#include "pipe/p_defines.h"
#include "pipe/p_state.h"
+#include "pipe/p_shader_tokens.h"
#include "util/u_debug.h"
#include "util/u_dump.h"
#include "util/u_memory.h"
#include "util/u_math.h"
#include "util/u_format.h"
+#include "util/u_cpu_detect.h"
#include "lp_bld_debug.h"
#include "lp_bld_type.h"
#include "lp_bld_const.h"
#include "lp_bld_conv.h"
#include "lp_bld_arit.h"
+#include "lp_bld_bitarit.h"
#include "lp_bld_logic.h"
+#include "lp_bld_printf.h"
#include "lp_bld_swizzle.h"
-#include "lp_bld_pack.h"
#include "lp_bld_flow.h"
#include "lp_bld_gather.h"
#include "lp_bld_format.h"
#include "lp_bld_sample.h"
+#include "lp_bld_sample_aos.h"
+#include "lp_bld_struct.h"
#include "lp_bld_quad.h"
-
-
-/**
- * Keep all information for sampling code generation in a single place.
- */
-struct lp_build_sample_context
-{
- LLVMBuilderRef builder;
-
- const struct lp_sampler_static_state *static_state;
-
- struct lp_sampler_dynamic_state *dynamic_state;
-
- 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;
-
- /** Unsigned integer coordinates */
- struct lp_type uint_coord_type;
- struct lp_build_context uint_coord_bld;
-
- /** Signed integer coordinates */
- struct lp_type int_coord_type;
- struct lp_build_context int_coord_bld;
-
- /** Output texels type and build context */
- struct lp_type texel_type;
- struct lp_build_context texel_bld;
-};
-
-
-/**
- * Does the given texture wrap mode allow sampling the texture border color?
- * XXX maybe move this into gallium util code.
- */
-static boolean
-wrap_mode_uses_border_color(unsigned mode)
-{
- switch (mode) {
- case PIPE_TEX_WRAP_REPEAT:
- case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
- case PIPE_TEX_WRAP_MIRROR_REPEAT:
- case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
- return FALSE;
- case PIPE_TEX_WRAP_CLAMP:
- case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
- case PIPE_TEX_WRAP_MIRROR_CLAMP:
- case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
- return TRUE;
- default:
- assert(0 && "unexpected wrap mode");
- return FALSE;
- }
-}
-
-
-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_RECT:
- case PIPE_TEXTURE_CUBE:
- return 2;
- case PIPE_TEXTURE_3D:
- return 3;
- default:
- assert(0 && "bad texture target in texture_dims()");
- return 2;
- }
-}
-
-
-static void
-apply_sampler_swizzle(struct lp_build_sample_context *bld,
- LLVMValueRef *texel)
-{
- unsigned char swizzles[4];
-
- swizzles[0] = bld->static_state->swizzle_r;
- swizzles[1] = bld->static_state->swizzle_g;
- swizzles[2] = bld->static_state->swizzle_b;
- swizzles[3] = bld->static_state->swizzle_a;
-
- lp_build_swizzle_soa_inplace(&bld->texel_bld, texel, swizzles);
-}
-
+#include "lp_bld_pack.h"
/**
* 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:
+ * The result, texel, will be float vectors:
* texel[0] = red values
* texel[1] = green values
* texel[2] = blue values
*/
static void
lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
+ unsigned sampler_unit,
LLVMValueRef width,
LLVMValueRef height,
LLVMValueRef depth,
LLVMValueRef y_stride,
LLVMValueRef z_stride,
LLVMValueRef data_ptr,
+ LLVMValueRef mipoffsets,
LLVMValueRef texel_out[4])
{
- const int dims = texture_dims(bld->static_state->target);
+ const struct lp_static_sampler_state *static_state = bld->static_sampler_state;
+ const unsigned dims = bld->dims;
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
+ LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef offset;
LLVMValueRef i, j;
LLVMValueRef use_border = NULL;
/* use_border = x < 0 || x >= width || y < 0 || y >= height */
- if (wrap_mode_uses_border_color(bld->static_state->wrap_s)) {
+ if (lp_sampler_wrap_mode_uses_border_color(static_state->wrap_s,
+ static_state->min_img_filter,
+ static_state->mag_img_filter)) {
LLVMValueRef b1, b2;
b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
- use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
+ use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
}
- if (dims >= 2 && wrap_mode_uses_border_color(bld->static_state->wrap_t)) {
+ if (dims >= 2 &&
+ lp_sampler_wrap_mode_uses_border_color(static_state->wrap_t,
+ static_state->min_img_filter,
+ static_state->mag_img_filter)) {
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);
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");
+ use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
+ use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
}
else {
- use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
+ use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
}
}
- if (dims == 3 && wrap_mode_uses_border_color(bld->static_state->wrap_r)) {
+ if (dims == 3 &&
+ lp_sampler_wrap_mode_uses_border_color(static_state->wrap_r,
+ static_state->min_img_filter,
+ static_state->mag_img_filter)) {
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");
+ use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
+ use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
}
else {
- use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
+ use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
}
}
/* convert x,y,z coords to linear offset from start of texture, in bytes */
- lp_build_sample_offset(&bld->uint_coord_bld,
+ lp_build_sample_offset(&bld->int_coord_bld,
bld->format_desc,
x, y, z, y_stride, z_stride,
&offset, &i, &j);
+ if (mipoffsets) {
+ offset = lp_build_add(&bld->int_coord_bld, offset, mipoffsets);
+ }
if (use_border) {
/* If we can sample the border color, it means that texcoords may
* coords which are out of bounds to become zero. Zero's guaranteed
* to be inside the texture image.
*/
- offset = lp_build_andc(&bld->uint_coord_bld, offset, use_border);
+ offset = lp_build_andnot(&bld->int_coord_bld, offset, use_border);
}
- lp_build_fetch_rgba_soa(bld->builder,
+ lp_build_fetch_rgba_soa(bld->gallivm,
bld->format_desc,
bld->texel_type,
data_ptr, offset,
i, j,
texel_out);
- apply_sampler_swizzle(bld, texel_out);
-
/*
* Note: if we find an app which frequently samples the texture border
* we might want to implement a true conditional here to avoid sampling
*/
if (use_border) {
- /* select texel color or border color depending on use_border */
+ /* select texel color or border color depending on use_border. */
+ LLVMValueRef border_color_ptr =
+ bld->dynamic_state->border_color(bld->dynamic_state,
+ bld->gallivm, sampler_unit);
+ const struct util_format_description *format_desc;
int chan;
+ format_desc = util_format_description(bld->static_texture_state->format);
+ /*
+ * Only replace channels which are actually present. The others should
+ * get optimized away eventually by sampler_view swizzle anyway but it's
+ * easier too as we'd need some extra logic for channels where we can't
+ * determine the format directly otherwise.
+ */
for (chan = 0; chan < 4; chan++) {
- LLVMValueRef border_chan =
- lp_build_const_vec(bld->texel_type,
- bld->static_state->border_color[chan]);
- texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
- border_chan, texel_out[chan]);
+ unsigned chan_s;
+ /* reverse-map channel... */
+ for (chan_s = 0; chan_s < 4; chan_s++) {
+ if (chan_s == format_desc->swizzle[chan]) {
+ break;
+ }
+ }
+ if (chan_s <= 3) {
+ LLVMValueRef border_chan =
+ lp_build_array_get(bld->gallivm, border_color_ptr,
+ lp_build_const_int32(bld->gallivm, chan));
+ LLVMValueRef border_chan_vec =
+ lp_build_broadcast_scalar(&bld->float_vec_bld, border_chan);
+
+ if (!bld->texel_type.floating) {
+ border_chan_vec = LLVMBuildBitCast(builder, border_chan_vec,
+ bld->texel_bld.vec_type, "");
+ }
+ else {
+ /*
+ * For normalized format need to clamp border color (technically
+ * probably should also quantize the data). Really sucks doing this
+ * here but can't avoid at least for now since this is part of
+ * sampler state and texture format is part of sampler_view state.
+ */
+ unsigned chan_type = format_desc->channel[chan_s].type;
+ unsigned chan_norm = format_desc->channel[chan_s].normalized;
+ if (chan_type == UTIL_FORMAT_TYPE_SIGNED && chan_norm) {
+ LLVMValueRef clamp_min;
+ clamp_min = lp_build_const_vec(bld->gallivm, bld->texel_type, -1.0F);
+ border_chan_vec = lp_build_clamp(&bld->texel_bld, border_chan_vec,
+ clamp_min,
+ bld->texel_bld.one);
+ }
+ else if (chan_type == UTIL_FORMAT_TYPE_UNSIGNED && chan_norm) {
+ border_chan_vec = lp_build_clamp(&bld->texel_bld, border_chan_vec,
+ bld->texel_bld.zero,
+ bld->texel_bld.one);
+ }
+ /* not exactly sure about all others but I think should be ok? */
+ }
+ texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
+ border_chan_vec, texel_out[chan]);
+ }
}
}
}
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
LLVMValueRef fract, flr, isOdd;
- /* fract = coord - floor(coord) */
- fract = lp_build_sub(coord_bld, coord, lp_build_floor(coord_bld, coord));
-
- /* flr = ifloor(coord); */
- flr = lp_build_ifloor(coord_bld, coord);
+ lp_build_ifloor_fract(coord_bld, coord, &flr, &fract);
/* isOdd = flr & 1 */
- isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
+ isOdd = LLVMBuildAnd(bld->gallivm->builder, flr, int_coord_bld->one, "");
/* make coord positive or negative depending on isOdd */
coord = lp_build_set_sign(coord_bld, fract, isOdd);
/**
- * We only support a few wrap modes in lp_build_sample_wrap_linear_int() at this time.
- * Return whether the given mode is supported by that function.
- */
-static boolean
-is_simple_wrap_mode(unsigned mode)
-{
- switch (mode) {
- case PIPE_TEX_WRAP_REPEAT:
- case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
- return TRUE;
- default:
- return FALSE;
- }
-}
-
-
-/**
- * Build LLVM code for texture wrap mode, for scaled integer texcoords.
- * \param coord the incoming texcoord (s,t,r or q) scaled to the texture size
- * \param length the texture size along one dimension
- * \param is_pot if TRUE, length is a power of two
- * \param wrap_mode one of PIPE_TEX_WRAP_x
- * \param i0 resulting sub-block pixel coordinate for coord0
- */
-static void
-lp_build_sample_wrap_nearest_int(struct lp_build_sample_context *bld,
- unsigned block_length,
- LLVMValueRef coord,
- LLVMValueRef length,
- LLVMValueRef stride,
- boolean is_pot,
- unsigned wrap_mode,
- LLVMValueRef *out_offset,
- LLVMValueRef *out_i)
-{
- struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
- struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
- LLVMValueRef length_minus_one;
-
- length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
-
- switch(wrap_mode) {
- case PIPE_TEX_WRAP_REPEAT:
- if(is_pot)
- coord = LLVMBuildAnd(bld->builder, coord, length_minus_one, "");
- else
- /* Signed remainder won't give the right results for negative
- * dividends but unsigned remainder does.*/
- coord = LLVMBuildURem(bld->builder, coord, length, "");
- break;
-
- case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
- coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
- coord = lp_build_min(int_coord_bld, coord, length_minus_one);
- break;
-
- case PIPE_TEX_WRAP_CLAMP:
- case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
- case PIPE_TEX_WRAP_MIRROR_REPEAT:
- case PIPE_TEX_WRAP_MIRROR_CLAMP:
- case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
- case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
- default:
- assert(0);
- }
-
- lp_build_sample_partial_offset(uint_coord_bld, block_length, coord, stride,
- out_offset, out_i);
-}
-
-
-/**
- * Build LLVM code for texture wrap mode, for scaled integer texcoords.
