* Texture sampling -- SoA.
*
* @author Jose Fonseca <jfonseca@vmware.com>
+ * @author Brian Paul <brianp@vmware.com>
*/
#include "pipe/p_defines.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_quad.h"
/**
}
+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);
+}
+
+
/**
* Generate code to fetch a texel from a texture at int coords (x, y, z).
LLVMValueRef y_stride,
LLVMValueRef z_stride,
LLVMValueRef data_ptr,
- LLVMValueRef *texel)
+ LLVMValueRef texel_out[4])
{
const int dims = texture_dims(bld->static_state->target);
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
}
}
- /*
- * Describe the coordinates in terms of pixel blocks.
- *
- * TODO: pixel blocks are power of two. LLVM should convert rem/div to
- * bit arithmetic. Verify this.
- */
-
- if (bld->format_desc->block.width == 1) {
- i = bld->uint_coord_bld.zero;
- }
- else {
- LLVMValueRef block_width = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.width);
- i = LLVMBuildURem(bld->builder, x, block_width, "");
- x = LLVMBuildUDiv(bld->builder, x, block_width, "");
- }
+ /* convert x,y,z coords to linear offset from start of texture, in bytes */
+ lp_build_sample_offset(&bld->uint_coord_bld,
+ bld->format_desc,
+ x, y, z, y_stride, z_stride,
+ &offset, &i, &j);
- if (bld->format_desc->block.height == 1) {
- j = bld->uint_coord_bld.zero;
- }
- else {
- LLVMValueRef block_height = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.height);
- j = LLVMBuildURem(bld->builder, y, block_height, "");
- y = LLVMBuildUDiv(bld->builder, y, block_height, "");
+ if (use_border) {
+ /* If we can sample the border color, it means that texcoords may
+ * lie outside the bounds of the texture image. We need to do
+ * something to prevent reading out of bounds and causing a segfault.
+ *
+ * Simply AND the texture coords with !use_border. This will cause
+ * 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);
}
- /* convert x,y,z coords to linear offset from start of texture, in bytes */
- offset = lp_build_sample_offset(&bld->uint_coord_bld,
- bld->format_desc,
- x, y, z, y_stride, z_stride);
-
lp_build_fetch_rgba_soa(bld->builder,
bld->format_desc,
bld->texel_type,
data_ptr, offset,
i, j,
- texel);
+ texel_out);
+
+ apply_sampler_swizzle(bld, texel_out);
/*
* Note: if we find an app which frequently samples the texture border
LLVMValueRef border_chan =
lp_build_const_vec(bld->texel_type,
bld->static_state->border_color[chan]);
- texel[chan] = lp_build_select(&bld->texel_bld, use_border,
- border_chan, texel[chan]);
+ texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
+ border_chan, texel_out[chan]);
}
}
}
+/**
+ * Fetch the texels as <4n x i8> in AoS form.
+ */
static LLVMValueRef
lp_build_sample_packed(struct lp_build_sample_context *bld,
LLVMValueRef x,
LLVMValueRef y_stride,
LLVMValueRef data_array)
{
- LLVMValueRef offset;
+ LLVMValueRef offset, i, j;
LLVMValueRef data_ptr;
+ LLVMValueRef res;
- offset = lp_build_sample_offset(&bld->uint_coord_bld,
- bld->format_desc,
- x, y, NULL, y_stride, NULL);
-
- assert(bld->format_desc->block.width == 1);
- assert(bld->format_desc->block.height == 1);
- assert(bld->format_desc->block.bits <= bld->texel_type.width);
+ /* convert x,y,z coords to linear offset from start of texture, in bytes */
+ lp_build_sample_offset(&bld->uint_coord_bld,
+ bld->format_desc,
+ x, y, NULL, y_stride, NULL,
+ &offset, &i, &j);
/* get pointer to mipmap level 0 data */
data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0);
- return lp_build_gather(bld->builder,
- bld->texel_type.length,
- bld->format_desc->block.bits,
- bld->texel_type.width,
- data_ptr, offset);
+ if (util_format_is_rgba8_variant(bld->format_desc)) {
+ /* Just fetch the data directly without swizzling */
+ assert(bld->format_desc->block.width == 1);
