*/
static void
lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
- unsigned sampler_unit,
LLVMValueRef width,
LLVMValueRef height,
LLVMValueRef depth,
*/
static void
lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
- unsigned sampler_unit,
LLVMValueRef size,
LLVMValueRef row_stride_vec,
LLVMValueRef img_stride_vec,
/*
* Get texture colors.
*/
- lp_build_sample_texel_soa(bld, sampler_unit,
+ lp_build_sample_texel_soa(bld,
width_vec, height_vec, depth_vec,
x, y, z,
row_stride_vec, img_stride_vec,
return lp_build_lerp(bld, weight1, val0, val1, 0);
}
+/*
+ * this is a bit excessive code for something OpenGL just recommends
+ * but does not require.
+ */
+#define ACCURATE_CUBE_CORNERS 1
+
/**
* Generate code to sample a mipmap level with linear filtering.
* If sampling a cube texture, r = cube face in [0,5].
+ * If linear_mask is present, only pixels having their mask set
+ * will receive linear filtering, the rest will use nearest.
*/
static void
lp_build_sample_image_linear(struct lp_build_sample_context *bld,
- unsigned sampler_unit,
LLVMValueRef size,
+ LLVMValueRef linear_mask,
LLVMValueRef row_stride_vec,
LLVMValueRef img_stride_vec,
LLVMValueRef data_ptr,
const LLVMValueRef *offsets,
LLVMValueRef colors_out[4])
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ struct lp_build_context *ivec_bld = &bld->int_coord_bld;
+ struct lp_build_context *coord_bld = &bld->coord_bld;
const unsigned dims = bld->dims;
LLVMValueRef width_vec;
LLVMValueRef height_vec;
LLVMValueRef flt_width_vec;
LLVMValueRef flt_height_vec;
LLVMValueRef flt_depth_vec;
- LLVMValueRef x0, y0 = NULL, z0 = NULL, x1, y1 = NULL, z1 = NULL;
+ LLVMValueRef fall_off[4], have_corners;
+ LLVMValueRef z1 = NULL;
+ LLVMValueRef z00 = NULL, z01 = NULL, z10 = NULL, z11 = NULL;
+ LLVMValueRef x00 = NULL, x01 = NULL, x10 = NULL, x11 = NULL;
+ LLVMValueRef y00 = NULL, y01 = NULL, y10 = NULL, y11 = NULL;
LLVMValueRef s_fpart, t_fpart = NULL, r_fpart = NULL;
+ LLVMValueRef xs[4], ys[4], zs[4];
LLVMValueRef neighbors[2][2][4];
- int chan;
+ int chan, texel_index;
+ boolean seamless_cube_filter, accurate_cube_corners;
+
+ seamless_cube_filter = bld->static_texture_state->target == PIPE_TEXTURE_CUBE &&
+ bld->static_sampler_state->seamless_cube_map;
+ accurate_cube_corners = ACCURATE_CUBE_CORNERS && seamless_cube_filter;
lp_build_extract_image_sizes(bld,
&bld->int_size_bld,
/*
* Compute integer texcoords.
*/
- lp_build_sample_wrap_linear(bld, coords[0], width_vec,
- flt_width_vec, offsets[0],
- bld->static_texture_state->pot_width,
- bld->static_sampler_state->wrap_s,
- &x0, &x1, &s_fpart);
- lp_build_name(x0, "tex.x0.wrapped");
- lp_build_name(x1, "tex.x1.wrapped");
- if (dims >= 2) {
- lp_build_sample_wrap_linear(bld, coords[1], height_vec,
- flt_height_vec, offsets[1],
- bld->static_texture_state->pot_height,
- bld->static_sampler_state->wrap_t,
- &y0, &y1, &t_fpart);
- lp_build_name(y0, "tex.y0.wrapped");
- lp_build_name(y1, "tex.y1.wrapped");
+ if (!seamless_cube_filter) {
+ lp_build_sample_wrap_linear(bld, coords[0], width_vec,
+ flt_width_vec, offsets[0],
+ bld->static_texture_state->pot_width,
+ bld->static_sampler_state->wrap_s,
+ &x00, &x01, &s_fpart);
+ lp_build_name(x00, "tex.x0.wrapped");
+ lp_build_name(x01, "tex.x1.wrapped");
+ x10 = x00;
+ x11 = x01;
- if (dims == 3) {
- lp_build_sample_wrap_linear(bld, coords[2], depth_vec,
- flt_depth_vec, offsets[2],
- bld->static_texture_state->pot_depth,
- bld->static_sampler_state->wrap_r,
- &z0, &z1, &r_fpart);
- lp_build_name(z0, "tex.z0.wrapped");
- lp_build_name(z1, "tex.z1.wrapped");
+ if (dims >= 2) {
+ lp_build_sample_wrap_linear(bld, coords[1], height_vec,
+ flt_height_vec, offsets[1],
+ bld->static_texture_state->pot_height,
+ bld->static_sampler_state->wrap_t,
+ &y00, &y10, &t_fpart);
+ lp_build_name(y00, "tex.y0.wrapped");
+ lp_build_name(y10, "tex.y1.wrapped");
+ y01 = y00;
+ y11 = y10;
+
+ if (dims == 3) {
+ lp_build_sample_wrap_linear(bld, coords[2], depth_vec,
+ flt_depth_vec, offsets[2],
+ bld->static_texture_state->pot_depth,
+ bld->static_sampler_state->wrap_r,
+ &z00, &z1, &r_fpart);
+ z01 = z10 = z11 = z00;
+ lp_build_name(z00, "tex.z0.wrapped");
+ lp_build_name(z1, "tex.z1.wrapped");
+ }
+ }
+ if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
+ bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
+ bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
+ z00 = z01 = z10 = z11 = z1 = coords[2]; /* cube face or layer */
+ lp_build_name(z00, "tex.z0.layer");
+ lp_build_name(z1, "tex.z1.layer");
}
}
- if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
- bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
- bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
- z0 = z1 = coords[2]; /* cube face or layer */
- lp_build_name(z0, "tex.z0.layer");
- lp_build_name(z1, "tex.z1.layer");
- }
+ else {
+ struct lp_build_if_state edge_if;
+ LLVMTypeRef int1t;
+ LLVMValueRef new_faces[4], new_xcoords[4][2], new_ycoords[4][2];
+ LLVMValueRef coord, have_edge, have_corner;
+ LLVMValueRef fall_off_ym_notxm, fall_off_ym_notxp, fall_off_x, fall_off_y;
+ LLVMValueRef fall_off_yp_notxm, fall_off_yp_notxp;
+ LLVMValueRef x0, x1, y0, y1, y0_clamped, y1_clamped;
+ LLVMValueRef face = coords[2];
+ LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5f);
+ LLVMValueRef length_minus_one = lp_build_sub(ivec_bld, width_vec, ivec_bld->one);
+ /* XXX drop height calcs. Could (should) do this without seamless filtering too */
+ height_vec = width_vec;
+ flt_height_vec = flt_width_vec;
+
+ /* XXX the overflow logic is actually sort of duplicated with trilinear,
+ * since an overflow in one mip should also have a corresponding overflow
+ * in another.
