#include "lp_bld_arit.h"
#include "lp_bld_const.h"
#include "lp_bld_debug.h"
+#include "lp_bld_printf.h"
#include "lp_bld_flow.h"
#include "lp_bld_sample.h"
#include "lp_bld_swizzle.h"
#include "lp_bld_type.h"
+/*
+ * Bri-linear factor. Use zero or any other number less than one to force
+ * tri-linear filtering.
+ */
+#define BRILINEAR_FACTOR 2
+
+
/**
* Does the given texture wrap mode allow sampling the texture border color?
* XXX maybe move this into gallium util code.
{
struct lp_build_context *float_size_bld = &bld->float_size_bld;
struct lp_build_context *float_bld = &bld->float_bld;
- const int dims = texture_dims(bld->static_state->target);
+ const unsigned dims = bld->dims;
LLVMTypeRef i32t = LLVMInt32Type();
LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
LLVMValueRef index1 = LLVMConstInt(i32t, 1, 0);
rho_vec = lp_build_max(float_size_bld, rho_x, rho_y);
- float_size = lp_build_int_to_float(float_size_bld, bld->uint_size);
+ float_size = lp_build_int_to_float(float_size_bld, bld->int_size);
rho_vec = lp_build_mul(float_size_bld, rho_vec, float_size);
}
+/*
+ * Bri-linear lod computation
+ *
+ * Use a piece-wise linear approximation of log2 such that:
+ * - round to nearest, for values in the neighborhood of -1, 0, 1, 2, etc.
+ * - linear approximation for values in the neighborhood of 0.5, 1.5., etc,
+ * with the steepness specified in 'factor'
+ * - exact result for 0.5, 1.5, etc.
+ *
+ *
+ * 1.0 - /----*
+ * /
+ * /
+ * /
+ * 0.5 - *
+ * /
+ * /
+ * /
+ * 0.0 - *----/
+ *
+ * | |
+ * 2^0 2^1
+ *
+ * This is a technique also commonly used in hardware:
+ * - http://ixbtlabs.com/articles2/gffx/nv40-rx800-3.html
+ *
+ * TODO: For correctness, this should only be applied when texture is known to
+ * have regular mipmaps, i.e., mipmaps derived from the base level.
+ *
+ * TODO: This could be done in fixed point, where applicable.
+ */
+static void
+lp_build_brilinear_lod(struct lp_build_context *bld,
+ LLVMValueRef lod,
+ double factor,
+ LLVMValueRef *out_lod_ipart,
+ LLVMValueRef *out_lod_fpart)
+{
+ LLVMValueRef lod_fpart;
+ double pre_offset = (factor - 0.5)/factor - 0.5;
+ double post_offset = 1 - factor;
+
+ if (0) {
+ lp_build_printf(bld->builder, "lod = %f\n", lod);
+ }
+
+ lod = lp_build_add(bld, lod,
+ lp_build_const_vec(bld->type, pre_offset));
+
+ lp_build_ifloor_fract(bld, lod, out_lod_ipart, &lod_fpart);
+
+ lod_fpart = lp_build_mul(bld, lod_fpart,
+ lp_build_const_vec(bld->type, factor));
+
+ lod_fpart = lp_build_add(bld, lod_fpart,
+ lp_build_const_vec(bld->type, post_offset));
+
+ /*
+ * It's not necessary to clamp lod_fpart since:
+ * - the above expression will never produce numbers greater than one.
+ * - the mip filtering branch is only taken if lod_fpart is positive
+ */
+
+ *out_lod_fpart = lod_fpart;
+
+ if (0) {
+ lp_build_printf(bld->builder, "lod_ipart = %i\n", *out_lod_ipart);
+ lp_build_printf(bld->builder, "lod_fpart = %f\n\n", *out_lod_fpart);
+ }
+}
+
+
+/*
+ * Combined log2 and brilinear lod computation.
+ *
+ * It's in all identical to calling lp_build_fast_log2() and
+ * lp_build_brilinear_lod() above, but by combining we can compute the interger
+ * and fractional part independently.