+ assert(bld->format_desc->block.height == 1);
+ assert(bld->format_desc->block.bits <= bld->texel_type.width);
+
+ res = lp_build_gather(bld->builder,
+ bld->texel_type.length,
+ bld->format_desc->block.bits,
+ bld->texel_type.width,
+ data_ptr, offset);
+ }
+ else {
+ struct lp_type type;
+
+ assert(bld->texel_type.width == 32);
+
+ memset(&type, 0, sizeof type);
+ type.width = 8;
+ type.length = bld->texel_type.length*4;
+ type.norm = TRUE;
+
+ res = lp_build_fetch_rgba_aos(bld->builder, bld->format_desc, type,
+ data_ptr, offset, i, j);
+ }
+
+ return res;
}
{
switch (mode) {
case PIPE_TEX_WRAP_REPEAT:
- case PIPE_TEX_WRAP_CLAMP:
case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
return TRUE;
- case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
default:
return FALSE;
}
coord = LLVMBuildURem(bld->builder, coord, length, "");
break;
- case PIPE_TEX_WRAP_CLAMP:
case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
- case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
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:
- /* FIXME */
- _debug_printf("llvmpipe: failed to translate texture wrap mode %s\n",
- util_dump_tex_wrap(wrap_mode, TRUE));
- coord = lp_build_max(uint_coord_bld, coord, uint_coord_bld->zero);
- coord = lp_build_min(uint_coord_bld, coord, length_minus_one);
- break;
-
default:
assert(0);
}
struct lp_build_context *coord_bld = &bld->coord_bld;
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
- LLVMValueRef two = lp_build_const_vec(coord_bld->type, 2.0);
LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
- LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
LLVMValueRef coord0, coord1, weight;
switch(wrap_mode) {
case PIPE_TEX_WRAP_CLAMP:
if (bld->static_state->normalized_coords) {
+ /* scale coord to length */
coord = lp_build_mul(coord_bld, coord, length_f);
}
+
+ /* 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_clamp(coord_bld, coord, coord_bld->zero,
- length_f_minus_one);
- coord1 = lp_build_add(coord_bld, coord, coord_bld->one);
- coord1 = lp_build_clamp(coord_bld, coord1, coord_bld->zero,
- length_f_minus_one);
- coord0 = lp_build_ifloor(coord_bld, coord0);
- coord1 = lp_build_ifloor(coord_bld, coord1);
+ coord0 = lp_build_ifloor(coord_bld, coord);
+ coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
break;
case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
else {
LLVMValueRef min, max;
/* clamp to [0.5, length - 0.5] */
- min = lp_build_const_vec(coord_bld->type, 0.5F);
+ min = half;
max = lp_build_sub(coord_bld, length_f, min);
coord = lp_build_clamp(coord_bld, coord, min, max);
}
{
LLVMValueRef min, max;
if (bld->static_state->normalized_coords) {
- /* min = -1.0 / (2 * length) = -0.5 / length */
- min = lp_build_mul(coord_bld,
- lp_build_const_vec(coord_bld->type, -0.5F),
- lp_build_rcp(coord_bld, length_f));
- /* max = 1.0 - min */
- max = lp_build_sub(coord_bld, coord_bld->one, min);
- /* coord = clamp(coord, min, max) */
- coord = lp_build_clamp(coord_bld, coord, min, max);
- /* scale coord to length (and sub 0.5?) */
+ /* scale coord to length */
coord = lp_build_mul(coord_bld, coord, length_f);
- coord = lp_build_sub(coord_bld, coord, half);
- }
- else {
- /* 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);
- coord = lp_build_sub(coord_bld, coord, half);
}
+ /* 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);
+ coord = lp_build_sub(coord_bld, coord, half);
/* compute lerp weight */
weight = lp_build_fract(coord_bld, coord);
/* convert to int */
break;
case PIPE_TEX_WRAP_MIRROR_CLAMP:
- {
- LLVMValueRef min, max;
- /* min = 1.0 / (2 * length) */
- min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
- /* max = 1.0 - min */
- max = lp_build_sub(coord_bld, coord_bld->one, min);
+ coord = lp_build_abs(coord_bld, coord);
