X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fauxiliary%2Fgallivm%2Flp_bld_sample_soa.c;h=e21933ffc851cc44dfc8481a86396fa720ca6dd6;hb=f72de6f3863049106288b7dd66efeb64c822fb17;hp=38a0eae0e9a92a001ef224a3a37bfefcf9c5de65;hpb=fb62388d6a9c22f81e56eb63b411abe3689eebb7;p=mesa.git diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c index 38a0eae0e9a..e21933ffc85 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c @@ -73,7 +73,6 @@ */ static void lp_build_sample_texel_soa(struct lp_build_sample_context *bld, - unsigned sampler_unit, LLVMValueRef width, LLVMValueRef height, LLVMValueRef depth, @@ -162,6 +161,7 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, bld->texel_type, data_ptr, offset, i, j, + bld->cache, texel_out); /* @@ -406,16 +406,17 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, break; case PIPE_TEX_WRAP_MIRROR_REPEAT: + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + offset = lp_build_div(coord_bld, offset, length_f); + coord = lp_build_add(coord_bld, coord, offset); + } /* compute mirror function */ coord = lp_build_coord_mirror(bld, coord); /* scale coord to length */ coord = lp_build_mul(coord_bld, coord, length_f); coord = lp_build_sub(coord_bld, coord, half); - if (offset) { - offset = lp_build_int_to_float(coord_bld, offset); - coord = lp_build_add(coord_bld, coord, offset); - } /* convert to int, compute lerp weight */ lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight); @@ -568,12 +569,13 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, coord = lp_build_mul(coord_bld, coord, length_f); } + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } /* floor */ /* use itrunc instead since we clamp to 0 anyway */ icoord = lp_build_itrunc(coord_bld, coord); - if (offset) { - icoord = lp_build_add(int_coord_bld, icoord, offset); - } /* clamp to [0, length - 1]. */ icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero, @@ -696,7 +698,6 @@ lp_build_sample_comparefunc(struct lp_build_sample_context *bld, */ 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, @@ -754,17 +755,21 @@ lp_build_sample_image_nearest(struct lp_build_sample_context *bld, lp_build_name(z, "tex.z.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) { - z = coords[2]; + if (has_layer_coord(bld->static_texture_state->target)) { + if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) { + /* add cube layer to face */ + z = lp_build_add(&bld->int_coord_bld, coords[2], coords[3]); + } + else { + z = coords[2]; + } lp_build_name(z, "tex.z.layer"); } /* * 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, @@ -824,14 +829,23 @@ lp_build_masklerp2d(struct lp_build_context *bld, 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, + boolean is_gather, LLVMValueRef size, + LLVMValueRef linear_mask, LLVMValueRef row_stride_vec, LLVMValueRef img_stride_vec, LLVMValueRef data_ptr, @@ -840,6 +854,10 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, 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; + struct lp_build_context *texel_bld = &bld->texel_bld; const unsigned dims = bld->dims; LLVMValueRef width_vec; LLVMValueRef height_vec; @@ -848,10 +866,30 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, 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_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) && + bld->static_sampler_state->seamless_cube_map; + /* + * XXX I don't know how this is really supposed to work with gather. From GL + * spec wording (not gather specific) it sounds like the 4th missing texel + * should be an average of the other 3, hence for gather could return this. + * This is however NOT how the code here works, which just fixes up the + * weights used for filtering instead. And of course for gather there is + * no filter to tweak... + */ + accurate_cube_corners = ACCURATE_CUBE_CORNERS && seamless_cube_filter && + !is_gather; lp_build_extract_image_sizes(bld, &bld->int_size_bld, @@ -870,62 +908,277 @@ lp_build_sample_image_linear(struct lp_build_sample_context *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 (has_layer_coord(bld->static_texture_state->target)) { + if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) { + /* add cube layer to face */ + z00 = z01 = z10 = z11 = z1 = + lp_build_add(&bld->int_coord_bld, coords[2], coords[3]); + } + else { + 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); + + if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) { + /* now can add cube layer to face (per sample) */ + z00 = lp_build_add(ivec_bld, z00, coords[3]); + z01 = lp_build_add(ivec_bld, z01, coords[3]); + z10 = lp_build_add(ivec_bld, z10, coords[3]); + z11 = lp_build_add(ivec_bld, z11, coords[3]); + } + + 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]); + if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) { + LLVMValueRef cube_layer = lp_build_add(ivec_bld, face, coords[3]); + LLVMBuildStore(builder, cube_layer, zs[0]); + LLVMBuildStore(builder, cube_layer, zs[1]); + LLVMBuildStore(builder, cube_layer, zs[2]); + LLVMBuildStore(builder, cube_layer, zs[3]); + } + else { + 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]); if (dims == 1) { + assert(!is_gather); if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) { /* Interpolate two samples from 1D image to produce one color */ for (chan = 0; chan < 4; chan++) { - colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart, + colors_out[chan] = lp_build_lerp(texel_bld, s_fpart, neighbors[0][0][chan], neighbors[0][1][chan], 0); @@ -936,37 +1189,165 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, cmpval0 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]); cmpval1 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]); /* simplified lerp, AND mask