diff options
Diffstat (limited to 'src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c')
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c | 1337 |
1 files changed, 1041 insertions, 296 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c index 9058f76c1d..995c016b9d 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c @@ -48,6 +48,7 @@ #include "lp_bld_logic.h" #include "lp_bld_swizzle.h" #include "lp_bld_pack.h" +#include "lp_bld_flow.h" #include "lp_bld_format.h" #include "lp_bld_sample.h" @@ -65,6 +66,14 @@ struct lp_build_sample_context const struct util_format_description *format_desc; + /** regular scalar float type */ + struct lp_type float_type; + struct lp_build_context float_bld; + + /** regular scalar float type */ + struct lp_type int_type; + struct lp_build_context int_bld; + /** Incoming coordinates type and build context */ struct lp_type coord_type; struct lp_build_context coord_bld; @@ -108,9 +117,78 @@ wrap_mode_uses_border_color(unsigned mode) } +static LLVMValueRef +lp_build_get_mipmap_level(struct lp_build_sample_context *bld, + LLVMValueRef data_array, LLVMValueRef level) +{ + LLVMValueRef indexes[2], data_ptr; + indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0); + indexes[1] = level; + data_ptr = LLVMBuildGEP(bld->builder, data_array, indexes, 2, ""); + data_ptr = LLVMBuildLoad(bld->builder, data_ptr, ""); + return data_ptr; +} + + +static LLVMValueRef +lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld, + LLVMValueRef data_array, int level) +{ + LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0); + return lp_build_get_mipmap_level(bld, data_array, lvl); +} + + +/** + * Dereference stride_array[mipmap_level] array to get a stride. + * Return stride as a vector. + */ +static LLVMValueRef +lp_build_get_level_stride_vec(struct lp_build_sample_context *bld, + LLVMValueRef stride_array, LLVMValueRef level) +{ + LLVMValueRef indexes[2], stride; + indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0); + indexes[1] = level; + stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, ""); + stride = LLVMBuildLoad(bld->builder, stride, ""); + stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride); + return stride; +} + + +/** Dereference stride_array[0] array to get a stride (as vector). */ +static LLVMValueRef +lp_build_get_const_level_stride_vec(struct lp_build_sample_context *bld, + LLVMValueRef stride_array, int level) +{ + LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0); + return lp_build_get_level_stride_vec(bld, stride_array, lvl); +} + + +static int +texture_dims(enum pipe_texture_target tex) +{ + switch (tex) { + case PIPE_TEXTURE_1D: + return 1; + case PIPE_TEXTURE_2D: + case PIPE_TEXTURE_CUBE: + return 2; + case PIPE_TEXTURE_3D: + return 3; + default: + assert(0 && "bad texture target in texture_dims()"); + return 2; + } +} + + /** - * Gen code to fetch a texel from a texture at int coords (x, y). + * Generate code to fetch a texel from a texture at int coords (x, y, z). + * The computation depends on whether the texture is 1D, 2D or 3D. * The result, texel, will be: * texel[0] = red values * texel[1] = green values @@ -121,12 +199,16 @@ static void lp_build_sample_texel_soa(struct lp_build_sample_context *bld, LLVMValueRef width, LLVMValueRef height, + LLVMValueRef depth, LLVMValueRef x, LLVMValueRef y, + LLVMValueRef z, LLVMValueRef y_stride, + LLVMValueRef z_stride, LLVMValueRef data_ptr, LLVMValueRef *texel) { + const int dims = texture_dims(bld->static_state->target); struct lp_build_context *int_coord_bld = &bld->int_coord_bld; LLVMValueRef offset; LLVMValueRef packed; @@ -140,7 +222,7 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2"); } - if (wrap_mode_uses_border_color(bld->static_state->wrap_t)) { + if (dims >= 2 && wrap_mode_uses_border_color(bld->static_state->wrap_t)) { LLVMValueRef b1, b2; b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero); b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height); @@ -153,6 +235,19 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, } } + if (dims == 3 && wrap_mode_uses_border_color(bld->static_state->wrap_r)) { + LLVMValueRef b1, b2; + b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero); + b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth); + if (use_border) { + use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1"); + use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2"); + } + else { + use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2"); + } + } + /* * Note: if we find an app which frequently samples the texture border * we might want to implement a true conditional here to avoid sampling @@ -168,11 +263,10 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, * the texel color results with the border color. */ - /* convert x,y coords to linear offset from start of texture, in bytes */ + /* convert x,y,z coords to linear offset from start of texture, in bytes */ offset = lp_build_sample_offset(&bld->uint_coord_bld, bld->format_desc, - x, y, y_stride, - data_ptr); + x, y, z, y_stride, z_stride); assert(bld->format_desc->block.width == 1); assert(bld->format_desc->block.height == 1); @@ -185,6 +279,8 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, bld->texel_type.width, data_ptr, offset); + texel[0] = texel[1] = texel[2] = texel[3] = NULL; + /* convert texels to float rgba */ lp_build_unpack_rgba_soa(bld->builder, bld->format_desc, @@ -196,7 +292,7 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, int chan; for (chan = 0; chan < 4; chan++) { LLVMValueRef border_chan = - lp_build_const_scalar(bld->texel_type, + lp_build_const_vec(bld->texel_type, bld->static_state->border_color[chan]); texel[chan] = lp_build_select(&bld->texel_bld, use_border, border_chan, texel[chan]); @@ -210,19 +306,22 @@ lp_build_sample_packed(struct lp_build_sample_context *bld, LLVMValueRef x, LLVMValueRef y, LLVMValueRef y_stride, - LLVMValueRef data_ptr) + LLVMValueRef data_array) { LLVMValueRef offset; + LLVMValueRef data_ptr; offset = lp_build_sample_offset(&bld->uint_coord_bld, bld->format_desc, - x, y, y_stride, - data_ptr); + x, y, NULL, y_stride, NULL); assert(bld->format_desc->block.width == 1); assert(bld->format_desc->block.height == 1); assert(bld->format_desc->block.bits <= bld->texel_type.width); + /* get pointer to mipmap level 0 data */ + data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0); + return lp_build_gather(bld->builder, bld->texel_type.length, bld->format_desc->block.bits, @@ -358,8 +457,8 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, struct lp_build_context *coord_bld = &bld->coord_bld; struct lp_build_context *int_coord_bld = &bld->int_coord_bld; struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld; - LLVMValueRef two = lp_build_const_scalar(coord_bld->type, 2.0); - LLVMValueRef half = lp_build_const_scalar(coord_bld->type, 0.5); + LLVMValueRef two = lp_build_const_vec(coord_bld->type, 2.0); + LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5); LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length); LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one); LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one); @@ -413,7 +512,7 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, else { LLVMValueRef min, max; /* clamp to [0.5, length - 0.5] */ - min = lp_build_const_scalar(coord_bld->type, 0.5F); + min = lp_build_const_vec(coord_bld->type, 0.5F); max = lp_build_sub(coord_bld, length_f, min); coord = lp_build_clamp(coord_bld, coord, min, max); } @@ -434,7 +533,7 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, if (bld->static_state->normalized_coords) { /* min = -1.0 / (2 * length) = -0.5 / length */ min = lp_build_mul(coord_bld, - lp_build_const_scalar(coord_bld->type, -0.5F), + lp_build_const_vec(coord_bld->type, -0.5F), lp_build_rcp(coord_bld, length_f)); /* max = 1.0 - min */ max = lp_build_sub(coord_bld, coord_bld->one, min); @@ -446,7 +545,7 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, } else { /* clamp to [-0.5, length + 0.5] */ - min = lp_build_const_scalar(coord_bld->type, -0.5F); + min = lp_build_const_vec(coord_bld->type, -0.5F); max = lp_build_sub(coord_bld, length_f, min); coord = lp_build_clamp(coord_bld, coord, min, max); coord = lp_build_sub(coord_bld, coord, half); @@ -521,7 +620,7 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, LLVMValueRef min, max; /* min = -1.0 / (2 * length) = -0.5 / length */ min = lp_build_mul(coord_bld, - lp_build_const_scalar(coord_bld->type, -0.5F), + lp_build_const_vec(coord_bld->type, -0.5F), lp_build_rcp(coord_bld, length_f)); /* max = 1.0 - min */ max = lp_build_sub(coord_bld, coord_bld->one, min); @@ -566,7 +665,7 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, struct lp_build_context *coord_bld = &bld->coord_bld; struct lp_build_context *int_coord_bld = &bld->int_coord_bld; struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld; - LLVMValueRef two = lp_build_const_scalar(coord_bld->type, 2.0); + LLVMValueRef two = lp_build_const_vec(coord_bld->type, 2.0); LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length); LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one); LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one); @@ -609,7 +708,7 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, } else { /* clamp to [0.5, length - 0.5] */ - min = lp_build_const_scalar(coord_bld->type, 0.5F); + min = lp_build_const_vec(coord_bld->type, 0.5F); max = lp_build_sub(coord_bld, length_f, min); } /* coord = clamp(coord, min, max) */ @@ -625,7 +724,7 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, if (bld->static_state->normalized_coords) { /* min = -1.0 / (2 * length) = -0.5 / length */ min = lp_build_mul(coord_bld, - lp_build_const_scalar(coord_bld->type, -0.5F), + lp_build_const_vec(coord_bld->type, -0.5F), lp_build_rcp(coord_bld, length_f)); /* max = length - min */ max = lp_build_sub(coord_bld, length_f, min); @@ -634,7 +733,7 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, } else { /* clamp to [-0.5, length + 0.5] */ - min = lp_build_const_scalar(coord_bld->type, -0.5F); + min = lp_build_const_vec(coord_bld->type, -0.5F); max = lp_build_sub(coord_bld, length_f, min); } /* coord = clamp(coord, min, max) */ @@ -711,83 +810,905 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, /** - * Sample 2D texture with nearest filtering. + * Codegen equivalent for u_minify(). + * Return max(1, base_size >> level); + */ +static LLVMValueRef +lp_build_minify(struct lp_build_sample_context *bld, + LLVMValueRef base_size, + LLVMValueRef level) +{ + LLVMValueRef size = LLVMBuildAShr(bld->builder, base_size, level, "minify"); + size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one); + return size; +} + + +/** + * Generate code to compute texture level of detail (lambda). + * \param s vector of texcoord s values + * \param t vector of texcoord t values + * \param r vector of texcoord r values + * \param width scalar int texture width + * \param height scalar int texture height + * \param depth scalar int texture depth + */ +static LLVMValueRef +lp_build_lod_selector(struct lp_build_sample_context *bld, + LLVMValueRef s, + LLVMValueRef t, + LLVMValueRef r, + LLVMValueRef width, + LLVMValueRef height, + LLVMValueRef depth) + +{ + if (bld->static_state->min_lod == bld->static_state->max_lod) { + /* User is forcing sampling from a particular mipmap level. + * This is hit during mipmap generation. + */ + return LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod); + } + else { + const int dims = texture_dims(bld->static_state->target); + struct lp_build_context *float_bld = &bld->float_bld; + LLVMValueRef lod_bias = LLVMConstReal(LLVMFloatType(), + bld->static_state->lod_bias); + LLVMValueRef min_lod = LLVMConstReal(LLVMFloatType(), + bld->static_state->min_lod); + LLVMValueRef max_lod = LLVMConstReal(LLVMFloatType(), + bld->static_state->max_lod); + + LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0); + LLVMValueRef index1 = LLVMConstInt(LLVMInt32Type(), 1, 0); + LLVMValueRef index2 = LLVMConstInt(LLVMInt32Type(), 2, 0); + + LLVMValueRef s0, s1, s2; + LLVMValueRef t0, t1, t2; + LLVMValueRef r0, r1, r2; + LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy; + LLVMValueRef rho, lod; + + /* + * dsdx = abs(s[1] - s[0]); + * dsdy = abs(s[2] - s[0]); + * dtdx = abs(t[1] - t[0]); + * dtdy = abs(t[2] - t[0]); + * drdx = abs(r[1] - r[0]); + * drdy = abs(r[2] - r[0]); + * XXX we're assuming a four-element quad in 2x2 layout here. + */ + s0 = LLVMBuildExtractElement(bld->builder, s, index0, "s0"); + s1 = LLVMBuildExtractElement(bld->builder, s, index1, "s1"); + s2 = LLVMBuildExtractElement(bld->builder, s, index2, "s2"); + dsdx = LLVMBuildSub(bld->builder, s1, s0, ""); + dsdx = lp_build_abs(float_bld, dsdx); + dsdy = LLVMBuildSub(bld->builder, s2, s0, ""); + dsdy = lp_build_abs(float_bld, dsdy); + if (dims > 1) { + t0 = LLVMBuildExtractElement(bld->builder, t, index0, "t0"); + t1 = LLVMBuildExtractElement(bld->builder, t, index1, "t1"); + t2 = LLVMBuildExtractElement(bld->builder, t, index2, "t2"); + dtdx = LLVMBuildSub(bld->builder, t1, t0, ""); + dtdx = lp_build_abs(float_bld, dtdx); + dtdy = LLVMBuildSub(bld->builder, t2, t0, ""); + dtdy = lp_build_abs(float_bld, dtdy); + if (dims > 2) { + r0 = LLVMBuildExtractElement(bld->builder, r, index0, "r0"); + r1 = LLVMBuildExtractElement(bld->builder, r, index1, "r1"); + r2 = LLVMBuildExtractElement(bld->builder, r, index2, "r2"); + drdx = LLVMBuildSub(bld->builder, r1, r0, ""); + drdx = lp_build_abs(float_bld, drdx); + drdy = LLVMBuildSub(bld->builder, r2, r0, ""); + drdy = lp_build_abs(float_bld, drdy); + } + } + + /* Compute rho = max of all partial derivatives scaled by texture size. + * XXX this could be vectorized somewhat + */ + rho = LLVMBuildMul(bld->builder, + lp_build_max(float_bld, dsdx, dsdy), + lp_build_int_to_float(float_bld, width), ""); + if (dims > 1) { + LLVMValueRef max; + max = LLVMBuildMul(bld->builder, + lp_build_max(float_bld, dtdx, dtdy), + lp_build_int_to_float(float_bld, height), ""); + rho = lp_build_max(float_bld, rho, max); + if (dims > 2) { + max = LLVMBuildMul(bld->builder, + lp_build_max(float_bld, drdx, drdy), + lp_build_int_to_float(float_bld, depth), ""); + rho = lp_build_max(float_bld, rho, max); + } + } + + /* compute lod = log2(rho) */ + lod = lp_build_log2(float_bld, rho); + + /* add lod bias */ + lod = LLVMBuildAdd(bld->builder, lod, lod_bias, "LOD bias"); + + /* clamp lod */ + lod = lp_build_clamp(float_bld, lod, min_lod, max_lod); + + return lod; + } +} + + +/** + * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer + * mipmap level index. + * Note: this is all scalar code. + * \param lod scalar float texture level of detail + * \param level_out returns integer */ static void -lp_build_sample_2d_nearest_soa(struct lp_build_sample_context *bld, - LLVMValueRef s, - LLVMValueRef t, - LLVMValueRef width, - LLVMValueRef height, - LLVMValueRef stride, - LLVMValueRef data_ptr, - LLVMValueRef *texel) +lp_build_nearest_mip_level(struct lp_build_sample_context *bld, + unsigned unit, + LLVMValueRef lod, + LLVMValueRef *level_out) { - LLVMValueRef x, y; + struct lp_build_context *float_bld = &bld->float_bld; + struct lp_build_context *int_bld = &bld->int_bld; + LLVMValueRef last_level, level; - x = lp_build_sample_wrap_nearest(bld, s, width, - bld->static_state->pot_width, - bld->static_state->wrap_s); - y = lp_build_sample_wrap_nearest(bld, t, height, - bld->static_state->pot_height, - bld->static_state->wrap_t); + LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0); - lp_build_name(x, "tex.x.wrapped"); - lp_build_name(y, "tex.y.wrapped"); + last_level = bld->dynamic_state->last_level(bld->dynamic_state, + bld->builder, unit); + + /* convert float lod to integer */ + level = lp_build_iround(float_bld, lod); - lp_build_sample_texel_soa(bld, width, height, x, y, stride, data_ptr, texel); + /* clamp level to legal range of levels */ + *level_out = lp_build_clamp(int_bld, level, zero, last_level); } /** - * Sample 2D texture with bilinear filtering. + * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to + * two (adjacent) mipmap level indexes. Later, we'll sample from those + * two mipmap levels and interpolate between them. */ static void -lp_build_sample_2d_linear_soa(struct lp_build_sample_context *bld, +lp_build_linear_mip_levels(struct lp_build_sample_context *bld, + unsigned unit, + LLVMValueRef lod, + LLVMValueRef *level0_out, + LLVMValueRef *level1_out, + LLVMValueRef *weight_out) +{ + struct lp_build_context *float_bld = &bld->float_bld; + struct lp_build_context *int_bld = &bld->int_bld; + LLVMValueRef last_level, level; + + last_level = bld->dynamic_state->last_level(bld->dynamic_state, + bld->builder, unit); + + /* convert float lod to integer */ + level = lp_build_ifloor(float_bld, lod); + + /* compute level 0 and clamp to legal range of levels */ + *level0_out = lp_build_clamp(int_bld, level, + int_bld->zero, + last_level); + /* compute level 1 and clamp to legal range of levels */ + *level1_out = lp_build_add(int_bld, *level0_out, int_bld->one); + *level1_out = lp_build_min(int_bld, *level1_out, last_level); + + *weight_out = lp_build_fract(float_bld, lod); +} + + +/** + * Generate code to sample a mipmap level with nearest filtering. + * If sampling a cube texture, r = cube face in [0,5]. + */ +static void +lp_build_sample_image_nearest(struct lp_build_sample_context *bld, + LLVMValueRef width_vec, + LLVMValueRef height_vec, + LLVMValueRef depth_vec, + LLVMValueRef row_stride_vec, + LLVMValueRef img_stride_vec, + LLVMValueRef data_ptr, LLVMValueRef s, LLVMValueRef t, - LLVMValueRef width, - LLVMValueRef height, - LLVMValueRef stride, - LLVMValueRef data_ptr, - LLVMValueRef *texel) + LLVMValueRef r, + LLVMValueRef colors_out[4]) { - LLVMValueRef s_fpart; - LLVMValueRef t_fpart; - LLVMValueRef x0, x1; - LLVMValueRef y0, y1; + const int dims = texture_dims(bld->static_state->target); + LLVMValueRef x, y, z; + + /* + * Compute integer texcoords. + */ + x = lp_build_sample_wrap_nearest(bld, s, width_vec, + bld->static_state->pot_width, + bld->static_state->wrap_s); + lp_build_name(x, "tex.x.wrapped"); + + if (dims >= 2) { + y = lp_build_sample_wrap_nearest(bld, t, height_vec, + bld->static_state->pot_height, + bld->static_state->wrap_t); + lp_build_name(y, "tex.y.wrapped"); + + if (dims == 3) { + z = lp_build_sample_wrap_nearest(bld, r, depth_vec, + bld->static_state->pot_height, + bld->static_state->wrap_r); + lp_build_name(z, "tex.z.wrapped"); + } + else if (bld->static_state->target == PIPE_TEXTURE_CUBE) { + z = r; + } + else { + z = NULL; + } + } + else { + y = z = NULL; + } + + /* + * Get texture colors. + */ + lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec, + x, y, z, + row_stride_vec, img_stride_vec, + data_ptr, colors_out); +} + + +/** + * Generate code to sample a mipmap level with linear filtering. + * If sampling a cube texture, r = cube face in [0,5]. + */ +static void +lp_build_sample_image_linear(struct lp_build_sample_context *bld, + LLVMValueRef width_vec, + LLVMValueRef height_vec, + LLVMValueRef