+ */
+ /* should always have normalized coords, and offsets are undefined */
+ assert(bld->static_sampler_state->normalized_coords);
+ coord = lp_build_mul(coord_bld, coords[0], flt_width_vec);
+ /* instead of clamp, build mask if overflowed */
+ coord = lp_build_sub(coord_bld, coord, half);
+ /* convert to int, compute lerp weight */
+ /* not ideal with AVX (and no AVX2) */
+ lp_build_ifloor_fract(coord_bld, coord, &x0, &s_fpart);
+ x1 = lp_build_add(ivec_bld, x0, ivec_bld->one);
+ coord = lp_build_mul(coord_bld, coords[1], flt_height_vec);
+ coord = lp_build_sub(coord_bld, coord, half);
+ lp_build_ifloor_fract(coord_bld, coord, &y0, &t_fpart);
+ y1 = lp_build_add(ivec_bld, y0, ivec_bld->one);
+
+ fall_off[0] = lp_build_cmp(ivec_bld, PIPE_FUNC_LESS, x0, ivec_bld->zero);
+ fall_off[1] = lp_build_cmp(ivec_bld, PIPE_FUNC_GREATER, x1, length_minus_one);
+ fall_off[2] = lp_build_cmp(ivec_bld, PIPE_FUNC_LESS, y0, ivec_bld->zero);
+ fall_off[3] = lp_build_cmp(ivec_bld, PIPE_FUNC_GREATER, y1, length_minus_one);
+
+ fall_off_x = lp_build_or(ivec_bld, fall_off[0], fall_off[1]);
+ fall_off_y = lp_build_or(ivec_bld, fall_off[2], fall_off[3]);
+ have_edge = lp_build_or(ivec_bld, fall_off_x, fall_off_y);
+ have_edge = lp_build_any_true_range(ivec_bld, ivec_bld->type.length, have_edge);
+
+ /* needed for accurate corner filtering branch later, rely on 0 init */
+ int1t = LLVMInt1TypeInContext(bld->gallivm->context);
+ have_corners = lp_build_alloca(bld->gallivm, int1t, "have_corner");
+
+ for (texel_index = 0; texel_index < 4; texel_index++) {
+ xs[texel_index] = lp_build_alloca(bld->gallivm, ivec_bld->vec_type, "xs");
+ ys[texel_index] = lp_build_alloca(bld->gallivm, ivec_bld->vec_type, "ys");
+ zs[texel_index] = lp_build_alloca(bld->gallivm, ivec_bld->vec_type, "zs");
+ }
+
+ lp_build_if(&edge_if, bld->gallivm, have_edge);
+
+ have_corner = lp_build_and(ivec_bld, fall_off_x, fall_off_y);
+ have_corner = lp_build_any_true_range(ivec_bld, ivec_bld->type.length, have_corner);
+ LLVMBuildStore(builder, have_corner, have_corners);
+
+ /*
+ * Need to feed clamped values here for cheap corner handling,
+ * but only for y coord (as when falling off both edges we only
+ * fall off the x one) - this should be sufficient.
+ */
+ y0_clamped = lp_build_max(ivec_bld, y0, ivec_bld->zero);
+ y1_clamped = lp_build_min(ivec_bld, y1, length_minus_one);
+ /*
+ * Get all possible new coords.
+ */
+ lp_build_cube_new_coords(ivec_bld, face,
+ x0, x1, y0_clamped, y1_clamped,
+ length_minus_one,
+ new_faces, new_xcoords, new_ycoords);
+
+ /* handle fall off x-, x+ direction */
+ /* determine new coords, face (not both fall_off vars can be true at same time) */
+ x00 = lp_build_select(ivec_bld, fall_off[0], new_xcoords[0][0], x0);
+ y00 = lp_build_select(ivec_bld, fall_off[0], new_ycoords[0][0], y0_clamped);
+ x10 = lp_build_select(ivec_bld, fall_off[0], new_xcoords[0][1], x0);
+ y10 = lp_build_select(ivec_bld, fall_off[0], new_ycoords[0][1], y1_clamped);
+ x01 = lp_build_select(ivec_bld, fall_off[1], new_xcoords[1][0], x1);
+ y01 = lp_build_select(ivec_bld, fall_off[1], new_ycoords[1][0], y0_clamped);
+ x11 = lp_build_select(ivec_bld, fall_off[1], new_xcoords[1][1], x1);
+ y11 = lp_build_select(ivec_bld, fall_off[1], new_ycoords[1][1], y1_clamped);
+
+ z00 = z10 = lp_build_select(ivec_bld, fall_off[0], new_faces[0], face);
+ z01 = z11 = lp_build_select(ivec_bld, fall_off[1], new_faces[1], face);
+
+ /* handle fall off y-, y+ direction */
+ /*
+ * Cheap corner logic: just hack up things so a texel doesn't fall
+ * off both sides (which means filter weights will be wrong but we'll only
+ * use valid texels in the filter).
+ * This means however (y) coords must additionally be clamped (see above).
+ * This corner handling should be fully OpenGL (but not d3d10) compliant.
+ */
+ fall_off_ym_notxm = lp_build_andnot(ivec_bld, fall_off[2], fall_off[0]);
+ fall_off_ym_notxp = lp_build_andnot(ivec_bld, fall_off[2], fall_off[1]);
+ fall_off_yp_notxm = lp_build_andnot(ivec_bld, fall_off[3], fall_off[0]);
+ fall_off_yp_notxp = lp_build_andnot(ivec_bld, fall_off[3], fall_off[1]);
+
+ x00 = lp_build_select(ivec_bld, fall_off_ym_notxm, new_xcoords[2][0], x00);
+ y00 = lp_build_select(ivec_bld, fall_off_ym_notxm, new_ycoords[2][0], y00);
+ x01 = lp_build_select(ivec_bld, fall_off_ym_notxp, new_xcoords[2][1], x01);
+ y01 = lp_build_select(ivec_bld, fall_off_ym_notxp, new_ycoords[2][1], y01);
+ x10 = lp_build_select(ivec_bld, fall_off_yp_notxm, new_xcoords[3][0], x10);
+ y10 = lp_build_select(ivec_bld, fall_off_yp_notxm, new_ycoords[3][0], y10);
+ x11 = lp_build_select(ivec_bld, fall_off_yp_notxp, new_xcoords[3][1], x11);
+ y11 = lp_build_select(ivec_bld, fall_off_yp_notxp, new_ycoords[3][1], y11);
+
+ z00 = lp_build_select(ivec_bld, fall_off_ym_notxm, new_faces[2], z00);
+ z01 = lp_build_select(ivec_bld, fall_off_ym_notxp, new_faces[2], z01);
+ z10 = lp_build_select(ivec_bld, fall_off_yp_notxm, new_faces[3], z10);
+ z11 = lp_build_select(ivec_bld, fall_off_yp_notxp, new_faces[3], z11);
+
+ LLVMBuildStore(builder, x00, xs[0]);
+ LLVMBuildStore(builder, x01, xs[1]);
+ LLVMBuildStore(builder, x10, xs[2]);
+ LLVMBuildStore(builder, x11, xs[3]);
+ LLVMBuildStore(builder, y00, ys[0]);
+ LLVMBuildStore(builder, y01, ys[1]);
+ LLVMBuildStore(builder, y10, ys[2]);
+ LLVMBuildStore(builder, y11, ys[3]);
+ LLVMBuildStore(builder, z00, zs[0]);
+ LLVMBuildStore(builder, z01, zs[1]);
+ LLVMBuildStore(builder, z10, zs[2]);
+ LLVMBuildStore(builder, z11, zs[3]);
+
+ lp_build_else(&edge_if);
+
+ LLVMBuildStore(builder, x0, xs[0]);
+ LLVMBuildStore(builder, x1, xs[1]);
+ LLVMBuildStore(builder, x0, xs[2]);
+ LLVMBuildStore(builder, x1, xs[3]);
+ LLVMBuildStore(builder, y0, ys[0]);
+ LLVMBuildStore(builder, y0, ys[1]);
+ LLVMBuildStore(builder, y1, ys[2]);
+ LLVMBuildStore(builder, y1, ys[3]);
+ LLVMBuildStore(builder, face, zs[0]);
+ LLVMBuildStore(builder, face, zs[1]);
+ LLVMBuildStore(builder, face, zs[2]);
+ LLVMBuildStore(builder, face, zs[3]);
+
+ lp_build_endif(&edge_if);
+
+ x00 = LLVMBuildLoad(builder, xs[0], "");
+ x01 = LLVMBuildLoad(builder, xs[1], "");
+ x10 = LLVMBuildLoad(builder, xs[2], "");
+ x11 = LLVMBuildLoad(builder, xs[3], "");
+ y00 = LLVMBuildLoad(builder, ys[0], "");
+ y01 = LLVMBuildLoad(builder, ys[1], "");
+ y10 = LLVMBuildLoad(builder, ys[2], "");
+ y11 = LLVMBuildLoad(builder, ys[3], "");
+ z00 = LLVMBuildLoad(builder, zs[0], "");
+ z01 = LLVMBuildLoad(builder, zs[1], "");
+ z10 = LLVMBuildLoad(builder, zs[2], "");
+ z11 = LLVMBuildLoad(builder, zs[3], "");
+ }
+
+ if (linear_mask) {
+ /*
+ * Whack filter weights into place. Whatever texel had more weight is
+ * the one which should have been selected by nearest filtering hence
+ * just use 100% weight for it.
