/**
* Do the one or two-sided stencil test comparison.
* \sa lp_build_stencil_test_single
* \param face an integer indicating front (+) or back (-) facing polygon.
* If NULL, assume front-facing.
*/
static LLVMValueRef
lp_build_stencil_test(struct lp_build_context *bld,
const struct pipe_stencil_state stencil[2],
LLVMValueRef stencilRefs[2],
LLVMValueRef stencilVals,
LLVMValueRef face)
{
LLVMValueRef res;
assert(stencil[0].enabled);
if (stencil[1].enabled && face) {
/* do two-sided test */
struct lp_build_flow_context *flow_ctx;
struct lp_build_if_state if_ctx;
LLVMValueRef front_facing;
LLVMValueRef zero = LLVMConstReal(LLVMFloatType(), 0.0);
LLVMValueRef result = bld->undef;
flow_ctx = lp_build_flow_create(bld->builder);
lp_build_flow_scope_begin(flow_ctx);
lp_build_flow_scope_declare(flow_ctx, &result);
/* front_facing = face > 0.0 */
front_facing = LLVMBuildFCmp(bld->builder, LLVMRealUGT, face, zero, "");
lp_build_if(&if_ctx, flow_ctx, bld->builder, front_facing);
{
result = lp_build_stencil_test_single(bld, &stencil[0],
stencilRefs[0], stencilVals);
}
lp_build_else(&if_ctx);
{
result = lp_build_stencil_test_single(bld, &stencil[1],
stencilRefs[1], stencilVals);
}
lp_build_endif(&if_ctx);
lp_build_flow_scope_end(flow_ctx);
lp_build_flow_destroy(flow_ctx);
res = result;
}
else {
/* do single-side test */
res = lp_build_stencil_test_single(bld, &stencil[0],
stencilRefs[0], stencilVals);
}
return res;
}
/**
* Generate code to do cube face selection and compute per-face texcoords.
*/
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;
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef rx, ry, rz;
LLVMValueRef arx, ary, arz;
LLVMValueRef c25 = lp_build_const_float(bld->gallivm, 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;
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(builder, LLVMRealUGE, arx, ary, "");
arx_ge_arz = LLVMBuildFCmp(builder, LLVMRealUGE, arx, arz, "");
ary_ge_arx = LLVMBuildFCmp(builder, LLVMRealUGE, ary, arx, "");
ary_ge_arz = LLVMBuildFCmp(builder, LLVMRealUGE, ary, arz, "");
arx_ge_ary_arz = LLVMBuildAnd(builder, arx_ge_ary, arx_ge_arz, "");
ary_ge_arx_arz = LLVMBuildAnd(builder, ary_ge_arx, ary_ge_arz, "");
{
struct lp_build_if_state if_ctx;
LLVMValueRef face_s_var;
LLVMValueRef face_t_var;
LLVMValueRef face_var;
face_s_var = lp_build_alloca(bld->gallivm, bld->coord_bld.vec_type, "face_s_var");
face_t_var = lp_build_alloca(bld->gallivm, bld->coord_bld.vec_type, "face_t_var");
face_var = lp_build_alloca(bld->gallivm, bld->int_bld.vec_type, "face_var");
lp_build_if(&if_ctx, bld->gallivm, 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);
LLVMBuildStore(builder, *face_s, face_s_var);
LLVMBuildStore(builder, *face_t, face_t_var);
LLVMBuildStore(builder, *face, face_var);
}
lp_build_else(&if_ctx);
{
struct lp_build_if_state if_ctx2;
lp_build_if(&if_ctx2, bld->gallivm, ary_ge_arx_arz);
{
/* +/- Y face */
LLVMValueRef sign = lp_build_sgn(float_bld, ry);
LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
*face_s = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
*face_t = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
*face = lp_build_cube_face(bld, ry,
PIPE_TEX_FACE_POS_Y,
PIPE_TEX_FACE_NEG_Y);
LLVMBuildStore(builder, *face_s, face_s_var);
LLVMBuildStore(builder, *face_t, face_t_var);
LLVMBuildStore(builder, *face, face_var);
}
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_s = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
*face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
*face = lp_build_cube_face(bld, rz,
PIPE_TEX_FACE_POS_Z,
PIPE_TEX_FACE_NEG_Z);
LLVMBuildStore(builder, *face_s, face_s_var);
LLVMBuildStore(builder, *face_t, face_t_var);
LLVMBuildStore(builder, *face, face_var);
}
lp_build_endif(&if_ctx2);
}
lp_build_endif(&if_ctx);
*face_s = LLVMBuildLoad(builder, face_s_var, "face_s");
*face_t = LLVMBuildLoad(builder, face_t_var, "face_t");
*face = LLVMBuildLoad(builder, face_var, "face");
}
}
/*
* Do a cached lookup.