depth_vec, + LLVMValueRef row_stride_vec, + LLVMValueRef img_stride_vec, + LLVMValueRef data_ptr, + LLVMValueRef s, + LLVMValueRef t, + LLVMValueRef r, + LLVMValueRef colors_out[4]) +{ + const int dims = texture_dims(bld->static_state->target); + LLVMValueRef x0, y0, z0, x1, y1, z1; + LLVMValueRef s_fpart, t_fpart, r_fpart; LLVMValueRef neighbors[2][2][4]; - unsigned chan; + int chan; - lp_build_sample_wrap_linear(bld, s, width, bld->static_state->pot_width, - bld->static_state->wrap_s, &x0, &x1, &s_fpart); - lp_build_sample_wrap_linear(bld, t, height, bld->static_state->pot_height, - bld->static_state->wrap_t, &y0, &y1, &t_fpart); + /* + * Compute integer texcoords. + */ + lp_build_sample_wrap_linear(bld, s, width_vec, + bld->static_state->pot_width, + bld->static_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, t, height_vec, + bld->static_state->pot_height, + bld->static_state->wrap_t, + &y0, &y1, &t_fpart); + lp_build_name(y0, "tex.y0.wrapped"); + lp_build_name(y1, "tex.y1.wrapped"); + + if (dims == 3) { + lp_build_sample_wrap_linear(bld, r, depth_vec, + bld->static_state->pot_depth, + bld->static_state->wrap_r, + &z0, &z1, &r_fpart); + lp_build_name(z0, "tex.z0.wrapped"); + lp_build_name(z1, "tex.z1.wrapped"); + } + else if (bld->static_state->target == PIPE_TEXTURE_CUBE) { + z0 = z1 = r; /* cube face */ + r_fpart = NULL; + } + else { + z0 = z1 = NULL; + r_fpart = NULL; + } + } + else { + y0 = y1 = t_fpart = NULL; + z0 = z1 = r_fpart = NULL; + } - lp_build_sample_texel_soa(bld, width, height, x0, y0, stride, data_ptr, neighbors[0][0]); - lp_build_sample_texel_soa(bld, width, height, x1, y0, stride, data_ptr, neighbors[0][1]); - lp_build_sample_texel_soa(bld, width, height, x0, y1, stride, data_ptr, neighbors[1][0]); - lp_build_sample_texel_soa(bld, width, height, x1, y1, stride, data_ptr, neighbors[1][1]); + /* + * Get texture colors. + */ + /* get x0/x1 texels */ + lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec, + x0, y0, z0, + row_stride_vec, img_stride_vec, + data_ptr, neighbors[0][0]); + lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec, + x1, y0, z0, + row_stride_vec, img_stride_vec, + data_ptr, neighbors[0][1]); + + if (dims == 1) { + /* 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, + neighbors[0][0][chan], + neighbors[0][1][chan]); + } + } + else { + /* 2D/3D texture */ + LLVMValueRef colors0[4]; + + /* get x0/x1 texels at y1 */ + lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec, + x0, y1, z0, + row_stride_vec, img_stride_vec, + data_ptr, neighbors[1][0]); + lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec, + x1, y1, z0, + row_stride_vec, img_stride_vec, + data_ptr, neighbors[1][1]); + + /* 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]); + } - /* TODO: Don't interpolate missing channels */ - for(chan = 0; chan < 4; ++chan) { - texel[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]); + 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, width_vec, height_vec, depth_vec, + x0, y0, z1, + row_stride_vec, img_stride_vec, + data_ptr, neighbors1[0][0]); + lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec, + x1, y0, z1, + row_stride_vec, img_stride_vec, + data_ptr, neighbors1[0][1]); + lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec, + x0, y1, z1, + row_stride_vec, img_stride_vec, + data_ptr, neighbors1[1][0]); + lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec, + x1, y1, z1, + row_stride_vec, img_stride_vec, + data_ptr, neighbors1[1][1]); + + /* 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, + s_fpart, t_fpart, + neighbors1[0][0][chan], + neighbors1[0][1][chan], + neighbors1[1][0][chan], + neighbors1[1][1][chan]); + } + + /* 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, + r_fpart, + colors0[chan], colors1[chan]); + } + } + else { + /* 2D tex */ + for (chan = 0; chan < 4; chan++) { + colors_out[chan] = colors0[chan]; + } + } + } +} + + +/** Helper used by lp_build_cube_lookup() */ +static LLVMValueRef +lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord) +{ + /* ima = -0.5 / abs(coord); */ + LLVMValueRef negHalf = lp_build_const_vec(coord_bld->type, -0.5); + LLVMValueRef absCoord = lp_build_abs(coord_bld, coord); + LLVMValueRef ima = lp_build_mul(coord_bld, negHalf, + lp_build_rcp(coord_bld, absCoord)); + return ima; +} + + +/** + * Helper used by lp_build_cube_lookup() + * \param sign scalar +1 or -1 + * \param coord float vector + * \param ima float vector + */ +static LLVMValueRef +lp_build_cube_coord(struct lp_build_context *coord_bld, + LLVMValueRef sign, int negate_coord, + LLVMValueRef coord, LLVMValueRef ima) +{ + /* return negate(coord) * ima * sign + 0.5; */ + LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5); + LLVMValueRef res; + + assert(negate_coord == +1 || negate_coord == -1); + + if (negate_coord == -1) { + coord = lp_build_negate(coord_bld, coord); + } + + res = lp_build_mul(coord_bld, coord, ima); + if (sign) { + sign = lp_build_broadcast_scalar(coord_bld, sign); + res = lp_build_mul(coord_bld, res, sign); + } + res = lp_build_add(coord_bld, res, half); + + return res; +} + + +/** Helper used by lp_build_cube_lookup() + * Return (major_coord >= 0) ? pos_face : neg_face; + */ +static LLVMValueRef +lp_build_cube_face(struct lp_build_sample_context *bld, + LLVMValueRef major_coord, + unsigned pos_face, unsigned neg_face) +{ + LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE, + major_coord, + bld->float_bld.zero, ""); + LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0); + LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0); + LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, ""); + return res; +} + + + +/** + * Generate code to do cube face selection and per-face texcoords. + */ +static void +lp_build_cube_lookup(struct lp_build_sample_context *bld, + LLVMValueRef s, + LLVMValueRef t, + LLVMValueRef r, + LLVMValueRef *face, + LLVMValueRef *face_s, + LLVMValueRef *face_t) +{ + struct lp_build_context *float_bld = &bld->float_bld; + struct lp_build_context *coord_bld = &bld->coord_bld; + LLVMValueRef rx, ry, rz; + LLVMValueRef arx, ary, arz; + LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25); + LLVMValueRef arx_ge_ary, arx_ge_arz; + LLVMValueRef ary_ge_arx, ary_ge_arz; + LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz; + LLVMValueRef rx_pos, ry_pos, rz_pos; + + assert(bld->coord_bld.type.length == 4); + + /* + * Use the average of the four pixel's texcoords to choose the face. + */ + rx = lp_build_mul(float_bld, c25, + lp_build_sum_vector(&bld->coord_bld, s)); + ry = lp_build_mul(float_bld, c25, + lp_build_sum_vector(&bld->coord_bld, t)); + rz = lp_build_mul(float_bld, c25, + lp_build_sum_vector(&bld->coord_bld, r)); + + arx = lp_build_abs(float_bld, rx); + ary = lp_build_abs(float_bld, ry); + arz = lp_build_abs(float_bld, rz); + + /* + * Compare sign/magnitude of rx,ry,rz to determine face + */ + arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, ""); + arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, ""); + ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, ""); + ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, ""); + + arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, ""); + ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, ""); + + rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, ""); + ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, ""); + rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, ""); + + { + struct lp_build_flow_context *flow_ctx; + struct lp_build_if_state if_ctx; + + flow_ctx = lp_build_flow_create(bld->builder); + lp_build_flow_scope_begin(flow_ctx); + + *face_s = bld->coord_bld.undef; + *face_t = bld->coord_bld.undef; + *face = bld->int_bld.undef; + + lp_build_name(*face_s, "face_s"); + lp_build_name(*face_t, "face_t"); + lp_build_name(*face, "face"); + + lp_build_flow_scope_declare(flow_ctx, face_s); + lp_build_flow_scope_declare(flow_ctx, face_t); + lp_build_flow_scope_declare(flow_ctx, face); + + lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz); + { + /* +/- X face */ + LLVMValueRef sign = lp_build_sgn(float_bld, rx); + LLVMValueRef ima = lp_build_cube_ima(coord_bld, s); + *face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima); + *face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima); + *face = lp_build_cube_face(bld, rx, + PIPE_TEX_FACE_POS_X, + PIPE_TEX_FACE_NEG_X); + } + lp_build_else(&if_ctx); + { + struct lp_build_flow_context *flow_ctx2; + struct lp_build_if_state if_ctx2; + + LLVMValueRef face_s2 = bld->coord_bld.undef; + LLVMValueRef face_t2 = bld->coord_bld.undef; + LLVMValueRef face2 = bld->int_bld.undef; + + flow_ctx2 = lp_build_flow_create(bld->builder); + lp_build_flow_scope_begin(flow_ctx2); + lp_build_flow_scope_declare(flow_ctx2, &face_s2); + lp_build_flow_scope_declare(flow_ctx2, &face_t2); + lp_build_flow_scope_declare(flow_ctx2, &face2); + + ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, ""); + + lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz); + { + /* +/- Y face */ + LLVMValueRef sign = lp_build_sgn(float_bld, ry); + LLVMValueRef ima = lp_build_cube_ima(coord_bld, t); + face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima); + face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima); + face2 = lp_build_cube_face(bld, ry, + PIPE_TEX_FACE_POS_Y, + PIPE_TEX_FACE_NEG_Y); + } + lp_build_else(&if_ctx2); + { + /* +/- Z face */ + LLVMValueRef sign = lp_build_sgn(float_bld, rz); + LLVMValueRef ima = lp_build_cube_ima(coord_bld, r); + face_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima); + face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima); + face2 = lp_build_cube_face(bld, rz, + PIPE_TEX_FACE_POS_Z, + PIPE_TEX_FACE_NEG_Z); + } + lp_build_endif(&if_ctx2); + lp_build_flow_scope_end(flow_ctx2); + lp_build_flow_destroy(flow_ctx2); + + *face_s = face_s2; + *face_t = face_t2; + *face = face2; + } + + lp_build_endif(&if_ctx); + lp_build_flow_scope_end(flow_ctx); + lp_build_flow_destroy(flow_ctx); + } +} + + + +/** + * 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. + * 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 img_filter, + unsigned mip_filter, + LLVMValueRef s, + LLVMValueRef t, + LLVMValueRef r, + LLVMValueRef lod_fpart, + LLVMValueRef width0_vec, + LLVMValueRef width1_vec, + LLVMValueRef height0_vec, + LLVMValueRef height1_vec, + LLVMValueRef depth0_vec, + LLVMValueRef depth1_vec, + LLVMValueRef row_stride0_vec, + LLVMValueRef row_stride1_vec, + LLVMValueRef img_stride0_vec, + LLVMValueRef img_stride1_vec, + LLVMValueRef data_ptr0, + LLVMValueRef data_ptr1, + LLVMValueRef *colors_out) +{ + LLVMValueRef colors0[4], colors1[4]; + int chan; + + if (img_filter == PIPE_TEX_FILTER_NEAREST) { + lp_build_sample_image_nearest(bld, + width0_vec, height0_vec, depth0_vec, + row_stride0_vec, img_stride0_vec, + data_ptr0, s, t, r, colors0); + + if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) { + /* sample the second mipmap level, and interp */ + lp_build_sample_image_nearest(bld, + width1_vec, height1_vec, depth1_vec, + row_stride1_vec, img_stride1_vec, + data_ptr1, s, t, r, colors1); + } + } + else { + assert(img_filter == PIPE_TEX_FILTER_LINEAR); + + lp_build_sample_image_linear(bld, + width0_vec, height0_vec, depth0_vec, + row_stride0_vec, img_stride0_vec, + data_ptr0, s, t, r, colors0); + + if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) { + /* sample the second mipmap level, and interp */ + lp_build_sample_image_linear(bld, + width1_vec, height1_vec, depth1_vec, + row_stride1_vec, img_stride1_vec, + data_ptr1, s, t, r, colors1); + } + } + + if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) { + /* interpolate samples from the two mipmap levels */ + for (chan = 0; chan < 4; chan++) { + colors_out[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart, + colors0[chan], colors1[chan]); + } + } + else { + /* use first/only level's colors */ + for (chan = 0; chan < 4; chan++) { + colors_out[chan] = colors0[chan]; + } } } + +/** + * General texture sampling codegen. + * This function handles texture sampling for all texture targets (1D, + * 2D, 3D, cube) and all filtering modes. + */ +static void +lp_build_sample_general(struct lp_build_sample_context *bld, + unsigned unit, + LLVMValueRef s, + LLVMValueRef t, + LLVMValueRef r, + LLVMValueRef width, + LLVMValueRef height, + LLVMValueRef depth, + LLVMValueRef width_vec, + LLVMValueRef height_vec, + LLVMValueRef depth_vec, + LLVMValueRef row_stride_array, + LLVMValueRef img_stride_vec, + LLVMValueRef data_array, + LLVMValueRef *colors_out) +{ + struct lp_build_context *float_bld = &bld->float_bld; + const unsigned mip_filter = bld->static_state->min_mip_filter; + const unsigned