- coord = lp_build_abs(coord_bld, coord);
- coord = lp_build_clamp(coord_bld, coord, min, max);
+ if (bld->static_state->normalized_coords) {
+ /* scale coord to length */
coord = lp_build_mul(coord_bld, coord, length_f);
- if(0)coord = lp_build_sub(coord_bld, coord, half);
- weight = lp_build_fract(coord_bld, coord);
- coord0 = lp_build_ifloor(coord_bld, coord);
- coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
}
+
+ /* 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);
+ coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
break;
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
{
LLVMValueRef min, max;
- /* min = 1.0 / (2 * length) */
- min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
- /* max = 1.0 - min */
- max = lp_build_sub(coord_bld, coord_bld->one, min);
coord = lp_build_abs(coord_bld, coord);
+
+ if (bld->static_state->normalized_coords) {
+ /* scale coord to length */
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ }
+
+ /* 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);
- coord = lp_build_mul(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);
coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
{
LLVMValueRef min, max;
- /* min = -1.0 / (2 * length) = -0.5 / length */
- min = lp_build_mul(coord_bld,
- lp_build_const_vec(coord_bld->type, -0.5F),
- lp_build_rcp(coord_bld, length_f));
- /* max = 1.0 - min */
- max = lp_build_sub(coord_bld, coord_bld->one, min);
coord = lp_build_abs(coord_bld, coord);
+
+ if (bld->static_state->normalized_coords) {
+ /* scale coord to length */
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ }
+
+ /* 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);
- coord = lp_build_mul(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);
coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
struct lp_build_context *coord_bld = &bld->coord_bld;
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
- LLVMValueRef two = lp_build_const_vec(coord_bld->type, 2.0);
LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
- LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
LLVMValueRef icoord;
switch(wrap_mode) {
break;
case PIPE_TEX_WRAP_CLAMP:
- /* mul by size */
+ case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
if (bld->static_state->normalized_coords) {
+ /* scale coord to length */
coord = lp_build_mul(coord_bld, coord, length_f);
}
+
/* floor */
icoord = lp_build_ifloor(coord_bld, coord);
- /* clamp to [0, size-1]. Note: int coord builder type */
+
+ /* clamp to [0, length - 1]. */
icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
length_minus_one);
break;
- case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
- {
- LLVMValueRef min, max;
- if (bld->static_state->normalized_coords) {
- /* min = 1.0 / (2 * length) */
- min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
- /* max = length - min */
- max = lp_build_sub(coord_bld, length_f, min);
- /* scale coord to length */
- coord = lp_build_mul(coord_bld, coord, length_f);
- }
- else {
- /* 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 = clamp(coord, min, max) */
- coord = lp_build_clamp(coord_bld, coord, min, max);
- icoord = lp_build_ifloor(coord_bld, coord);
- }
- 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) {
- /* min = -1.0 / (2 * length) = -0.5 / length */
- min = lp_build_mul(coord_bld,
- lp_build_const_vec(coord_bld->type, -0.5F),
- lp_build_rcp(coord_bld, length_f));
- /* max = length - min */
- max = lp_build_sub(coord_bld, length_f, min);
/* scale coord to length */
coord = lp_build_mul(coord_bld, coord, length_f);
}
- else {
- /* 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 = clamp(coord, min, max) */
- coord = lp_build_clamp(coord_bld, coord, min, max);
+
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);
}
break;