+ */
+ struct lp_build_context *c_bld = &bld->coord_bld;
+ LLVMValueRef w1_mask, w1_weight;
+ LLVMValueRef half = lp_build_const_vec(bld->gallivm, c_bld->type, 0.5f);
+
+ w1_mask = lp_build_cmp(c_bld, PIPE_FUNC_GREATER, s_fpart, half);
+ /* this select is really just a "and" */
+ w1_weight = lp_build_select(c_bld, w1_mask, c_bld->one, c_bld->zero);
+ s_fpart = lp_build_select(c_bld, linear_mask, s_fpart, w1_weight);
+ if (dims >= 2) {
+ w1_mask = lp_build_cmp(c_bld, PIPE_FUNC_GREATER, t_fpart, half);
+ w1_weight = lp_build_select(c_bld, w1_mask, c_bld->one, c_bld->zero);
+ t_fpart = lp_build_select(c_bld, linear_mask, t_fpart, w1_weight);
+ if (dims == 3) {
+ w1_mask = lp_build_cmp(c_bld, PIPE_FUNC_GREATER, r_fpart, half);
+ w1_weight = lp_build_select(c_bld, w1_mask, c_bld->one, c_bld->zero);
+ r_fpart = lp_build_select(c_bld, linear_mask, r_fpart, w1_weight);
+ }
+ }
+ }
/*
* Get texture colors.
*/
/* get x0/x1 texels */
- lp_build_sample_texel_soa(bld, sampler_unit,
+ lp_build_sample_texel_soa(bld,
width_vec, height_vec, depth_vec,
- x0, y0, z0,
+ x00, y00, z00,
row_stride_vec, img_stride_vec,
data_ptr, mipoffsets, neighbors[0][0]);
- lp_build_sample_texel_soa(bld, sampler_unit,
+ lp_build_sample_texel_soa(bld,
width_vec, height_vec, depth_vec,
- x1, y0, z0,
+ x01, y01, z01,
row_stride_vec, img_stride_vec,
data_ptr, mipoffsets, neighbors[0][1]);
}
else {
/* 2D/3D texture */
- LLVMValueRef colors0[4];
+ struct lp_build_if_state corner_if;
+ LLVMValueRef colors0[4], colorss[4];
/* get x0/x1 texels at y1 */
- lp_build_sample_texel_soa(bld, sampler_unit,
+ lp_build_sample_texel_soa(bld,
width_vec, height_vec, depth_vec,
- x0, y1, z0,
+ x10, y10, z10,
row_stride_vec, img_stride_vec,
data_ptr, mipoffsets, neighbors[1][0]);
- lp_build_sample_texel_soa(bld, sampler_unit,
+ lp_build_sample_texel_soa(bld,
width_vec, height_vec, depth_vec,
- x1, y1, z0,
+ x11, y11, z11,
row_stride_vec, img_stride_vec,
data_ptr, mipoffsets, neighbors[1][1]);
+ /*
+ * To avoid having to duplicate linear_mask / fetch code use
+ * another branch (with corner condition though edge would work
+ * as well) here.
+ */
+ if (accurate_cube_corners) {
+ LLVMValueRef w00, w01, w10, w11, wx0, wy0;
+ LLVMValueRef c_weight, c00, c01, c10, c11;
+ LLVMValueRef have_corner, one_third, tmp;
+
+ colorss[0] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs");
+ colorss[1] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs");
+ colorss[2] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs");
+ colorss[3] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs");
+
+ have_corner = LLVMBuildLoad(builder, have_corners, "");
+
+ lp_build_if(&corner_if, bld->gallivm, have_corner);
+
+ /*
+ * we can't use standard 2d lerp as we need per-element weight
+ * in case of corners, so just calculate bilinear result as
+ * w00*s00 + w01*s01 + w10*s10 + w11*s11.
+ * (This is actually less work than using 2d lerp, 7 vs. 9 instructions,
+ * however calculating the weights needs another 6, so actually probably
+ * not slower than 2d lerp only for 4 channels as weights only need
+ * to be calculated once - of course fixing the weights has additional cost.)