- * \param coord0 the incoming texcoord (s,t,r or q) scaled to the texture size
- * \param length the texture size along one dimension
- * \param stride pixel stride along the coordinate axis
- * \param block_length is the length of the pixel block along the
- * coordinate axis
- * \param is_pot if TRUE, length is a power of two
- * \param wrap_mode one of PIPE_TEX_WRAP_x
- * \param offset0 resulting relative offset for coord0
- * \param offset1 resulting relative offset for coord0 + 1
- * \param i0 resulting sub-block pixel coordinate for coord0
- * \param i1 resulting sub-block pixel coordinate for coord0 + 1
+ * Helper to compute the first coord and the weight for
+ * linear wrap repeat npot textures
*/
-static void
-lp_build_sample_wrap_linear_int(struct lp_build_sample_context *bld,
- unsigned block_length,
- LLVMValueRef coord0,
- LLVMValueRef length,
- LLVMValueRef stride,
- boolean is_pot,
- unsigned wrap_mode,
- LLVMValueRef *offset0,
- LLVMValueRef *offset1,
- LLVMValueRef *i0,
- LLVMValueRef *i1)
+void
+lp_build_coord_repeat_npot_linear(struct lp_build_sample_context *bld,
+ LLVMValueRef coord_f,
+ LLVMValueRef length_i,
+ LLVMValueRef length_f,
+ LLVMValueRef *coord0_i,
+ LLVMValueRef *weight_f)
{
- struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
+ struct lp_build_context *coord_bld = &bld->coord_bld;
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
- LLVMValueRef length_minus_one;
- LLVMValueRef lmask, umask, mask;
-
- if (block_length != 1) {
- /*
- * If the pixel block covers more than one pixel then there is no easy
- * way to calculate offset1 relative to offset0. Instead, compute them
- * independently.
- */
-
- LLVMValueRef coord1;
-
- lp_build_sample_wrap_nearest_int(bld,
- block_length,
- coord0,
- length,
- stride,
- is_pot,
- wrap_mode,
- offset0, i0);
-
- coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
-
- lp_build_sample_wrap_nearest_int(bld,
- block_length,
- coord1,
- length,
- stride,
- is_pot,
- wrap_mode,
- offset1, i1);
-
- return;
- }
-
+ LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
+ LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length_i,
+ int_coord_bld->one);
+ LLVMValueRef mask;
+ /* wrap with normalized floats is just fract */
+ coord_f = lp_build_fract(coord_bld, coord_f);
+ /* mul by size and subtract 0.5 */
+ coord_f = lp_build_mul(coord_bld, coord_f, length_f);
+ coord_f = lp_build_sub(coord_bld, coord_f, half);
/*
- * Scalar pixels -- try to compute offset0 and offset1 with a single stride
- * multiplication.
+ * we avoided the 0.5/length division before the repeat wrap,
+ * now need to fix up edge cases with selects
*/
-
- *i0 = uint_coord_bld->zero;
- *i1 = uint_coord_bld->zero;
-
- length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
-
- switch(wrap_mode) {
- case PIPE_TEX_WRAP_REPEAT:
- if (is_pot) {
- coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
- }
- else {
- /* Signed remainder won't give the right results for negative
- * dividends but unsigned remainder does.*/
- coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
- }
-
- mask = lp_build_compare(bld->builder, int_coord_bld->type,
- PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
-
- *offset0 = lp_build_mul(uint_coord_bld, coord0, stride);
- *offset1 = LLVMBuildAnd(bld->builder,
- lp_build_add(uint_coord_bld, *offset0, stride),
- mask, "");
- break;
-
- case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
- lmask = lp_build_compare(int_coord_bld->builder, int_coord_bld->type,
- PIPE_FUNC_GEQUAL, coord0, int_coord_bld->zero);
- umask = lp_build_compare(int_coord_bld->builder, int_coord_bld->type,
- PIPE_FUNC_LESS, coord0, length_minus_one);
-
- coord0 = lp_build_select(int_coord_bld, lmask, coord0, int_coord_bld->zero);
- coord0 = lp_build_select(int_coord_bld, umask, coord0, length_minus_one);
-
- mask = LLVMBuildAnd(bld->builder, lmask, umask, "");
-
- *offset0 = lp_build_mul(uint_coord_bld, coord0, stride);
- *offset1 = lp_build_add(uint_coord_bld,
- *offset0,
- LLVMBuildAnd(bld->builder, stride, mask, ""));
- break;
-
- case PIPE_TEX_WRAP_CLAMP:
- case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
- case PIPE_TEX_WRAP_MIRROR_REPEAT:
- case PIPE_TEX_WRAP_MIRROR_CLAMP:
- case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
- case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
- default:
- assert(0);
- *offset0 = uint_coord_bld->zero;
- *offset1 = uint_coord_bld->zero;
- break;
- }
+ /* convert to int, compute lerp weight */
+ lp_build_ifloor_fract(coord_bld, coord_f, coord0_i, weight_f);
+ mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
+ PIPE_FUNC_LESS, *coord0_i, int_coord_bld->zero);
+ *coord0_i = lp_build_select(int_coord_bld, mask, length_minus_one, *coord0_i);
}
lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
LLVMValueRef coord,
LLVMValueRef length,
+ LLVMValueRef length_f,
+ LLVMValueRef offset,
boolean is_pot,
unsigned wrap_mode,
LLVMValueRef *x0_out,
{
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 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);
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
+ LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
LLVMValueRef coord0, coord1, weight;
switch(wrap_mode) {
case PIPE_TEX_WRAP_REPEAT:
- /* mul by size and subtract 0.5 */
- coord = lp_build_mul(coord_bld, coord, length_f);
- coord = lp_build_sub(coord_bld, coord, half);
- /* convert to int */
- coord0 = lp_build_ifloor(coord_bld, coord);
- coord1 = lp_build_add(uint_coord_bld, coord0, uint_coord_bld->one);
- /* compute lerp weight */
- weight = lp_build_fract(coord_bld, coord);
- /* repeat wrap */
if (is_pot) {
- coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
- coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
+ /* mul by size and subtract 0.5 */
+ 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);
+ coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
+ /* repeat wrap */
+ coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, "");
+ coord1 = LLVMBuildAnd(builder, coord1, length_minus_one, "");
}
else {
- /* Signed remainder won't give the right results for negative
- * dividends but unsigned remainder does.*/
- coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
- coord1 = LLVMBuildURem(bld->builder, coord1, length, "");
+ LLVMValueRef mask;
+ 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);
+ }
+ lp_build_coord_repeat_npot_linear(bld, coord,
+ length, length_f,
+ &coord0, &weight);
+ mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
+ PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
+ coord1 = LLVMBuildAnd(builder,
+ lp_build_add(int_coord_bld, coord0, int_coord_bld->one),
+ mask, "");
}
break;
case PIPE_TEX_WRAP_CLAMP:
- if (bld->static_state->normalized_coords) {
+ if (bld->static_sampler_state->normalized_coords) {
/* scale coord to length */
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);
+ }
/* clamp to [0, length] */
coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
coord = lp_build_sub(coord_bld, coord, half);
- weight = lp_build_fract(coord_bld, coord);
- coord0 = lp_build_ifloor(coord_bld, coord);
+ /* convert to int, compute lerp weight */
+ lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
break;
case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
- if (bld->static_state->normalized_coords) {
- /* clamp to [0,1] */
- coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, coord_bld->one);
- /* mul by tex size and subtract 0.5 */
- coord = lp_build_mul(coord_bld, coord, length_f);
- coord = lp_build_sub(coord_bld, coord, half);
- }
- else {
- LLVMValueRef min, max;
- /* clamp to [0.5, length - 0.5] */
- min = half;
- max = lp_build_sub(coord_bld, length_f, min);
- coord = lp_build_clamp(coord_bld, coord, min, max);
- }
- /* compute lerp weight */
- weight = lp_build_fract(coord_bld, coord);
- /* coord0 = floor(coord); */
- coord0 = lp_build_ifloor(coord_bld, coord);
- coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
- /* coord0 = max(coord0, 0) */
- coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
- /* coord1 = min(coord1, length-1) */
- coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
- break;
-
- case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
{
- LLVMValueRef min, max;
- if (bld->static_state->normalized_coords) {
- /* scale coord to length */
+ struct lp_build_context abs_coord_bld = bld->coord_bld;
+ abs_coord_bld.type.sign = FALSE;
+
+ if (bld->static_sampler_state->normalized_coords) {
+ /* mul by tex size */
coord = lp_build_mul(coord_bld, coord, length_f);
}
- /* clamp to [-0.5, length + 0.5] */
- 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);
+ if (offset) {
+ offset = lp_build_int_to_float(coord_bld, offset);
+ coord = lp_build_add(coord_bld, coord, offset);
+ }
+
+ /* clamp to length max */
+ coord = lp_build_min(coord_bld, coord, length_f);
+ /* subtract 0.5 */
coord = lp_build_sub(coord_bld, coord, half);
- /* compute lerp weight */
- weight = lp_build_fract(coord_bld, coord);
- /* convert to int */
- coord0 = lp_build_ifloor(coord_bld, coord);
+ /* clamp to [0, length - 0.5] */
+ coord = lp_build_max(coord_bld, coord, coord_bld->zero);
+ /* convert to int, compute lerp weight */
+ lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
+ /* coord1 = min(coord1, length-1) */
+ coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
+ break;
+ }
+
+ case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
+ if (bld->static_sampler_state->normalized_coords) {
+ /* scale coord to length */
+ 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);
}
+ /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */
+ /* can skip clamp (though might not work for very large coord values */
+ coord = lp_build_sub(coord_bld, coord, half);
+ /* convert to int, compute lerp weight */
+ lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
+ coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
break;
case PIPE_TEX_WRAP_MIRROR_REPEAT:
/* 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);
+ }
- /* compute lerp weight */
- weight = lp_build_fract(coord_bld, coord);
-
- /* convert to int coords */
- coord0 = lp_build_ifloor(coord_bld, coord);
+ /* convert to int, compute lerp weight */
+ lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
/* coord0 = max(coord0, 0) */
break;
case PIPE_TEX_WRAP_MIRROR_CLAMP:
- coord = lp_build_abs(coord_bld, coord);
-
- if (bld->static_state->normalized_coords) {
+ if (bld->static_sampler_state->normalized_coords) {
/* scale coord to length */
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);
+ }
+ coord = lp_build_abs(coord_bld, coord);
/* clamp to [0, length] */
coord = lp_build_min(coord_bld, coord, length_f);
coord = lp_build_sub(coord_bld, coord, half);
- weight = lp_build_fract(coord_bld, coord);
- coord0 = lp_build_ifloor(coord_bld, coord);
+ /* convert to int, compute lerp weight */
+ lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
break;
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
{
- LLVMValueRef min, max;
+ struct lp_build_context abs_coord_bld = bld->coord_bld;
+ abs_coord_bld.type.sign = FALSE;
- coord = lp_build_abs(coord_bld, coord);
-
- if (bld->static_state->normalized_coords) {
+ if (bld->static_sampler_state->normalized_coords) {
/* scale coord to length */
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);
+ }
+ coord = lp_build_abs(coord_bld, coord);
- /* clamp to [0.5, length - 0.5] */
- min = half;
- max = lp_build_sub(coord_bld, length_f, min);
- coord = lp_build_clamp(coord_bld, coord, min, max);
-
+ /* clamp to length max */
+ coord = lp_build_min(coord_bld, coord, length_f);
+ /* subtract 0.5 */
coord = lp_build_sub(coord_bld, coord, half);
+ /* clamp to [0, length - 0.5] */
+ coord = lp_build_max(coord_bld, coord, coord_bld->zero);
- weight = lp_build_fract(coord_bld, coord);
- coord0 = lp_build_ifloor(coord_bld, coord);
+ /* convert to int, compute lerp weight */
+ lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
+ /* coord1 = min(coord1, length-1) */
+ coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
}
break;
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
{
- LLVMValueRef min, max;
-
- coord = lp_build_abs(coord_bld, coord);
-
- if (bld->static_state->normalized_coords) {
+ if (bld->static_sampler_state->normalized_coords) {
/* scale coord to length */
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);
+ }
+ coord = lp_build_abs(coord_bld, coord);
- /* clamp to [-0.5, length + 0.5] */
- min = lp_build_negate(coord_bld, half);
- max = lp_build_sub(coord_bld, length_f, min);
- coord = lp_build_clamp(coord_bld, coord, min, max);
-
+ /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
+ /* skip clamp - always positive, and other side
+ only potentially matters for very large coords */
coord = lp_build_sub(coord_bld, coord, half);
- weight = lp_build_fract(coord_bld, coord);
- coord0 = lp_build_ifloor(coord_bld, coord);
+ /* convert to int, compute lerp weight */
+ lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
}
break;
/**
* Build LLVM code for texture wrap mode for nearest filtering.