case PIPE_TEX_WRAP_MIRROR_REPEAT:
- {
- LLVMValueRef min, max;
- /* min = 1.0 / (2 * length) */
- min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
- /* max = length - min */
- max = lp_build_sub(coord_bld, length_f, min);
+ /* compute mirror function */
+ coord = lp_build_coord_mirror(bld, coord);
- /* compute mirror function */
- coord = lp_build_coord_mirror(bld, coord);
+ /* scale coord to length */
+ assert(bld->static_state->normalized_coords);
+ coord = lp_build_mul(coord_bld, coord, length_f);
- /* scale coord to length */
- coord = lp_build_mul(coord_bld, coord, length_f);
+ icoord = lp_build_ifloor(coord_bld, coord);
- /* coord = clamp(coord, min, max) */
- coord = lp_build_clamp(coord_bld, coord, min, max);
- icoord = lp_build_ifloor(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:
- coord = lp_build_abs(coord_bld, coord);
- coord = lp_build_mul(coord_bld, coord, length_f);
- coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f_minus_one);
- icoord = lp_build_ifloor(coord_bld, coord);
- break;
-
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
- {
- LLVMValueRef min, max;
- /* min = 1.0 / (2 * length) */
- min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
- /* max = length - min */
- max = lp_build_sub(coord_bld, length_f, min);
+ coord = lp_build_abs(coord_bld, coord);
- coord = lp_build_abs(coord_bld, coord);
+ if (bld->static_state->normalized_coords) {
+ /* scale coord to length */
coord = lp_build_mul(coord_bld, coord, length_f);
- coord = lp_build_clamp(coord_bld, coord, min, max);
- icoord = lp_build_ifloor(coord_bld, coord);
}
+
+ icoord = lp_build_ifloor(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:
- {
- LLVMValueRef min, max;
- /* min = 1.0 / (2 * length) */
- min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
- min = lp_build_negate(coord_bld, min);
- /* max = length - min */
- max = lp_build_sub(coord_bld, length_f, min);
+ coord = lp_build_abs(coord_bld, coord);
- coord = lp_build_abs(coord_bld, coord);
+ if (bld->static_state->normalized_coords) {
+ /* scale coord to length */
coord = lp_build_mul(coord_bld, coord, length_f);
- coord = lp_build_clamp(coord_bld, coord, min, max);
- icoord = lp_build_ifloor(coord_bld, coord);
}
+
+ icoord = lp_build_ifloor(coord_bld, coord);
+
+ /* clamp to [0, length] */
+ icoord = lp_build_min(int_coord_bld, icoord, length);
break;
default:
/**
* Generate code to compute texture level of detail (lambda).
- * \param s vector of texcoord s values
- * \param t vector of texcoord t values
- * \param r vector of texcoord r values
+ * \param 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,
- LLVMValueRef s,
- LLVMValueRef t,
- LLVMValueRef r,
+ const LLVMValueRef ddx[4],
+ const LLVMValueRef ddy[4],
+ LLVMValueRef lod_bias, /* optional */
+ LLVMValueRef explicit_lod, /* optional */
LLVMValueRef width,
LLVMValueRef height,
LLVMValueRef depth)
return LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod);
}
else {
- const int dims = texture_dims(bld->static_state->target);
struct lp_build_context *float_bld = &bld->float_bld;
- LLVMValueRef lod_bias = LLVMConstReal(LLVMFloatType(),
- bld->static_state->lod_bias);
+ LLVMValueRef 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 index1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
- LLVMValueRef index2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
-
- LLVMValueRef s0, s1, s2;
- LLVMValueRef t0, t1, t2;
- LLVMValueRef r0, r1, r2;
- LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy;
- LLVMValueRef rho, lod;
-
- /*
- * dsdx = abs(s[1] - s[0]);
- * dsdy = abs(s[2] - s[0]);
- * dtdx = abs(t[1] - t[0]);
- * dtdy = abs(t[2] - t[0]);
- * drdx = abs(r[1] - r[0]);
- * drdy = abs(r[2] - r[0]);
- * XXX we're assuming a four-element quad in 2x2 layout here.