+ */
+ wx0 = lp_build_sub(coord_bld, coord_bld->one, s_fpart);
+ wy0 = lp_build_sub(coord_bld, coord_bld->one, t_fpart);
+ w00 = lp_build_mul(coord_bld, wx0, wy0);
+ w01 = lp_build_mul(coord_bld, s_fpart, wy0);
+ w10 = lp_build_mul(coord_bld, wx0, t_fpart);
+ w11 = lp_build_mul(coord_bld, s_fpart, t_fpart);
+
+ /* find corner weight */
+ c00 = lp_build_and(ivec_bld, fall_off[0], fall_off[2]);
+ c_weight = lp_build_select(coord_bld, c00, w00, coord_bld->zero);
+ c01 = lp_build_and(ivec_bld, fall_off[1], fall_off[2]);
+ c_weight = lp_build_select(coord_bld, c01, w01, c_weight);
+ c10 = lp_build_and(ivec_bld, fall_off[0], fall_off[3]);
+ c_weight = lp_build_select(coord_bld, c10, w10, c_weight);
+ c11 = lp_build_and(ivec_bld, fall_off[1], fall_off[3]);
+ c_weight = lp_build_select(coord_bld, c11, w11, c_weight);
+
+ /*
+ * add 1/3 of the corner weight to each of the 3 other samples
+ * and null out corner weight
+ */
+ one_third = lp_build_const_vec(bld->gallivm, coord_bld->type, 1.0f/3.0f);
+ c_weight = lp_build_mul(coord_bld, c_weight, one_third);
+ w00 = lp_build_add(coord_bld, w00, c_weight);
+ c00 = LLVMBuildBitCast(builder, c00, coord_bld->vec_type, "");
+ w00 = lp_build_andnot(coord_bld, w00, c00);
+ w01 = lp_build_add(coord_bld, w01, c_weight);
+ c01 = LLVMBuildBitCast(builder, c01, coord_bld->vec_type, "");
+ w01 = lp_build_andnot(coord_bld, w01, c01);
+ w10 = lp_build_add(coord_bld, w10, c_weight);
+ c10 = LLVMBuildBitCast(builder, c10, coord_bld->vec_type, "");
+ w10 = lp_build_andnot(coord_bld, w10, c10);
+ w11 = lp_build_add(coord_bld, w11, c_weight);
+ c11 = LLVMBuildBitCast(builder, c11, coord_bld->vec_type, "");
+ w11 = lp_build_andnot(coord_bld, w11, c11);
+
+ if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
+ for (chan = 0; chan < 4; chan++) {
+ colors0[chan] = lp_build_mul(coord_bld, w00, neighbors[0][0][chan]);
+ tmp = lp_build_mul(coord_bld, w01, neighbors[0][1][chan]);
+ colors0[chan] = lp_build_add(coord_bld, tmp, colors0[chan]);
+ tmp = lp_build_mul(coord_bld, w10, neighbors[1][0][chan]);
+ colors0[chan] = lp_build_add(coord_bld, tmp, colors0[chan]);
+ tmp = lp_build_mul(coord_bld, w11, neighbors[1][1][chan]);
+ colors0[chan] = lp_build_add(coord_bld, tmp, colors0[chan]);
+ }
+ }
+ else {
+ LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
+ cmpval00 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]);
+ cmpval01 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]);
+ cmpval10 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][0][0]);
+ cmpval11 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][1][0]);
+ /* inputs to interpolation are just masks so just add masked weights together */
+ cmpval00 = LLVMBuildBitCast(builder, cmpval00, coord_bld->vec_type, "");
+ cmpval01 = LLVMBuildBitCast(builder, cmpval01, coord_bld->vec_type, "");
+ cmpval10 = LLVMBuildBitCast(builder, cmpval10, coord_bld->vec_type, "");
+ cmpval11 = LLVMBuildBitCast(builder, cmpval11, coord_bld->vec_type, "");
+ colors0[0] = lp_build_and(coord_bld, w00, cmpval00);
+ tmp = lp_build_and(coord_bld, w01, cmpval01);
+ colors0[0] = lp_build_add(coord_bld, tmp, colors0[0]);
+ tmp = lp_build_and(coord_bld, w10, cmpval10);
+ colors0[0] = lp_build_add(coord_bld, tmp, colors0[0]);
+ tmp = lp_build_and(coord_bld, w11, cmpval11);
+ colors0[0] = lp_build_add(coord_bld, tmp, colors0[0]);
+ colors0[1] = colors0[2] = colors0[3] = colors0[0];
+ }
+
+ LLVMBuildStore(builder, colors0[0], colorss[0]);
+ LLVMBuildStore(builder, colors0[1], colorss[1]);
+ LLVMBuildStore(builder, colors0[2], colorss[2]);
+ LLVMBuildStore(builder, colors0[3], colorss[3]);
+
+ lp_build_else(&corner_if);
+ }
+
if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
/* Bilinear interpolate the four samples from the 2D image / 3D slice */
for (chan = 0; chan < 4; chan++) {
colors0[1] = colors0[2] = colors0[3] = colors0[0];
}
+ if (accurate_cube_corners) {
+ LLVMBuildStore(builder, colors0[0], colorss[0]);
+ LLVMBuildStore(builder, colors0[1], colorss[1]);
+ LLVMBuildStore(builder, colors0[2], colorss[2]);
+ LLVMBuildStore(builder, colors0[3], colorss[3]);
+
+ lp_build_endif(&corner_if);
+
+ colors0[0] = LLVMBuildLoad(builder, colorss[0], "");
+ colors0[1] = LLVMBuildLoad(builder, colorss[1], "");
+ colors0[2] = LLVMBuildLoad(builder, colorss[2], "");
+ colors0[3] = LLVMBuildLoad(builder, colorss[3], "");
+ }
+
if (dims == 3) {
LLVMValueRef neighbors1[2][2][4];
LLVMValueRef colors1[4];
/* get x0/x1/y0/y1 texels at z1 */
- lp_build_sample_texel_soa(bld, sampler_unit,
+ lp_build_sample_texel_soa(bld,
width_vec, height_vec, depth_vec,
- x0, y0, z1,
+ x00, y00, z1,
row_stride_vec, img_stride_vec,
data_ptr, mipoffsets, neighbors1[0][0]);
- lp_build_sample_texel_soa(bld, sampler_unit,
+ lp_build_sample_texel_soa(bld,
width_vec, height_vec, depth_vec,
- x1, y0, z1,
+ x01, y01, z1,
row_stride_vec, img_stride_vec,
data_ptr, mipoffsets, neighbors1[0][1]);
- lp_build_sample_texel_soa(bld, sampler_unit,
+ lp_build_sample_texel_soa(bld,
width_vec, height_vec, depth_vec,
- x0, y1, z1,
+ x10, y10, z1,
row_stride_vec, img_stride_vec,
data_ptr, mipoffsets, neighbors1[1][0]);
- lp_build_sample_texel_soa(bld, sampler_unit,
+ lp_build_sample_texel_soa(bld,
width_vec, height_vec, depth_vec,
- x1, y1, z1,
+ x11, y11, z1,
row_stride_vec, img_stride_vec,
data_ptr, mipoffsets, neighbors1[1][1]);
/**
* Sample the texture/mipmap using given image filter and mip filter.
- * data0_ptr and data1_ptr point to the two mipmap levels to sample
- * from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
+ * ilevel0 and ilevel1 indicate the two mipmap levels to sample
+ * from (vectors or scalars).
* If we're using nearest miplevel sampling the '1' values will be null/unused.
*/
static void
lp_build_sample_mipmap(struct lp_build_sample_context *bld,
- unsigned sampler_unit,
unsigned img_filter,
unsigned mip_filter,
LLVMValueRef *coords,
lp_build_mipmap_level_sizes(bld, ilevel0,
&size0,
&row_stride0_vec, &img_stride0_vec);
- if (bld->num_lods == 1) {
+ if (bld->num_mips == 1) {
data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
}
else {
mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
}
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
- lp_build_sample_image_nearest(bld, sampler_unit,
- size0,
+ lp_build_sample_image_nearest(bld, size0,
row_stride0_vec, img_stride0_vec,
data_ptr0, mipoff0, coords, offsets,
colors0);
}
else {
assert(img_filter == PIPE_TEX_FILTER_LINEAR);
- lp_build_sample_image_linear(bld, sampler_unit,
- size0,
+ lp_build_sample_image_linear(bld, size0, NULL,
row_stride0_vec, img_stride0_vec,
data_ptr0, mipoff0, coords, offsets,
colors0);
/* need_lerp = lod_fpart > 0 */
if (bld->num_lods == 1) {
need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
- lod_fpart, bld->levelf_bld.zero,
+ lod_fpart, bld->lodf_bld.zero,
"need_lerp");
}
else {
* We'll do mip filtering if any of the quads (or individual
* pixel in case of per-pixel lod) need it.
* It might be better to split the vectors here and only fetch/filter
- * quads which need it.
- */
- /*
- * We unfortunately need to clamp lod_fpart here since we can get
- * negative values which would screw up filtering if not all
- * lod_fpart values have same sign.
+ * quads which need it (if there's one lod per quad).
*/
- lod_fpart = lp_build_max(&bld->levelf_bld, lod_fpart,
- bld->levelf_bld.zero);
- need_lerp = lp_build_compare(bld->gallivm, bld->levelf_bld.type,
+ need_lerp = lp_build_compare(bld->gallivm, bld->lodf_bld.type,
PIPE_FUNC_GREATER,
- lod_fpart, bld->levelf_bld.zero);
- need_lerp = lp_build_any_true_range(&bld->leveli_bld, bld->num_lods, need_lerp);
+ lod_fpart, bld->lodf_bld.zero);
+ need_lerp = lp_build_any_true_range(&bld->lodi_bld, bld->num_lods, need_lerp);
}
lp_build_if(&if_ctx, bld->gallivm, need_lerp);
{
+ /*
+ * We unfortunately need to clamp lod_fpart here since we can get
+ * negative values which would screw up filtering if not all
+ * lod_fpart values have same sign.
