/**
* Non-interleaved pack and saturate.
*
* Same as lp_build_pack2 but will saturate values so that they fit into the
* destination type.
*/
LLVMValueRef
lp_build_packs2(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
LLVMValueRef lo,
LLVMValueRef hi)
{
boolean clamp;
assert(!src_type.floating);
assert(!dst_type.floating);
assert(src_type.sign == dst_type.sign);
assert(src_type.width == dst_type.width * 2);
assert(src_type.length * 2 == dst_type.length);
clamp = TRUE;
/* All X86 SSE non-interleaved pack instructions take signed inputs and
* saturate them, so no need to clamp for those cases. */
if(util_cpu_caps.has_sse2 &&
src_type.width * src_type.length >= 128 &&
src_type.sign &&
(src_type.width == 32 || src_type.width == 16))
clamp = FALSE;
if(clamp) {
struct lp_build_context bld;
unsigned dst_bits = dst_type.sign ? dst_type.width - 1 : dst_type.width;
LLVMValueRef dst_max = lp_build_const_int_vec(gallivm, src_type, ((unsigned long long)1 << dst_bits) - 1);
lp_build_context_init(&bld, gallivm, src_type);
lo = lp_build_min(&bld, lo, dst_max);
hi = lp_build_min(&bld, hi, dst_max);
/* FIXME: What about lower bound? */
}
return lp_build_pack2(gallivm, src_type, dst_type, lo, hi);
}
/**
* 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];
}
}
/**
* Generic type conversion.
*
* TODO: Take a precision argument, or even better, add a new precision member
* to the lp_type union.
*/
void
lp_build_conv(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
const LLVMValueRef *src, unsigned num_srcs,
LLVMValueRef *dst, unsigned num_dsts)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_type tmp_type;
LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
unsigned num_tmps;
unsigned i;
/* We must not loose or gain channels. Only precision */
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
assert(num_srcs <= LP_MAX_VECTOR_LENGTH);
assert(num_dsts <= LP_MAX_VECTOR_LENGTH);
tmp_type = src_type;
for(i = 0; i < num_srcs; ++i) {
assert(lp_check_value(src_type, src[i]));
tmp[i] = src[i];
}
num_tmps = num_srcs;
/* Special case 4x4f --> 1x16ub
*/
if (src_type.floating == 1 &&
src_type.fixed == 0 &&
src_type.sign == 1 &&
src_type.norm == 0 &&
src_type.width == 32 &&
src_type.length == 4 &&
dst_type.floating == 0 &&
dst_type.fixed == 0 &&
dst_type.sign == 0 &&
dst_type.norm == 1 &&
dst_type.width == 8 &&
dst_type.length == 16 &&
4 * num_dsts == num_srcs &&
util_cpu_caps.has_sse2)
{
struct lp_build_context bld;
struct lp_type int16_type = dst_type;
struct lp_type int32_type = dst_type;
LLVMValueRef const_255f;
unsigned i, j;
lp_build_context_init(&bld, gallivm, src_type);
int16_type.width *= 2;
int16_type.length /= 2;
int16_type.sign = 1;
int32_type.width *= 4;
int32_type.length /= 4;
int32_type.sign = 1;
const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);
for (i = 0; i < num_dsts; ++i, src += 4) {
LLVMValueRef lo, hi;
for (j = 0; j < 4; ++j) {
tmp[j] = LLVMBuildFMul(builder, src[j], const_255f, "");
tmp[j] = lp_build_iround(&bld, tmp[j]);
}
/* relying on clamping behavior of sse2 intrinsics here */
lo = lp_build_pack2(gallivm, int32_type, int16_type, tmp[0], tmp[1]);
hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
}
return;
}
/* Special case 2x8f --> 1x16ub
*/
else if (src_type.floating == 1 &&
src_type.fixed == 0 &&
src_type.sign == 1 &&
src_type.norm == 0 &&
src_type.width == 32 &&
src_type.length == 8 &&
dst_type.floating == 0 &&
dst_type.fixed == 0 &&
dst_type.sign == 0 &&
dst_type.norm == 1 &&
dst_type.width == 8 &&
dst_type.length == 16 &&
2 * num_dsts == num_srcs &&
util_cpu_caps.has_avx) {
struct lp_build_context bld;
struct lp_type int16_type = dst_type;
struct lp_type int32_type = dst_type;
