LLVMValueRef
lp_build_lerp_2d(struct lp_build_context *bld,
LLVMValueRef x,
LLVMValueRef y,
LLVMValueRef v00,
LLVMValueRef v01,
LLVMValueRef v10,
LLVMValueRef v11)
{
LLVMValueRef v0 = lp_build_lerp(bld, x, v00, v01);
LLVMValueRef v1 = lp_build_lerp(bld, x, v10, v11);
return lp_build_lerp(bld, y, v0, v1);
}
/**
* Performs optimisations and blending independent of SoA/AoS
*
* @param func the blend function
* @param factor_src PIPE_BLENDFACTOR_xxx
* @param factor_dst PIPE_BLENDFACTOR_xxx
* @param src source rgba
* @param dst dest rgba
* @param src_factor src factor computed value
* @param dst_factor dst factor computed value
* @param not_alpha_dependent same factors accross all channels of src/dst
*
* not_alpha_dependent should be:
* SoA: always true as it is only one channel at a time
* AoS: rgb_src_factor == alpha_src_factor && rgb_dst_factor == alpha_dst_factor
*
* Note that pretty much every possible optimisation can only be done on non-unorm targets
* due to unorm values not going above 1.0 meaning factorisation can change results.
* e.g. (0.9 * 0.9) + (0.9 * 0.9) != 0.9 * (0.9 + 0.9) as result of + is always <= 1.
*/
LLVMValueRef
lp_build_blend(struct lp_build_context *bld,
unsigned func,
unsigned factor_src,
unsigned factor_dst,
LLVMValueRef src,
LLVMValueRef dst,
LLVMValueRef src_factor,
LLVMValueRef dst_factor,
boolean not_alpha_dependent,
boolean optimise_only)
{
LLVMValueRef result, src_term, dst_term;
/* If we are not alpha dependent we can mess with the src/dst factors */
if (not_alpha_dependent) {
if (lp_build_blend_factor_complementary(factor_src, factor_dst)) {
if (func == PIPE_BLEND_ADD) {
if (factor_src < factor_dst) {
return lp_build_lerp(bld, src_factor, dst, src, 0);
} else {
return lp_build_lerp(bld, dst_factor, src, dst, 0);
}
} else if(bld->type.floating && func == PIPE_BLEND_SUBTRACT) {
result = lp_build_add(bld, src, dst);
if (factor_src < factor_dst) {
result = lp_build_mul(bld, result, src_factor);
return lp_build_sub(bld, result, dst);
} else {
result = lp_build_mul(bld, result, dst_factor);
return lp_build_sub(bld, src, result);
}
} else if(bld->type.floating && func == PIPE_BLEND_REVERSE_SUBTRACT) {
result = lp_build_add(bld, src, dst);
if (factor_src < factor_dst) {
result = lp_build_mul(bld, result, src_factor);
return lp_build_sub(bld, dst, result);
} else {
result = lp_build_mul(bld, result, dst_factor);
return lp_build_sub(bld, result, src);
}
}
}
if (bld->type.floating && factor_src == factor_dst) {
if (func == PIPE_BLEND_ADD ||
func == PIPE_BLEND_SUBTRACT ||
func == PIPE_BLEND_REVERSE_SUBTRACT) {
LLVMValueRef result;
result = lp_build_blend_func(bld, func, src, dst);
return lp_build_mul(bld, result, src_factor);
}
}
}
if (optimise_only)
return NULL;
src_term = lp_build_mul(bld, src, src_factor);
dst_term = lp_build_mul(bld, dst, dst_factor);
return lp_build_blend_func(bld, func, src_term, dst_term);
}
/**
* Sample the texture/mipmap using given image filter and mip filter.
* data0_ptr and data1_ptr point to the two mipmap levels to sample
* from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
* If we're using nearest miplevel sampling the '1' values will be null/unused.