+ */
+ lod_fpart = lp_build_max(&bld->lodf_bld, lod_fpart,
+ bld->lodf_bld.zero);
/* sample the second mipmap level */
lp_build_mipmap_level_sizes(bld, ilevel1,
&size1,
&row_stride1_vec, &img_stride1_vec);
- if (bld->num_lods == 1) {
+ if (bld->num_mips == 1) {
data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
}
else {
mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
}
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
- lp_build_sample_image_nearest(bld, sampler_unit,
- size1,
+ lp_build_sample_image_nearest(bld, size1,
row_stride1_vec, img_stride1_vec,
data_ptr1, mipoff1, coords, offsets,
colors1);
}
else {
- lp_build_sample_image_linear(bld, sampler_unit,
- size1,
+ lp_build_sample_image_linear(bld, size1, NULL,
row_stride1_vec, img_stride1_vec,
data_ptr1, mipoff1, coords, offsets,
colors1);
if (bld->num_lods != bld->coord_type.length)
lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
- bld->levelf_bld.type,
+ bld->lodf_bld.type,
+ bld->texel_bld.type,
+ lod_fpart);
+
+ for (chan = 0; chan < 4; chan++) {
+ colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
+ colors0[chan], colors1[chan],
+ 0);
+ LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
+ }
+ }
+ lp_build_endif(&if_ctx);
+ }
+}
+
+
+/**
+ * Sample the texture/mipmap using given mip filter, and using
+ * both nearest and linear filtering at the same time depending
+ * on linear_mask.
+ * lod can be per quad but linear_mask is always per pixel.
+ * ilevel0 and ilevel1 indicate the two mipmap levels to sample
+ * from (vectors or scalars).
+ * If we're using nearest miplevel sampling the '1' values will be null/unused.
+ */
+static void
+lp_build_sample_mipmap_both(struct lp_build_sample_context *bld,
+ LLVMValueRef linear_mask,
+ unsigned mip_filter,
+ LLVMValueRef *coords,
+ const LLVMValueRef *offsets,
+ LLVMValueRef ilevel0,
+ LLVMValueRef ilevel1,
+ LLVMValueRef lod_fpart,
+ LLVMValueRef lod_positive,
+ LLVMValueRef *colors_out)
+{
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ LLVMValueRef size0 = NULL;
+ LLVMValueRef size1 = NULL;
+ LLVMValueRef row_stride0_vec = NULL;
+ LLVMValueRef row_stride1_vec = NULL;
+ LLVMValueRef img_stride0_vec = NULL;
+ LLVMValueRef img_stride1_vec = NULL;
+ LLVMValueRef data_ptr0 = NULL;
+ LLVMValueRef data_ptr1 = NULL;
+ LLVMValueRef mipoff0 = NULL;
+ LLVMValueRef mipoff1 = NULL;
+ LLVMValueRef colors0[4], colors1[4];
+ unsigned chan;
+
+ /* sample the first mipmap level */
+ lp_build_mipmap_level_sizes(bld, ilevel0,
+ &size0,
+ &row_stride0_vec, &img_stride0_vec);
+ if (bld->num_mips == 1) {
+ data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
+ }
+ else {
+ /* This path should work for num_lods 1 too but slightly less efficient */
+ data_ptr0 = bld->base_ptr;
+ mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
+ }
+
+ lp_build_sample_image_linear(bld, size0, linear_mask,
+ row_stride0_vec, img_stride0_vec,
+ data_ptr0, mipoff0, coords, offsets,
+ colors0);
+
+ /* Store the first level's colors in the output variables */
+ for (chan = 0; chan < 4; chan++) {
+ LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
+ }
+
+ if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
+ struct lp_build_if_state if_ctx;
+ LLVMValueRef need_lerp;
+
+ /*
+ * We'll do mip filtering if any of the quads (or individual
+ * pixel in case of per-pixel lod) need it.
+ * Note using lod_positive here not lod_fpart since it may be the same
+ * condition as that used in the outer "if" in the caller hence llvm
+ * should be able to merge the branches in this case.
+ */
+ need_lerp = lp_build_any_true_range(&bld->lodi_bld, bld->num_lods, lod_positive);
+
+ lp_build_if(&if_ctx, bld->gallivm, need_lerp);
+ {
+ /*
+ * We unfortunately need to clamp lod_fpart here since we can get
+ * negative values which would screw up filtering if not all
+ * lod_fpart values have same sign.
+ */
+ lod_fpart = lp_build_max(&bld->lodf_bld, lod_fpart,
+ bld->lodf_bld.zero);
+ /* sample the second mipmap level */
+ lp_build_mipmap_level_sizes(bld, ilevel1,
+ &size1,
+ &row_stride1_vec, &img_stride1_vec);
+ if (bld->num_mips == 1) {
+ data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
+ }
+ else {
+ data_ptr1 = bld->base_ptr;
+ mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
+ }
+
+ lp_build_sample_image_linear(bld, size1, linear_mask,
+ row_stride1_vec, img_stride1_vec,
+ data_ptr1, mipoff1, coords, offsets,
+ colors1);
+
+ /* interpolate samples from the two mipmap levels */
+
+ if (bld->num_lods != bld->coord_type.length)
+ lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
+ bld->lodf_bld.type,
bld->texel_bld.type,
lod_fpart);
const struct lp_derivatives *derivs, /* optional */
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
- LLVMValueRef *lod_ipart,
+ LLVMValueRef *lod_pos_or_zero,
LLVMValueRef *lod_fpart,
LLVMValueRef *ilevel0,
LLVMValueRef *ilevel1)
const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
const unsigned target = bld->static_texture_state->target;
LLVMValueRef first_level, cube_rho = NULL;
+ LLVMValueRef lod_ipart = NULL;
+ struct lp_derivatives cube_derivs;
/*
printf("%s mip %d min %d mag %d\n", __FUNCTION__,
mip_filter != PIPE_TEX_MIPFILTER_NONE) &&
!bld->static_sampler_state->min_max_lod_equal &&
!explicit_lod);
- lp_build_cube_lookup(bld, coords, derivs, &cube_rho, need_derivs);
+ lp_build_cube_lookup(bld, coords, derivs, &cube_rho, &cube_derivs, need_derivs);
+ derivs = &cube_derivs;
}
else if (target == PIPE_TEXTURE_1D_ARRAY ||
target == PIPE_TEXTURE_2D_ARRAY) {
coords[0], coords[1], coords[2], cube_rho,
derivs, lod_bias, explicit_lod,
mip_filter,
- lod_ipart, lod_fpart);
+ &lod_ipart, lod_fpart, lod_pos_or_zero);
} else {
- *lod_ipart = bld->leveli_bld.zero;
+ lod_ipart = bld->lodi_bld.zero;
+ *lod_pos_or_zero = bld->lodi_bld.zero;
+ }
+
+ if (bld->num_lods != bld->num_mips) {
+ /* only makes sense if there's just a single mip level */
+ assert(bld->num_mips == 1);
+ lod_ipart = lp_build_extract_range(bld->gallivm, lod_ipart, 0, 1);
}
/*
* We should be able to set ilevel0 = const(0) but that causes
* bad x86 code to be emitted.