LLVMValueRef const_255f;
unsigned i;
lp_build_context_init(&bld, gallivm, src_type);
int16_type.width *= 2;
int16_type.length /= 2;
int16_type.sign = 1;
int32_type.width *= 4;
int32_type.length /= 4;
int32_type.sign = 1;
const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);
for (i = 0; i < num_dsts; ++i, src += 2) {
LLVMValueRef lo, hi, a, b;
a = LLVMBuildFMul(builder, src[0], const_255f, "");
b = LLVMBuildFMul(builder, src[1], const_255f, "");
a = lp_build_iround(&bld, a);
b = lp_build_iround(&bld, b);
tmp[0] = lp_build_extract_range(gallivm, a, 0, 4);
tmp[1] = lp_build_extract_range(gallivm, a, 4, 4);
tmp[2] = lp_build_extract_range(gallivm, b, 0, 4);
tmp[3] = lp_build_extract_range(gallivm, b, 4, 4);
/* relying on clamping behavior of sse2 intrinsics here */
lo = lp_build_pack2(gallivm, int32_type, int16_type, tmp[0], tmp[1]);
hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
}
return;
}
/* Pre convert half-floats to floats
*/
else if (src_type.floating && src_type.width == 16)
{
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_half_to_float(gallivm, src_type, tmp[i]);
tmp_type.width = 32;
}
/*
* Clamp if necessary
*/
if(memcmp(&src_type, &dst_type, sizeof src_type) != 0) {
struct lp_build_context bld;
double src_min = lp_const_min(src_type);
double dst_min = lp_const_min(dst_type);
double src_max = lp_const_max(src_type);
double dst_max = lp_const_max(dst_type);
LLVMValueRef thres;
lp_build_context_init(&bld, gallivm, tmp_type);
if(src_min < dst_min) {
if(dst_min == 0.0)
thres = bld.zero;
else
thres = lp_build_const_vec(gallivm, src_type, dst_min);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_max(&bld, tmp[i], thres);
}
if(src_max > dst_max) {
if(dst_max == 1.0)
thres = bld.one;
else
thres = lp_build_const_vec(gallivm, src_type, dst_max);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_min(&bld, tmp[i], thres);
}
}
/*
* Scale to the narrowest range
*/
if(dst_type.floating) {
/* Nothing to do */
}
else if(tmp_type.floating) {
if(!dst_type.fixed && !dst_type.sign && dst_type.norm) {
for(i = 0; i < num_tmps; ++i) {
tmp[i] = lp_build_clamped_float_to_unsigned_norm(gallivm,
tmp_type,
dst_type.width,
tmp[i]);
}
tmp_type.floating = FALSE;
}
else {
double dst_scale = lp_const_scale(dst_type);
LLVMTypeRef tmp_vec_type;
if (dst_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, dst_scale);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = FALSE;
tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
else
tmp[i] = LLVMBuildFPToUI(builder, tmp[i], tmp_vec_type, "");
#else
/* FIXME: there is no SSE counterpart for LLVMBuildFPToUI */
tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
#endif
}
}
}
else {
unsigned src_shift = lp_const_shift(src_type);
unsigned dst_shift = lp_const_shift(dst_type);
unsigned src_offset = lp_const_offset(src_type);
unsigned dst_offset = lp_const_offset(dst_type);
/* Compensate for different offsets */
if (dst_offset > src_offset && src_type.width > dst_type.width) {
for (i = 0; i < num_tmps; ++i) {
LLVMValueRef shifted;
LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, src_shift - 1);
if(src_type.sign)
shifted = LLVMBuildAShr(builder, tmp[i], shift, "");
else
shifted = LLVMBuildLShr(builder, tmp[i], shift, "");
tmp[i] = LLVMBuildSub(builder, tmp[i], shifted, "");
}
}
if(src_shift > dst_shift) {
LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type,
src_shift - dst_shift);
for(i = 0; i < num_tmps; ++i)
if(src_type.sign)
tmp[i] = LLVMBuildAShr(builder, tmp[i], shift, "");
else
tmp[i] = LLVMBuildLShr(builder, tmp[i], shift, "");
}
}
/*
* Truncate or expand bit width
*
* No data conversion should happen here, although the sign bits are
* crucial to avoid bad clamping.