*/
static void
lp_build_sample_mipmap(struct lp_build_sample_context *bld,
unsigned img_filter,
unsigned mip_filter,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
LLVMValueRef ilevel0,
LLVMValueRef ilevel1,
LLVMValueRef lod_fpart,
LLVMValueRef colors_lo_var,
LLVMValueRef colors_hi_var)
{
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef size0;
LLVMValueRef size1;
LLVMValueRef row_stride0_vec;
LLVMValueRef row_stride1_vec;
LLVMValueRef img_stride0_vec;
LLVMValueRef img_stride1_vec;
LLVMValueRef data_ptr0;
LLVMValueRef data_ptr1;
LLVMValueRef colors0_lo, colors0_hi;
LLVMValueRef colors1_lo, colors1_hi;
/* sample the first mipmap level */
lp_build_mipmap_level_sizes(bld, ilevel0,
&size0,
&row_stride0_vec, &img_stride0_vec);
data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
lp_build_sample_image_nearest(bld,
size0,
row_stride0_vec, img_stride0_vec,
data_ptr0, s, t, r,
&colors0_lo, &colors0_hi);
}
else {
assert(img_filter == PIPE_TEX_FILTER_LINEAR);
lp_build_sample_image_linear(bld,
size0,
row_stride0_vec, img_stride0_vec,
data_ptr0, s, t, r,
&colors0_lo, &colors0_hi);
}
/* Store the first level's colors in the output variables */
LLVMBuildStore(builder, colors0_lo, colors_lo_var);
LLVMBuildStore(builder, colors0_hi, colors_hi_var);
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
LLVMValueRef h16_scale = lp_build_const_float(bld->gallivm, 256.0);
LLVMTypeRef i32_type = LLVMIntTypeInContext(bld->gallivm->context, 32);
struct lp_build_if_state if_ctx;
LLVMValueRef need_lerp;
lod_fpart = LLVMBuildFMul(builder, lod_fpart, h16_scale, "");
lod_fpart = LLVMBuildFPToSI(builder, lod_fpart, i32_type, "lod_fpart.fixed16");
/* need_lerp = lod_fpart > 0 */
need_lerp = LLVMBuildICmp(builder, LLVMIntSGT,
lod_fpart, LLVMConstNull(i32_type),
"need_lerp");
lp_build_if(&if_ctx, bld->gallivm, need_lerp);
{
struct lp_build_context h16_bld;
lp_build_context_init(&h16_bld, bld->gallivm, lp_type_ufixed(16));
/* sample the second mipmap level */
lp_build_mipmap_level_sizes(bld, ilevel1,
&size1,
&row_stride1_vec, &img_stride1_vec);
data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
lp_build_sample_image_nearest(bld,
size1,
row_stride1_vec, img_stride1_vec,
data_ptr1, s, t, r,
&colors1_lo, &colors1_hi);
}
else {
lp_build_sample_image_linear(bld,
size1,
row_stride1_vec, img_stride1_vec,
data_ptr1, s, t, r,
&colors1_lo, &colors1_hi);
}
/* interpolate samples from the two mipmap levels */
lod_fpart = LLVMBuildTrunc(builder, lod_fpart, h16_bld.elem_type, "");
lod_fpart = lp_build_broadcast_scalar(&h16_bld, lod_fpart);
#if HAVE_LLVM == 0x208
/* This is a work-around for a bug in LLVM 2.8.
* Evidently, something goes wrong in the construction of the
* lod_fpart short[8] vector. Adding this no-effect shuffle seems
* to force the vector to be properly constructed.
* Tested with mesa-demos/src/tests/mipmap_limits.c (press t, f).
*/
{
LLVMValueRef shuffles[8], shuffle;
int i;
assert(h16_bld.type.length <= Elements(shuffles));
for (i = 0; i < h16_bld.type.length; i++)
shuffles[i] = lp_build_const_int32(bld->gallivm, 2 * (i & 1));
shuffle = LLVMConstVector(shuffles, h16_bld.type.length);
lod_fpart = LLVMBuildShuffleVector(builder,
lod_fpart, lod_fpart,
shuffle, "");
}
#endif
colors0_lo = lp_build_lerp(&h16_bld, lod_fpart,
colors0_lo, colors1_lo);
colors0_hi = lp_build_lerp(&h16_bld, lod_fpart,
colors0_hi, colors1_hi);
LLVMBuildStore(builder, colors0_lo, colors_lo_var);
LLVMBuildStore(builder, colors0_hi, colors_hi_var);
}
lp_build_endif(&if_ctx);
}
}
/**
* Sample a single texture image with (bi-)(tri-)linear sampling.
* Return filtered color as two vectors of 16-bit fixed point values.