*/
- assert(*lod_ipart);
- lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0, NULL);
+ assert(lod_ipart);
+ lp_build_nearest_mip_level(bld, texture_index, lod_ipart, ilevel0, NULL);
}
else {
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
}
break;
case PIPE_TEX_MIPFILTER_NEAREST:
- assert(*lod_ipart);
- lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0, NULL);
+ assert(lod_ipart);
+ lp_build_nearest_mip_level(bld, texture_index, lod_ipart, ilevel0, NULL);
break;
case PIPE_TEX_MIPFILTER_LINEAR:
- assert(*lod_ipart);
+ assert(lod_ipart);
assert(*lod_fpart);
lp_build_linear_mip_levels(bld, texture_index,
- *lod_ipart, lod_fpart,
+ lod_ipart, lod_fpart,
ilevel0, ilevel1);
break;
}
unsigned sampler_unit,
LLVMValueRef *coords,
const LLVMValueRef *offsets,
- LLVMValueRef lod_ipart,
+ LLVMValueRef lod_positive,
LLVMValueRef lod_fpart,
LLVMValueRef ilevel0,
LLVMValueRef ilevel1,
LLVMValueRef *colors_out)
{
- struct lp_build_context *int_bld = &bld->int_bld;
LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_static_sampler_state *sampler_state = bld->static_sampler_state;
const unsigned mip_filter = sampler_state->min_mip_filter;
if (min_filter == mag_filter) {
/* no need to distinguish between minification and magnification */
- lp_build_sample_mipmap(bld, sampler_unit,
- min_filter, mip_filter,
+ lp_build_sample_mipmap(bld, min_filter, mip_filter,
coords, offsets,
ilevel0, ilevel1, lod_fpart,
texels);
}
else {
- /* Emit conditional to choose min image filter or mag image filter
- * depending on the lod being > 0 or <= 0, respectively.
- */
- struct lp_build_if_state if_ctx;
- LLVMValueRef minify;
-
/*
- * XXX this should take all lods into account, if some are min
- * some max probably could hack up the coords/weights in the linear
- * path with selects to work for nearest.
- * If that's just two quads sitting next to each other it seems
- * quite ok to do the same filtering method on both though, at
- * least unless we have explicit lod (and who uses different
- * min/mag filter with that?)
+ * Could also get rid of the if-logic and always use mipmap_both, both
+ * for the single lod and multi-lod case if nothing really uses this.
*/
- if (bld->num_lods > 1)
- lod_ipart = LLVMBuildExtractElement(builder, lod_ipart,
- lp_build_const_int32(bld->gallivm, 0), "");
+ if (bld->num_lods == 1) {
+ /* Emit conditional to choose min image filter or mag image filter
+ * depending on the lod being > 0 or <= 0, respectively.
+ */
+ struct lp_build_if_state if_ctx;
+
+ lod_positive = LLVMBuildTrunc(builder, lod_positive,
+ LLVMInt1TypeInContext(bld->gallivm->context), "");
+
+ lp_build_if(&if_ctx, bld->gallivm, lod_positive);
+ {
+ /* Use the minification filter */
+ lp_build_sample_mipmap(bld, min_filter, mip_filter,
+ coords, offsets,
+ ilevel0, ilevel1, lod_fpart,
+ texels);
+ }
+ lp_build_else(&if_ctx);
+ {
+ /* Use the magnification filter */
+ lp_build_sample_mipmap(bld, mag_filter, PIPE_TEX_MIPFILTER_NONE,
+ coords, offsets,
+ ilevel0, NULL, NULL,
+ texels);
+ }
+ lp_build_endif(&if_ctx);
+ }
+ else {
+ LLVMValueRef need_linear, linear_mask;
+ unsigned mip_filter_for_nearest;
+ struct lp_build_if_state if_ctx;
- /* minify = lod >= 0.0 */
- minify = LLVMBuildICmp(builder, LLVMIntSGE,
- lod_ipart, int_bld->zero, "");
+ if (min_filter == PIPE_TEX_FILTER_LINEAR) {
+ linear_mask = lod_positive;
+ mip_filter_for_nearest = PIPE_TEX_MIPFILTER_NONE;
+ }
+ else {
+ linear_mask = lp_build_not(&bld->lodi_bld, lod_positive);
+ mip_filter_for_nearest = mip_filter;
+ }
+ need_linear = lp_build_any_true_range(&bld->lodi_bld, bld->num_lods,
+ linear_mask);
+
+ if (bld->num_lods != bld->coord_type.length) {
+ linear_mask = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
+ bld->lodi_type,
+ bld->int_coord_type,
+ linear_mask);
+ }
- lp_build_if(&if_ctx, bld->gallivm, minify);
- {
- /* Use the minification filter */
- lp_build_sample_mipmap(bld, sampler_unit,
- min_filter, mip_filter,
- coords, offsets,
- ilevel0, ilevel1, lod_fpart,
- texels);
- }
- lp_build_else(&if_ctx);
- {
- /* Use the magnification filter */
- lp_build_sample_mipmap(bld, sampler_unit,
- mag_filter, PIPE_TEX_MIPFILTER_NONE,
- coords, offsets,
- ilevel0, NULL, NULL,
- texels);
+ lp_build_if(&if_ctx, bld->gallivm, need_linear);
+ {
+ /*
+ * Do sampling with both filters simultaneously. This means using
+ * a linear filter and doing some tricks (with weights) for the pixels
+ * which need nearest filter.
+ * Note that it's probably rare some pixels need nearest and some
+ * linear filter but the fixups required for the nearest pixels
+ * aren't all that complicated so just always run a combined path
+ * if at least some pixels require linear.
+ */
+ lp_build_sample_mipmap_both(bld, linear_mask, mip_filter,
+ coords, offsets,
+ ilevel0, ilevel1,
+ lod_fpart, lod_positive,
+ texels);
+ }
+ lp_build_else(&if_ctx);
+ {
+ /*
+ * All pixels require just nearest filtering, which is way
+ * cheaper than linear, hence do a separate path for that.
+ */
+ lp_build_sample_mipmap(bld, PIPE_TEX_FILTER_NEAREST,
+ mip_filter_for_nearest,
+ coords, offsets,
+ ilevel0, ilevel1, lod_fpart,
+ texels);
+ }
+ lp_build_endif(&if_ctx);
}
- lp_build_endif(&if_ctx);
}
for (chan = 0; chan < 4; ++chan) {
const LLVMValueRef *offsets,
LLVMValueRef *colors_out)
{
- struct lp_build_context *perquadi_bld = &bld->leveli_bld;
+ struct lp_build_context *perquadi_bld = &bld->lodi_bld;
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
unsigned dims = bld->dims, chan;
unsigned target = bld->static_texture_state->target;
out_of_bounds = int_coord_bld->zero;
if (explicit_lod && bld->static_texture_state->target != PIPE_BUFFER) {
- if (bld->num_lods != int_coord_bld->type.length) {
+ if (bld->num_mips != int_coord_bld->type.length) {
ilevel = lp_build_pack_aos_scalars(bld->gallivm, int_coord_bld->type,
perquadi_bld->type, explicit_lod, 0);
}
out_of_bound_ret_zero ? &out_of_bounds : NULL);
}
else {
- assert(bld->num_lods == 1);
+ assert(bld->num_mips == 1);
if (bld->static_texture_state->target != PIPE_BUFFER) {
ilevel = bld->dynamic_state->first_level(bld->dynamic_state,
bld->gallivm, texture_unit);
unsigned target = static_texture_state->target;
unsigned dims = texture_dims(target);
unsigned num_quads = type.length / 4;
- unsigned mip_filter, i;
+ unsigned mip_filter, min_img_filter, mag_img_filter, i;
struct lp_build_sample_context bld;
struct lp_static_sampler_state derived_sampler_state = *static_sampler_state;
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
debug_printf("Sample from %s\n", util_format_name(fmt));
}
+ if (static_texture_state->format == PIPE_FORMAT_NONE) {
+ /*
+ * If there's nothing bound, format is NONE, and we must return
+ * all zero as mandated by d3d10 in this case.