*
* Returns (vectors of) 4x8 rgba aos value
*/
LLVMValueRef
lp_build_fetch_cached_texels(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
unsigned n,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j,
LLVMValueRef cache)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned count, low_bit, log2size;
LLVMValueRef color, offset_stored, addr, ptr_addrtrunc, tmp;
LLVMValueRef ij_index, hash_index, hash_mask, block_index;
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef i64t = LLVMInt64TypeInContext(gallivm->context);
struct lp_type type;
struct lp_build_context bld32;
memset(&type, 0, sizeof type);
type.width = 32;
type.length = n;
assert(format_desc->block.width == 4);
assert(format_desc->block.height == 4);
lp_build_context_init(&bld32, gallivm, type);
/*
* compute hash - we use direct mapped cache, the hash function could
* be better but it needs to be simple
* per-element:
* compare offset with offset stored at tag (hash)
* if not equal decode/store block, update tag
* extract color from cache
* assemble result vector
*/
/* TODO: not ideal with 32bit pointers... */
low_bit = util_logbase2(format_desc->block.bits / 8);
log2size = util_logbase2(LP_BUILD_FORMAT_CACHE_SIZE);
addr = LLVMBuildPtrToInt(builder, base_ptr, i64t, "");
ptr_addrtrunc = LLVMBuildPtrToInt(builder, base_ptr, i32t, "");
ptr_addrtrunc = lp_build_broadcast_scalar(&bld32, ptr_addrtrunc);
/* For the hash function, first mask off the unused lowest bits. Then just
do some xor with address bits - only use lower 32bits */
ptr_addrtrunc = LLVMBuildAdd(builder, offset, ptr_addrtrunc, "");
ptr_addrtrunc = LLVMBuildLShr(builder, ptr_addrtrunc,
lp_build_const_int_vec(gallivm, type, low_bit), "");
/* This only really makes sense for size 64,128,256 */
hash_index = ptr_addrtrunc;
ptr_addrtrunc = LLVMBuildLShr(builder, ptr_addrtrunc,
lp_build_const_int_vec(gallivm, type, 2*log2size), "");
hash_index = LLVMBuildXor(builder, ptr_addrtrunc, hash_index, "");
tmp = LLVMBuildLShr(builder, hash_index,
lp_build_const_int_vec(gallivm, type, log2size), "");
hash_index = LLVMBuildXor(builder, hash_index, tmp, "");
hash_mask = lp_build_const_int_vec(gallivm, type, LP_BUILD_FORMAT_CACHE_SIZE - 1);
hash_index = LLVMBuildAnd(builder, hash_index, hash_mask, "");
ij_index = LLVMBuildShl(builder, i, lp_build_const_int_vec(gallivm, type, 2), "");
ij_index = LLVMBuildAdd(builder, ij_index, j, "");
block_index = LLVMBuildShl(builder, hash_index,
lp_build_const_int_vec(gallivm, type, 4), "");
block_index = LLVMBuildAdd(builder, ij_index, block_index, "");
if (n > 1) {
color = LLVMGetUndef(LLVMVectorType(i32t, n));
for (count = 0; count < n; count++) {
LLVMValueRef index, cond, colorx;
LLVMValueRef block_indexx, hash_indexx, addrx, offsetx, ptr_addrx;
struct lp_build_if_state if_ctx;
index = lp_build_const_int32(gallivm, count);
offsetx = LLVMBuildExtractElement(builder, offset, index, "");
addrx = LLVMBuildZExt(builder, offsetx, i64t, "");
addrx = LLVMBuildAdd(builder, addrx, addr, "");
block_indexx = LLVMBuildExtractElement(builder, block_index, index, "");
hash_indexx = LLVMBuildLShr(builder, block_indexx,
lp_build_const_int32(gallivm, 4), "");
offset_stored = lookup_tag_data(gallivm, cache, hash_indexx);
cond = LLVMBuildICmp(builder, LLVMIntNE, offset_stored, addrx, "");
lp_build_if(&if_ctx, gallivm, cond);
{
ptr_addrx = LLVMBuildIntToPtr(builder, addrx,
LLVMPointerType(i8t, 0), "");
update_cached_block(gallivm, format_desc, ptr_addrx, hash_indexx, cache);
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, 1,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
#endif
}
lp_build_endif(&if_ctx);
colorx = lookup_cached_pixel(gallivm, cache, block_indexx);
color = LLVMBuildInsertElement(builder, color, colorx,
lp_build_const_int32(gallivm, count), "");
}
}
else {
LLVMValueRef cond;
struct lp_build_if_state if_ctx;
tmp = LLVMBuildZExt(builder, offset, i64t, "");
addr = LLVMBuildAdd(builder, tmp, addr, "");
offset_stored = lookup_tag_data(gallivm, cache, hash_index);
cond = LLVMBuildICmp(builder, LLVMIntNE, offset_stored, addr, "");
lp_build_if(&if_ctx, gallivm, cond);
{
tmp = LLVMBuildIntToPtr(builder, addr, LLVMPointerType(i8t, 0), "");
update_cached_block(gallivm, format_desc, tmp, hash_index, cache);
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, 1,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
#endif
}
lp_build_endif(&if_ctx);
color = lookup_cached_pixel(gallivm, cache, block_index);
}
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, n,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_TOTAL);
#endif
return LLVMBuildBitCast(builder, color, LLVMVectorType(i8t, n * 4), "");
}
/**
* Generate the code to do inside/outside triangle testing for the
* four pixels in a 2x2 quad. This will set the four elements of the
* quad mask vector to 0 or ~0.
