e584..630da4f3c36 100644
--- a/src/gallium/drivers/d3d12/d3d12_lower_int_cubemap_to_array.c
+++ b/src/gallium/drivers/d3d12/d3d12_lower_int_cubemap_to_array.c
@@ -45,6 +45,7 @@ lower_int_cubmap_to_array_filter(const nir_instr *instr,
case nir_texop_txl:
case nir_texop_txs:
case nir_texop_lod:
+ case nir_texop_tg4:
break;
default:
return false;
@@ -121,12 +122,12 @@ evaluate_face_z(nir_builder *b, coord_t *coord)
}
static nir_ssa_def *
-create_array_tex_from_cube_tex(nir_builder *b, nir_tex_instr *tex, nir_ssa_def *coord)
+create_array_tex_from_cube_tex(nir_builder *b, nir_tex_instr *tex, nir_ssa_def *coord, nir_texop op)
{
nir_tex_instr *array_tex;
array_tex = nir_tex_instr_create(b->shader, tex->num_srcs);
- array_tex->op = tex->op;
+ array_tex->op = op;
array_tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
array_tex->is_array = true;
array_tex->is_shadow = tex->is_shadow;
@@ -151,6 +152,164 @@ create_array_tex_from_cube_tex(nir_builder *b, nir_tex_instr *tex, nir_ssa_def *
return &array_tex->dest.ssa;
}
+static nir_ssa_def *
+handle_cube_edge(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y, nir_ssa_def *face, nir_ssa_def *array_slice_cube_base, nir_ssa_def *tex_size)
+{
+ enum cube_remap
+ {
+ cube_remap_zero = 0,
+ cube_remap_x,
+ cube_remap_y,
+ cube_remap_tex_size,
+ cube_remap_tex_size_minus_x,
+ cube_remap_tex_size_minus_y,
+
+ cube_remap_size,
+ };
+
+ struct cube_remap_table
+ {
+ enum cube_remap remap_x;
+ enum cube_remap remap_y;
+ uint32_t remap_face;
+ };
+
+ static const struct cube_remap_table cube_remap_neg_x[6] =
+ {
+ {cube_remap_tex_size, cube_remap_y, 4},
+ {cube_remap_tex_size, cube_remap_y, 5},
+ {cube_remap_y, cube_remap_zero, 1},
+ {cube_remap_tex_size_minus_y, cube_remap_tex_size, 1},
+ {cube_remap_tex_size, cube_remap_y, 1},
+ {cube_remap_tex_size, cube_remap_y, 0},
+ };
+
+ static const struct cube_remap_table cube_remap_pos_x[6] =
+ {
+ {cube_remap_zero, cube_remap_y, 5},
+ {cube_remap_zero, cube_remap_y, 4},
+ {cube_remap_tex_size_minus_y, cube_remap_zero, 0},
+ {cube_remap_y, cube_remap_tex_size, 0},
+ {cube_remap_zero, cube_remap_y, 0},
+ {cube_remap_zero, cube_remap_y, 1},
+ };
+
+ static const struct cube_remap_table cube_remap_neg_y[6] =
+ {
+ {cube_remap_tex_size, cube_remap_tex_size_minus_x, 2},
+ {cube_remap_zero, cube_remap_x, 2},
+ {cube_remap_tex_size_minus_x, cube_remap_zero, 5},
+ {cube_remap_x, cube_remap_tex_size, 4},
+ {cube_remap_x, cube_remap_tex_size, 2},
+ {cube_remap_tex_size_minus_x, cube_remap_zero, 2},
+ };
+
+ static const struct cube_remap_table cube_remap_pos_y[6] =
+ {
+ {cube_remap_tex_size, cube_remap_x, 3},
+ {cube_remap_zero, cube_remap_tex_size_minus_x, 3},
+ {cube_remap_x, cube_remap_zero, 4},
+ {cube_remap_tex_size_minus_x, cube_remap_tex_size, 5},
+ {cube_remap_x, cube_remap_zero, 3},
+ {cube_remap_tex_size_minus_x, cube_remap_tex_size, 3},
+ };
+
+ static const struct cube_remap_table* remap_tables[4] = {
+ cube_remap_neg_x,
+ cube_remap_pos_x,
+ cube_remap_neg_y,
+ cube_remap_pos_y
+ };
+
+ nir_ssa_def *zero = nir_imm_int(b, 0);
+
+ /* Doesn't matter since the texture is square */
+ tex_size = nir_channel(b, tex_size, 0);
+
+ nir_ssa_def *x_on = nir_iand(b, nir_ige(b, x, zero), nir_ige(b, tex_size, x));
+ nir_ssa_def *y_on = nir_iand(b, nir_ige(b, y, zero), nir_ige(b, tex_size, y));
+ nir_ssa_def *one_on = nir_ixor(b, x_on, y_on);
+
+ /* If the sample did not fall off the face in either dimension, then set output = input */
+ nir_ssa_def *x_result = x;
+ nir_ssa_def *y_result = y;
+ nir_ssa_def *face_result = face;
+
+ /* otherwise, if the sample fell off the face in either the X or the Y direction, remap to the new face */
+ nir_ssa_def *remap_predicates[4] =