- */
- s0 = LLVMBuildExtractElement(bld->builder, s, index0, "s0");
- s1 = LLVMBuildExtractElement(bld->builder, s, index1, "s1");
- s2 = LLVMBuildExtractElement(bld->builder, s, index2, "s2");
- dsdx = LLVMBuildSub(bld->builder, s1, s0, "");
- dsdx = lp_build_abs(float_bld, dsdx);
- dsdy = LLVMBuildSub(bld->builder, s2, s0, "");
- dsdy = lp_build_abs(float_bld, dsdy);
- if (dims > 1) {
- t0 = LLVMBuildExtractElement(bld->builder, t, index0, "t0");
- t1 = LLVMBuildExtractElement(bld->builder, t, index1, "t1");
- t2 = LLVMBuildExtractElement(bld->builder, t, index2, "t2");
- dtdx = LLVMBuildSub(bld->builder, t1, t0, "");
- dtdx = lp_build_abs(float_bld, dtdx);
- dtdy = LLVMBuildSub(bld->builder, t2, t0, "");
- dtdy = lp_build_abs(float_bld, dtdy);
- if (dims > 2) {
- r0 = LLVMBuildExtractElement(bld->builder, r, index0, "r0");
- r1 = LLVMBuildExtractElement(bld->builder, r, index1, "r1");
- r2 = LLVMBuildExtractElement(bld->builder, r, index2, "r2");
- drdx = LLVMBuildSub(bld->builder, r1, r0, "");
- drdx = lp_build_abs(float_bld, drdx);
- drdy = LLVMBuildSub(bld->builder, r2, r0, "");
- drdy = lp_build_abs(float_bld, drdy);
- }
+ 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 = LLVMBuildMul(bld->builder,
- lp_build_max(float_bld, dsdx, dsdy),
- lp_build_int_to_float(float_bld, width), "");
- if (dims > 1) {
- LLVMValueRef max;
- max = LLVMBuildMul(bld->builder,
- lp_build_max(float_bld, dtdx, dtdy),
- lp_build_int_to_float(float_bld, height), "");
- rho = lp_build_max(float_bld, rho, max);
- if (dims > 2) {
+ /* Compute rho = max of all partial derivatives scaled by texture size.
+ * XXX this could be vectorized somewhat
+ */
+ rho = LLVMBuildMul(bld->builder,
+ lp_build_max(float_bld, dsdx, dsdy),
+ lp_build_int_to_float(float_bld, width), "");
+ if (dims > 1) {
+ LLVMValueRef max;
max = LLVMBuildMul(bld->builder,
- lp_build_max(float_bld, drdx, drdy),
- lp_build_int_to_float(float_bld, depth), "");
+ lp_build_max(float_bld, dtdx, dtdy),
+ lp_build_int_to_float(float_bld, height), "");
rho = lp_build_max(float_bld, rho, max);
+ if (dims > 2) {
+ max = LLVMBuildMul(bld->builder,
+ lp_build_max(float_bld, drdx, drdy),
+ lp_build_int_to_float(float_bld, depth), "");
+ rho = lp_build_max(float_bld, rho, max);
+ }
}
- }
- /* compute lod = log2(rho) */
- lod = lp_build_log2(float_bld, rho);
+ /* compute lod = log2(rho) */
+ lod = lp_build_log2(float_bld, rho);
- /* add lod bias */
- lod = LLVMBuildAdd(bld->builder, lod, lod_bias, "LOD bias");
+ /* add shader lod bias */
+ if (lod_bias) {
+ lod_bias = LLVMBuildExtractElement(bld->builder, lod_bias,
+ index0, "");
+ lod = LLVMBuildAdd(bld->builder, lod, lod_bias, "shader_lod_bias");
+ }
+ }
+
+ /* add sampler lod bias */
+ lod = LLVMBuildAdd(bld->builder, lod, sampler_lod_bias, "sampler_lod_bias");
/* clamp lod */
lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
int_bld->zero,
last_level);
/* compute level 1 and clamp to legal range of levels */
- *level1_out = lp_build_add(int_bld, *level0_out, int_bld->one);
- *level1_out = lp_build_min(int_bld, *level1_out, last_level);
+ 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 do cube face selection and per-face texcoords.
+ * Generate code to do cube face selection and compute per-face texcoords.