+ */
+ unsigned chan;
+ LLVMValueRef zero = lp_build_const_vec(gallivm, type, 0.0F);
+ for (chan = 0; chan < 4; chan++) {
+ texel_out[chan] = zero;
+ }
+ return;
+ }
+
assert(type.floating);
/* Setup our build context */
debug_printf(" .min_mip_filter = %u\n", derived_sampler_state.min_mip_filter);
}
+ if (static_texture_state->target == PIPE_TEXTURE_CUBE ||
+ static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY)
+ {
+ /*
+ * Seamless filtering ignores wrap modes.
+ * Setting to CLAMP_TO_EDGE is correct for nearest filtering, for
+ * bilinear it's not correct but way better than using for instance repeat.
+ * Note we even set this for non-seamless. Technically GL allows any wrap
+ * mode, which made sense when supporting true borders (can get seamless
+ * effect with border and CLAMP_TO_BORDER), but gallium doesn't support
+ * borders and d3d9 requires wrap modes to be ignored and it's a pain to fix
+ * up the sampler state (as it makes it texture dependent).
+ */
+ derived_sampler_state.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
+ derived_sampler_state.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
+ }
+
+ min_img_filter = derived_sampler_state.min_img_filter;
+ mag_img_filter = derived_sampler_state.mag_img_filter;
+
+
/*
* This is all a bit complicated different paths are chosen for performance
* reasons.
* There are other situations where at least the multiple int lods could be
* avoided like min and max lod being equal.
*/
- if (lod_property == LP_SAMPLER_LOD_PER_ELEMENT &&
- (explicit_lod || lod_bias ||
- (derivs && static_texture_state->target != PIPE_TEXTURE_CUBE)) &&
- ((is_fetch && target != PIPE_BUFFER) ||
- (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)))
+ bld.num_mips = bld.num_lods = 1;
+
+ if ((gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) &&
+ (gallivm_debug & GALLIVM_DEBUG_NO_RHO_APPROX) &&
+ (static_texture_state->target == PIPE_TEXTURE_CUBE) &&
+ (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
+ /*
+ * special case for using per-pixel lod even for implicit lod,
+ * which is generally never required (ok by APIs) except to please
+ * some (somewhat broken imho) tests (because per-pixel face selection
+ * can cause derivatives to be different for pixels outside the primitive
+ * due to the major axis division even if pre-project derivatives are
+ * looking normal).
+ */
+ bld.num_mips = type.length;
bld.num_lods = type.length;
+ }
+ else if (lod_property == LP_SAMPLER_LOD_PER_ELEMENT ||
+ (explicit_lod || lod_bias || derivs)) {
+ if ((is_fetch && target != PIPE_BUFFER) ||
+ (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
+ bld.num_mips = type.length;
+ bld.num_lods = type.length;
+ }
+ else if (!is_fetch && min_img_filter != mag_img_filter) {
+ bld.num_mips = 1;
+ bld.num_lods = type.length;
+ }
+ }
/* TODO: for true scalar_lod should only use 1 lod value */
- else if ((is_fetch && explicit_lod && target != PIPE_BUFFER ) ||
+ else if ((is_fetch && explicit_lod && target != PIPE_BUFFER) ||
(!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
+ bld.num_mips = num_quads;
bld.num_lods = num_quads;
}
- else {
- bld.num_lods = 1;
+ else if (!is_fetch && min_img_filter != mag_img_filter) {
+ bld.num_mips = 1;
+ bld.num_lods = num_quads;
}
- bld.levelf_type = type;
+
+ bld.lodf_type = type;
/* we want native vector size to be able to use our intrinsics */
if (bld.num_lods != type.length) {
- bld.levelf_type.length = type.length > 4 ? ((type.length + 15) / 16) * 4 : 1;
+ /* TODO: this currently always has to be per-quad or per-element */
+ bld.lodf_type.length = type.length > 4 ? ((type.length + 15) / 16) * 4 : 1;
+ }
+ bld.lodi_type = lp_int_type(bld.lodf_type);
+ bld.levelf_type = bld.lodf_type;
+ if (bld.num_mips == 1) {
+ bld.levelf_type.length = 1;
}
bld.leveli_type = lp_int_type(bld.levelf_type);
bld.float_size_type = bld.float_size_in_type;
/* Note: size vectors may not be native. They contain minified w/h/d/_ values,
* with per-element lod that is w0/h0/d0/_/w1/h1/d1_/... so up to 8x4f32 */
- if (bld.num_lods > 1) {
- bld.float_size_type.length = bld.num_lods == type.length ?
- bld.num_lods * bld.float_size_in_type.length :
+ if (bld.num_mips > 1) {
+ bld.float_size_type.length = bld.num_mips == type.length ?
+ bld.num_mips * bld.float_size_in_type.length :
type.length;
}
bld.int_size_type = lp_int_type(bld.float_size_type);
lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type);
lp_build_context_init(&bld.levelf_bld, gallivm, bld.levelf_type);
lp_build_context_init(&bld.leveli_bld, gallivm, bld.leveli_type);
+ lp_build_context_init(&bld.lodf_bld, gallivm, bld.lodf_type);
+ lp_build_context_init(&bld.lodi_bld, gallivm, bld.lodi_type);
/* Get the dynamic state */
tex_width = dynamic_state->width(dynamic_state, gallivm, texture_index);
}
else {
- LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
+ LLVMValueRef lod_fpart = NULL, lod_positive = NULL;
LLVMValueRef ilevel0 = NULL, ilevel1 = NULL;
boolean use_aos = util_format_fits_8unorm(bld.format_desc) &&
/* not sure this is strictly needed or simply impossible */
- static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE &&
- lp_is_simple_wrap_mode(static_sampler_state->wrap_s);
+ derived_sampler_state.compare_mode == PIPE_TEX_COMPARE_NONE &&
+ lp_is_simple_wrap_mode(derived_sampler_state.wrap_s);
+
+ use_aos &= bld.num_lods <= num_quads ||
+ derived_sampler_state.min_img_filter ==
+ derived_sampler_state.mag_img_filter;
if (dims > 1) {
- use_aos &= lp_is_simple_wrap_mode(static_sampler_state->wrap_t);
+ use_aos &= lp_is_simple_wrap_mode(derived_sampler_state.wrap_t);
if (dims > 2) {
- use_aos &= lp_is_simple_wrap_mode(static_sampler_state->wrap_r);
+ use_aos &= lp_is_simple_wrap_mode(derived_sampler_state.wrap_r);
}
}
+ if (static_texture_state->target == PIPE_TEXTURE_CUBE &&
+ derived_sampler_state.seamless_cube_map &&
+ (derived_sampler_state.min_img_filter == PIPE_TEX_FILTER_LINEAR ||
+ derived_sampler_state.mag_img_filter == PIPE_TEX_FILTER_LINEAR)) {
+ /* theoretically possible with AoS filtering but not implemented (complex!) */
+ use_aos = 0;
+ }
if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
!use_aos && util_format_fits_8unorm(bld.format_desc)) {
debug_printf("%s: using floating point linear filtering for %s\n",
__FUNCTION__, bld.format_desc->short_name);
- debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d wrapr %d\n",
- static_sampler_state->min_img_filter,
- static_sampler_state->mag_img_filter,
- static_sampler_state->min_mip_filter,
- static_sampler_state->wrap_s,
- static_sampler_state->wrap_t,
- static_sampler_state->wrap_r);
+ debug_printf(" min_img %d mag_img %d mip %d target %d seamless %d"
+ " wraps %d wrapt %d wrapr %d\n",
+ derived_sampler_state.min_img_filter,
+ derived_sampler_state.mag_img_filter,
+ derived_sampler_state.min_mip_filter,
+ static_texture_state->target,
+ derived_sampler_state.seamless_cube_map,
+ derived_sampler_state.wrap_s,
+ derived_sampler_state.wrap_t,
+ derived_sampler_state.wrap_r);
}
lp_build_sample_common(&bld, texture_index, sampler_index,
newcoords,
derivs, lod_bias, explicit_lod,
- &lod_ipart, &lod_fpart,
+ &lod_positive, &lod_fpart,
&ilevel0, &ilevel1);
/*
* we only try 8-wide sampling with soa as it appears to
- * be a loss with aos with AVX (but it should work).