* \param coord the incoming texcoord (nominally in [0,1])
- * \param length the texture size along one dimension, as int
+ * \param length the texture size along one dimension, as int vector
+ * \param length_f the texture size along one dimension, as float vector
+ * \param offset texel offset along one dimension (as int vector)
* \param is_pot if TRUE, length is a power of two
* \param wrap_mode one of PIPE_TEX_WRAP_x
*/
lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
LLVMValueRef coord,
LLVMValueRef length,
+ LLVMValueRef length_f,
+ LLVMValueRef offset,
boolean is_pot,
unsigned wrap_mode)
{
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 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);
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
LLVMValueRef icoord;
switch(wrap_mode) {
case PIPE_TEX_WRAP_REPEAT:
- coord = lp_build_mul(coord_bld, coord, length_f);
- icoord = lp_build_ifloor(coord_bld, coord);
- if (is_pot)
- icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
- else
- /* Signed remainder won't give the right results for negative
- * dividends but unsigned remainder does.*/
- icoord = LLVMBuildURem(bld->builder, icoord, length, "");
+ if (is_pot) {
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ icoord = lp_build_ifloor(coord_bld, coord);
+ if (offset) {
+ icoord = lp_build_add(int_coord_bld, icoord, offset);
+ }
+ icoord = LLVMBuildAnd(builder, icoord, length_minus_one, "");
+ }
+ else {
+ 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);
+ }
+ /* take fraction, unnormalize */
+ coord = lp_build_fract_safe(coord_bld, coord);
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ icoord = lp_build_itrunc(coord_bld, coord);
+ }
break;
case PIPE_TEX_WRAP_CLAMP:
case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
- if (bld->static_state->normalized_coords) {
+ if (bld->static_sampler_state->normalized_coords) {
/* scale coord to length */
coord = lp_build_mul(coord_bld, coord, length_f);
}
/* floor */
- icoord = lp_build_ifloor(coord_bld, coord);
+ /* 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,
break;
case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
- /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
- {
- LLVMValueRef min, max;
-
- if (bld->static_state->normalized_coords) {
- /* scale coord to length */
- coord = lp_build_mul(coord_bld, coord, length_f);
- }
-
- icoord = lp_build_ifloor(coord_bld, coord);
-
- /* clamp to [-1, length] */
- min = lp_build_negate(int_coord_bld, int_coord_bld->one);
- max = length;
- icoord = lp_build_clamp(int_coord_bld, icoord, min, max);
+ if (bld->static_sampler_state->normalized_coords) {
+ /* scale coord to length */
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ }
+ /* no clamp necessary, border masking will handle this */
+ icoord = lp_build_ifloor(coord_bld, coord);
+ if (offset) {
+ icoord = lp_build_add(int_coord_bld, icoord, offset);
}
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 */
- assert(bld->static_state->normalized_coords);
+ assert(bld->static_sampler_state->normalized_coords);
coord = lp_build_mul(coord_bld, coord, length_f);
- icoord = lp_build_ifloor(coord_bld, coord);
+ /* itrunc == ifloor here */
+ icoord = lp_build_itrunc(coord_bld, coord);
/* clamp to [0, length - 1] */
icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
case PIPE_TEX_WRAP_MIRROR_CLAMP:
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
- coord = lp_build_abs(coord_bld, coord);
-
- if (bld->static_state->normalized_coords) {
+ if (bld->static_sampler_state->normalized_coords) {
/* scale coord to length */
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);
+ }
+ coord = lp_build_abs(coord_bld, coord);
- icoord = lp_build_ifloor(coord_bld, coord);
+ /* itrunc == ifloor here */
+ icoord = lp_build_itrunc(coord_bld, coord);
/* clamp to [0, length - 1] */
icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
break;
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
- coord = lp_build_abs(coord_bld, coord);
-
- if (bld->static_state->normalized_coords) {
+ if (bld->static_sampler_state->normalized_coords) {
/* scale coord to length */
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);
+ }
+ coord = lp_build_abs(coord_bld, coord);
- icoord = lp_build_ifloor(coord_bld, coord);
-
- /* clamp to [0, length] */
- icoord = lp_build_min(int_coord_bld, icoord, length);
+ /* itrunc == ifloor here */
+ icoord = lp_build_itrunc(coord_bld, coord);
break;
default:
/**
- * 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 = LLVMBuildLShr(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 ddx partial derivatives of (s, t, r, q) with respect to X
- * \param ddy partial derivatives of (s, t, r, q) with respect to Y
- * \param lod_bias optional float vector with the shader lod bias
- * \param explicit_lod optional float vector with the explicit lod
- * \param width scalar int texture width
- * \param height scalar int texture height
- * \param depth scalar int texture depth
- *
- * XXX: The resulting lod is scalar, so ignore all but the first element of
- * derivatives, lod_bias, etc that are passed by the shader.
- */
-static LLVMValueRef
-lp_build_lod_selector(struct lp_build_sample_context *bld,
- const LLVMValueRef ddx[4],
- const LLVMValueRef ddy[4],
- LLVMValueRef lod_bias, /* optional */
- LLVMValueRef explicit_lod, /* optional */
- 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 {
- struct lp_build_context *float_bld = &bld->float_bld;
- LLVMValueRef sampler_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 lod;
-
- if (explicit_lod) {
- lod = LLVMBuildExtractElement(bld->builder, explicit_lod,
- index0, "");
- }
- else {
- const int dims = texture_dims(bld->static_state->target);
- LLVMValueRef dsdx, dsdy;
- LLVMValueRef dtdx = NULL, dtdy = NULL, drdx = NULL, drdy = NULL;
- LLVMValueRef rho;
-
- dsdx = LLVMBuildExtractElement(bld->builder, ddx[0], index0, "dsdx");
- dsdx = lp_build_abs(float_bld, dsdx);
- dsdy = LLVMBuildExtractElement(bld->builder, ddy[0], index0, "dsdy");
- dsdy = lp_build_abs(float_bld, dsdy);
- if (dims > 1) {
- dtdx = LLVMBuildExtractElement(bld->builder, ddx[1], index0, "dtdx");
- dtdx = lp_build_abs(float_bld, dtdx);
- dtdy = LLVMBuildExtractElement(bld->builder, ddy[1], index0, "dtdy");
- dtdy = lp_build_abs(float_bld, dtdy);
- if (dims > 2) {
- drdx = LLVMBuildExtractElement(bld->builder, ddx[2], index0, "drdx");
- drdx = lp_build_abs(float_bld, drdx);
- drdy = LLVMBuildExtractElement(bld->builder, ddy[2], index0, "drdy");
- 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 = LLVMBuildFMul(bld->builder,
- lp_build_max(float_bld, dsdx, dsdy),
- lp_build_int_to_float(float_bld, width), "");
- if (dims > 1) {
- LLVMValueRef max;
- max = LLVMBuildFMul(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 = LLVMBuildFMul(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 shader lod bias */
- if (lod_bias) {
- lod_bias = LLVMBuildExtractElement(bld->builder, lod_bias,
- index0, "");
- lod = LLVMBuildFAdd(bld->builder, lod, lod_bias, "shader_lod_bias");
- }
- }
-
- /* add sampler lod bias */
- lod = LLVMBuildFAdd(bld->builder, lod, sampler_lod_bias, "sampler_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_nearest_mip_level(struct lp_build_sample_context *bld,
- unsigned unit,
- LLVMValueRef lod,
- LLVMValueRef *level_out)
-{
- struct lp_build_context *float_bld = &bld->float_bld;
- struct lp_build_context *int_bld = &bld->int_bld;
- LLVMValueRef last_level, level;
-
- LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
-
- 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);
-
- /* clamp level to legal range of levels */
- *level_out = lp_build_clamp(int_bld, level, 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 *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 */
- level = lp_build_add(int_bld, level, int_bld->one);
- *level1_out = lp_build_clamp(int_bld, level,
- int_bld->zero,
- 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].
+ * 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,
+ unsigned sampler_unit,
+ LLVMValueRef size,
LLVMValueRef row_stride_vec,
LLVMValueRef img_stride_vec,
LLVMValueRef data_ptr,
+ LLVMValueRef mipoffsets,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
+ const LLVMValueRef *offsets,
LLVMValueRef colors_out[4])
{
- const int dims = texture_dims(bld->static_state->target);
- LLVMValueRef x, y, z;
+ const unsigned dims = bld->dims;
+ LLVMValueRef width_vec;
+ LLVMValueRef height_vec;
+ LLVMValueRef depth_vec;
+ LLVMValueRef flt_size;
+ LLVMValueRef flt_width_vec;
+ LLVMValueRef flt_height_vec;
+ LLVMValueRef flt_depth_vec;
+ LLVMValueRef x, y = NULL, z = NULL;
+
+ lp_build_extract_image_sizes(bld,
+ &bld->int_size_bld,
+ bld->int_coord_type,
+ size,
+ &width_vec, &height_vec, &depth_vec);
+
+ flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
+
+ lp_build_extract_image_sizes(bld,
+ &bld->float_size_bld,
+ bld->coord_type,
+ flt_size,
+ &flt_width_vec, &flt_height_vec, &flt_depth_vec);
/*
* Compute integer texcoords.