with weight and add */ - colors_out[0] = lp_build_masklerp(&bld->texel_bld, s_fpart, + colors_out[0] = lp_build_masklerp(texel_bld, s_fpart, cmpval0, cmpval1); colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0]; } } 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[chan] = lp_build_lerp_2d(&bld->texel_bld, - s_fpart, t_fpart, - neighbors[0][0][chan], - neighbors[0][1][chan], - neighbors[1][0][chan], - neighbors[1][1][chan], - 0); + if (is_gather) { + /* + * Just assign the red channel (no component selection yet). + * This is a bit hackish, we usually do the swizzle at the + * end of sampling (much less values to swizzle), but this + * obviously cannot work when using gather. + */ + unsigned chan_swiz = bld->static_texture_state->swizzle_r; + colors0[0] = lp_build_swizzle_soa_channel(texel_bld, + neighbors[1][0], + chan_swiz); + colors0[1] = lp_build_swizzle_soa_channel(texel_bld, + neighbors[1][1], + chan_swiz); + colors0[2] = lp_build_swizzle_soa_channel(texel_bld, + neighbors[0][1], + chan_swiz); + colors0[3] = lp_build_swizzle_soa_channel(texel_bld, + neighbors[0][0], + chan_swiz); + } + else { + /* Bilinear interpolate the four samples from the 2D image / 3D slice */ + for (chan = 0; chan < 4; chan++) { + colors0[chan] = lp_build_lerp_2d(texel_bld, + s_fpart, t_fpart, + neighbors[0][0][chan], + neighbors[0][1][chan], + neighbors[1][0][chan], + neighbors[1][1][chan], + 0); + } } } else { @@ -975,41 +1356,82 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, 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]); - colors0[0] = lp_build_masklerp2d(&bld->texel_bld, s_fpart, t_fpart, - cmpval00, cmpval01, cmpval10, cmpval11); - colors0[1] = colors0[2] = colors0[3] = colors0[0]; + + if (is_gather) { + /* more hacks for swizzling, should be X, ONE or ZERO... */ + unsigned chan_swiz = bld->static_texture_state->swizzle_r; + if (chan_swiz <= PIPE_SWIZZLE_ALPHA) { + colors0[0] = lp_build_select(texel_bld, cmpval10, + texel_bld->one, texel_bld->zero); + colors0[1] = lp_build_select(texel_bld, cmpval11, + texel_bld->one, texel_bld->zero); + colors0[2] = lp_build_select(texel_bld, cmpval01, + texel_bld->one, texel_bld->zero); + colors0[3] = lp_build_select(texel_bld, cmpval00, + texel_bld->one, texel_bld->zero); + } + else if (chan_swiz == PIPE_SWIZZLE_ZERO) { + colors0[0] = colors0[1] = colors0[2] = colors0[3] = + texel_bld->zero; + } + else { + colors0[0] = colors0[1] = colors0[2] = colors0[3] = + texel_bld->one; + } + } + else { + colors0[0] = lp_build_masklerp2d(texel_bld, s_fpart, t_fpart, + cmpval00, cmpval01, cmpval10, cmpval11); + 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]; + assert(!is_gather); + /* 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]); if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) { /* Bilinear interpolate the four samples from the second Z slice */ for (chan = 0; chan < 4; chan++) { - colors1[chan] = lp_build_lerp_2d(&bld->texel_bld, + colors1[chan] = lp_build_lerp_2d(texel_bld, s_fpart, t_fpart, neighbors1[0][0][chan], neighbors1[0][1][chan], @@ -1019,7 +1441,7 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, } /* Linearly interpolate the two samples from the two 3D slices */ for (chan = 0; chan < 4; chan++) { - colors_out[chan] = lp_build_lerp(&bld->texel_bld, + colors_out[chan] = lp_build_lerp(texel_bld, r_fpart, colors0[chan], colors1[chan], 0); @@ -1031,13 +1453,13 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, 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]); - colors1[0] = lp_build_masklerp2d(&bld->texel_bld, s_fpart, t_fpart, + colors1[0] = lp_build_masklerp2d(texel_bld, s_fpart, t_fpart, cmpval00, cmpval01, cmpval10, cmpval11); /* Linearly interpolate the two samples from the two 3D slices */ - colors_out[0] = lp_build_lerp(&bld->texel_bld, - r_fpart, - colors0[0], colors1[0], - 0); + colors_out[0] = lp_build_lerp(texel_bld, + r_fpart, + colors0[0], colors1[0], + 0); colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0]; } } @@ -1053,15 +1475,15 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, /** * 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, + boolean is_gather, LLVMValueRef *coords, const LLVMValueRef *offsets, LLVMValueRef ilevel0, @@ -1087,7 +1509,7 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, 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 { @@ -1096,16 +1518,14 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, 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, is_gather, size0, NULL, row_stride0_vec, img_stride0_vec, data_ptr0, mipoff0, coords, offsets, colors0); @@ -1123,7 +1543,7 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, /* 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 { @@ -1131,28 +1551,28 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, * 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 { @@ -1160,15 +1580,13 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, 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, FALSE, size1, NULL, row_stride1_vec, img_stride1_vec, data_ptr1, mipoff1, coords, offsets, colors1); @@ -1178,7 +1596,124 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, 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, FALSE, 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, FALSE, 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); @@ -1202,17 +1737,19 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, static LLVMValueRef lp_build_layer_coord(struct lp_build_sample_context *bld, unsigned texture_unit, + boolean is_cube_array, LLVMValueRef layer, LLVMValueRef *out_of_bounds) { LLVMValueRef num_layers; struct lp_build_context *int_coord_bld = &bld->int_coord_bld; - num_layers = bld->dynamic_state->depth(bld->dynamic_state, - bld->gallivm, texture_unit); + num_layers = bld->dynamic_state->depth(bld->dynamic_state, bld->gallivm, + bld->context_ptr, texture_unit); if (out_of_bounds) { LLVMValueRef out1, out; + assert(!is_cube_array); num_layers = lp_build_broadcast_scalar(int_coord_bld, num_layers); out = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, layer, int_coord_bld->zero); out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, layer, num_layers); @@ -1221,7 +1758,9 @@ lp_build_layer_coord(struct lp_build_sample_context *bld, } else { LLVMValueRef maxlayer; - maxlayer = lp_build_sub(&bld->int_bld, num_layers, bld->int_bld.one); + LLVMValueRef s = is_cube_array ? lp_build_const_int32(bld->gallivm, 6) : + bld->int_bld.one; + maxlayer = lp_build_sub(&bld->int_bld, num_layers, s); maxlayer = lp_build_broadcast_scalar(int_coord_bld, maxlayer); return lp_build_clamp(int_coord_bld, layer, int_coord_bld->zero, maxlayer); } @@ -1239,7 +1778,7 @@ lp_build_sample_common(struct lp_build_sample_context *bld, 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) @@ -1249,6 +1788,8 @@ lp_build_sample_common(struct lp_build_sample_context *bld, 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__, @@ -1259,18 +1800,27 @@ lp_build_sample_common(struct lp_build_sample_context *bld, * Choose cube face, recompute texcoords for the chosen face and * compute rho here too (as it requires transform of derivatives). */ - if (target == PIPE_TEXTURE_CUBE) { + if (target == PIPE_TEXTURE_CUBE || target == PIPE_TEXTURE_CUBE_ARRAY) { boolean need_derivs; need_derivs = ((min_filter != mag_filter || 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; + if (target == PIPE_TEXTURE_CUBE_ARRAY) { + /* calculate cube layer coord now */ + LLVMValueRef layer = lp_build_iround(&bld->coord_bld, coords[3]); + LLVMValueRef six = lp_build_const_int_vec(bld->gallivm, bld->int_coord_type, 6); + layer = lp_build_mul(&bld->int_coord_bld, layer, six); + coords[3] = lp_build_layer_coord(bld, texture_index, TRUE, layer, NULL); + /* because of seamless filtering can't add it to face (coords[2]) here. */ + } } else if (target == PIPE_TEXTURE_1D_ARRAY || target == PIPE_TEXTURE_2D_ARRAY) { coords[2] = lp_build_iround(&bld->coord_bld, coords[2]); - coords[2] = lp_build_layer_coord(bld, texture_index, coords[2], NULL); + coords[2] = lp_build_layer_coord(bld, texture_index, FALSE, coords[2], NULL); } if (bld->static_sampler_state->compare_mode != PIPE_TEX_COMPARE_NONE) { @@ -1309,9 +1859,16 @@ lp_build_sample_common(struct lp_build_sample_context *bld, 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); } /* @@ -1323,30 +1880,21 @@ lp_build_sample_common(struct lp_build_sample_context *bld, /* fall-through */ case PIPE_TEX_MIPFILTER_NONE: /* always use mip level 0 */ - if (HAVE_LLVM == 0x0207 && target == PIPE_TEXTURE_CUBE) { - /* XXX this is a work-around for an apparent bug in LLVM 2.7. - * 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); - } - else { - first_level = bld->dynamic_state->first_level(bld->dynamic_state, - bld->gallivm, texture_index); - first_level = lp_build_broadcast_scalar(&bld->leveli_bld, first_level); - *ilevel0 = first_level; - } + first_level = bld->dynamic_state->first_level(bld->dynamic_state, + bld->gallivm, bld->context_ptr, + texture_index); + first_level = lp_build_broadcast_scalar(&bld->leveli_bld, first_level); + *ilevel0 = first_level; 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; } @@ -1359,8 +1907,8 @@ lp_build_clamp_border_color(struct lp_build_sample_context *bld, struct gallivm_state *gallivm = bld->gallivm; LLVMBuilderRef builder = gallivm->builder; LLVMValueRef border_color_ptr = - bld->dynamic_state->border_color(bld->dynamic_state, - gallivm, sampler_unit); + bld->dynamic_state->border_color(bld->dynamic_state, gallivm, + bld->context_ptr, sampler_unit); LLVMValueRef border_color; const struct util_format_description *format_desc = bld->format_desc; struct lp_type vec4_type = bld->texel_type; @@ -1423,7 +1971,7 @@ lp_build_clamp_border_color(struct lp_build_sample_context *bld, * Border color was stored as int, hence need min/max clamp * only if chan has less than 32 bits.. */ - unsigned chan_size = format_desc->channel[chan].size < 32; + unsigned chan_size = format_desc->channel[chan].size; if (chan_size < 32) { min_clamp = lp_build_const_int_vec(gallivm, vec4_type, 0 - (1 << (chan_size - 1))); @@ -1451,7 +1999,7 @@ lp_build_clamp_border_color(struct lp_build_sample_context *bld, * Border color was stored as uint, hence never need min * clamp, and only need max clamp if chan has less than 32 bits. */ - unsigned chan_size = format_desc->channel[chan].size < 32; + unsigned chan_size = format_desc->channel[chan].size; if (chan_size < 32) { max_clamp = lp_build_const_int_vec(gallivm, vec4_type, (1 << chan_size) - 1); @@ -1579,15 +2127,15 @@ lp_build_clamp_border_color(struct lp_build_sample_context *bld, static void lp_build_sample_general(struct lp_build_sample_context *bld, unsigned sampler_unit, + boolean is_gather, 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; @@ -1623,55 +2171,99 @@ lp_build_sample_general(struct lp_build_sample_context *bld, 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, + is_gather, 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, FALSE, + 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, + FALSE, + 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, FALSE, + 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) { @@ -1697,7 +2289,7 @@ lp_build_fetch_texel(struct lp_build_sample_context *bld, 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; @@ -1711,7 +2303,7 @@ lp_build_fetch_texel(struct lp_build_sample_context *bld, 