+ */
+static void
+lp_build_brilinear_rho(struct lp_build_context *bld,
+ LLVMValueRef rho,
+ double factor,
+ LLVMValueRef *out_lod_ipart,
+ LLVMValueRef *out_lod_fpart)
+{
+ LLVMValueRef lod_ipart;
+ LLVMValueRef lod_fpart;
+
+ const double pre_factor = (2*factor - 0.5)/(M_SQRT2*factor);
+ const double post_offset = 1 - 2*factor;
+
+ assert(bld->type.floating);
+
+ assert(lp_check_value(bld->type, rho));
+
+ /*
+ * The pre factor will make the intersections with the exact powers of two
+ * happen precisely where we want then to be, which means that the integer
+ * part will not need any post adjustments.
+ */
+ rho = lp_build_mul(bld, rho,
+ lp_build_const_vec(bld->type, pre_factor));
+
+ /* ipart = ifloor(log2(rho)) */
+ lod_ipart = lp_build_extract_exponent(bld, rho, 0);
+
+ /* fpart = rho / 2**ipart */
+ lod_fpart = lp_build_extract_mantissa(bld, rho);
+
+ lod_fpart = lp_build_mul(bld, lod_fpart,
+ lp_build_const_vec(bld->type, factor));
+
+ lod_fpart = lp_build_add(bld, lod_fpart,
+ lp_build_const_vec(bld->type, post_offset));
+
+ /*
+ * Like lp_build_brilinear_lod, it's not necessary to clamp lod_fpart since:
+ * - the above expression will never produce numbers greater than one.
+ * - the mip filtering branch is only taken if lod_fpart is positive
+ */
+
+ *out_lod_ipart = lod_ipart;
+ *out_lod_fpart = lod_fpart;
+}
+
+
/**
* Generate code to compute texture level of detail (lambda).
* \param ddx partial derivatives of (s, t, r, q) with respect to X
const LLVMValueRef ddy[4],
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
- LLVMValueRef width,
- LLVMValueRef height,
- LLVMValueRef depth,
unsigned mip_filter,
LLVMValueRef *out_lod_ipart,
LLVMValueRef *out_lod_fpart)
rho = lp_build_rho(bld, ddx, ddy);
- /* compute lod = log2(rho) */
- if ((mip_filter == PIPE_TEX_MIPFILTER_NONE ||
- mip_filter == PIPE_TEX_MIPFILTER_NEAREST) &&
- !lod_bias &&
+ /*
+ * Compute lod = log2(rho)
+ */
+
+ if (!lod_bias &&
!bld->static_state->lod_bias_non_zero &&
!bld->static_state->apply_max_lod &&
!bld->static_state->apply_min_lod) {
- *out_lod_ipart = lp_build_ilog2(float_bld, rho);
- *out_lod_fpart = bld->float_bld.zero;
- return;
+ /*
+ * Special case when there are no post-log2 adjustments, which
+ * saves instructions but keeping the integer and fractional lod
+ * computations separate from the start.
+ */
+
+ if (mip_filter == PIPE_TEX_MIPFILTER_NONE ||
+ mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
+ *out_lod_ipart = lp_build_ilog2(float_bld, rho);
+ *out_lod_fpart = bld->float_bld.zero;
+ return;
+ }
+ if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR &&
+ BRILINEAR_FACTOR > 1.0) {
+ lp_build_brilinear_rho(float_bld, rho, BRILINEAR_FACTOR,
+ out_lod_ipart, out_lod_fpart);
+ return;
+ }
}
if (0) {
}
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
- LLVMValueRef ipart = lp_build_ifloor(float_bld, lod);
- lp_build_name(ipart, "lod_ipart");
- *out_lod_ipart = ipart;
- ipart = LLVMBuildSIToFP(bld->builder, ipart, float_bld->vec_type, "");
- *out_lod_fpart = LLVMBuildFSub(bld->builder, lod, ipart, "lod_fpart");
+ if (BRILINEAR_FACTOR > 1.0) {
+ lp_build_brilinear_lod(float_bld, lod, BRILINEAR_FACTOR,
+ out_lod_ipart, out_lod_fpart);
+ }
+ else {
+ lp_build_ifloor_fract(float_bld, lod, out_lod_ipart, out_lod_fpart);
+ }
+
+ lp_build_name(*out_lod_fpart, "lod_fpart");
}
else {
*out_lod_ipart = lp_build_iround(float_bld, lod);
}
+ lp_build_name(*out_lod_ipart, "lod_ipart");
+
return;
}
lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
unsigned unit,
LLVMValueRef lod_ipart,
+ LLVMValueRef *lod_fpart_inout,
LLVMValueRef *level0_out,
LLVMValueRef *level1_out)
{
+ LLVMBuilderRef builder = bld->builder;
struct lp_build_context *int_bld = &bld->int_bld;
- LLVMValueRef last_level, level;
+ struct lp_build_context *float_bld = &bld->float_bld;
+ LLVMValueRef last_level;
+ LLVMValueRef clamp_min;
+ LLVMValueRef clamp_max;
+
+ *level0_out = lod_ipart;
+ *level1_out = lp_build_add(int_bld, lod_ipart, int_bld->one);
last_level = bld->dynamic_state->last_level(bld->dynamic_state,
bld->builder, unit);
- /* convert float lod to integer */
- level = lod_ipart;
+ /*
+ * Clamp both lod_ipart and lod_ipart + 1 to [0, last_level], with the
+ * minimum number of comparisons, and zeroing lod_fpart in the extreme
+ * ends in the process.