*/
- x = lp_build_sample_wrap_nearest(bld, s, width_vec,
- bld->static_state->pot_width,
- bld->static_state->wrap_s);
+ x = lp_build_sample_wrap_nearest(bld, s, width_vec, flt_width_vec, offsets[0],
+ bld->static_texture_state->pot_width,
+ bld->static_sampler_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);
+ y = lp_build_sample_wrap_nearest(bld, t, height_vec, flt_height_vec, offsets[1],
+ bld->static_texture_state->pot_height,
+ bld->static_sampler_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);
+ z = lp_build_sample_wrap_nearest(bld, r, depth_vec, flt_depth_vec, offsets[2],
+ bld->static_texture_state->pot_depth,
+ bld->static_sampler_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;
+ 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 = r;
+ lp_build_name(z, "tex.z.layer");
}
/*
* Get texture colors.
*/
- lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ lp_build_sample_texel_soa(bld, sampler_unit,
+ width_vec, height_vec, depth_vec,
x, y, z,
row_stride_vec, img_stride_vec,
- data_ptr, colors_out);
+ data_ptr, mipoffsets, colors_out);
}
*/
static void
lp_build_sample_image_linear(struct lp_build_sample_context *bld,
- LLVMValueRef width_vec,
- LLVMValueRef height_vec,
- LLVMValueRef depth_vec,
+ unsigned sampler_unit,
+ LLVMValueRef size,
LLVMValueRef row_stride_vec,
LLVMValueRef img_stride_vec,
LLVMValueRef data_ptr,
+ LLVMValueRef mipoffsets,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
+ const LLVMValueRef *offsets,
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;
+ const unsigned dims = bld->dims;
+ LLVMValueRef width_vec;
+ LLVMValueRef height_vec;
+ LLVMValueRef depth_vec;
+ LLVMValueRef flt_size;
+ LLVMValueRef flt_width_vec;
+ LLVMValueRef flt_height_vec;
+ LLVMValueRef flt_depth_vec;
+ LLVMValueRef x0, y0 = NULL, z0 = NULL, x1, y1 = NULL, z1 = NULL;
+ LLVMValueRef s_fpart, t_fpart = NULL, r_fpart = NULL;
LLVMValueRef neighbors[2][2][4];
int chan;
+ lp_build_extract_image_sizes(bld,
+ &bld->int_size_bld,
+ bld->int_coord_type,
+ size,
+ &width_vec, &height_vec, &depth_vec);
+
+ flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
+
+ lp_build_extract_image_sizes(bld,
+ &bld->float_size_bld,
+ bld->coord_type,
+ flt_size,
+ &flt_width_vec, &flt_height_vec, &flt_depth_vec);
+
/*
* Compute integer texcoords.
*/
- lp_build_sample_wrap_linear(bld, s, width_vec,
- bld->static_state->pot_width,
- bld->static_state->wrap_s,
+ lp_build_sample_wrap_linear(bld, s, 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, t, height_vec,
- bld->static_state->pot_height,
- bld->static_state->wrap_t,
+ lp_build_sample_wrap_linear(bld, t, 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 (dims == 3) {
- lp_build_sample_wrap_linear(bld, r, depth_vec,
- bld->static_state->pot_depth,
- bld->static_state->wrap_r,
+ lp_build_sample_wrap_linear(bld, r, 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");
}
- 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;
+ 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 = r; /* cube face or array layer */
+ lp_build_name(z0, "tex.z0.layer");
+ lp_build_name(z1, "tex.z1.layer");
}
+
/*
* Get texture colors.
*/
/* get x0/x1 texels */
- lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ lp_build_sample_texel_soa(bld, sampler_unit,
+ 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,
+ data_ptr, mipoffsets, neighbors[0][0]);
+ lp_build_sample_texel_soa(bld, sampler_unit,
+ width_vec, height_vec, depth_vec,
x1, y0, z0,
row_stride_vec, img_stride_vec,
- data_ptr, neighbors[0][1]);
+ data_ptr, mipoffsets, 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]);
+ neighbors[0][1][chan],
+ 0);
}
}
else {
LLVMValueRef colors0[4];
/* get x0/x1 texels at y1 */
- lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ lp_build_sample_texel_soa(bld, sampler_unit,
+ 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,
+ data_ptr, mipoffsets, neighbors[1][0]);
+ lp_build_sample_texel_soa(bld, sampler_unit,
+ width_vec, height_vec, depth_vec,
x1, y1, z0,
row_stride_vec, img_stride_vec,
- data_ptr, neighbors[1][1]);
+ data_ptr, mipoffsets, neighbors[1][1]);
/* Bilinear interpolate the four samples from the 2D image / 3D slice */
for (chan = 0; chan < 4; chan++) {
neighbors[0][0][chan],
neighbors[0][1][chan],
neighbors[1][0][chan],
- neighbors[1][1][chan]);
+ neighbors[1][1][chan],
+ 0);
}
if (dims == 3) {
LLVMValueRef colors1[4];
/* get x0/x1/y0/y1 texels at z1 */
- lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ lp_build_sample_texel_soa(bld, sampler_unit,
+ 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,
+ data_ptr, mipoffsets, neighbors1[0][0]);
+ lp_build_sample_texel_soa(bld, sampler_unit,
+ 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,
+ data_ptr, mipoffsets, neighbors1[0][1]);
+ lp_build_sample_texel_soa(bld, sampler_unit,
+ 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,
+ data_ptr, mipoffsets, neighbors1[1][0]);
+ lp_build_sample_texel_soa(bld, sampler_unit,
+ width_vec, height_vec, depth_vec,
x1, y1, z1,
row_stride_vec, img_stride_vec,
- data_ptr, neighbors1[1][1]);
+ data_ptr, mipoffsets, neighbors1[1][1]);
/* Bilinear interpolate the four samples from the second Z slice */
for (chan = 0; chan < 4; chan++) {
neighbors1[0][0][chan],
neighbors1[0][1][chan],
neighbors1[1][0][chan],
- neighbors1[1][1][chan]);
+ neighbors1[1][1][chan],
+ 0);
}
/* 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]);
+ colors0[chan], colors1[chan],
+ 0);
}
}
else {
}
-/** 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_div(coord_bld, negHalf, 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 compute 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
*/
static void
lp_build_sample_mipmap(struct lp_build_sample_context *bld,
+ unsigned sampler_unit,
unsigned img_filter,
unsigned mip_filter,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
+ const LLVMValueRef *offsets,
+ LLVMValueRef ilevel0,
+ LLVMValueRef ilevel1,
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)
{
+ 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];
- int chan;
+ unsigned chan;
+ /* sample the first mipmap level */
+ lp_build_mipmap_level_sizes(bld, ilevel0,
+ &size0,
+ &row_stride0_vec, &img_stride0_vec);
+ if (bld->num_lods == 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);
+ }
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
- /* sample the first mipmap level */
- lp_build_sample_image_nearest(bld,
- width0_vec, height0_vec, depth0_vec,
+ lp_build_sample_image_nearest(bld, sampler_unit,
+ size0,
row_stride0_vec, img_stride0_vec,
- data_ptr0, s, t, r, colors0);
-
- if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
- /* sample the second mipmap level */
- lp_build_sample_image_nearest(bld,
- width1_vec, height1_vec, depth1_vec,
- row_stride1_vec, img_stride1_vec,
- data_ptr1, s, t, r, colors1);
- }
+ data_ptr0, mipoff0, s, t, r, offsets,
+ colors0);
}
else {
assert(img_filter == PIPE_TEX_FILTER_LINEAR);
-
- /* sample the first mipmap level */
- lp_build_sample_image_linear(bld,
- width0_vec, height0_vec, depth0_vec,
+ lp_build_sample_image_linear(bld, sampler_unit,
+ size0,
row_stride0_vec, img_stride0_vec,
- data_ptr0, s, t, r, colors0);
+ data_ptr0, mipoff0, s, t, r, offsets,
+ colors0);
+ }
- if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
- /* sample the second mipmap level */
- lp_build_sample_image_linear(bld,
- width1_vec, height1_vec, depth1_vec,
- row_stride1_vec, img_stride1_vec,
- data_ptr1, s, t, r, colors1);
- }
+ /* 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) {
- /* 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]);
+ struct lp_build_if_state if_ctx;
+ LLVMValueRef need_lerp;
+
+ /* need_lerp = lod_fpart > 0 */
+ if (bld->num_lods == 1) {
+ need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
+ lod_fpart, bld->levelf_bld.zero,
+ "need_lerp");
}
- }
- else {
- /* use first/only level's colors */
- for (chan = 0; chan < 4; chan++) {
- colors_out[chan] = colors0[chan];
+ else {
+ /*
+ * We'll do mip filtering if any of the quads (or individual
+ * pixel in case of per-pixel lod) need it.
+ * It might be better to split the vectors here and only fetch/filter
+ * quads which need it.
+ */
+ /*
+ * We unfortunately need to clamp lod_fpart here since we can get
+ * negative values which would screw up filtering if not all
+ * lod_fpart values have same sign.
+ */
+ 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,
+ PIPE_FUNC_GREATER,
+ lod_fpart, bld->levelf_bld.zero);
+ need_lerp = lp_build_any_true_range(&bld->leveli_bld, bld->num_lods, need_lerp);
}
+
+ lp_build_if(&if_ctx, bld->gallivm, need_lerp);
+ {
+ /* sample the second mipmap level */
+ lp_build_mipmap_level_sizes(bld, ilevel1,
+ &size1,
+ &row_stride1_vec, &img_stride1_vec);
+ if (bld->num_lods == 1) {
+ data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
+ }
+ else {
+ data_ptr1 = bld->base_ptr;
+ mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
+ }
+ if (img_filter == PIPE_TEX_FILTER_NEAREST) {
+ lp_build_sample_image_nearest(bld, sampler_unit,
+ size1,
+ row_stride1_vec, img_stride1_vec,
+ data_ptr1, mipoff1, s, t, r, offsets,
+ colors1);
+ }
+ else {
+ lp_build_sample_image_linear(bld, sampler_unit,
+ size1,
+ row_stride1_vec, img_stride1_vec,
+ data_ptr1, mipoff1, s, t, r, 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->levelf_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);
}
}
+/**
+ * Build (per-coord) layer value.
+ * Either clamp layer to valid values or fill in optional out_of_bounds
+ * value and just return value unclamped.
+ */
+static LLVMValueRef
+lp_build_layer_coord(struct lp_build_sample_context *bld,
+ unsigned texture_unit,
+ 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);
+
+ if (out_of_bounds) {
+ LLVMValueRef out1, out;
+ 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);
+ *out_of_bounds = lp_build_or(int_coord_bld, out, out1);
+ return layer;
+ }
+ else {
+ LLVMValueRef maxlayer;
+ maxlayer = lp_build_sub(&bld->int_bld, num_layers, bld->int_bld.one);
+ maxlayer = lp_build_broadcast_scalar(int_coord_bld, maxlayer);
+ return lp_build_clamp(int_coord_bld, layer, int_coord_bld->zero, maxlayer);
+ }
+}
+
/**
- * General texture sampling codegen.