*/
static void
lp_build_cube_lookup(struct lp_build_sample_context *bld,
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;
int chan;
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
+ /* sample the first mipmap level */
lp_build_sample_image_nearest(bld,
width0_vec, height0_vec, depth0_vec,
row_stride0_vec, img_stride0_vec,
data_ptr0, s, t, r, colors0);
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
- /* sample the second mipmap level, and interp */
+ /* sample the second mipmap level */
lp_build_sample_image_nearest(bld,
width1_vec, height1_vec, depth1_vec,
row_stride1_vec, img_stride1_vec,
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,
row_stride0_vec, img_stride0_vec,
data_ptr0, s, t, r, colors0);
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
- /* sample the second mipmap level, and interp */
+ /* sample the second mipmap level */
lp_build_sample_image_linear(bld,
width1_vec, height1_vec, depth1_vec,
row_stride1_vec, img_stride1_vec,
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 height_vec,
LLVMValueRef depth_vec,
LLVMValueRef row_stride_array,
- LLVMValueRef img_stride_vec,
+ LLVMValueRef img_stride_array,
LLVMValueRef data_array,
LLVMValueRef *colors_out)
{
const unsigned min_filter = bld->static_state->min_img_filter;
const unsigned mag_filter = bld->static_state->mag_img_filter;
const int dims = texture_dims(bld->static_state->target);
- LLVMValueRef lod, lod_fpart;
- LLVMValueRef ilevel0, ilevel1, ilevel0_vec, ilevel1_vec;
+ LLVMValueRef 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;
+ LLVMValueRef data_ptr0, data_ptr1 = NULL;
+ LLVMValueRef face_ddx[4], face_ddy[4];
/*
printf("%s mip %d min %d mag %d\n", __FUNCTION__,
mip_filter, min_filter, mag_filter);
*/
+ /*
+ * Choose cube face, recompute texcoords and derivatives for the chosen face.
+ */
+ if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
+ LLVMValueRef face, face_s, face_t;
+ lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
+ s = face_s; /* vec */
+ t = face_t; /* vec */
+ /* use 'r' to indicate cube face */
+ r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
+
+ /* 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;
+ }
+
/*
* Compute the level of detail (float).
*/
/* Need to compute lod either to choose mipmap levels or to
* distinguish between minification/magnification with one mipmap level.
*/
- lod = lp_build_lod_selector(bld, s, t, r, width, height, depth);
+ lod = lp_build_lod_selector(bld, ddx, ddy,
+ lod_bias, explicit_lod,
+ width, height, depth);
}
/*
*/
if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
/* always use mip level 0 */
- ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
+ if (bld->static_state->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);
}
row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
ilevel0);
if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
- img_stride0_vec = lp_build_mul(&bld->int_coord_bld,
- row_stride0_vec, height0_vec);
+ 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);
}
row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
ilevel1);
if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
- img_stride1_vec = lp_build_mul(&bld->int_coord_bld,
- row_stride1_vec, height1_vec);
+ 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);
}
}
}
- /*
- * Choose cube face, recompute per-face texcoords.
- */
- if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
- LLVMValueRef face, face_s, face_t;
- lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
- s = face_s; /* vec */
- t = face_t; /* vec */
- /* use 'r' to indicate cube face */
- r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
- }
-
/*
* Get pointer(s) to image data for mipmap level(s).
*/
-static void
-lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder,
- struct lp_type dst_type,
- LLVMValueRef packed,
- LLVMValueRef *rgba)
-{
- LLVMValueRef mask = lp_build_const_int_vec(dst_type, 0xff);
- unsigned chan;
-
- /* Decode the input vector components */
- for (chan = 0; chan < 4; ++chan) {
- unsigned start = chan*8;
- unsigned stop = start + 8;
- LLVMValueRef input;
-
- input = packed;
-
- if(start)
- input = LLVMBuildLShr(builder, input, lp_build_const_int_vec(dst_type, start), "");
-
- if(stop < 32)
- input = LLVMBuildAnd(builder, input, mask, "");
-
- input = lp_build_unsigned_norm_to_float(builder, 8, dst_type, input);
-
- rgba[chan] = input;
- }
-}
-
-
static void
lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
LLVMValueRef s,
LLVMValueRef height,
LLVMValueRef stride_array,
LLVMValueRef data_array,
- LLVMValueRef *texel)
+ LLVMValueRef texel_out[4])
{
LLVMBuilderRef builder = bld->builder;
struct lp_build_context i32, h16, u8n;
LLVMValueRef unswizzled[4];
LLVMValueRef stride;
+ assert(bld->static_state->target == PIPE_TEXTURE_2D);
+ 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));
* Convert to SoA and swizzle.