+ */
+
+ /* lod_ipart < 0 */
+ clamp_min = LLVMBuildICmp(builder, LLVMIntSLT,
+ lod_ipart, int_bld->zero,
+ "clamp_lod_to_zero");
- /* 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 */
- level = lp_build_add(int_bld, level, int_bld->one);
- *level1_out = lp_build_clamp(int_bld, level,
- int_bld->zero,
- last_level);
+ *level0_out = LLVMBuildSelect(builder, clamp_min,
+ int_bld->zero, *level0_out, "");
+
+ *level1_out = LLVMBuildSelect(builder, clamp_min,
+ int_bld->zero, *level1_out, "");
+
+ *lod_fpart_inout = LLVMBuildSelect(builder, clamp_min,
+ float_bld->zero, *lod_fpart_inout, "");
+
+ /* lod_ipart >= last_level */
+ clamp_max = LLVMBuildICmp(builder, LLVMIntSGE,
+ lod_ipart, last_level,
+ "clamp_lod_to_last");
+
+ *level0_out = LLVMBuildSelect(builder, clamp_max,
+ last_level, *level0_out, "");
+
+ *level1_out = LLVMBuildSelect(builder, clamp_max,
+ last_level, *level1_out, "");
+
+ *lod_fpart_inout = LLVMBuildSelect(builder, clamp_max,
+ float_bld->zero, *lod_fpart_inout, "");
+
+ lp_build_name(*level0_out, "sampler%u_miplevel0", unit);
+ lp_build_name(*level1_out, "sampler%u_miplevel1", unit);
+ lp_build_name(*lod_fpart_inout, "sampler%u_mipweight", unit);
}
*/
LLVMValueRef
lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
- LLVMValueRef data_array, LLVMValueRef level)
+ 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 = LLVMBuildGEP(bld->builder, bld->data_array, indexes, 2, "");
data_ptr = LLVMBuildLoad(bld->builder, data_ptr, "");
return data_ptr;
}
LLVMValueRef
lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
- LLVMValueRef data_array, int level)
+ int level)
{
LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
- return lp_build_get_mipmap_level(bld, data_array, lvl);
+ return lp_build_get_mipmap_level(bld, lvl);
}
* Return max(1, base_size >> level);
*/
static LLVMValueRef
-lp_build_minify(struct lp_build_sample_context *bld,
+lp_build_minify(struct lp_build_context *bld,
LLVMValueRef base_size,
LLVMValueRef level)
{
- if (level == bld->int_coord_bld.zero) {
+ assert(lp_check_value(bld->type, base_size));
+ assert(lp_check_value(bld->type, level));
+
+ if (level == bld->zero) {
/* if we're using mipmap level zero, no minification is needed */
return base_size;
}
else {
LLVMValueRef size =
LLVMBuildLShr(bld->builder, base_size, level, "minify");
- size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
+ assert(bld->type.sign);
+ size = lp_build_max(bld, size, bld->one);
return size;
}
}
*/
void
lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
- unsigned dims,
- LLVMValueRef width_vec,
- LLVMValueRef height_vec,
- LLVMValueRef depth_vec,
- LLVMValueRef ilevel0,
- LLVMValueRef ilevel1,
- LLVMValueRef row_stride_array,
- LLVMValueRef img_stride_array,
- 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 ilevel,
+ LLVMValueRef *out_size,
+ LLVMValueRef *row_stride_vec,
+ LLVMValueRef *img_stride_vec)
{
- const unsigned mip_filter = bld->static_state->min_mip_filter;
- LLVMValueRef ilevel0_vec, ilevel1_vec;