- * This function handles texture sampling for all texture targets (1D,
- * 2D, 3D, cube) and all filtering modes.
+ * Calculate cube face, lod, mip levels.
*/
static void
-lp_build_sample_general(struct lp_build_sample_context *bld,
- unsigned unit,
- LLVMValueRef s,
- LLVMValueRef t,
- LLVMValueRef r,
- const LLVMValueRef *ddx,
- const LLVMValueRef *ddy,
- LLVMValueRef lod_bias, /* optional */
- LLVMValueRef explicit_lod, /* optional */
- LLVMValueRef width,
- LLVMValueRef height,
- LLVMValueRef depth,
- LLVMValueRef width_vec,
- LLVMValueRef height_vec,
- LLVMValueRef depth_vec,
- LLVMValueRef row_stride_array,
- LLVMValueRef img_stride_array,
- LLVMValueRef data_array,
- LLVMValueRef *colors_out)
+lp_build_sample_common(struct lp_build_sample_context *bld,
+ unsigned texture_index,
+ unsigned sampler_index,
+ LLVMValueRef *s,
+ LLVMValueRef *t,
+ LLVMValueRef *r,
+ const struct lp_derivatives *derivs, /* optional */
+ LLVMValueRef lod_bias, /* optional */
+ LLVMValueRef explicit_lod, /* optional */
+ LLVMValueRef *lod_ipart,
+ LLVMValueRef *lod_fpart,
+ LLVMValueRef *ilevel0,
+ LLVMValueRef *ilevel1)
{
- 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 = NULL, lod_fpart = NULL;
- LLVMValueRef ilevel0, ilevel1 = NULL, ilevel0_vec, ilevel1_vec = NULL;
- 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 = NULL;
- LLVMValueRef face_ddx[4], face_ddy[4];
+ const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
+ const unsigned min_filter = bld->static_sampler_state->min_img_filter;
+ 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;
/*
printf("%s mip %d min %d mag %d\n", __FUNCTION__,
*/
/*
- * Choose cube face, recompute texcoords and derivatives for the chosen face.
+ * Choose cube face, recompute texcoords for the chosen face and
+ * compute rho here too (as it requires transform of derivatives).
*/
- if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
+ if (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 */
+ 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, *s, *t, *r, derivs, &face, &face_s, &face_t,
+ &cube_rho, need_derivs);
+ *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 */
-
- /* recompute ddx, ddy using the new (s,t) face texcoords */
- face_ddx[0] = lp_build_ddx(&bld->coord_bld, s);
- face_ddx[1] = lp_build_ddx(&bld->coord_bld, t);
- face_ddx[2] = NULL;
- face_ddx[3] = NULL;
- face_ddy[0] = lp_build_ddy(&bld->coord_bld, s);
- face_ddy[1] = lp_build_ddy(&bld->coord_bld, t);
- face_ddy[2] = NULL;
- face_ddy[3] = NULL;
- ddx = face_ddx;
- ddy = face_ddy;
+ *r = face; /* vec */
+ }
+ else if (target == PIPE_TEXTURE_1D_ARRAY) {
+ *r = lp_build_iround(&bld->coord_bld, *t);
+ *r = lp_build_layer_coord(bld, texture_index, *r, NULL);
+ }
+ else if (target == PIPE_TEXTURE_2D_ARRAY) {
+ *r = lp_build_iround(&bld->coord_bld, *r);
+ *r = lp_build_layer_coord(bld, texture_index, *r, NULL);
}
/*
/* 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, ddx, ddy,
- lod_bias, explicit_lod,
- width, height, depth);
+ lp_build_lod_selector(bld, texture_index, sampler_index,
+ *s, *t, *r, cube_rho,
+ derivs, lod_bias, explicit_lod,
+ mip_filter,
+ lod_ipart, lod_fpart);
+ } else {
+ *lod_ipart = bld->leveli_bld.zero;
}
/*
- * Compute integer mipmap level(s) to fetch texels from.
+ * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
*/
- if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
+ switch (mip_filter) {
+ default:
+ assert(0 && "bad mip_filter value in lp_build_sample_soa()");
+ /* fall-through */
+ case PIPE_TEX_MIPFILTER_NONE:
/* always use mip level 0 */
- if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
+ 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.
*/
- lod = lp_build_const_elem(bld->coord_bld.type, 0.0);
- lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
- }
- else {
- ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
- }
- }
- else {
- assert(lod);
- if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
- lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
+ assert(*lod_ipart);
+ lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0, NULL);
}
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);
+ 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;
}
+ break;
+ case PIPE_TEX_MIPFILTER_NEAREST:
+ 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_fpart);
+ lp_build_linear_mip_levels(bld, texture_index,
+ *lod_ipart, lod_fpart,
+ ilevel0, ilevel1);
+ break;
}
+}
- /*
- * 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);
+/**
+ * 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 sampler_unit,
+ LLVMValueRef s,
+ LLVMValueRef t,
+ LLVMValueRef r,
+ const LLVMValueRef *offsets,
+ LLVMValueRef lod_ipart,
+ 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 unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
+ const unsigned min_filter = bld->static_sampler_state->min_img_filter;
+ const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
+ LLVMValueRef texels[4];
+ unsigned chan;
/*
- * Compute width, height at mipmap level 'ilevel0'
+ * Get/interpolate texture colors.
*/
- 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_get_level_stride_vec(bld,
- img_stride_array,
- ilevel0);
- 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_get_level_stride_vec(bld,
- img_stride_array,
- ilevel1);
- if (dims ==3) {
- depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_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);
+ for (chan = 0; chan < 4; ++chan) {
+ texels[chan] = lp_build_alloca(bld->gallivm, bld->texel_bld.vec_type, "");
+ lp_build_name(texels[chan], "sampler%u_texel_%c_var", sampler_unit, "xyzw"[chan]);
}
- /*
- * 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);
+ /* no need to distinguish between minification and magnification */
+ lp_build_sample_mipmap(bld, sampler_unit,
+ min_filter, mip_filter,
+ s, t, r, 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.
+ * 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]);
+ /*
+ * XXX this should to 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?)
+ */
+ if (bld->num_lods > 1)
+ lod_ipart = LLVMBuildExtractElement(builder, lod_ipart,
+ lp_build_const_int32(bld->gallivm, 0), "");
- /* minify = lod > 0.0 */
- minify = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
- lod, float_bld->zero, "");
+ /* minify = lod >= 0.0 */
+ minify = LLVMBuildICmp(builder, LLVMIntSGE,
+ lod_ipart, int_bld->zero, "");
- lp_build_if(&if_ctx, flow_ctx, bld->builder, minify);
+ lp_build_if(&if_ctx, bld->gallivm, 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_sample_mipmap(bld, sampler_unit,
+ min_filter, mip_filter,
+ s, t, r, offsets,
+ ilevel0, ilevel1, lod_fpart,
+ texels);
}
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_sample_mipmap(bld, sampler_unit,
+ mag_filter, PIPE_TEX_MIPFILTER_NONE,
+ s, t, r, offsets,
+ ilevel0, NULL, NULL,
+ texels);
}
lp_build_endif(&if_ctx);
+ }
- lp_build_flow_scope_end(flow_ctx);
- lp_build_flow_destroy(flow_ctx);
+ for (chan = 0; chan < 4; ++chan) {
+ colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
+ lp_build_name(colors_out[chan], "sampler%u_texel_%c", sampler_unit, "xyzw"[chan]);
}
}
-
+/**
+ * Texel fetch function.
+ * In contrast to general sampling there is no filtering, no coord minification,
+ * lod (if any) is always explicit uint, coords are uints (in terms of texel units)
+ * directly to be applied to the selected mip level (after adding texel offsets).
+ * This function handles texel fetch for all targets where texel fetch is supported
+ * (no cube maps, but 1d, 2d, 3d are supported, arrays and buffers should be too).
+ */
static void
-lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
- LLVMValueRef s,
- LLVMValueRef t,
- LLVMValueRef width,
- LLVMValueRef height,
- LLVMValueRef stride_array,
- LLVMValueRef data_array,
- LLVMValueRef texel_out[4])
+lp_build_fetch_texel(struct lp_build_sample_context *bld,
+ unsigned texture_unit,
+ const LLVMValueRef *coords,
+ LLVMValueRef explicit_lod,
+ const LLVMValueRef *offsets,
+ LLVMValueRef *colors_out)
{
- LLVMBuilderRef builder = bld->builder;
- struct lp_build_context i32, h16, u8n;
- LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
- LLVMValueRef i32_c8, i32_c128, i32_c255;
- LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
- LLVMValueRef t_ipart, t_fpart, t_fpart_lo, t_fpart_hi;
- LLVMValueRef data_ptr;
- LLVMValueRef x_stride, y_stride;
- LLVMValueRef x_offset0, x_offset1;
- LLVMValueRef y_offset0, y_offset1;
- LLVMValueRef offset[2][2];
- LLVMValueRef x_subcoord[2], y_subcoord[2];
- LLVMValueRef neighbors_lo[2][2];
- LLVMValueRef neighbors_hi[2][2];
- LLVMValueRef packed, packed_lo, packed_hi;
- LLVMValueRef unswizzled[4];
- const unsigned level = 0;
- unsigned i, j;
-
- assert(bld->static_state->target == PIPE_TEXTURE_2D
- || bld->static_state->target == PIPE_TEXTURE_RECT);
- assert(bld->static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR);
- assert(bld->static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR);
- assert(bld->static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE);
-
- 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));
-
- i32_vec_type = lp_build_vec_type(i32.type);
- h16_vec_type = lp_build_vec_type(h16.type);
- u8n_vec_type = lp_build_vec_type(u8n.type);
-
- if (bld->static_state->normalized_coords) {
- LLVMTypeRef coord_vec_type = lp_build_vec_type(bld->coord_type);
- LLVMValueRef fp_width = LLVMBuildSIToFP(bld->builder, width, coord_vec_type, "");
- LLVMValueRef fp_height = LLVMBuildSIToFP(bld->builder, height, coord_vec_type, "");
- s = lp_build_mul(&bld->coord_bld, s, fp_width);
- t = lp_build_mul(&bld->coord_bld, t, fp_height);
- }
-
- /* scale coords by 256 (8 fractional bits) */
- s = lp_build_mul_imm(&bld->coord_bld, s, 256);
- t = lp_build_mul_imm(&bld->coord_bld, t, 256);
-
- /* convert float to int */
- s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
- t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
-
- /* subtract 0.5 (add -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_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_const_int_vec(i32.type, 255);
- s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
- t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
-
- x_stride = lp_build_const_vec(bld->uint_coord_bld.type,
- bld->format_desc->block.bits/8);
-
- y_stride = lp_build_get_const_level_stride_vec(bld, stride_array, level);
-
- lp_build_sample_wrap_linear_int(bld,
- bld->format_desc->block.width,
- s_ipart, width, x_stride,
- bld->static_state->pot_width,
- bld->static_state->wrap_s,
- &x_offset0, &x_offset1,
- &x_subcoord[0], &x_subcoord[1]);
- lp_build_sample_wrap_linear_int(bld,
- bld->format_desc->block.height,
- t_ipart, height, y_stride,
- bld->static_state->pot_height,
- bld->static_state->wrap_t,
- &y_offset0, &y_offset1,
- &y_subcoord[0], &y_subcoord[1]);
-
- offset[0][0] = lp_build_add(&bld->uint_coord_bld, x_offset0, y_offset0);
- offset[0][1] = lp_build_add(&bld->uint_coord_bld, x_offset1, y_offset0);
- offset[1][0] = lp_build_add(&bld->uint_coord_bld, x_offset0, y_offset1);
- offset[1][1] = lp_build_add(&bld->uint_coord_bld, x_offset1, y_offset1);
-
- /*
- * Transform 4 x i32 in
- *
- * s_fpart = {s0, s1, s2, s3}
- *
- * into 8 x i16
- *
- * s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
- *
- * into two 8 x i16
- *
- * s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
- * s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
- *
- * and likewise for t_fpart. There is no risk of loosing precision here
- * since the fractional parts only use the lower 8bits.