* \param i which quad of the quad group to test, in [0,3]
*/
static void
generate_tri_edge_mask(LLVMBuilderRef builder,
unsigned i,
LLVMValueRef *mask, /* ivec4, out */
LLVMValueRef c0, /* int32 */
LLVMValueRef c1, /* int32 */
LLVMValueRef c2, /* int32 */
LLVMValueRef step0_ptr, /* ivec4 */
LLVMValueRef step1_ptr, /* ivec4 */
LLVMValueRef step2_ptr) /* ivec4 */
{
#define OPTIMIZE_IN_OUT_TEST 0
#if OPTIMIZE_IN_OUT_TEST
struct lp_build_if_state ifctx;
LLVMValueRef not_draw_all;
#endif
struct lp_build_flow_context *flow;
struct lp_type i32_type;
LLVMTypeRef i32vec4_type, mask_type;
LLVMValueRef c0_vec, c1_vec, c2_vec;
LLVMValueRef in_out_mask;
assert(i < 4);
/* int32 vector type */
memset(&i32_type, 0, sizeof i32_type);
i32_type.floating = FALSE; /* values are integers */
i32_type.sign = TRUE; /* values are signed */
i32_type.norm = FALSE; /* values are not normalized */
i32_type.width = 32; /* 32-bit int values */
i32_type.length = 4; /* 4 elements per vector */
i32vec4_type = lp_build_int32_vec4_type();
mask_type = LLVMIntType(32 * 4);
/*
* Use a conditional here to do detailed pixel in/out testing.
* We only have to do this if c0 != INT_MIN.
*/
flow = lp_build_flow_create(builder);
lp_build_flow_scope_begin(flow);
{
#if OPTIMIZE_IN_OUT_TEST
/* not_draw_all = (c0 != INT_MIN) */
not_draw_all = LLVMBuildICmp(builder,
LLVMIntNE,
c0,
LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
"");
in_out_mask = lp_build_int_const_scalar(i32_type, ~0);
lp_build_flow_scope_declare(flow, &in_out_mask);
/* if (not_draw_all) {... */
lp_build_if(&ifctx, flow, builder, not_draw_all);
#endif
{
LLVMValueRef step0_vec, step1_vec, step2_vec;
LLVMValueRef m0_vec, m1_vec, m2_vec;
LLVMValueRef index, m;
/* c0_vec = {c0, c0, c0, c0}
* Note that we emit this code four times but LLVM optimizes away
* three instances of it.
*/
c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
lp_build_name(c0_vec, "edgeconst0vec");
lp_build_name(c1_vec, "edgeconst1vec");
lp_build_name(c2_vec, "edgeconst2vec");
/* load step0vec, step1, step2 vec from memory */
index = LLVMConstInt(LLVMInt32Type(), i, 0);
step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
lp_build_name(step0_vec, "step0vec");
lp_build_name(step1_vec, "step1vec");
lp_build_name(step2_vec, "step2vec");
/* m0_vec = step0_ptr[i] > c0_vec */
m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
/* in_out_mask = m0_vec & m1_vec & m2_vec */
m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
lp_build_name(in_out_mask, "inoutmaskvec");
}
#if OPTIMIZE_IN_OUT_TEST
lp_build_endif(&ifctx);
#endif
}
lp_build_flow_scope_end(flow);
lp_build_flow_destroy(flow);
/* This is the initial alive/dead pixel mask for a quad of four pixels.
* It's an int[4] vector with each word set to 0 or ~0.
* Words will get cleared when pixels faile the Z test, etc.
*/
*mask = in_out_mask;
}
/**
* Texture sampling in AoS format. Used when sampling common 32-bit/texel
* formats. 1D/2D/3D/cube texture supported. All mipmap sampling modes
* but only limited texture coord wrap modes.