+ * be a loss with aos with AVX (but it should work, except
+ * for conformance if min_filter != mag_filter if num_lods > 1).
* (It should be faster if we'd support avx2)
*/
if (num_quads == 1 || !use_aos) {
-
- if (num_quads > 1) {
- if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
- LLVMValueRef index0 = lp_build_const_int32(gallivm, 0);
- /*
- * These parameters are the same for all quads,
- * could probably simplify.
- */
- lod_ipart = LLVMBuildExtractElement(builder, lod_ipart, index0, "");
- ilevel0 = LLVMBuildExtractElement(builder, ilevel0, index0, "");
- }
- }
if (use_aos) {
/* do sampling/filtering with fixed pt arithmetic */
lp_build_sample_aos(&bld, sampler_index,
newcoords[0], newcoords[1],
newcoords[2],
- offsets, lod_ipart, lod_fpart,
+ offsets, lod_positive, lod_fpart,
ilevel0, ilevel1,
texel_out);
}
else {
lp_build_sample_general(&bld, sampler_index,
newcoords, offsets,
- lod_ipart, lod_fpart,
+ lod_positive, lod_fpart,
ilevel0, ilevel1,
texel_out);
}
bld4.int_size_in_type = lp_int_type(bld4.float_size_in_type);
bld4.texel_type = bld.texel_type;
bld4.texel_type.length = 4;
- bld4.levelf_type = type4;
- /* we want native vector size to be able to use our intrinsics */
- bld4.levelf_type.length = 1;
- bld4.leveli_type = lp_int_type(bld4.levelf_type);
- if (lod_property == LP_SAMPLER_LOD_PER_ELEMENT &&
- (explicit_lod || lod_bias ||
- (derivs && static_texture_state->target != PIPE_TEXTURE_CUBE)) &&
- ((is_fetch && target != PIPE_BUFFER) ||
- (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)))
+ bld4.num_mips = bld4.num_lods = 1;
+ if ((gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) &&
+ (gallivm_debug & GALLIVM_DEBUG_NO_RHO_APPROX) &&
+ (static_texture_state->target == PIPE_TEXTURE_CUBE) &&
+ (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
+ bld4.num_mips = type4.length;
bld4.num_lods = type4.length;
- else
- bld4.num_lods = 1;
+ }
+ if (lod_property == LP_SAMPLER_LOD_PER_ELEMENT &&
+ (explicit_lod || lod_bias || derivs)) {
+ if ((is_fetch && target != PIPE_BUFFER) ||
+ (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
+ bld4.num_mips = type4.length;
+ bld4.num_lods = type4.length;
+ }
+ else if (!is_fetch && min_img_filter != mag_img_filter) {
+ bld4.num_mips = 1;
+ bld4.num_lods = type4.length;
+ }
+ }
- bld4.levelf_type = type4;
/* we want native vector size to be able to use our intrinsics */
+ bld4.lodf_type = type4;
if (bld4.num_lods != type4.length) {
+ bld4.lodf_type.length = 1;
+ }
+ bld4.lodi_type = lp_int_type(bld4.lodf_type);
+ bld4.levelf_type = type4;
+ if (bld4.num_mips != type4.length) {
bld4.levelf_type.length = 1;
}
bld4.leveli_type = lp_int_type(bld4.levelf_type);
bld4.float_size_type = bld4.float_size_in_type;
- if (bld4.num_lods > 1) {
- bld4.float_size_type.length = bld4.num_lods == type4.length ?
- bld4.num_lods * bld4.float_size_in_type.length :
+ if (bld4.num_mips > 1) {
+ bld4.float_size_type.length = bld4.num_mips == type4.length ?
+ bld4.num_mips * bld4.float_size_in_type.length :
type4.length;
}
bld4.int_size_type = lp_int_type(bld4.float_size_type);
lp_build_context_init(&bld4.texel_bld, gallivm, bld4.texel_type);
lp_build_context_init(&bld4.levelf_bld, gallivm, bld4.levelf_type);
lp_build_context_init(&bld4.leveli_bld, gallivm, bld4.leveli_type);
+ lp_build_context_init(&bld4.lodf_bld, gallivm, bld4.lodf_type);
+ lp_build_context_init(&bld4.lodi_bld, gallivm, bld4.lodi_type);
for (i = 0; i < num_quads; i++) {
LLVMValueRef s4, t4, r4;
- LLVMValueRef lod_ipart4, lod_fpart4 = NULL;
+ LLVMValueRef lod_positive4, lod_fpart4 = NULL;
LLVMValueRef ilevel04, ilevel14 = NULL;
LLVMValueRef offsets4[4] = { NULL };
unsigned num_lods = bld4.num_lods;
}
}
}
- lod_ipart4 = lp_build_extract_range(gallivm, lod_ipart, num_lods * i, num_lods);
- ilevel04 = lp_build_extract_range(gallivm, ilevel0, num_lods * i, num_lods);
+ lod_positive4 = lp_build_extract_range(gallivm, lod_positive, num_lods * i, num_lods);
+ ilevel04 = bld.num_mips == 1 ? ilevel0 :
+ lp_build_extract_range(gallivm, ilevel0, num_lods * i, num_lods);
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
ilevel14 = lp_build_extract_range(gallivm, ilevel1, num_lods * i, num_lods);
lod_fpart4 = lp_build_extract_range(gallivm, lod_fpart, num_lods * i, num_lods);
/* do sampling/filtering with fixed pt arithmetic */
lp_build_sample_aos(&bld4, sampler_index,
s4, t4, r4, offsets4,
- lod_ipart4, lod_fpart4,
+ lod_positive4, lod_fpart4,
ilevel04, ilevel14,
texelout4);
}
lp_build_sample_general(&bld4, sampler_index,
newcoords4, offsets4,
- lod_ipart4, lod_fpart4,
+ lod_positive4, lod_fpart4,
ilevel04, ilevel14,
texelout4);
}
unsigned num_lods = 1;
struct lp_build_context bld_int_vec4;
+ if (static_state->format == PIPE_FORMAT_NONE) {
+ /*
+ * If there's nothing bound, format is NONE, and we must return
+ * all zero as mandated by d3d10 in this case.
+ */
+ unsigned chan;
+ LLVMValueRef zero = lp_build_const_vec(gallivm, int_type, 0.0F);
+ for (chan = 0; chan < 4; chan++) {
+ sizes_out[chan] = zero;
+ }
+ return;
+ }
+
/*
* Do some sanity verification about bound texture and shader dcl target.
* Not entirely sure what's possible but assume array/non-array
lp_build_const_int32(gallivm, 2), "");
}
- size = lp_build_minify(&bld_int_vec4, size, lod);
+ size = lp_build_minify(&bld_int_vec4, size, lod, TRUE);
if (has_array)
size = LLVMBuildInsertElement(gallivm->builder, size,