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); } @@ -1722,10 +2314,10 @@ lp_build_fetch_texel(struct lp_build_sample_context *bld, 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); + ilevel = bld->dynamic_state->first_level(bld->dynamic_state, bld->gallivm, + bld->context_ptr, texture_unit); } else { ilevel = lp_build_const_int32(bld->gallivm, 0); @@ -1740,11 +2332,11 @@ lp_build_fetch_texel(struct lp_build_sample_context *bld, if (target == PIPE_TEXTURE_1D_ARRAY || target == PIPE_TEXTURE_2D_ARRAY) { if (out_of_bound_ret_zero) { - z = lp_build_layer_coord(bld, texture_unit, z, &out1); + z = lp_build_layer_coord(bld, texture_unit, FALSE, z, &out1); out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1); } else { - z = lp_build_layer_coord(bld, texture_unit, z, NULL); + z = lp_build_layer_coord(bld, texture_unit, FALSE, z, NULL); } } @@ -1798,6 +2390,7 @@ lp_build_fetch_texel(struct lp_build_sample_context *bld, bld->texel_type, bld->base_ptr, offset, i, j, + bld->cache, colors_out); if (out_of_bound_ret_zero) { @@ -1834,50 +2427,98 @@ lp_build_sample_nop(struct gallivm_state *gallivm, /** - * Build texture sampling code. + * Build the actual texture sampling code. * 'texel' will return a vector of four LLVMValueRefs corresponding to * R, G, B, A. * \param type vector float type to use for coords, etc. - * \param is_fetch if this is a texel fetch instruction. + * \param sample_key * \param derivs partial derivatives of (s,t,r,q) with respect to x and y */ -void -lp_build_sample_soa(struct gallivm_state *gallivm, - const struct lp_static_texture_state *static_texture_state, - const struct lp_static_sampler_state *static_sampler_state, - struct lp_sampler_dynamic_state *dynamic_state, - struct lp_type type, - boolean is_fetch, - unsigned texture_index, - unsigned sampler_index, - const LLVMValueRef *coords, - const LLVMValueRef *offsets, - const struct lp_derivatives *derivs, /* optional */ - LLVMValueRef lod_bias, /* optional */ - LLVMValueRef explicit_lod, /* optional */ - enum lp_sampler_lod_property lod_property, - LLVMValueRef texel_out[4]) +static void +lp_build_sample_soa_code(struct gallivm_state *gallivm, + const struct lp_static_texture_state *static_texture_state, + const struct lp_static_sampler_state *static_sampler_state, + struct lp_sampler_dynamic_state *dynamic_state, + struct lp_type type, + unsigned sample_key, + unsigned texture_index, + unsigned sampler_index, + LLVMValueRef context_ptr, + LLVMValueRef thread_data_ptr, + const LLVMValueRef *coords, + const LLVMValueRef *offsets, + const struct lp_derivatives *derivs, /* optional */ + LLVMValueRef lod, /* optional */ + LLVMValueRef texel_out[4]) { 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); LLVMBuilderRef builder = gallivm->builder; LLVMValueRef tex_width, newcoords[5]; + enum lp_sampler_lod_property lod_property; + enum lp_sampler_lod_control lod_control; + enum lp_sampler_op_type op_type; + LLVMValueRef lod_bias = NULL; + LLVMValueRef explicit_lod = NULL; + boolean op_is_tex; if (0) { enum pipe_format fmt = static_texture_state->format; debug_printf("Sample from %s\n", util_format_name(fmt)); } + lod_property = (sample_key & LP_SAMPLER_LOD_PROPERTY_MASK) >> + LP_SAMPLER_LOD_PROPERTY_SHIFT; + lod_control = (sample_key & LP_SAMPLER_LOD_CONTROL_MASK) >> + LP_SAMPLER_LOD_CONTROL_SHIFT; + op_type = (sample_key & LP_SAMPLER_OP_TYPE_MASK) >> + LP_SAMPLER_OP_TYPE_SHIFT; + + op_is_tex = op_type == LP_SAMPLER_OP_TEXTURE; + + if (lod_control == LP_SAMPLER_LOD_BIAS) { + lod_bias = lod; + assert(lod); + assert(derivs == NULL); + } + else if (lod_control == LP_SAMPLER_LOD_EXPLICIT) { + explicit_lod = lod; + assert(lod); + assert(derivs == NULL); + } + else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) { + assert(derivs); + assert(lod == NULL); + } + else { + assert(derivs == NULL); + assert(lod == NULL); + } + + 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_zero(gallivm, type); + for (chan = 0; chan < 4; chan++) { + texel_out[chan] = zero; + } + return; + } + assert(type.floating); /* Setup our build context */ memset(&bld, 0, sizeof bld); bld.gallivm = gallivm; + bld.context_ptr = context_ptr; bld.static_sampler_state = &derived_sampler_state; bld.static_texture_state = static_texture_state; bld.dynamic_state = dynamic_state; @@ -1918,12 +2559,47 @@ lp_build_sample_soa(struct gallivm_state *gallivm, } else { derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE; } + if (op_type == LP_SAMPLER_OP_GATHER) { + /* + * gather4 is exactly like GL_LINEAR filtering but in the end skipping + * the actual filtering. Using mostly the same paths, so cube face + * selection, coord wrapping etc. all naturally uses the same code. + */ + derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE; + derived_sampler_state.min_img_filter = PIPE_TEX_FILTER_LINEAR; + derived_sampler_state.mag_img_filter = PIPE_TEX_FILTER_LINEAR; + } mip_filter = derived_sampler_state.min_mip_filter; if (0) { 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; + } + /* + * We could force CLAMP to CLAMP_TO_EDGE here if min/mag filter is nearest, + * so AoS path could be used. Not sure it's worth the trouble... + */ + + 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. @@ -1942,31 +2618,66 @@ lp_build_sample_soa(struct gallivm_state *gallivm, * There are other situations where at least the multiple int lods could be * avoided like min and max lod being equal. */ - if (explicit_lod && lod_property == LP_SAMPLER_LOD_PER_ELEMENT && - ((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 || + static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) && + (op_is_tex && 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 ((!op_is_tex && target != PIPE_BUFFER) || + (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) { + bld.num_mips = type.length; + bld.num_lods = type.length; + } + else if (op_is_tex && 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 ) || - (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) { + else if ((!op_is_tex && explicit_lod && target != PIPE_BUFFER) || + (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) { + bld.num_mips = num_quads; bld.num_lods = num_quads; } - else { - bld.num_lods = 1; + else if (op_is_tex && 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); @@ -1983,32 +2694,49 @@ lp_build_sample_soa(struct gallivm_state *gallivm, 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); - bld.row_stride_array = dynamic_state->row_stride(dynamic_state, gallivm, texture_index); - bld.img_stride_array = dynamic_state->img_stride(dynamic_state, gallivm, texture_index); - bld.base_ptr = dynamic_state->base_ptr(dynamic_state, gallivm, texture_index); - bld.mip_offsets = dynamic_state->mip_offsets(dynamic_state, gallivm, texture_index); + tex_width = dynamic_state->width(dynamic_state, gallivm, + context_ptr, texture_index); + bld.row_stride_array = dynamic_state->row_stride(dynamic_state, gallivm, + context_ptr, texture_index); + bld.img_stride_array = dynamic_state->img_stride(dynamic_state, gallivm, + context_ptr, texture_index); + bld.base_ptr = dynamic_state->base_ptr(dynamic_state, gallivm, + context_ptr, texture_index); + bld.mip_offsets = dynamic_state->mip_offsets(dynamic_state, gallivm, + context_ptr, texture_index); /* Note that mip_offsets is an array[level] of offsets to texture images */ + if (dynamic_state->cache_ptr && thread_data_ptr) { + bld.cache = dynamic_state->cache_ptr(dynamic_state, gallivm, + thread_data_ptr, texture_index); + } + /* width, height, depth as single int vector */ if (dims <= 1) { bld.int_size = tex_width; } else { bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef, - tex_width, LLVMConstInt(i32t, 0, 0), ""); + tex_width, + LLVMConstInt(i32t, 0, 0), ""); if (dims >= 2) { LLVMValueRef tex_height = - dynamic_state->height(dynamic_state, gallivm, texture_index); + dynamic_state->height(dynamic_state, gallivm, + context_ptr, texture_index); bld.int_size = LLVMBuildInsertElement(builder, bld.int_size, - tex_height, LLVMConstInt(i32t, 1, 0), ""); + tex_height, + LLVMConstInt(i32t, 1, 0), ""); if (dims >= 3) { LLVMValueRef tex_depth = - dynamic_state->depth(dynamic_state, gallivm, texture_index); + dynamic_state->depth(dynamic_state, gallivm, context_ptr, + texture_index); bld.int_size = LLVMBuildInsertElement(builder, bld.int_size, - tex_depth, LLVMConstInt(i32t, 2, 0), ""); + tex_depth, + LLVMConstInt(i32t, 2, 0), ""); } } } @@ -2025,77 +2753,116 @@ lp_build_sample_soa(struct gallivm_state *gallivm, texel_out); } - else if (is_fetch) { + else if (op_type == LP_SAMPLER_OP_FETCH) { lp_build_fetch_texel(&bld, texture_index, newcoords, - explicit_lod, offsets, + lod, offsets, texel_out); } 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); + boolean use_aos; + + if (util_format_is_pure_integer(static_texture_state->format) && + !util_format_has_depth(bld.format_desc) && + (static_sampler_state->min_mip_filter == PIPE_TEX_MIPFILTER_LINEAR || + static_sampler_state->min_img_filter == PIPE_TEX_FILTER_LINEAR || + static_sampler_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR)) { + /* + * Bail if impossible filtering is specified (the awkard additional + * depth check is because it is legal in gallium to have things like S8Z24 + * here which would say it's pure int despite such formats should sample + * the depth component). + * In GL such filters make the texture incomplete, this makes it robust + * against state trackers which set this up regardless (we'd crash in the + * lerp later (except for gather)). + * Must do this after fetch_texel code since with GL state tracker we'll + * get some junk sampler for buffer textures. + */ + unsigned chan; + LLVMValueRef zero = lp_build_zero(gallivm, type); + for (chan = 0; chan < 4; chan++) { + texel_out[chan] = zero; + } + return; + } + + use_aos = util_format_fits_8unorm(bld.format_desc) && + op_is_tex && + /* not sure this is strictly needed or simply impossible */ + 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 || + static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) && + 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); + if (use_aos && static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) { + /* The aos path doesn't do seamless filtering so simply add cube layer + * to face now. + */ + newcoords[2] = lp_build_add(&bld.int_coord_bld, newcoords[2], newcoords[3]); + } + /* * 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, + op_type == LP_SAMPLER_OP_GATHER, newcoords, offsets, - lod_ipart, lod_fpart, + lod_positive, lod_fpart, ilevel0, ilevel1, texel_out); } @@ -2113,6 +2880,7 @@ lp_build_sample_soa(struct gallivm_state *gallivm, /* Setup our build context */ memset(&bld4, 0, sizeof bld4); bld4.gallivm = bld.gallivm; + bld4.context_ptr = bld.context_ptr; bld4.static_texture_state = bld.static_texture_state; bld4.static_sampler_state = bld.static_sampler_state; bld4.dynamic_state = bld.dynamic_state; @@ -2123,6 +2891,7 @@ lp_build_sample_soa(struct gallivm_state *gallivm, bld4.base_ptr = bld.base_ptr; bld4.mip_offsets = bld.mip_offsets; bld4.int_size = bld.int_size; + bld4.cache = bld.cache; bld4.vector_width = lp_type_width(type4); @@ -2135,28 +2904,44 @@ lp_build_sample_soa(struct gallivm_state *gallivm, 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 (explicit_lod && lod_property == LP_SAMPLER_LOD_PER_ELEMENT && - ((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 || + static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) && + (op_is_tex && 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 ((!op_is_tex && target != PIPE_BUFFER) || + (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) { + bld4.num_mips = type4.length; + bld4.num_lods = type4.length; + } + else if (op_is_tex && 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); @@ -2173,10 +2958,12 @@ lp_build_sample_soa(struct gallivm_state *gallivm, 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; @@ -2194,8 +2981,9 @@ lp_build_sample_soa(struct gallivm_state *gallivm, } } } - 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); @@ -2205,7 +2993,7 @@ lp_build_sample_soa(struct gallivm_state *gallivm, /* 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); } @@ -2220,8 +3008,9 @@ lp_build_sample_soa(struct gallivm_state *gallivm, newcoords4[4] = lp_build_extract_range(gallivm, newcoords[4], 4*i, 4); lp_build_sample_general(&bld4, sampler_index, + op_type == LP_SAMPLER_OP_GATHER, newcoords4, offsets4, - lod_ipart4, lod_fpart4, + lod_positive4, lod_fpart4, ilevel04, ilevel14, texelout4); } @@ -2236,7 +3025,7 @@ lp_build_sample_soa(struct gallivm_state *gallivm, } } - if (target != PIPE_BUFFER) { + if (target != PIPE_BUFFER && op_type != LP_SAMPLER_OP_GATHER) { apply_sampler_swizzle(&bld, texel_out); } @@ -2253,6 +3042,399 @@ lp_build_sample_soa(struct gallivm_state *gallivm, } } + +#define USE_TEX_FUNC_CALL 1 + +#define LP_MAX_TEX_FUNC_ARGS 32 + +static inline void +get_target_info(enum pipe_texture_target target, + unsigned *num_coords, unsigned *num_derivs, + unsigned *num_offsets, unsigned *layer) +{ + unsigned dims = texture_dims(target); + *num_coords = dims; + *num_offsets = dims; + *num_derivs = (target == PIPE_TEXTURE_CUBE || + target == PIPE_TEXTURE_CUBE_ARRAY) ? 3 : dims; + *layer = has_layer_coord(target) ? 2: 0; + if (target == PIPE_TEXTURE_CUBE_ARRAY) { + /* + * dims doesn't include r coord for cubes - this is handled + * by layer instead, but need to fix up for cube arrays... + */ + *layer = 3; + *num_coords = 3; + } +} + + +/** + * Generate the function body for a texture sampling function. + */ +static void +lp_build_sample_gen_func(struct gallivm_state *gallivm, + const struct lp_static_texture_state *static_texture_state, + const struct lp_static_sampler_state *static_sampler_state, + struct lp_sampler_dynamic_state *dynamic_state, + struct lp_type type, + unsigned texture_index, + unsigned sampler_index, + LLVMValueRef function, + unsigned num_args, + unsigned sample_key) +{ + LLVMBuilderRef old_builder; + LLVMBasicBlockRef block; + LLVMValueRef coords[5]; + LLVMValueRef offsets[3] = { NULL }; + LLVMValueRef lod = NULL; + LLVMValueRef context_ptr; + LLVMValueRef thread_data_ptr = NULL; + LLVMValueRef texel_out[4]; + struct lp_derivatives derivs; + struct lp_derivatives *deriv_ptr = NULL; + unsigned num_param = 0; + unsigned i, num_coords, num_derivs, num_offsets, layer; + enum lp_sampler_lod_control lod_control; + boolean need_cache = FALSE; + + lod_control = (sample_key & LP_SAMPLER_LOD_CONTROL_MASK) >> + LP_SAMPLER_LOD_CONTROL_SHIFT; + + get_target_info(static_texture_state->target, + &num_coords, &num_derivs, &num_offsets, &layer); + + if (dynamic_state->cache_ptr) { + const struct util_format_description *format_desc; + format_desc = util_format_description(static_texture_state->format); + if (format_desc && format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) { + need_cache = TRUE; + } + } + + /* "unpack" arguments */ + context_ptr = LLVMGetParam(function, num_param++); + if (need_cache) { + thread_data_ptr = LLVMGetParam(function, num_param++); + } + for (i = 0; i < num_coords; i++) { + coords[i] = LLVMGetParam(function, num_param++); + } + for (i = num_coords; i < 5; i++) { + /* This is rather unfortunate... */ + coords[i] = lp_build_undef(gallivm, type); + } + if (layer) { + coords[layer] = LLVMGetParam(function, num_param++); + } + if (sample_key & LP_SAMPLER_SHADOW) { + coords[4] = LLVMGetParam(function, num_param++); + } + if (sample_key & LP_SAMPLER_OFFSETS) { + for (i = 0; i < num_offsets; i++) { + offsets[i] = LLVMGetParam(function, num_param++); + } + } + if (lod_control == LP_SAMPLER_LOD_BIAS || + lod_control == LP_SAMPLER_LOD_EXPLICIT) { + lod = LLVMGetParam(function, num_param++); + } + else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) { + for (i = 0; i < num_derivs; i++) { + derivs.ddx[i] = LLVMGetParam(function, num_param++); + derivs.ddy[i] = LLVMGetParam(function, num_param++); + } + deriv_ptr = &derivs; + } + + assert(num_args == num_param); + + /* + * Function body + */ + + old_builder = gallivm->builder; + block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry"); + gallivm->builder = LLVMCreateBuilderInContext(gallivm->context); + LLVMPositionBuilderAtEnd(gallivm->builder, block); + + lp_build_sample_soa_code(gallivm, + static_texture_state, + static_sampler_state, + dynamic_state, + type, + sample_key, + texture_index, + sampler_index, + context_ptr, + thread_data_ptr, + coords, + offsets, + deriv_ptr, + lod, + texel_out); + + LLVMBuildAggregateRet(gallivm->builder, texel_out, 4); + + LLVMDisposeBuilder(gallivm->builder); + gallivm->builder = old_builder; + + gallivm_verify_function(gallivm, function); +} + + +/** + * Call the matching function for texture sampling. + * If there's no match, generate a new one. + */ +static void +lp_build_sample_soa_func(struct gallivm_state *gallivm, + const struct lp_static_texture_state *static_texture_state, + const struct lp_static_sampler_state *static_sampler_state, + struct lp_sampler_dynamic_state *dynamic_state, + const struct lp_sampler_params *params) +{ + LLVMBuilderRef builder = gallivm->builder; + LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent( + LLVMGetInsertBlock(builder))); + LLVMValueRef function, inst; + LLVMValueRef args[LP_MAX_TEX_FUNC_ARGS]; + LLVMBasicBlockRef bb; + LLVMValueRef tex_ret; + unsigned num_args = 0; + char func_name[64]; + unsigned i, num_coords, num_derivs, num_offsets, layer; + unsigned texture_index = params->texture_index; + unsigned sampler_index = params->sampler_index; + unsigned sample_key = params->sample_key; + const LLVMValueRef *coords = params->coords; + const LLVMValueRef *offsets = params->offsets; + const struct lp_derivatives *derivs = params->derivs; + enum lp_sampler_lod_control lod_control; + boolean need_cache = FALSE; + + lod_control = (sample_key & LP_SAMPLER_LOD_CONTROL_MASK) >> + LP_SAMPLER_LOD_CONTROL_SHIFT; + + get_target_info(static_texture_state->target, + &num_coords, &num_derivs, &num_offsets, &layer); + + if (dynamic_state->cache_ptr) { + const struct util_format_description *format_desc; + format_desc = util_format_description(static_texture_state->format); + if (format_desc && format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) { + /* + * This is not 100% correct, if we have cache but the + * util_format_s3tc_prefer is true the cache won't get used + * regardless (could hook up the block decode there...) */ + need_cache = TRUE; + } + } + /* + * texture function matches are found by name. + * Thus the name has to include both the texture and sampler unit + * (which covers all static state) plus the actual texture function + * (including things like offsets, shadow coord, lod control). + * Additionally lod_property has to be included too. + */ + + util_snprintf(func_name, sizeof(func_name), "texfunc_res_%d_sam_%d_%x", + texture_index, sampler_index, sample_key); + + function = LLVMGetNamedFunction(module, func_name); + + if(!function) { + LLVMTypeRef arg_types[LP_MAX_TEX_FUNC_ARGS]; + LLVMTypeRef ret_type; + LLVMTypeRef function_type; + LLVMTypeRef val_type[4]; + unsigned num_param = 0; + + /* + * Generate the function prototype. + */ + + arg_types[num_param++] = LLVMTypeOf(params->context_ptr); + if (need_cache) { + arg_types[num_param++] = LLVMTypeOf(params->thread_data_ptr); + } + for (i = 0; i < num_coords; i++) { + arg_types[num_param++] = LLVMTypeOf(coords[0]); + assert(LLVMTypeOf(coords[0]) == LLVMTypeOf(coords[i])); + } + if (layer) { + arg_types[num_param++] = LLVMTypeOf(coords[layer]); + assert(LLVMTypeOf(coords[0]) == LLVMTypeOf(coords[layer])); + } + if (sample_key & LP_SAMPLER_SHADOW) { + arg_types[num_param++] = LLVMTypeOf(coords[0]); + } + if (sample_key & LP_SAMPLER_OFFSETS) { + for (i = 0; i < num_offsets; i++) { + arg_types[num_param++] = LLVMTypeOf(offsets[0]); + assert(LLVMTypeOf(offsets[0]) == LLVMTypeOf(offsets[i])); + } + } + if (lod_control == LP_SAMPLER_LOD_BIAS || + lod_control == LP_SAMPLER_LOD_EXPLICIT) { + arg_types[num_param++] = LLVMTypeOf(params->lod); + } + else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) { + for (i = 0; i < num_derivs; i++) { + arg_types[num_param++] = LLVMTypeOf(derivs->ddx[i]); + arg_types[num_param++] = LLVMTypeOf(derivs->ddy[i]); + assert(LLVMTypeOf(derivs->ddx[0]) == LLVMTypeOf(derivs->ddx[i])); + assert(LLVMTypeOf(derivs->ddy[0]) == LLVMTypeOf(derivs->ddy[i])); + } + } + + val_type[0] = val_type[1] = val_type[2] = val_type[3] = + lp_build_vec_type(gallivm, params->type); + ret_type = LLVMStructTypeInContext(gallivm->context, val_type, 4, 0); + function_type = LLVMFunctionType(ret_type, arg_types, num_param, 0); + function = LLVMAddFunction(module, func_name, function_type); + + for (i = 0; i < num_param; ++i) { + if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) { + LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute); + } + } + + LLVMSetFunctionCallConv(function, LLVMFastCallConv); + LLVMSetLinkage(function, LLVMPrivateLinkage); + + lp_build_sample_gen_func(gallivm, + static_texture_state, + static_sampler_state, + dynamic_state, + params->type, + texture_index, + sampler_index, + function, + num_param, + sample_key); + } + + num_args = 0; + args[num_args++] = params->context_ptr; + if (need_cache) { + args[num_args++] = params->thread_data_ptr; + } + for (i = 0; i < num_coords; i++) { + args[num_args++] = coords[i]; + } + if (layer) { + args[num_args++] = coords[layer]; + } + if (sample_key & LP_SAMPLER_SHADOW) { + args[num_args++] = coords[4]; + } + if (sample_key & LP_SAMPLER_OFFSETS) { + for (i = 0; i < num_offsets; i++) { + args[num_args++] = offsets[i]; + } + } + if (lod_control == LP_SAMPLER_LOD_BIAS || + lod_control == LP_SAMPLER_LOD_EXPLICIT) { + args[num_args++] = params->lod; + } + else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) { + for (i = 0; i < num_derivs; i++) { + args[num_args++] = derivs->ddx[i]; + args[num_args++] = derivs->ddy[i]; + } + } + + assert(num_args <= LP_MAX_TEX_FUNC_ARGS); + + tex_ret = LLVMBuildCall(builder, function, args, num_args, ""); + bb = LLVMGetInsertBlock(builder); + inst = LLVMGetLastInstruction(bb); + LLVMSetInstructionCallConv(inst, LLVMFastCallConv); + + for (i = 0; i < 4; i++) { + params->texel[i] = LLVMBuildExtractValue(gallivm->builder, tex_ret, i, ""); + } +} + + +/** + * Build texture sampling code. + * Either via a function call or inline it directly. + */ +void +lp_build_sample_soa(const struct lp_static_texture_state *static_texture_state, + const struct lp_static_sampler_state *static_sampler_state, + struct lp_sampler_dynamic_state *dynamic_state, + struct gallivm_state *gallivm, + const struct lp_sampler_params *params) +{ + boolean use_tex_func = FALSE; + + /* + * Do not use a function call if the sampling is "simple enough". + * We define this by + * a) format + * b) no mips (either