*/
void
lp_build_sample_aos(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 texel_out[4])
{
struct lp_build_context *int_bld = &bld->int_bld;
LLVMBuilderRef builder = bld->gallivm->builder;
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 unsigned dims = bld->dims;
LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
LLVMValueRef ilevel0, ilevel1 = NULL;
LLVMValueRef packed, packed_lo, packed_hi;
LLVMValueRef unswizzled[4];
LLVMValueRef face_ddx[4], face_ddy[4];
struct lp_build_context h16_bld;
LLVMValueRef first_level;
LLVMValueRef i32t_zero = lp_build_const_int32(bld->gallivm, 0);
/* we only support the common/simple wrap modes at this time */
assert(lp_is_simple_wrap_mode(bld->static_state->wrap_s));
if (dims >= 2)
assert(lp_is_simple_wrap_mode(bld->static_state->wrap_t));
if (dims >= 3)
assert(lp_is_simple_wrap_mode(bld->static_state->wrap_r));
/* make 16-bit fixed-pt builder context */
lp_build_context_init(&h16_bld, bld->gallivm, lp_type_ufixed(16));
/* cube face selection, compute pre-face coords, etc. */
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_scalar_ddx(&bld->coord_bld, s);
face_ddx[1] = lp_build_scalar_ddx(&bld->coord_bld, t);
face_ddx[2] = NULL;
face_ddx[3] = NULL;
face_ddy[0] = lp_build_scalar_ddy(&bld->coord_bld, s);
face_ddy[1] = lp_build_scalar_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).
*/
if (min_filter != mag_filter ||
mip_filter != PIPE_TEX_MIPFILTER_NONE) {
/* Need to compute lod either to choose mipmap levels or to
* distinguish between minification/magnification with one mipmap level.
*/
lp_build_lod_selector(bld, unit, ddx, ddy,
lod_bias, explicit_lod,
mip_filter,
&lod_ipart, &lod_fpart);
} else {
lod_ipart = i32t_zero;
}
/*
* Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
*/
switch (mip_filter) {
default:
assert(0 && "bad mip_filter value in lp_build_sample_aos()");
/* fall-through */
case PIPE_TEX_MIPFILTER_NONE:
/* always use mip level 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.
*/
assert(lod_ipart);
lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
}
else {
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
bld->gallivm, unit);
ilevel0 = first_level;
}
break;
case PIPE_TEX_MIPFILTER_NEAREST:
assert(lod_ipart);
lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
break;
case PIPE_TEX_MIPFILTER_LINEAR:
assert(lod_ipart);
assert(lod_fpart);
lp_build_linear_mip_levels(bld, unit,
lod_ipart, &lod_fpart,
&ilevel0, &ilevel1);
break;
}
/*
* Get/interpolate texture colors.
*/
packed_lo = lp_build_alloca(bld->gallivm, h16_bld.vec_type, "packed_lo");
packed_hi = lp_build_alloca(bld->gallivm, h16_bld.vec_type, "packed_hi");
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,
ilevel0, ilevel1, lod_fpart,
packed_lo, packed_hi);
}
else {
/* Emit conditional to choose min image filter or mag image filter
* depending on the lod being > 0 or <= 0, respectively.
*/
struct lp_build_if_state if_ctx;
LLVMValueRef minify;
/* minify = lod >= 0.0 */
minify = LLVMBuildICmp(builder, LLVMIntSGE,
lod_ipart, int_bld->zero, "");
lp_build_if(&if_ctx, bld->gallivm, minify);
{
/* Use the minification filter */
lp_build_sample_mipmap(bld,
min_filter, mip_filter,
s, t, r,
ilevel0, ilevel1, lod_fpart,
packed_lo, packed_hi);
}
lp_build_else(&if_ctx);
{
/* Use the magnification filter */
lp_build_sample_mipmap(bld,
mag_filter, PIPE_TEX_MIPFILTER_NONE,
s, t, r,
ilevel0, NULL, NULL,
packed_lo, packed_hi);
}
lp_build_endif(&if_ctx);
}
/*
* combine the values stored in 'packed_lo' and 'packed_hi' variables
* into 'packed'
*/
packed = lp_build_pack2(bld->gallivm,
h16_bld.type, lp_type_unorm(8),
LLVMBuildLoad(builder, packed_lo, ""),
LLVMBuildLoad(builder, packed_hi, ""));
/*
* Convert to SoA and swizzle.
*/
lp_build_rgba8_to_f32_soa(bld->gallivm,
bld->texel_type,
packed, unswizzled);
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];
}
}
/**
* 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);
}
}