*/
{
struct lp_type new_type;
new_type = tmp_type;
new_type.sign = dst_type.sign;
new_type.width = dst_type.width;
new_type.length = dst_type.length;
lp_build_resize(gallivm, tmp_type, new_type, tmp, num_srcs, tmp, num_dsts);
tmp_type = new_type;
num_tmps = num_dsts;
}
/*
* Scale to the widest range
*/
if(src_type.floating) {
/* Nothing to do */
}
else if(!src_type.floating && dst_type.floating) {
if(!src_type.fixed && !src_type.sign && src_type.norm) {
for(i = 0; i < num_tmps; ++i) {
tmp[i] = lp_build_unsigned_norm_to_float(gallivm,
src_type.width,
dst_type,
tmp[i]);
}
tmp_type.floating = TRUE;
}
else {
double src_scale = lp_const_scale(src_type);
LLVMTypeRef tmp_vec_type;
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = TRUE;
tmp_type.sign = TRUE;
tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
else
tmp[i] = LLVMBuildUIToFP(builder, tmp[i], tmp_vec_type, "");
#else
/* FIXME: there is no SSE counterpart for LLVMBuildUIToFP */
tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
#endif
}
if (src_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, 1.0/src_scale);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
}
}
else {
unsigned src_shift = lp_const_shift(src_type);
unsigned dst_shift = lp_const_shift(dst_type);
unsigned src_offset = lp_const_offset(src_type);
unsigned dst_offset = lp_const_offset(dst_type);
if (src_shift < dst_shift) {
LLVMValueRef pre_shift[LP_MAX_VECTOR_LENGTH];
LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, dst_shift - src_shift);
for (i = 0; i < num_tmps; ++i) {
pre_shift[i] = tmp[i];
tmp[i] = LLVMBuildShl(builder, tmp[i], shift, "");
}
/* Compensate for different offsets */
if (dst_offset > src_offset) {
for (i = 0; i < num_tmps; ++i) {
tmp[i] = LLVMBuildSub(builder, tmp[i], pre_shift[i], "");
}
}
}
}
for(i = 0; i < num_dsts; ++i) {
dst[i] = tmp[i];
assert(lp_check_value(dst_type, dst[i]));
}
}
/**
* Generate a * b
*/
LLVMValueRef
lp_build_mul(struct lp_build_context *bld,
LLVMValueRef a,
LLVMValueRef b)
{
const struct lp_type type = bld->type;
LLVMValueRef shift;
LLVMValueRef res;
if(a == bld->zero)
return bld->zero;
if(a == bld->one)
return b;
if(b == bld->zero)
return bld->zero;
if(b == bld->one)
return a;
if(a == bld->undef || b == bld->undef)
return bld->undef;
if(!type.floating && !type.fixed && type.norm) {
if(type.width == 8) {
struct lp_type i16_type = lp_wider_type(type);
LLVMValueRef al, ah, bl, bh, abl, abh, ab;
lp_build_unpack2(bld->builder, type, i16_type, a, &al, &ah);
lp_build_unpack2(bld->builder, type, i16_type, b, &bl, &bh);
/* PMULLW, PSRLW, PADDW */
abl = lp_build_mul_u8n(bld->builder, i16_type, al, bl);
abh = lp_build_mul_u8n(bld->builder, i16_type, ah, bh);
ab = lp_build_pack2(bld->builder, i16_type, type, abl, abh);
return ab;
}
/* FIXME */
assert(0);
}
if(type.fixed)
shift = lp_build_int_const_scalar(type, type.width/2);
else
shift = NULL;
if(LLVMIsConstant(a) && LLVMIsConstant(b)) {
res = LLVMConstMul(a, b);
if(shift) {
if(type.sign)
res = LLVMConstAShr(res, shift);
else
res = LLVMConstLShr(res, shift);
}
}
else {
res = LLVMBuildMul(bld->builder, a, b, "");
if(shift) {
if(type.sign)
res = LLVMBuildAShr(bld->builder, res, shift, "");
else
res = LLVMBuildLShr(bld->builder, res, shift, "");
}
}
return res;
}
