/**
* Interleave vector elements.
*
* Matches the PUNPCKLxx and PUNPCKHxx SSE instructions
* (but not for 256bit AVX vectors).
*/
LLVMValueRef
lp_build_interleave2(struct gallivm_state *gallivm,
struct lp_type type,
LLVMValueRef a,
LLVMValueRef b,
unsigned lo_hi)
{
LLVMValueRef shuffle;
if (type.length == 2 && type.width == 128 && util_cpu_caps.has_avx) {
/*
* XXX: This is a workaround for llvm code generation deficiency. Strangely
* enough, while this needs vinsertf128/vextractf128 instructions (hence
* a natural match when using 2x128bit vectors) the "normal" unpack shuffle
* generates code ranging from atrocious (llvm 3.1) to terrible (llvm 3.2, 3.3).
* So use some different shuffles instead (the exact shuffles don't seem to
* matter, as long as not using 128bit wide vectors, works with 8x32 or 4x64).
*/
struct lp_type tmp_type = type;
LLVMValueRef srchalf[2], tmpdst;
tmp_type.length = 4;
tmp_type.width = 64;
a = LLVMBuildBitCast(gallivm->builder, a, lp_build_vec_type(gallivm, tmp_type), "");
b = LLVMBuildBitCast(gallivm->builder, b, lp_build_vec_type(gallivm, tmp_type), "");
srchalf[0] = lp_build_extract_range(gallivm, a, lo_hi * 2, 2);
srchalf[1] = lp_build_extract_range(gallivm, b, lo_hi * 2, 2);
tmp_type.length = 2;
tmpdst = lp_build_concat(gallivm, srchalf, tmp_type, 2);
return LLVMBuildBitCast(gallivm->builder, tmpdst, lp_build_vec_type(gallivm, type), "");
}
shuffle = lp_build_const_unpack_shuffle(gallivm, type.length, lo_hi);
return LLVMBuildShuffleVector(gallivm->builder, a, b, shuffle, "");
}
/**
* Truncate or expand the bitwidth.
*
* NOTE: Getting the right sign flags is crucial here, as we employ some
* intrinsics that do saturation.
*/
void
lp_build_resize(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;
LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
unsigned i;
/*
* We don't support float <-> int conversion here. That must be done
* before/after calling this function.
*/
assert(src_type.floating == dst_type.floating);
/*
* We don't support double <-> float conversion yet, although it could be
* added with little effort.
*/
assert((!src_type.floating && !dst_type.floating) ||
src_type.width == dst_type.width);
/* We must not loose or gain channels. Only precision */
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
/* We don't support M:N conversion, only 1:N, M:1, or 1:1 */
assert(num_srcs == 1 || num_dsts == 1);
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);
if (src_type.width > dst_type.width) {
/*
* Truncate bit width.
*/
assert(num_dsts == 1);
if (src_type.width * src_type.length == dst_type.width * dst_type.length) {
/*
* Register width remains constant -- use vector packing intrinsics
*/
tmp[0] = lp_build_pack(gallivm, src_type, dst_type, TRUE, src, num_srcs);
}
else {
if (src_type.width / dst_type.width > num_srcs) {
/*
* First change src vectors size (with shuffle) so they have the
* same size as the destination vector, then pack normally.
* Note: cannot use cast/extract because llvm generates atrocious code.
*/
unsigned size_ratio = (src_type.width * src_type.length) /
(dst_type.length * dst_type.width);
unsigned new_length = src_type.length / size_ratio;
for (i = 0; i < size_ratio * num_srcs; i++) {
unsigned start_index = (i % size_ratio) * new_length;
tmp[i] = lp_build_extract_range(gallivm, src[i / size_ratio],
start_index, new_length);
}
num_srcs *= size_ratio;
src_type.length = new_length;
tmp[0] = lp_build_pack(gallivm, src_type, dst_type, TRUE, tmp, num_srcs);
}
else {
/*
* Truncate bit width but expand vector size - first pack
* then expand simply because this should be more AVX-friendly
* for the cases we probably hit.
*/
unsigned size_ratio = (dst_type.width * dst_type.length) /
(src_type.length * src_type.width);
unsigned num_pack_srcs = num_srcs / size_ratio;
dst_type.length = dst_type.length / size_ratio;
for (i = 0; i < size_ratio; i++) {
tmp[i] = lp_build_pack(gallivm, src_type, dst_type, TRUE,
&src[i*num_pack_srcs], num_pack_srcs);
}
tmp[0] = lp_build_concat(gallivm, tmp, dst_type, size_ratio);
}
}
}
else if (src_type.width < dst_type.width) {
/*
* Expand bit width.
*/
assert(num_srcs == 1);
if (src_type.width * src_type.length == dst_type.width * dst_type.length) {
/*
* Register width remains constant -- use vector unpack intrinsics
*/
lp_build_unpack(gallivm, src_type, dst_type, src[0], tmp, num_dsts);
}
else {
/*
* Do it element-wise.
*/
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
for (i = 0; i < num_dsts; i++) {
tmp[i] = lp_build_undef(gallivm, dst_type);
}
for (i = 0; i < src_type.length; ++i) {
unsigned j = i / dst_type.length;
LLVMValueRef srcindex = lp_build_const_int32(gallivm, i);
LLVMValueRef dstindex = lp_build_const_int32(gallivm, i % dst_type.length);
LLVMValueRef val = LLVMBuildExtractElement(builder, src[0], srcindex, "");
if (src_type.sign && dst_type.sign) {
val = LLVMBuildSExt(builder, val, lp_build_elem_type(gallivm, dst_type), "");
} else {
val = LLVMBuildZExt(builder, val, lp_build_elem_type(gallivm, dst_type), "");
}
tmp[j] = LLVMBuildInsertElement(builder, tmp[j], val, dstindex, "");
}
}
}
else {
/*
* No-op
*/
assert(num_srcs == 1);
assert(num_dsts == 1);
tmp[0] = src[0];
}
for(i = 0; i < num_dsts; ++i)
dst[i] = tmp[i];
}
/**
* Non-interleaved pack.