- */
-
- s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
- t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
-
- {
- LLVMTypeRef elem_type = LLVMInt32Type();
- LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
- LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
- LLVMValueRef shuffle_lo;
- LLVMValueRef shuffle_hi;
-
- for(j = 0; j < h16.type.length; j += 4) {
-#ifdef PIPE_ARCH_LITTLE_ENDIAN
- unsigned subindex = 0;
-#else
- unsigned subindex = 1;
-#endif
- LLVMValueRef index;
-
- index = LLVMConstInt(elem_type, j/2 + subindex, 0);
- for(i = 0; i < 4; ++i)
- shuffles_lo[j + i] = index;
-
- index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
- for(i = 0; i < 4; ++i)
- shuffles_hi[j + i] = index;
+ struct lp_build_context *perquadi_bld = &bld->leveli_bld;
+ struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
+ unsigned dims = bld->dims, chan;
+ unsigned target = bld->static_texture_state->target;
+ boolean out_of_bound_ret_zero = TRUE;
+ LLVMValueRef size, ilevel;
+ LLVMValueRef row_stride_vec = NULL, img_stride_vec = NULL;
+ LLVMValueRef x = coords[0], y = coords[1], z = coords[2];
+ LLVMValueRef width, height, depth, i, j;
+ LLVMValueRef offset, out_of_bounds, out1;
+
+ 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) {
+ ilevel = lp_build_pack_aos_scalars(bld->gallivm, int_coord_bld->type,
+ perquadi_bld->type, explicit_lod, 0);
+ }
+ else {
+ ilevel = explicit_lod;
+ }
+ lp_build_nearest_mip_level(bld, texture_unit, ilevel, &ilevel,
+ out_of_bound_ret_zero ? &out_of_bounds : NULL);
+ }
+ else {
+ assert(bld->num_lods == 1);
+ if (bld->static_texture_state->target != PIPE_BUFFER) {
+ ilevel = bld->dynamic_state->first_level(bld->dynamic_state,
+ bld->gallivm, texture_unit);
+ }
+ else {
+ ilevel = lp_build_const_int32(bld->gallivm, 0);
+ }
+ }
+ lp_build_mipmap_level_sizes(bld, ilevel,
+ &size,
+ &row_stride_vec, &img_stride_vec);
+ lp_build_extract_image_sizes(bld, &bld->int_size_bld, int_coord_bld->type,
+ size, &width, &height, &depth);
+
+ if (target == PIPE_TEXTURE_1D_ARRAY ||
+ target == PIPE_TEXTURE_2D_ARRAY) {
+ if (target == PIPE_TEXTURE_1D_ARRAY) {
+ z = y;
+ }
+ if (out_of_bound_ret_zero) {
+ z = lp_build_layer_coord(bld, texture_unit, 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);
}
-
- shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
- shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
-
- s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_lo, "");
- t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_lo, "");
- s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_hi, "");
- t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, "");
}
- /*
- * get pointer to mipmap level 0 data
- */
- data_ptr = lp_build_get_const_mipmap_level(bld, data_array, level);
-
- /*
- * Fetch the pixels as 4 x 32bit (rgba order might differ):
- *
- * rgba0 rgba1 rgba2 rgba3
- *
- * bit cast them into 16 x u8
- *
- * r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
- *
- * unpack them into two 8 x i16:
- *
- * r0 g0 b0 a0 r1 g1 b1 a1
- * r2 g2 b2 a2 r3 g3 b3 a3
- *
- * The higher 8 bits of the resulting elements will be zero.
- */
-
- for (j = 0; j < 2; ++j) {
- for (i = 0; i < 2; ++i) {
- LLVMValueRef rgba8;
-
- if (util_format_is_rgba8_variant(bld->format_desc)) {
- /*
- * Given the format is a rgba8, just read the pixels as is,
- * without any swizzling. Swizzling will be done later.
- */
- rgba8 = lp_build_gather(bld->builder,
- bld->texel_type.length,
- bld->format_desc->block.bits,
- bld->texel_type.width,
- data_ptr, offset[j][i]);
-
- rgba8 = LLVMBuildBitCast(builder, rgba8, u8n_vec_type, "");
+ /* This is a lot like border sampling */
+ if (offsets[0]) {
+ /*
+ * coords are really unsigned, offsets are signed, but I don't think
+ * exceeding 31 bits is possible
+ */
+ x = lp_build_add(int_coord_bld, x, offsets[0]);
+ }
+ out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
+ out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
+ out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
+ out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
+ if (dims >= 2) {
+ if (offsets[1]) {
+ y = lp_build_add(int_coord_bld, y, offsets[1]);
+ }
+ out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
+ out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
+ out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
+ out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
+
+ if (dims >= 3) {
+ if (offsets[2]) {
+ z = lp_build_add(int_coord_bld, z, offsets[2]);
}
- else {
- rgba8 = lp_build_fetch_rgba_aos(bld->builder,
- bld->format_desc,
- u8n.type,
- data_ptr, offset[j][i],
- x_subcoord[i],
- y_subcoord[j]);
- }
-
- lp_build_unpack2(builder, u8n.type, h16.type,
- rgba8,
- &neighbors_lo[j][i], &neighbors_hi[j][i]);
+ out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
+ out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
+ out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
+ out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
}
}
- /*
- * Linear interpolate with 8.8 fixed point.
- */
-
- packed_lo = lp_build_lerp_2d(&h16,
- s_fpart_lo, t_fpart_lo,
- neighbors_lo[0][0],
- neighbors_lo[0][1],
- neighbors_lo[1][0],
- neighbors_lo[1][1]);
-
- packed_hi = lp_build_lerp_2d(&h16,
- s_fpart_hi, t_fpart_hi,
- neighbors_hi[0][0],
- neighbors_hi[0][1],
- neighbors_hi[1][0],
- neighbors_hi[1][1]);
+ lp_build_sample_offset(int_coord_bld,
+ bld->format_desc,
+ x, y, z, row_stride_vec, img_stride_vec,
+ &offset, &i, &j);
- packed = lp_build_pack2(builder, h16.type, u8n.type, packed_lo, packed_hi);
+ if (bld->static_texture_state->target != PIPE_BUFFER) {
+ offset = lp_build_add(int_coord_bld, offset,
+ lp_build_get_mip_offsets(bld, ilevel));
+ }
- /*
- * Convert to SoA and swizzle.
- */
+ offset = lp_build_andnot(int_coord_bld, offset, out_of_bounds);
- lp_build_rgba8_to_f32_soa(bld->builder,
- bld->texel_type,
- packed, unswizzled);
+ lp_build_fetch_rgba_soa(bld->gallivm,
+ bld->format_desc,
+ bld->texel_type,
+ bld->base_ptr, offset,
+ i, j,
+ colors_out);
- if (util_format_is_rgba8_variant(bld->format_desc)) {
- lp_build_format_swizzle_soa(bld->format_desc,
- &bld->texel_bld,
- unswizzled, texel_out);
- } else {
- texel_out[0] = unswizzled[0];
- texel_out[1] = unswizzled[1];
- texel_out[2] = unswizzled[2];
- texel_out[3] = unswizzled[3];
+ if (out_of_bound_ret_zero) {
+ /*
+ * Only needed for ARB_robust_buffer_access_behavior and d3d10.
+ * Could use min/max above instead of out-of-bounds comparisons
+ * if we don't care about the result returned for out-of-bounds.
+ */
+ for (chan = 0; chan < 4; chan++) {
+ colors_out[chan] = lp_build_select(&bld->texel_bld, out_of_bounds,
+ bld->texel_bld.zero, colors_out[chan]);
+ }
}
-
- apply_sampler_swizzle(bld, texel_out);
}
+/**
+ * Do shadow test/comparison.
+ * \param coords incoming texcoords
+ * \param texel the texel to compare against (use the X channel)
+ * Ideally this should really be done per-sample.
+ */
static void
lp_build_sample_compare(struct lp_build_sample_context *bld,
- LLVMValueRef p,
+ const LLVMValueRef *coords,
LLVMValueRef texel[4])
{
struct lp_build_context *texel_bld = &bld->texel_bld;
- LLVMValueRef res;
- unsigned chan;
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ LLVMValueRef res, p;
+ const unsigned chan = 0;
+ unsigned chan_type;
+ const struct util_format_description *format_desc;
- if(bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
+ if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE)
return;
- /* TODO: Compare before swizzling, to avoid redundant computations */
- res = NULL;
- for(chan = 0; chan < 4; ++chan) {
- LLVMValueRef cmp;
- cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]);
- cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero);
+ if (bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY ||
+ bld->static_texture_state->target == PIPE_TEXTURE_CUBE) {
+ p = coords[3];
+ }
+ else {
+ p = coords[2];
+ }
- if(res)
- res = lp_build_add(texel_bld, res, cmp);
- else
- res = cmp;
+ /* debug code */
+ if (0) {
+ LLVMValueRef indx = lp_build_const_int32(bld->gallivm, 0);
+ LLVMValueRef coord = LLVMBuildExtractElement(builder, p, indx, "");
+ LLVMValueRef tex = LLVMBuildExtractElement(builder, texel[chan], indx, "");
+ lp_build_printf(bld->gallivm, "shadow compare coord %f to texture %f\n",
+ coord, tex);
}
- assert(res);
- res = lp_build_mul(texel_bld, res, lp_build_const_vec(texel_bld->type, 0.25));
+ /* Clamp p coords to [0,1] for fixed function depth texture format */
+ format_desc = util_format_description(bld->static_texture_state->format);
+ /* not entirely sure we couldn't end up with non-valid swizzle here */
+ chan_type = format_desc->swizzle[0] <= UTIL_FORMAT_SWIZZLE_W ?