+ {
+ nir_iand(b, one_on, nir_ilt(b, x, zero)),
+ nir_iand(b, one_on, nir_ilt(b, tex_size, x)),
+ nir_iand(b, one_on, nir_ilt(b, y, zero)),
+ nir_iand(b, one_on, nir_ilt(b, tex_size, y)),
+ };
+
+ nir_ssa_def *remap_array[cube_remap_size];
+
+ remap_array[cube_remap_zero] = zero;
+ remap_array[cube_remap_x] = x;
+ remap_array[cube_remap_y] = y;
+ remap_array[cube_remap_tex_size] = tex_size;
+ remap_array[cube_remap_tex_size_minus_x] = nir_isub(b, tex_size, x);
+ remap_array[cube_remap_tex_size_minus_y] = nir_isub(b, tex_size, y);
+
+ /* For each possible way the sample could have fallen off */
+ for (unsigned i = 0; i < 4; i++) {
+ const struct cube_remap_table* remap_table = remap_tables[i];
+
+ /* For each possible original face */
+ for (unsigned j = 0; j < 6; j++) {
+ nir_ssa_def *predicate = nir_iand(b, remap_predicates[i], nir_ieq(b, face, nir_imm_int(b, j)));
+
+ x_result = nir_bcsel(b, predicate, remap_array[remap_table[j].remap_x], x_result);
+ y_result = nir_bcsel(b, predicate, remap_array[remap_table[j].remap_y], y_result);
+ face_result = nir_bcsel(b, predicate, remap_array[remap_table[j].remap_face], face_result);
+ }
+ }
+
+ return nir_vec3(b, x_result, y_result, nir_iadd(b, face_result, array_slice_cube_base));
+}
+
+static nir_ssa_def *
+handle_cube_gather(nir_builder *b, nir_tex_instr *tex, nir_ssa_def *coord)
+{
+ nir_ssa_def *tex_size = nir_get_texture_size(b, tex);
+
+ /* nir_get_texture_size puts the cursor before the tex op */
+ b->cursor = nir_after_instr(coord->parent_instr);
+
+ nir_ssa_def *const_05 = nir_imm_float(b, 0.5f);
+ nir_ssa_def *texel_coords = nir_fmul(b, nir_channels(b, coord, 3),
+ nir_i2f32(b, nir_channels(b, tex_size, 3)));
+
+ nir_ssa_def *x_orig = nir_channel(b, texel_coords, 0);
+ nir_ssa_def *y_orig = nir_channel(b, texel_coords, 1);
+
+ nir_ssa_def *x_pos = nir_f2i32(b, nir_fadd(b, x_orig, const_05));
+ nir_ssa_def *x_neg = nir_f2i32(b, nir_fsub(b, x_orig, const_05));
+ nir_ssa_def *y_pos = nir_f2i32(b, nir_fadd(b, y_orig, const_05));
+ nir_ssa_def *y_neg = nir_f2i32(b, nir_fsub(b, y_orig, const_05));
+ nir_ssa_def *coords[4][2] = {
+ { x_neg, y_pos },
+ { x_pos, y_pos },
+ { x_pos, y_neg },
+ { x_neg, y_neg },
+ };
+
+ nir_ssa_def *array_slice_2d = nir_f2i32(b, nir_channel(b, coord, 2));
+ nir_ssa_def *face = nir_imod(b, array_slice_2d, nir_imm_int(b, 6));
+ nir_ssa_def *array_slice_cube_base = nir_isub(b, array_slice_2d, face);
+
+ nir_ssa_def *channels[4];
+ for (unsigned i = 0; i < 4; ++i) {
+ nir_ssa_def *final_coord = handle_cube_edge(b, coords[i][0], coords[i][1], face, array_slice_cube_base, tex_size);
+ nir_ssa_def *sampled_val = create_array_tex_from_cube_tex(b, tex, final_coord, nir_texop_txf);
+ channels[i] = nir_channel(b, sampled_val, tex->component);
+ }
+
+ return nir_vec(b, channels, 4);
+}
+
static nir_ssa_def *
lower_cube_sample(nir_builder *b, nir_tex_instr *tex)
{
@@ -196,7 +355,10 @@ lower_cube_sample(nir_builder *b, nir_tex_instr *tex)
// This contains in xy the normalized sample coordinates, and in z the face index
nir_ssa_def *coord_and_face = nir_if_phi(b, face_x_coord, face_y_or_z_coord);
- return create_array_tex_from_cube_tex(b, tex, coord_and_face);
+ if (tex->op == nir_texop_tg4)
+ return handle_cube_gather(b, tex, coord_and_face);
+ else
+ return create_array_tex_from_cube_tex(b, tex, coord_and_face, tex->op);
}
static nir_ssa_def *
@@ -255,6 +417,7 @@ lower_int_cubmap_to_array_impl(nir_builder *b, nir_instr *instr,
case nir_texop_txd:
case nir_texop_txl:
case nir_texop_lod:
+ case nir_texop_tg4:
return lower_cube_sample(b, tex);
case nir_texop_txs:
return lower_cube_txs(b, tex);
More information about the mesa-commit
mailing list