min_filter = bld->static_state->min_img_filter; + const unsigned mag_filter = bld->static_state->mag_img_filter; + const int dims = texture_dims(bld->static_state->target); + LLVMValueRef lod, lod_fpart; + LLVMValueRef ilevel0, ilevel1, ilevel0_vec, ilevel1_vec; + LLVMValueRef width0_vec = NULL, height0_vec = NULL, depth0_vec = NULL; + LLVMValueRef width1_vec = NULL, height1_vec = NULL, depth1_vec = NULL; + LLVMValueRef row_stride0_vec = NULL, row_stride1_vec = NULL; + LLVMValueRef img_stride0_vec = NULL, img_stride1_vec = NULL; + LLVMValueRef data_ptr0, data_ptr1; + + /* + printf("%s mip %d min %d mag %d\n", __FUNCTION__, + mip_filter, min_filter, mag_filter); + */ + + /* + * Compute the level of detail (float). + */ + if (min_filter != mag_filter || + mip_filter != PIPE_TEX_MIPFILTER_NONE) { + /* Need to compute lod either to choose mipmap levels or to + * distinguish between minification/magnification with one mipmap level. + */ + lod = lp_build_lod_selector(bld, s, t, r, width, height, depth); + } + + /* + * Compute integer mipmap level(s) to fetch texels from. + */ + if (mip_filter == PIPE_TEX_MIPFILTER_NONE) { + /* always use mip level 0 */ + ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0); + } + else { + if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) { + lp_build_nearest_mip_level(bld, unit, lod, &ilevel0); + } + else { + assert(mip_filter == PIPE_TEX_MIPFILTER_LINEAR); + lp_build_linear_mip_levels(bld, unit, lod, &ilevel0, &ilevel1, + &lod_fpart); + lod_fpart = lp_build_broadcast_scalar(&bld->coord_bld, lod_fpart); + } + } + + /* + * Convert scalar integer mipmap levels into vectors. + */ + ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0); + if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) + ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1); + + /* + * Compute width, height at mipmap level 'ilevel0' + */ + width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec); + if (dims >= 2) { + height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec); + row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array, + ilevel0); + if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) { + img_stride0_vec = lp_build_mul(&bld->int_coord_bld, + row_stride0_vec, height0_vec); + if (dims == 3) { + depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec); + } + } + } + if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) { + /* compute width, height, depth for second mipmap level at 'ilevel1' */ + width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec); + if (dims >= 2) { + height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec); + row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array, + ilevel1); + if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) { + img_stride1_vec = lp_build_mul(&bld->int_coord_bld, + row_stride1_vec, height1_vec); + if (dims ==3) { + depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec); + } + } + } + } + + /* + * Choose cube face, recompute per-face texcoords. + */ + if (bld->static_state->target == PIPE_TEXTURE_CUBE) { + LLVMValueRef face, face_s, face_t; + lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t); + s = face_s; /* vec */ + t = face_t; /* vec */ + /* use 'r' to indicate cube face */ + r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */ + } + + /* + * Get pointer(s) to image data for mipmap level(s). + */ + data_ptr0 = lp_build_get_mipmap_level(bld, data_array, ilevel0); + if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) { + data_ptr1 = lp_build_get_mipmap_level(bld, data_array, ilevel1); + } + + /* + * Get/interpolate texture colors. + */ + if (min_filter == mag_filter) { + /* no need to distinquish between minification and magnification */ + lp_build_sample_mipmap(bld, min_filter, mip_filter, s, t, r, lod_fpart, + width0_vec, width1_vec, + height0_vec, height1_vec, + depth0_vec, depth1_vec, + row_stride0_vec, row_stride1_vec, + img_stride0_vec, img_stride1_vec, + data_ptr0, data_ptr1, + colors_out); + } + else { + /* Emit conditional to choose min image filter or mag image filter + * depending on the lod being >0 or <= 0, respectively. + */ + struct lp_build_flow_context *flow_ctx; + struct lp_build_if_state if_ctx; + LLVMValueRef minify; + + flow_ctx = lp_build_flow_create(bld->builder); + lp_build_flow_scope_begin(flow_ctx); + + lp_build_flow_scope_declare(flow_ctx, &colors_out[0]); + lp_build_flow_scope_declare(flow_ctx, &colors_out[1]); + lp_build_flow_scope_declare(flow_ctx, &colors_out[2]); + lp_build_flow_scope_declare(flow_ctx, &colors_out[3]); + + /* minify = lod > 0.0 */ + minify = LLVMBuildFCmp(bld->builder, LLVMRealUGE, + lod, float_bld->zero, ""); + + lp_build_if(&if_ctx, flow_ctx, bld->builder, minify); + { + /* Use the minification filter */ + lp_build_sample_mipmap(bld, min_filter, mip_filter, + s, t, r, lod_fpart, + width0_vec, width1_vec, + height0_vec, height1_vec, + depth0_vec, depth1_vec, + row_stride0_vec, row_stride1_vec, + img_stride0_vec, img_stride1_vec, + data_ptr0, data_ptr1, + colors_out); + } + lp_build_else(&if_ctx); + { + /* Use the magnification filter */ + lp_build_sample_mipmap(bld, mag_filter, mip_filter, + s, t, r, lod_fpart, + width0_vec, width1_vec, + height0_vec, height1_vec, + depth0_vec, depth1_vec, + row_stride0_vec, row_stride1_vec, + img_stride0_vec, img_stride1_vec, + data_ptr0, data_ptr1, + colors_out); + } + lp_build_endif(&if_ctx); + + lp_build_flow_scope_end(flow_ctx); + lp_build_flow_destroy(flow_ctx); + } +} + + + static void lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder, struct lp_type dst_type, LLVMValueRef packed, LLVMValueRef *rgba) { - LLVMValueRef mask = lp_build_int_const_scalar(dst_type, 0xff); + LLVMValueRef mask = lp_build_const_int_vec(dst_type, 0xff); unsigned chan; /* Decode the input vector components */ @@ -799,7 +1720,7 @@ lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder, input = packed; if(start) - input = LLVMBuildLShr(builder, input, lp_build_int_const_scalar(dst_type, start), ""); + input = LLVMBuildLShr(builder, input, lp_build_const_int_vec(dst_type, start), ""); if(stop < 32) input = LLVMBuildAnd(builder, input, mask, ""); @@ -817,8 +1738,8 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld, LLVMValueRef t, LLVMValueRef width, LLVMValueRef height, - LLVMValueRef stride, - LLVMValueRef data_ptr, + LLVMValueRef