+ const unsigned dims = bld->dims;
+ LLVMValueRef ilevel_vec;
- 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);
+ ilevel_vec = lp_build_broadcast_scalar(&bld->int_size_bld, ilevel);
/*
- * Compute width, height, depth at mipmap level 'ilevel0'
+ * Compute width, height, depth at mipmap level 'ilevel'
*/
- *width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
+ *out_size = lp_build_minify(&bld->int_size_bld, bld->int_size, ilevel_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);
+ *row_stride_vec = lp_build_get_level_stride_vec(bld,
+ bld->row_stride_array,
+ ilevel);
if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
- *img_stride0_vec = lp_build_get_level_stride_vec(bld,
- img_stride_array,
- ilevel0);
- if (dims == 3) {
- *depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
- }
+ *img_stride_vec = lp_build_get_level_stride_vec(bld,
+ bld->img_stride_array,
+ ilevel);
}
}
- 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_get_level_stride_vec(bld,
- img_stride_array,
- ilevel1);
- if (dims == 3) {
- *depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
- }
- }
+}
+
+
+/**
+ * Extract and broadcast texture size.
+ *
+ * @param size_type type of the texture size vector (either
+ * bld->int_size_type or bld->float_size_type)
+ * @param coord_type type of the texture size vector (either
+ * bld->int_coord_type or bld->coord_type)
+ * @param int_size vector with the integer texture size (width, height,
+ * depth)
+ */
+void
+lp_build_extract_image_sizes(struct lp_build_sample_context *bld,
+ struct lp_type size_type,
+ struct lp_type coord_type,
+ LLVMValueRef size,
+ LLVMValueRef *out_width,
+ LLVMValueRef *out_height,
+ LLVMValueRef *out_depth)
+{
+ const unsigned dims = bld->dims;
+ LLVMTypeRef i32t = LLVMInt32Type();
+
+ *out_width = lp_build_extract_broadcast(bld->builder,
+ size_type,
+ coord_type,
+ size,
+ LLVMConstInt(i32t, 0, 0));
+ if (dims >= 2) {
+ *out_height = lp_build_extract_broadcast(bld->builder,
+ size_type,
+ coord_type,
+ size,
+ LLVMConstInt(i32t, 1, 0));
+ if (dims == 3) {
+ *out_depth = lp_build_extract_broadcast(bld->builder,
+ size_type,
+ coord_type,
+ size,
+ LLVMConstInt(i32t, 2, 0));
}
}
}
+/**
+ * Unnormalize coords.
+ *
+ * @param int_size vector with the integer texture size (width, height, depth)
+ */
+void
+lp_build_unnormalized_coords(struct lp_build_sample_context *bld,
+ LLVMValueRef flt_size,
+ LLVMValueRef *s,
+ LLVMValueRef *t,
+ LLVMValueRef *r)
+{
+ const unsigned dims = bld->dims;
+ LLVMValueRef width;
+ LLVMValueRef height;
+ LLVMValueRef depth;
+
+ lp_build_extract_image_sizes(bld,
+ bld->float_size_type,
+ bld->coord_type,
+ flt_size,
+ &width,
+ &height,
+ &depth);
+
+ /* s = s * width, t = t * height */
+ *s = lp_build_mul(&bld->coord_bld, *s, width);
+ if (dims >= 2) {
+ *t = lp_build_mul(&bld->coord_bld, *t, height);
+ if (dims >= 3) {
+ *r = lp_build_mul(&bld->coord_bld, *r, depth);
+ }
+ }
+}
+
/** Helper used by lp_build_cube_lookup() */
static LLVMValueRef