*/
static void
lp_build_sample_image_linear(struct lp_build_sample_context *bld,
LLVMValueRef int_size,
LLVMValueRef row_stride_vec,
LLVMValueRef img_stride_vec,
LLVMValueRef data_ptr,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
LLVMValueRef *colors_lo,
LLVMValueRef *colors_hi)
{
const unsigned dims = bld->dims;
LLVMBuilderRef builder = bld->gallivm->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 width_vec, height_vec, depth_vec;
LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
LLVMValueRef t_ipart = NULL, t_fpart = NULL, t_fpart_lo = NULL, t_fpart_hi = NULL;
LLVMValueRef r_ipart = NULL, r_fpart = NULL, r_fpart_lo = NULL, r_fpart_hi = NULL;
LLVMValueRef x_stride, y_stride, z_stride;
LLVMValueRef x_offset0, x_offset1;
LLVMValueRef y_offset0, y_offset1;
LLVMValueRef z_offset0, z_offset1;
LLVMValueRef offset[2][2][2]; /* [z][y][x] */
LLVMValueRef x_subcoord[2], y_subcoord[2], z_subcoord[2];
LLVMValueRef neighbors_lo[2][2][2]; /* [z][y][x] */
LLVMValueRef neighbors_hi[2][2][2]; /* [z][y][x] */
LLVMValueRef packed_lo, packed_hi;
unsigned x, y, z;
unsigned i, j, k;
unsigned numj, numk;
lp_build_context_init(&i32, bld->gallivm, lp_type_int_vec(32));
lp_build_context_init(&h16, bld->gallivm, lp_type_ufixed(16));
lp_build_context_init(&u8n, bld->gallivm, lp_type_unorm(8));
i32_vec_type = lp_build_vec_type(bld->gallivm, i32.type);
h16_vec_type = lp_build_vec_type(bld->gallivm, h16.type);
u8n_vec_type = lp_build_vec_type(bld->gallivm, u8n.type);
lp_build_extract_image_sizes(bld,
bld->int_size_type,
bld->int_coord_type,
int_size,
&width_vec,
&height_vec,
&depth_vec);
if (bld->static_state->normalized_coords) {
LLVMValueRef scaled_size;
LLVMValueRef flt_size;
/* scale size by 256 (8 fractional bits) */
scaled_size = lp_build_shl_imm(&bld->int_size_bld, int_size, 8);
flt_size = lp_build_int_to_float(&bld->float_size_bld, scaled_size);
lp_build_unnormalized_coords(bld, flt_size, &s, &t, &r);
}
else {
/* scale coords by 256 (8 fractional bits) */
s = lp_build_mul_imm(&bld->coord_bld, s, 256);
if (dims >= 2)
t = lp_build_mul_imm(&bld->coord_bld, t, 256);
if (dims >= 3)
r = lp_build_mul_imm(&bld->coord_bld, r, 256);
}
/* convert float to int */
s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
if (dims >= 2)
t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
if (dims >= 3)
r = LLVMBuildFPToSI(builder, r, i32_vec_type, "");
/* subtract 0.5 (add -128) */
i32_c128 = lp_build_const_int_vec(bld->gallivm, i32.type, -128);
s = LLVMBuildAdd(builder, s, i32_c128, "");
if (dims >= 2) {
t = LLVMBuildAdd(builder, t, i32_c128, "");
}
if (dims >= 3) {
r = LLVMBuildAdd(builder, r, i32_c128, "");
}
/* compute floor (shift right 8) */
i32_c8 = lp_build_const_int_vec(bld->gallivm, i32.type, 8);
s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
if (dims >= 2)
t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
if (dims >= 3)
r_ipart = LLVMBuildAShr(builder, r, i32_c8, "");
/* compute fractional part (AND with 0xff) */
i32_c255 = lp_build_const_int_vec(bld->gallivm, i32.type, 255);
s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
if (dims >= 2)
t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
if (dims >= 3)
r_fpart = LLVMBuildAnd(builder, r, i32_c255, "");
/* get pixel, row and image strides */
x_stride = lp_build_const_vec(bld->gallivm, bld->int_coord_bld.type,
bld->format_desc->block.bits/8);
y_stride = row_stride_vec;
z_stride = img_stride_vec;
/* do texcoord wrapping and compute texel offsets */
lp_build_sample_wrap_linear_int(bld,
bld->format_desc->block.width,
s_ipart, width_vec, x_stride,
bld->static_state->pot_width,
bld->static_state->wrap_s,
&x_offset0, &x_offset1,
&x_subcoord[0], &x_subcoord[1]);
for (z = 0; z < 2; z++) {
for (y = 0; y < 2; y++) {
offset[z][y][0] = x_offset0;
offset[z][y][1] = x_offset1;
}
}
if (dims >= 2) {
lp_build_sample_wrap_linear_int(bld,
bld->format_desc->block.height,