one level only or no mip filter) + * No mips will definitely make the code smaller, though + * the format requirement is a bit iffy - there's some (SoA) formats + * which definitely generate less code. This does happen to catch + * some important cases though which are hurt quite a bit by using + * a call (though not really because of the call overhead but because + * they are reusing the same texture unit with some of the same + * parameters). + * Ideally we'd let llvm recognize this stuff by doing IPO passes. + */ + + if (USE_TEX_FUNC_CALL) { + const struct util_format_description *format_desc; + boolean simple_format; + boolean simple_tex; + enum lp_sampler_op_type op_type; + format_desc = util_format_description(static_texture_state->format); + simple_format = !format_desc || + (util_format_is_rgba8_variant(format_desc) && + format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB); + + op_type = (params->sample_key & LP_SAMPLER_OP_TYPE_MASK) >> + LP_SAMPLER_OP_TYPE_SHIFT; + simple_tex = + op_type != LP_SAMPLER_OP_TEXTURE || + ((static_sampler_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE || + static_texture_state->level_zero_only == TRUE) && + static_sampler_state->min_img_filter == static_sampler_state->mag_img_filter); + + use_tex_func = format_desc && !(simple_format && simple_tex); + } + + if (use_tex_func) { + lp_build_sample_soa_func(gallivm, + static_texture_state, + static_sampler_state, + dynamic_state, + params); + } + else { + lp_build_sample_soa_code(gallivm, + static_texture_state, + static_sampler_state, + dynamic_state, + params->type, + params->sample_key, + params->texture_index, + params->sampler_index, + params->context_ptr, + params->thread_data_ptr, + params->coords, + params->offsets, + params->derivs, + params->lod, + params->texel); + } +} + + void lp_build_size_query_soa(struct gallivm_state *gallivm, const struct lp_static_texture_state *static_state, @@ -2260,6 +3442,7 @@ lp_build_size_query_soa(struct gallivm_state *gallivm, struct lp_type int_type, unsigned texture_unit, unsigned target, + LLVMValueRef context_ptr, boolean is_sviewinfo, enum lp_sampler_lod_property lod_property, LLVMValueRef explicit_lod, @@ -2272,6 +3455,19 @@ lp_build_size_query_soa(struct gallivm_state *gallivm, 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 @@ -2304,6 +3500,7 @@ lp_build_size_query_soa(struct gallivm_state *gallivm, switch (target) { case PIPE_TEXTURE_1D_ARRAY: case PIPE_TEXTURE_2D_ARRAY: + case PIPE_TEXTURE_CUBE_ARRAY: has_array = TRUE; break; default: @@ -2317,8 +3514,10 @@ lp_build_size_query_soa(struct gallivm_state *gallivm, if (explicit_lod) { /* FIXME: this needs to honor per-element lod */ - lod = LLVMBuildExtractElement(gallivm->builder, explicit_lod, lp_build_const_int32(gallivm, 0), ""); - first_level = dynamic_state->first_level(dynamic_state, gallivm, texture_unit); + lod = LLVMBuildExtractElement(gallivm->builder, explicit_lod, + lp_build_const_int32(gallivm, 0), ""); + first_level = dynamic_state->first_level(dynamic_state, gallivm, + context_ptr, texture_unit); level = LLVMBuildAdd(gallivm->builder, lod, first_level, "level"); lod = lp_build_broadcast_scalar(&bld_int_vec4, level); } else { @@ -2328,27 +3527,41 @@ lp_build_size_query_soa(struct gallivm_state *gallivm, size = bld_int_vec4.undef; size = LLVMBuildInsertElement(gallivm->builder, size, - dynamic_state->width(dynamic_state, gallivm, texture_unit), + dynamic_state->width(dynamic_state, gallivm, + context_ptr, texture_unit), lp_build_const_int32(gallivm, 0), ""); if (dims >= 2) { size = LLVMBuildInsertElement(gallivm->builder, size, - dynamic_state->height(dynamic_state, gallivm, texture_unit), + dynamic_state->height(dynamic_state, gallivm, + context_ptr, texture_unit), lp_build_const_int32(gallivm, 1), ""); } if (dims >= 3) { size = LLVMBuildInsertElement(gallivm->builder, size, - dynamic_state->depth(dynamic_state, gallivm, texture_unit), + dynamic_state->depth(dynamic_state, gallivm, + context_ptr, texture_unit), 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, - dynamic_state->depth(dynamic_state, gallivm, texture_unit), + if (has_array) { + LLVMValueRef layers = dynamic_state->depth(dynamic_state, gallivm, + context_ptr, texture_unit); + if (target == PIPE_TEXTURE_CUBE_ARRAY) { + /* + * It looks like GL wants number of cubes, d3d10.1 has it undefined? + * Could avoid this by passing in number of cubes instead of total + * number of layers (might make things easier elsewhere too). + */ + LLVMValueRef six = lp_build_const_int32(gallivm, 6); + layers = LLVMBuildSDiv(gallivm->builder, layers, six, ""); + } + size = LLVMBuildInsertElement(gallivm->builder, size, layers, lp_build_const_int32(gallivm, dims), ""); + } /* * d3d10 requires zero for x/y/z values (but not w, i.e. mip levels) @@ -2361,7 +3574,8 @@ lp_build_size_query_soa(struct gallivm_state *gallivm, /* everything is scalar for now */ lp_build_context_init(&leveli_bld, gallivm, lp_type_int_vec(32, 32)); - last_level = dynamic_state->last_level(dynamic_state, gallivm, texture_unit); + last_level = dynamic_state->last_level(dynamic_state, gallivm, + context_ptr, texture_unit); out = lp_build_cmp(&leveli_bld, PIPE_FUNC_LESS, level, first_level); out1 = lp_build_cmp(&leveli_bld, PIPE_FUNC_GREATER, level, last_level); @@ -2401,7 +3615,8 @@ lp_build_size_query_soa(struct gallivm_state *gallivm, else { LLVMValueRef last_level; - last_level = dynamic_state->last_level(dynamic_state, gallivm, texture_unit); + last_level = dynamic_state->last_level(dynamic_state, gallivm, + context_ptr, texture_unit); num_levels = lp_build_sub(&bld_int_scalar, last_level, first_level); num_levels = lp_build_add(&bld_int_scalar, num_levels, bld_int_scalar.one); }