*/
- packed = LLVMBuildBitCast(builder, packed, i32_vec_type, "");
-
lp_build_rgba8_to_f32_soa(bld->builder,
bld->texel_type,
packed, unswizzled);
- lp_build_format_swizzle_soa(bld->format_desc,
- bld->texel_type, unswizzled,
- texel);
+ 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];
+ }
+
+ apply_sampler_swizzle(bld, texel_out);
}
static void
lp_build_sample_compare(struct lp_build_sample_context *bld,
LLVMValueRef p,
- LLVMValueRef *texel)
+ LLVMValueRef texel[4])
{
struct lp_build_context *texel_bld = &bld->texel_bld;
LLVMValueRef res;
}
+/**
+ * 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,
+ LLVMValueRef texel_out[4])
+{
+ struct lp_build_context *texel_bld = &bld->texel_bld;
+ unsigned chan;
+
+ for (chan = 0; chan < 4; chan++) {
+ /*lp_bld_mov(texel_bld, texel, texel_bld->one);*/
+ texel_out[chan] = texel_bld->one;
+ }
+}
+
+
/**
* Build texture sampling code.
* 'texel' will return a vector of four LLVMValueRefs corresponding to
* R, G, B, A.
* \param type vector float type to use for coords, etc.
+ * \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
*/
void
lp_build_sample_soa(LLVMBuilderRef builder,
unsigned unit,
unsigned num_coords,
const LLVMValueRef *coords,
- LLVMValueRef lodbias,
- LLVMValueRef *texel)
+ const LLVMValueRef ddx[4],
+ const LLVMValueRef ddy[4],
+ LLVMValueRef lod_bias, /* optional */
+ LLVMValueRef explicit_lod, /* optional */
+ LLVMValueRef texel_out[4])
{
struct lp_build_sample_context bld;
LLVMValueRef width, width_vec;
LLVMValueRef height, height_vec;
LLVMValueRef depth, depth_vec;
- LLVMValueRef stride_array;
+ LLVMValueRef row_stride_array, img_stride_array;
LLVMValueRef data_array;
LLVMValueRef s;
LLVMValueRef t;
LLVMValueRef r;
- (void) lp_build_lod_selector; /* temporary to silence warning */
- (void) lp_build_nearest_mip_level;
- (void) lp_build_linear_mip_levels;
- (void) lp_build_minify;
+ if (0) {
+ enum pipe_format fmt = static_state->format;
+ debug_printf("Sample from %s\n", util_format_name(fmt));
+ }
/* Setup our build context */
memset(&bld, 0, sizeof bld);
width = dynamic_state->width(dynamic_state, builder, unit);
height = dynamic_state->height(dynamic_state, builder, unit);
depth = dynamic_state->depth(dynamic_state, builder, unit);
- stride_array = dynamic_state->row_stride(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 */
height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth);
- if (util_format_is_rgba8_variant(bld.format_desc) &&
- static_state->target == PIPE_TEXTURE_2D &&
- static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
- static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
- static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
- is_simple_wrap_mode(static_state->wrap_s) &&
- is_simple_wrap_mode(static_state->wrap_t)) {
+ 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->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
+ static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
+ static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
+ is_simple_wrap_mode(static_state->wrap_s) &&
+ is_simple_wrap_mode(static_state->wrap_t)) {
/* special case */
lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec,
- stride_array, data_array, texel);
+ row_stride_array, data_array, texel_out);
}
else {
- lp_build_sample_general(&bld, unit, s, t, r,
+ lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy,
+ lod_bias, explicit_lod,
width, height, depth,
width_vec, height_vec, depth_vec,
- stride_array, NULL, data_array,
- texel);
+ row_stride_array, img_stride_array,
+ data_array,
+ texel_out);
}
- lp_build_sample_compare(&bld, r, texel);
+ lp_build_sample_compare(&bld, r, texel_out);
}
projects / mesa.git / blobdiff