t_ipart, height_vec, y_stride,
bld->static_state->pot_height,
bld->static_state->wrap_t,
&y_offset0, &y_offset1,
&y_subcoord[0], &y_subcoord[1]);
for (z = 0; z < 2; z++) {
for (x = 0; x < 2; x++) {
offset[z][0][x] = lp_build_add(&bld->int_coord_bld,
offset[z][0][x], y_offset0);
offset[z][1][x] = lp_build_add(&bld->int_coord_bld,
offset[z][1][x], y_offset1);
}
}
}
if (dims >= 3) {
lp_build_sample_wrap_linear_int(bld,
bld->format_desc->block.height,
r_ipart, depth_vec, z_stride,
bld->static_state->pot_depth,
bld->static_state->wrap_r,
&z_offset0, &z_offset1,
&z_subcoord[0], &z_subcoord[1]);
for (y = 0; y < 2; y++) {
for (x = 0; x < 2; x++) {
offset[0][y][x] = lp_build_add(&bld->int_coord_bld,
offset[0][y][x], z_offset0);
offset[1][y][x] = lp_build_add(&bld->int_coord_bld,
offset[1][y][x], z_offset1);
}
}
}
else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
LLVMValueRef z_offset;
z_offset = lp_build_mul(&bld->int_coord_bld, r, img_stride_vec);
for (y = 0; y < 2; y++) {
for (x = 0; x < 2; x++) {
/* The r coord is the cube face in [0,5] */
offset[0][y][x] = lp_build_add(&bld->int_coord_bld,
offset[0][y][x], z_offset);
}
}
}
/*
* 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, "");
if (dims >= 2)
t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
if (dims >= 3)
r_fpart = LLVMBuildBitCast(builder, r_fpart, h16_vec_type, "");
{
LLVMTypeRef elem_type = LLVMInt32TypeInContext(bld->gallivm->context);
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;
}
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, "");
s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef,
shuffle_hi, "");
if (dims >= 2) {
t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef,
shuffle_lo, "");
t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef,
shuffle_hi, "");
}
if (dims >= 3) {
r_fpart_lo = LLVMBuildShuffleVector(builder, r_fpart, h16.undef,
shuffle_lo, "");
r_fpart_hi = LLVMBuildShuffleVector(builder, r_fpart, h16.undef,
shuffle_hi, "");
}
}
/*
* 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.
*/
numj = 1 + (dims >= 2);
numk = 1 + (dims >= 3);
for (k = 0; k < numk; k++) {
for (j = 0; j < numj; 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->gallivm,
bld->texel_type.length,
bld->format_desc->block.bits,
bld->texel_type.width,
data_ptr, offset[k][j][i]);
rgba8 = LLVMBuildBitCast(builder, rgba8, u8n_vec_type, "");
}
else {
rgba8 = lp_build_fetch_rgba_aos(bld->gallivm,
bld->format_desc,
u8n.type,
data_ptr, offset[k][j][i],
x_subcoord[i],
y_subcoord[j]);
}
/* Expand one 4*rgba8 to two 2*rgba16 */
lp_build_unpack2(bld->gallivm, u8n.type, h16.type,
rgba8,
&neighbors_lo[k][j][i], &neighbors_hi[k][j][i]);
}
}
}
/*
* Linear interpolation with 8.8 fixed point.
*/
if (dims == 1) {
/* 1-D lerp */
packed_lo = lp_build_lerp(&h16,
s_fpart_lo,
neighbors_lo[0][0][0],
neighbors_lo[0][0][1]);
packed_hi = lp_build_lerp(&h16,
s_fpart_hi,
neighbors_hi[0][0][0],
neighbors_hi[0][0][1]);
}
else {
/* 2-D lerp */
packed_lo = lp_build_lerp_2d(&h16,
s_fpart_lo, t_fpart_lo,
neighbors_lo[0][0][0],
neighbors_lo[0][0][1],
neighbors_lo[0][1][0],
neighbors_lo[0][1][1]);
packed_hi = lp_build_lerp_2d(&h16,
s_fpart_hi, t_fpart_hi,
neighbors_hi[0][0][0],
neighbors_hi[0][0][1],
neighbors_hi[0][1][0],
neighbors_hi[0][1][1]);
if (dims >= 3) {
LLVMValueRef packed_lo2, packed_hi2;
/* lerp in the second z slice */
packed_lo2 = lp_build_lerp_2d(&h16,
s_fpart_lo, t_fpart_lo,
neighbors_lo[1][0][0],
neighbors_lo[1][0][1],
neighbors_lo[1][1][0],
neighbors_lo[1][1][1]);
packed_hi2 = lp_build_lerp_2d(&h16,
s_fpart_hi, t_fpart_hi,
neighbors_hi[1][0][0],
neighbors_hi[1][0][1],
neighbors_hi[1][1][0],
neighbors_hi[1][1][1]);
/* interp between two z slices */
packed_lo = lp_build_lerp(&h16, r_fpart_lo,
packed_lo, packed_lo2);
packed_hi = lp_build_lerp(&h16, r_fpart_hi,
packed_hi, packed_hi2);
}
}
*colors_lo = packed_lo;
*colors_hi = packed_hi;
}