+ format_desc->channel[format_desc->swizzle[0]].type :
+ UTIL_FORMAT_TYPE_FLOAT;
+ if (chan_type != UTIL_FORMAT_TYPE_FLOAT) {
+ p = lp_build_clamp(&bld->coord_bld, p,
+ bld->coord_bld.zero, bld->coord_bld.one);
+ }
+
+ /*
+ * technically this is not entirely correct for unorm depth as the ref value
+ * should be converted to the depth format (quantization!) and comparison
+ * then done in texture format.
+ */
+
+ /* result = (p FUNC texel) ? 1 : 0 */
+ /*
+ * honor d3d10 floating point rules here, which state that comparisons
+ * are ordered except NOT_EQUAL which is unordered.
+ */
+ if (bld->static_sampler_state->compare_func != PIPE_FUNC_NOTEQUAL) {
+ res = lp_build_cmp_ordered(texel_bld, bld->static_sampler_state->compare_func,
+ p, texel[chan]);
+ }
+ else {
+ res = lp_build_cmp(texel_bld, bld->static_sampler_state->compare_func,
+ p, texel[chan]);
+ }
+ res = lp_build_select(texel_bld, res, texel_bld->one, texel_bld->zero);
- /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
- for(chan = 0; chan < 3; ++chan)
- texel[chan] = res;
+ /*
+ * returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE.
+ * This should be ok because sampler swizzle is applied on top of it.
+ */
+ texel[0] =
+ texel[1] =
+ texel[2] = res;
texel[3] = texel_bld->one;
}
* Just set texels to white instead of actually sampling the texture.
* For debugging.
*/
-static void
-lp_build_sample_nop(struct lp_build_sample_context *bld,
+void
+lp_build_sample_nop(struct gallivm_state *gallivm,
+ struct lp_type type,
+ const LLVMValueRef *coords,
LLVMValueRef texel_out[4])
{
- struct lp_build_context *texel_bld = &bld->texel_bld;
+ LLVMValueRef one = lp_build_one(gallivm, type);
unsigned chan;
for (chan = 0; chan < 4; chan++) {
- /*lp_bld_mov(texel_bld, texel, texel_bld->one);*/
- texel_out[chan] = texel_bld->one;
+ texel_out[chan] = one;
}
}
* '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 ddx partial derivatives of (s,t,r,q) with respect to x
- * \param ddy partial derivatives of (s,t,r,q) with respect to y
+ * \param is_fetch if this is a texel fetch instruction.
+ * \param derivs partial derivatives of (s,t,r,q) with respect to x and y
*/
void
-lp_build_sample_soa(LLVMBuilderRef builder,
- const struct lp_sampler_static_state *static_state,
+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,
- unsigned unit,
- unsigned num_coords,
+ boolean is_fetch,
+ unsigned texture_index,
+ unsigned sampler_index,
const LLVMValueRef *coords,
- const LLVMValueRef ddx[4],
- const LLVMValueRef ddy[4],
+ const LLVMValueRef *offsets,
+ const struct lp_derivatives *derivs, /* optional */
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
+ boolean scalar_lod,
LLVMValueRef texel_out[4])
{
+ unsigned dims = texture_dims(static_texture_state->target);
+ unsigned num_quads = type.length / 4;
+ unsigned mip_filter;
struct lp_build_sample_context bld;
- LLVMValueRef width, width_vec;
- LLVMValueRef height, height_vec;
- LLVMValueRef depth, depth_vec;
- LLVMValueRef row_stride_array, img_stride_array;
- LLVMValueRef data_array;
+ struct lp_static_sampler_state derived_sampler_state = *static_sampler_state;
+ LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef tex_width;
LLVMValueRef s;
LLVMValueRef t;
LLVMValueRef r;
if (0) {
- enum pipe_format fmt = static_state->format;
+ enum pipe_format fmt = static_texture_state->format;
debug_printf("Sample from %s\n", util_format_name(fmt));
}
/* Setup our build context */
memset(&bld, 0, sizeof bld);
- bld.builder = builder;
- bld.static_state = static_state;
+ bld.gallivm = gallivm;
+ bld.static_sampler_state = &derived_sampler_state;
+ bld.static_texture_state = static_texture_state;
bld.dynamic_state = dynamic_state;
- bld.format_desc = util_format_description(static_state->format);
+ bld.format_desc = util_format_description(static_texture_state->format);
+ bld.dims = dims;
+
+ bld.vector_width = lp_type_width(type);
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.float_size_in_type = lp_type_float(32);
+ bld.float_size_in_type.length = dims > 1 ? 4 : 1;
+ bld.int_size_in_type = lp_int_type(bld.float_size_in_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);
- lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
+ /* always using the first channel hopefully should be safe,
+ * if not things WILL break in other places anyway.
+ */
+ if (bld.format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
+ bld.format_desc->channel[0].pure_integer) {
+ if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
+ bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
+ }
+ else if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
+ bld.texel_type = lp_type_uint_vec(type.width, type.width * type.length);
+ }
+ }
+ else if (util_format_has_stencil(bld.format_desc) &&
+ !util_format_has_depth(bld.format_desc)) {
+ /* for stencil only formats, sample stencil (uint) */
+ bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
+ }
+
+ if (!static_texture_state->level_zero_only) {
+ derived_sampler_state.min_mip_filter = static_sampler_state->min_mip_filter;
+ } else {
+ derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
+ }
+ mip_filter = derived_sampler_state.min_mip_filter;
+
+ if (0) {
+ debug_printf(" .min_mip_filter = %u\n", derived_sampler_state.min_mip_filter);
+ }
+
+ /*
+ * This is all a bit complicated different paths are chosen for performance
+ * reasons.
+ * Essentially, there can be 1 lod per element, 1 lod per quad or 1 lod for
+ * everything (the last two options are equivalent for 4-wide case).
+ * If there's per-quad lod but we split to 4-wide so we can use AoS, per-quad
+ * lod is calculated then the lod value extracted afterwards so making this
+ * case basically the same as far as lod handling is concerned for the
+ * further sample/filter code as the 1 lod for everything case.
+ * Different lod handling mostly shows up when building mipmap sizes
+ * (lp_build_mipmap_level_sizes() and friends) and also in filtering
+ * (getting the fractional part of the lod to the right texels).
+ */
+
+ /*
+ * There are other situations where at least the multiple int lods could be
+ * avoided like min and max lod being equal.
+ */
+ if (explicit_lod && !scalar_lod &&
+ ((is_fetch && bld.static_texture_state->target != PIPE_BUFFER) ||
+ (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)))
+ bld.num_lods = type.length;
+ /* TODO: for true scalar_lod should only use 1 lod value */
+ else if ((is_fetch && explicit_lod && bld.static_texture_state->target != PIPE_BUFFER ) ||
+ (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
+ bld.num_lods = num_quads;
+ }
+ else {
+ bld.num_lods = 1;
+ }
+
+ bld.levelf_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;
+ }
+ 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 :
+ type.length;
+ }
+ bld.int_size_type = lp_int_type(bld.float_size_type);
+
+ lp_build_context_init(&bld.float_bld, gallivm, bld.float_type);
+ lp_build_context_init(&bld.float_vec_bld, gallivm, type);
+ lp_build_context_init(&bld.int_bld, gallivm, bld.int_type);
+ lp_build_context_init(&bld.coord_bld, gallivm, bld.coord_type);
+ lp_build_context_init(&bld.int_coord_bld, gallivm, bld.int_coord_type);
+ lp_build_context_init(&bld.int_size_in_bld, gallivm, bld.int_size_in_type);
+ lp_build_context_init(&bld.float_size_in_bld, gallivm, bld.float_size_in_type);
+ lp_build_context_init(&bld.int_size_bld, gallivm, bld.int_size_type);
+ lp_build_context_init(&bld.float_size_bld, gallivm, bld.float_size_type);
+ lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type);
+ lp_build_context_init(&bld.levelf_bld, gallivm, bld.levelf_type);
+ lp_build_context_init(&bld.leveli_bld, gallivm, bld.leveli_type);
/* Get the dynamic state */
- width = dynamic_state->width(dynamic_state, builder, unit);
- height = dynamic_state->height(dynamic_state, builder, unit);
- depth = dynamic_state->depth(dynamic_state, builder, unit);
- row_stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
- img_stride_array = dynamic_state->img_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 */
+ 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);
+ /* Note that mip_offsets is an array[level] of offsets to texture images */
s = coords[0];
t = coords[1];
r = coords[2];
- 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);
+ /* 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), "");
+ if (dims >= 2) {
+ LLVMValueRef tex_height =
+ dynamic_state->height(dynamic_state, gallivm, texture_index);
+ bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
+ tex_height, LLVMConstInt(i32t, 1, 0), "");
+ if (dims >= 3) {
+ LLVMValueRef tex_depth =
+ dynamic_state->depth(dynamic_state, gallivm, texture_index);
+ bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
+ tex_depth, LLVMConstInt(i32t, 2, 0), "");
+ }
+ }
+ }
if (0) {
/* For debug: no-op texture sampling */
- lp_build_sample_nop(&bld, texel_out);
- }
- else if (util_format_fits_8unorm(bld.format_desc) &&
- bld.format_desc->nr_channels > 1 &&
- (static_state->target == PIPE_TEXTURE_2D ||
- static_state->target == PIPE_TEXTURE_RECT) &&
- 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,
- row_stride_array, data_array, texel_out);
+ lp_build_sample_nop(gallivm,
+ bld.texel_type,
+ coords,
+ texel_out);
}
+
+ else if (is_fetch) {
+ lp_build_fetch_texel(&bld, texture_index, coords,
+ explicit_lod, offsets,
+ texel_out);
+ }
+
else {
- lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy,
- lod_bias, explicit_lod,
- width, height, depth,
- width_vec, height_vec, depth_vec,
- row_stride_array, img_stride_array,
- data_array,
- texel_out);
+ LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
+ LLVMValueRef ilevel0 = NULL, ilevel1 = NULL;
+ boolean use_aos = util_format_fits_8unorm(bld.format_desc) &&
+ lp_is_simple_wrap_mode(static_sampler_state->wrap_s) &&
+ lp_is_simple_wrap_mode(static_sampler_state->wrap_t);
+
+ 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\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);
+ }
+
+ lp_build_sample_common(&bld, texture_index, sampler_index,
+ &s, &t, &r,
+ derivs, lod_bias, explicit_lod,
+ &lod_ipart, &lod_fpart,
+ &ilevel0, &ilevel1);
+
+ /*
+ * we only try 8-wide sampling with soa as it appears to
+ * be a loss with aos with AVX (but it should work).