stride_array, + LLVMValueRef data_array, LLVMValueRef *texel) { LLVMBuilderRef builder = bld->builder; @@ -834,8 +1755,9 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld, LLVMValueRef neighbors_hi[2][2]; LLVMValueRef packed, packed_lo, packed_hi; LLVMValueRef unswizzled[4]; + LLVMValueRef stride; - lp_build_context_init(&i32, builder, lp_type_int(32)); + lp_build_context_init(&i32, builder, lp_type_int_vec(32)); lp_build_context_init(&h16, builder, lp_type_ufixed(16)); lp_build_context_init(&u8n, builder, lp_type_unorm(8)); @@ -860,17 +1782,17 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld, t = LLVMBuildFPToSI(builder, t, i32_vec_type, ""); /* subtract 0.5 (add -128) */ - i32_c128 = lp_build_int_const_scalar(i32.type, -128); + i32_c128 = lp_build_const_int_vec(i32.type, -128); s = LLVMBuildAdd(builder, s, i32_c128, ""); t = LLVMBuildAdd(builder, t, i32_c128, ""); /* compute floor (shift right 8) */ - i32_c8 = lp_build_int_const_scalar(i32.type, 8); + i32_c8 = lp_build_const_int_vec(i32.type, 8); s_ipart = LLVMBuildAShr(builder, s, i32_c8, ""); t_ipart = LLVMBuildAShr(builder, t, i32_c8, ""); /* compute fractional part (AND with 0xff) */ - i32_c255 = lp_build_int_const_scalar(i32.type, 255); + i32_c255 = lp_build_const_int_vec(i32.type, 255); s_fpart = LLVMBuildAnd(builder, s, i32_c255, ""); t_fpart = LLVMBuildAnd(builder, t, i32_c255, ""); @@ -941,6 +1863,8 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld, t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, ""); } + stride = lp_build_get_const_level_stride_vec(bld, stride_array, 0); + /* * Fetch the pixels as 4 x 32bit (rgba order might differ): * @@ -958,10 +1882,10 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld, * The higher 8 bits of the resulting elements will be zero. */ - neighbors[0][0] = lp_build_sample_packed(bld, x0, y0, stride, data_ptr); - neighbors[0][1] = lp_build_sample_packed(bld, x1, y0, stride, data_ptr); - neighbors[1][0] = lp_build_sample_packed(bld, x0, y1, stride, data_ptr); - neighbors[1][1] = lp_build_sample_packed(bld, x1, y1, stride, data_ptr); + neighbors[0][0] = lp_build_sample_packed(bld, x0, y0, stride, data_array); + neighbors[0][1] = lp_build_sample_packed(bld, x1, y0, stride, data_array); + neighbors[1][0] = lp_build_sample_packed(bld, x0, y1, stride, data_array); + neighbors[1][1] = lp_build_sample_packed(bld, x1, y1, stride, data_array); neighbors[0][0] = LLVMBuildBitCast(builder, neighbors[0][0], u8n_vec_type, ""); neighbors[0][1] = LLVMBuildBitCast(builder, neighbors[0][1], u8n_vec_type, ""); @@ -1035,7 +1959,7 @@ lp_build_sample_compare(struct lp_build_sample_context *bld, } assert(res); - res = lp_build_mul(texel_bld, res, lp_build_const_scalar(texel_bld->type, 0.25)); + res = lp_build_mul(texel_bld, res, lp_build_const_vec(texel_bld->type, 0.25)); /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */ for(chan = 0; chan < 3; ++chan) @@ -1044,194 +1968,11 @@ lp_build_sample_compare(struct lp_build_sample_context *bld, } -static int -texture_dims(enum pipe_texture_target tex) -{ - switch (tex) { - case PIPE_TEXTURE_1D: - return 1; - case PIPE_TEXTURE_2D: - case PIPE_TEXTURE_CUBE: - return 2; - case PIPE_TEXTURE_3D: - return 3; - default: - assert(0 && "bad texture target in texture_dims()"); - return 2; - } -} - - -/** - * Generate code to compute texture level of detail (lambda). - * \param s vector of texcoord s values - * \param t vector of texcoord t values - * \param r vector of texcoord r values - * \param width scalar int texture width - * \param height scalar int texture height - * \param depth scalar int texture depth - */ -static LLVMValueRef -lp_build_lod_selector(struct lp_build_sample_context *bld, - LLVMValueRef s, - LLVMValueRef t, - LLVMValueRef r, - LLVMValueRef width, - LLVMValueRef height, - LLVMValueRef depth) - -{ - const int dims = texture_dims(bld->static_state->target); - struct lp_build_context *coord_bld = &bld->coord_bld; - - LLVMValueRef lod_bias = lp_build_const_scalar(bld->coord_bld.type, - bld->static_state->lod_bias); - LLVMValueRef min_lod = lp_build_const_scalar(bld->coord_bld.type, - bld->static_state->min_lod); - LLVMValueRef max_lod = lp_build_const_scalar(bld->coord_bld.type, - bld->static_state->max_lod); - - LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0); - LLVMValueRef index1 = LLVMConstInt(LLVMInt32Type(), 1, 0); - LLVMValueRef index2 = LLVMConstInt(LLVMInt32Type(), 2, 0); - - LLVMValueRef s0, s1, s2; - LLVMValueRef t0, t1, t2; - LLVMValueRef r0, r1, r2; - LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy; - LLVMValueRef rho, lod; - - /* - * dsdx = abs(s[1] - s[0]); - * dsdy = abs(s[2] - s[0]); - * dtdx = abs(t[1] - t[0]); - * dtdy = abs(t[2] - t[0]); - * drdx = abs(r[1] - r[0]); - * drdy = abs(r[2] - r[0]); - * XXX we're assuming a four-element quad in 2x2 layout here. - */ - s0 = LLVMBuildExtractElement(bld->builder, s, index0, "s0"); - s1 = LLVMBuildExtractElement(bld->builder, s, index1, "s1"); - s2 = LLVMBuildExtractElement(bld->builder, s, index2, "s2"); - dsdx = lp_build_abs(coord_bld, lp_build_sub(coord_bld, s1, s0)); - dsdy = lp_build_abs(coord_bld, lp_build_sub(coord_bld, s2, s0)); - if (dims > 1) { - t0 = LLVMBuildExtractElement(bld->builder, t, index0, "t0"); - t1 = LLVMBuildExtractElement(bld->builder, t, index1, "t1"); - t2 = LLVMBuildExtractElement(bld->builder, t, index2, "t2"); - dtdx = lp_build_abs(coord_bld, lp_build_sub(coord_bld, t1, t0)); - dtdy = lp_build_abs(coord_bld, lp_build_sub(coord_bld, t2, t0)); - if (dims > 2) { - r0 = LLVMBuildExtractElement(bld->builder, r, index0, "r0"); - r1 = LLVMBuildExtractElement(bld->builder, r, index1, "r1"); - r2 = LLVMBuildExtractElement(bld->builder, r, index2, "r2"); - drdx = lp_build_abs(coord_bld, lp_build_sub(coord_bld, r1, r0)); - drdy = lp_build_abs(coord_bld, lp_build_sub(coord_bld, r2, r0)); - } - } - - /* Compute rho = max of all partial derivatives scaled by texture size. - * XXX this can be vectorized somewhat - */ - rho = lp_build_mul(coord_bld, - lp_build_max(coord_bld, dsdx, dsdy), - lp_build_int_to_float(coord_bld, width)); - if (dims > 1) { - LLVMValueRef max; - max = lp_build_mul(coord_bld, - lp_build_max(coord_bld, dtdx, dtdy), - lp_build_int_to_float(coord_bld, height)); - rho = lp_build_max(coord_bld, rho, max); - if (dims > 2) { - max = lp_build_mul(coord_bld, - lp_build_max(coord_bld, drdx, drdy), - lp_build_int_to_float(coord_bld, depth)); - rho = lp_build_max(coord_bld, rho, max); - } - } - - /* compute lod = log2(rho) */ - lod = lp_build_log2(coord_bld, rho); - - /* add lod bias */ - lod = lp_build_add(coord_bld, lod, lod_bias); - - /* clamp lod */ - lod = lp_build_clamp(coord_bld, lod, min_lod, max_lod); - - return lod; -} - - -/** - * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer - * mipmap level index. - * \param lod scalar float texture level of detail - * \param level_out returns integer - */ -static void -lp_build_nearest_mip_level(struct lp_build_sample_context *bld, - unsigned unit, - LLVMValueRef lod, - LLVMValueRef *level_out) -{ - struct lp_build_context *coord_bld = &bld->coord_bld; - struct lp_build_context *int_coord_bld = &bld->int_coord_bld; - LLVMValueRef last_level, level; - - last_level = bld->dynamic_state->last_level(bld->dynamic_state, - bld->builder, unit); - - /* convert float lod to integer */ - level = lp_build_iround(coord_bld, lod); - - /* clamp level to legal range of levels */ - *level_out = lp_build_clamp(int_coord_bld, level, - int_coord_bld->zero, - last_level); -} - - -/** - * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to - * two (adjacent) mipmap level indexes. Later, we'll sample from those - * two mipmap levels and interpolate between them. - */ -static void -lp_build_linear_mip_levels(struct lp_build_sample_context *bld, - unsigned unit, - LLVMValueRef lod, - LLVMValueRef *level0_out, - LLVMValueRef *level1_out, - LLVMValueRef *weight_out) -{ - struct lp_build_context *coord_bld = &bld->coord_bld; - struct lp_build_context *int_coord_bld = &bld->int_coord_bld; - LLVMValueRef last_level, level; - - last_level = bld->dynamic_state->last_level(bld->dynamic_state, - bld->builder, unit); - - /* convert float lod to integer */ - level = lp_build_ifloor(coord_bld, lod); - - /* compute level 0 and clamp to legal range of levels */ - *level0_out = lp_build_clamp(int_coord_bld, level, - int_coord_bld->zero, - last_level); - /* compute level 1 and clamp to legal range of levels */ - *level1_out = lp_build_add(int_coord_bld, *level0_out, int_coord_bld->one); - *level1_out = lp_build_min(int_coord_bld, *level1_out, int_coord_bld->zero); - - *weight_out = lp_build_fract(coord_bld, lod); -} - - - /** * Build 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. */ void lp_build_sample_soa(LLVMBuilderRef builder, @@ -1245,10 +1986,11 @@ lp_build_sample_soa(LLVMBuilderRef builder, LLVMValueRef *texel) { struct lp_build_sample_context bld; - LLVMValueRef width; - LLVMValueRef height; - LLVMValueRef stride; - LLVMValueRef data_ptr; + LLVMValueRef width, width_vec; + LLVMValueRef height, height_vec; + LLVMValueRef depth, depth_vec; + LLVMValueRef stride_array; + LLVMValueRef data_array; LLVMValueRef s; LLVMValueRef t; LLVMValueRef r; @@ -1256,6 +1998,7 @@ lp_build_sample_soa(LLVMBuilderRef builder, (void) lp_build_lod_selector; /* temporary to silence warning */ (void) lp_build_nearest_mip_level; (void) lp_build_linear_mip_levels; + (void) lp_build_minify; /* Setup our build context */ memset(&bld, 0, sizeof bld); @@ -1263,10 +2006,16 @@ lp_build_sample_soa(LLVMBuilderRef builder, bld.static_state = static_state; bld.dynamic_state = dynamic_state; bld.format_desc = util_format_description(static_state->format); + + bld.float_type = lp_type_float(32); + bld.int_type = lp_type_int(32); bld.coord_type = type; bld.uint_coord_type = lp_uint_type(type); bld.int_coord_type = lp_int_type(type); bld.texel_type = type; + + lp_build_context_init(&bld.float_bld, builder, bld.float_type); + lp_build_context_init(&bld.int_bld, builder, bld.int_type); lp_build_context_init(&bld.coord_bld, builder, bld.coord_type); lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type); lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type); @@ -1275,41 +2024,37 @@ lp_build_sample_soa(LLVMBuilderRef builder, /* Get the dynamic state */ width = dynamic_state->width(dynamic_state, builder, unit); height = dynamic_state->height(dynamic_state, builder, unit); - stride = dynamic_state->stride(dynamic_state, builder, unit); - data_ptr = dynamic_state->data_ptr(dynamic_state, builder, unit); + depth = dynamic_state->depth(dynamic_state, builder, unit); + stride_array = dynamic_state->row_stride(dynamic_state, builder, unit); + data_array = dynamic_state->data_ptr(dynamic_state, builder, unit); + /* Note that data_array is an array[level] of pointers to texture images */ s = coords[0]; t = coords[1]; r = coords[2]; - width = lp_build_broadcast_scalar(&bld.uint_coord_bld, width); - height = lp_build_broadcast_scalar(&bld.uint_coord_bld, height); - stride = lp_build_broadcast_scalar(&bld.uint_coord_bld, stride); - - if(static_state->target == PIPE_TEXTURE_1D) - t = bld.coord_bld.zero; - - switch (static_state->min_img_filter) { - case PIPE_TEX_FILTER_NEAREST: - lp_build_sample_2d_nearest_soa(&bld, s, t, width, height, - stride, data_ptr, texel); - break; - case PIPE_TEX_FILTER_LINEAR: - if(lp_format_is_rgba8(bld.format_desc) && - is_simple_wrap_mode(static_state->wrap_s) && - is_simple_wrap_mode(static_state->wrap_t)) - lp_build_sample_2d_linear_aos(&bld, s, t, width, height, - stride, data_ptr, texel); - else - lp_build_sample_2d_linear_soa(&bld, s, t, width, height, - stride, data_ptr, texel); - break; - default: - assert(0); + width_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, width); + height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height); + depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth); + + if (lp_format_is_rgba8(bld.format_desc) && + static_state->target == PIPE_TEXTURE_2D && + static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR && + static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR && + static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE && + is_simple_wrap_mode(static_state->wrap_s) && + is_simple_wrap_mode(static_state->wrap_t)) { + /* special case */ + lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec, + stride_array, data_array, texel); + } + else { + lp_build_sample_general(&bld, unit, s, t, r, + width, height, depth, + width_vec, height_vec, depth_vec, + stride_array, NULL, data_array, + texel); } - - /* FIXME: respect static_state->min_mip_filter */; - /* FIXME: respect static_state->mag_img_filter */; lp_build_sample_compare(&bld, r, texel); } |