+ * (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,
+ s, t, r, offsets,
+ lod_ipart, lod_fpart,
+ ilevel0, ilevel1,
+ texel_out);
+ }
+
+ else {
+ lp_build_sample_general(&bld, sampler_index,
+ s, t, r, offsets,
+ lod_ipart, lod_fpart,
+ ilevel0, ilevel1,
+ texel_out);
+ }
+ }
+ else {
+ unsigned j;
+ struct lp_build_sample_context bld4;
+ struct lp_type type4 = type;
+ unsigned i;
+ LLVMValueRef texelout4[4];
+ LLVMValueRef texelouttmp[4][LP_MAX_VECTOR_LENGTH/16];
+
+ type4.length = 4;
+
+ /* Setup our build context */
+ memset(&bld4, 0, sizeof bld4);
+ bld4.gallivm = bld.gallivm;
+ bld4.static_texture_state = bld.static_texture_state;
+ bld4.static_sampler_state = bld.static_sampler_state;
+ bld4.dynamic_state = bld.dynamic_state;
+ bld4.format_desc = bld.format_desc;
+ bld4.dims = bld.dims;
+ bld4.row_stride_array = bld.row_stride_array;
+ bld4.img_stride_array = bld.img_stride_array;
+ bld4.base_ptr = bld.base_ptr;
+ bld4.mip_offsets = bld.mip_offsets;
+ bld4.int_size = bld.int_size;
+
+ bld4.vector_width = lp_type_width(type4);
+
+ bld4.float_type = lp_type_float(32);
+ bld4.int_type = lp_type_int(32);
+ bld4.coord_type = type4;
+ bld4.int_coord_type = lp_int_type(type4);
+ bld4.float_size_in_type = lp_type_float(32);
+ bld4.float_size_in_type.length = dims > 1 ? 4 : 1;
+ 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 && !scalar_lod &&
+ ((is_fetch && bld.static_texture_state->target != PIPE_BUFFER) ||
+ (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)))
+ bld4.num_lods = type4.length;
+ else
+ bld4.num_lods = 1;
+
+ bld4.levelf_type = type4;
+ /* we want native vector size to be able to use our intrinsics */
+ if (bld4.num_lods != 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 :
+ type4.length;
+ }
+ bld4.int_size_type = lp_int_type(bld4.float_size_type);
+
+ lp_build_context_init(&bld4.float_bld, gallivm, bld4.float_type);
+ lp_build_context_init(&bld4.float_vec_bld, gallivm, type4);
+ lp_build_context_init(&bld4.int_bld, gallivm, bld4.int_type);
+ lp_build_context_init(&bld4.coord_bld, gallivm, bld4.coord_type);
+ lp_build_context_init(&bld4.int_coord_bld, gallivm, bld4.int_coord_type);
+ lp_build_context_init(&bld4.int_size_in_bld, gallivm, bld4.int_size_in_type);
+ lp_build_context_init(&bld4.float_size_in_bld, gallivm, bld4.float_size_in_type);
+ lp_build_context_init(&bld4.int_size_bld, gallivm, bld4.int_size_type);
+ lp_build_context_init(&bld4.float_size_bld, gallivm, bld4.float_size_type);
+ lp_build_context_init(&bld4.texel_bld, gallivm, bld4.texel_type);
+ lp_build_context_init(&bld4.levelf_bld, gallivm, bld4.levelf_type);
+ lp_build_context_init(&bld4.leveli_bld, gallivm, bld4.leveli_type);
+
+ for (i = 0; i < num_quads; i++) {
+ LLVMValueRef s4, t4, r4;
+ LLVMValueRef lod_ipart4, lod_fpart4 = NULL;
+ LLVMValueRef ilevel04, ilevel14 = NULL;
+ LLVMValueRef offsets4[4] = { NULL };
+ unsigned num_lods = bld4.num_lods;
+
+ s4 = lp_build_extract_range(gallivm, s, 4*i, 4);
+ t4 = lp_build_extract_range(gallivm, t, 4*i, 4);
+ r4 = lp_build_extract_range(gallivm, r, 4*i, 4);
+
+ if (offsets[0]) {
+ offsets4[0] = lp_build_extract_range(gallivm, offsets[0], 4*i, 4);
+ if (dims > 1) {
+ offsets4[1] = lp_build_extract_range(gallivm, offsets[1], 4*i, 4);
+ if (dims > 2) {
+ offsets4[2] = lp_build_extract_range(gallivm, offsets[2], 4*i, 4);
+ }
+ }
+ }
+ 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);
+ 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);
+ }
+
+ if (use_aos) {
+ /* do sampling/filtering with fixed pt arithmetic */
+ lp_build_sample_aos(&bld4, sampler_index,
+ s4, t4, r4, offsets4,
+ lod_ipart4, lod_fpart4,
+ ilevel04, ilevel14,
+ texelout4);
+ }
+
+ else {
+ lp_build_sample_general(&bld4, sampler_index,
+ s4, t4, r4, offsets4,
+ lod_ipart4, lod_fpart4,
+ ilevel04, ilevel14,
+ texelout4);
+ }
+ for (j = 0; j < 4; j++) {
+ texelouttmp[j][i] = texelout4[j];
+ }
+ }
+
+ for (j = 0; j < 4; j++) {
+ texel_out[j] = lp_build_concat(gallivm, texelouttmp[j], type4, num_quads);
+ }
+ }
+
+ lp_build_sample_compare(&bld, coords, texel_out);
}
- lp_build_sample_compare(&bld, r, texel_out);
+ if (static_texture_state->target != PIPE_BUFFER) {
+ apply_sampler_swizzle(&bld, texel_out);
+ }
+
+ /*
+ * texel type can be a (32bit) int/uint (for pure int formats only),
+ * however we are expected to always return floats (storage is untyped).
+ */
+ if (!bld.texel_type.floating) {
+ unsigned chan;
+ for (chan = 0; chan < 4; chan++) {
+ texel_out[chan] = LLVMBuildBitCast(builder, texel_out[chan],
+ lp_build_vec_type(gallivm, type), "");
+ }
+ }
+}
+
+void
+lp_build_size_query_soa(struct gallivm_state *gallivm,
+ const struct lp_static_texture_state *static_state,
+ struct lp_sampler_dynamic_state *dynamic_state,
+ struct lp_type int_type,
+ unsigned texture_unit,
+ unsigned target,
+ boolean is_sviewinfo,
+ boolean scalar_lod,
+ LLVMValueRef explicit_lod,
+ LLVMValueRef *sizes_out)
+{
+ LLVMValueRef lod, level, size;
+ LLVMValueRef first_level = NULL;
+ int dims, i;
+ boolean has_array;
+ unsigned num_lods = 1;
+ struct lp_build_context bld_int_vec4;
+
+ /*
+ * Do some sanity verification about bound texture and shader dcl target.
+ * Not entirely sure what's possible but assume array/non-array
+ * always compatible (probably not ok for OpenGL but d3d10 has no
+ * distinction of arrays at the resource level).
+ * Everything else looks bogus (though not entirely sure about rect/2d).
+ * Currently disabled because it causes assertion failures if there's
+ * nothing bound (or rather a dummy texture, not that this case would
+ * return the right values).
+ */
+ if (0 && static_state->target != target) {
+ if (static_state->target == PIPE_TEXTURE_1D)
+ assert(target == PIPE_TEXTURE_1D_ARRAY);
+ else if (static_state->target == PIPE_TEXTURE_1D_ARRAY)
+ assert(target == PIPE_TEXTURE_1D);
+ else if (static_state->target == PIPE_TEXTURE_2D)
+ assert(target == PIPE_TEXTURE_2D_ARRAY);
+ else if (static_state->target == PIPE_TEXTURE_2D_ARRAY)
+ assert(target == PIPE_TEXTURE_2D);
+ else if (static_state->target == PIPE_TEXTURE_CUBE)
+ assert(target == PIPE_TEXTURE_CUBE_ARRAY);
+ else if (static_state->target == PIPE_TEXTURE_CUBE_ARRAY)
+ assert(target == PIPE_TEXTURE_CUBE);
+ else
+ assert(0);
+ }
+
+ dims = texture_dims(target);
+
+ switch (target) {
+ case PIPE_TEXTURE_1D_ARRAY:
+ case PIPE_TEXTURE_2D_ARRAY:
+ has_array = TRUE;
+ break;
+ default:
+ has_array = FALSE;
+ break;
+ }
+
+ assert(!int_type.floating);
+
+ lp_build_context_init(&bld_int_vec4, gallivm, lp_type_int_vec(32, 128));
+
+ 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);
+ level = LLVMBuildAdd(gallivm->builder, lod, first_level, "level");
+ lod = lp_build_broadcast_scalar(&bld_int_vec4, level);
+ } else {
+ lod = bld_int_vec4.zero;
+ }
+
+ size = bld_int_vec4.undef;
+
+ size = LLVMBuildInsertElement(gallivm->builder, size,
+ dynamic_state->width(dynamic_state, gallivm, texture_unit),
+ lp_build_const_int32(gallivm, 0), "");
+
+ if (dims >= 2) {
+ size = LLVMBuildInsertElement(gallivm->builder, size,
+ dynamic_state->height(dynamic_state, gallivm, texture_unit),
+ lp_build_const_int32(gallivm, 1), "");
+ }
+
+ if (dims >= 3) {
+ size = LLVMBuildInsertElement(gallivm->builder, size,
+ dynamic_state->depth(dynamic_state, gallivm, texture_unit),
+ lp_build_const_int32(gallivm, 2), "");
+ }
+
+ size = lp_build_minify(&bld_int_vec4, size, lod);
+
+ if (has_array)
+ size = LLVMBuildInsertElement(gallivm->builder, size,
+ dynamic_state->depth(dynamic_state, gallivm, texture_unit),
+ lp_build_const_int32(gallivm, dims), "");
+
+ /*
+ * d3d10 requires zero for x/y/z values (but not w, i.e. mip levels)
+ * if level is out of bounds (note this can't cover unbound texture
+ * here, which also requires returning zero).
+ */
+ if (explicit_lod && is_sviewinfo) {
+ LLVMValueRef last_level, out, out1;
+ struct lp_build_context leveli_bld;
+
+ /* 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);
+
+ out = lp_build_cmp(&leveli_bld, PIPE_FUNC_LESS, level, first_level);
+ out1 = lp_build_cmp(&leveli_bld, PIPE_FUNC_GREATER, level, last_level);
+ out = lp_build_or(&leveli_bld, out, out1);
+ if (num_lods == 1) {
+ out = lp_build_broadcast_scalar(&bld_int_vec4, out);
+ }
+ else {
+ /* TODO */
+ assert(0);
+ }
+ size = lp_build_andnot(&bld_int_vec4, size, out);
+ }
+ for (i = 0; i < dims + (has_array ? 1 : 0); i++) {
+ sizes_out[i] = lp_build_extract_broadcast(gallivm, bld_int_vec4.type, int_type,
+ size,
+ lp_build_const_int32(gallivm, i));
+ }
+ if (is_sviewinfo) {
+ for (; i < 4; i++) {
+ sizes_out[i] = lp_build_const_vec(gallivm, int_type, 0.0);
+ }
+ }
+
+ /*
+ * if there's no explicit_lod (buffers, rects) queries requiring nr of
+ * mips would be illegal.
+ */
+ if (is_sviewinfo && explicit_lod) {
+ struct lp_build_context bld_int_scalar;
+ LLVMValueRef num_levels;
+ lp_build_context_init(&bld_int_scalar, gallivm, lp_type_int(32));
+
+ if (static_state->level_zero_only) {
+ num_levels = bld_int_scalar.one;
+ }
+ else {
+ LLVMValueRef last_level;
+
+ last_level = dynamic_state->last_level(dynamic_state, gallivm, 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);
+ }
+ sizes_out[3] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, int_type),
+ num_levels);
+ }
}