*
* This will move values as
* (LSB) (MSB)
* lo = l0 __ l1 __ l2 __.. __ ln __
* hi = h0 __ h1 __ h2 __.. __ hn __
* res = l0 l1 l2 .. ln h0 h1 h2 .. hn
*
* This will only change the number of bits the values are represented, not the
* values themselves.
*
* It is assumed the values are already clamped into the destination type range.
* Values outside that range will produce undefined results. Use
* lp_build_packs2 instead.
*/
LLVMValueRef
lp_build_pack2(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
LLVMValueRef lo,
LLVMValueRef hi)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef dst_vec_type = lp_build_vec_type(gallivm, dst_type);
LLVMValueRef shuffle;
LLVMValueRef res = NULL;
struct lp_type intr_type = dst_type;
#if HAVE_LLVM < 0x0207
intr_type = src_type;
#endif
assert(!src_type.floating);
assert(!dst_type.floating);
assert(src_type.width == dst_type.width * 2);
assert(src_type.length * 2 == dst_type.length);
/* Check for special cases first */
if(util_cpu_caps.has_sse2 && src_type.width * src_type.length >= 128) {
const char *intrinsic = NULL;
switch(src_type.width) {
case 32:
if(dst_type.sign) {
intrinsic = "llvm.x86.sse2.packssdw.128";
}
else {
if (util_cpu_caps.has_sse4_1) {
intrinsic = "llvm.x86.sse41.packusdw";
#if HAVE_LLVM < 0x0207
/* llvm < 2.7 has inconsistent signatures except for packusdw */
intr_type = dst_type;
#endif
}
}
break;
case 16:
if (dst_type.sign) {
intrinsic = "llvm.x86.sse2.packsswb.128";
}
else {
intrinsic = "llvm.x86.sse2.packuswb.128";
}
break;
/* default uses generic shuffle below */
}
if (intrinsic) {
if (src_type.width * src_type.length == 128) {
LLVMTypeRef intr_vec_type = lp_build_vec_type(gallivm, intr_type);
res = lp_build_intrinsic_binary(builder, intrinsic, intr_vec_type, lo, hi);
if (dst_vec_type != intr_vec_type) {
res = LLVMBuildBitCast(builder, res, dst_vec_type, "");
}
}
else {
int num_split = src_type.width * src_type.length / 128;
int i;
int nlen = 128 / src_type.width;
struct lp_type ndst_type = lp_type_unorm(dst_type.width, 128);
struct lp_type nintr_type = lp_type_unorm(intr_type.width, 128);
LLVMValueRef tmpres[LP_MAX_VECTOR_WIDTH / 128];
LLVMValueRef tmplo, tmphi;
LLVMTypeRef ndst_vec_type = lp_build_vec_type(gallivm, ndst_type);
LLVMTypeRef nintr_vec_type = lp_build_vec_type(gallivm, nintr_type);
assert(num_split <= LP_MAX_VECTOR_WIDTH / 128);
for (i = 0; i < num_split / 2; i++) {
tmplo = lp_build_extract_range(gallivm,
lo, i*nlen*2, nlen);
tmphi = lp_build_extract_range(gallivm,
lo, i*nlen*2 + nlen, nlen);
tmpres[i] = lp_build_intrinsic_binary(builder, intrinsic,
nintr_vec_type, tmplo, tmphi);
if (ndst_vec_type != nintr_vec_type) {
tmpres[i] = LLVMBuildBitCast(builder, tmpres[i], ndst_vec_type, "");
}
}
for (i = 0; i < num_split / 2; i++) {
tmplo = lp_build_extract_range(gallivm,
hi, i*nlen*2, nlen);
tmphi = lp_build_extract_range(gallivm,
hi, i*nlen*2 + nlen, nlen);
tmpres[i+num_split/2] = lp_build_intrinsic_binary(builder, intrinsic,
nintr_vec_type,
tmplo, tmphi);
if (ndst_vec_type != nintr_vec_type) {
tmpres[i+num_split/2] = LLVMBuildBitCast(builder, tmpres[i+num_split/2],
ndst_vec_type, "");
}
}
res = lp_build_concat(gallivm, tmpres, ndst_type, num_split);
}
return res;
}
}
/* generic shuffle */
lo = LLVMBuildBitCast(builder, lo, dst_vec_type, "");
hi = LLVMBuildBitCast(builder, hi, dst_vec_type, "");
shuffle = lp_build_const_pack_shuffle(gallivm, dst_type.length);
res = LLVMBuildShuffleVector(builder, lo, hi, shuffle, "");
return res;
}
/**
* 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]));
}
}
/**
* Store depth/stencil values.
* Incoming values are swizzled (typically n 2x2 quads), stored linear.
* If there's a mask it will do select/store otherwise just store.
*
* \param type the data type of the fragment depth/stencil values
* \param format_desc description of the depth/stencil surface
* \param mask the alive/dead pixel mask for the quad (vector)
* \param z_fb z values read from fb (with padding)
* \param s_fb s values read from fb (with padding)
* \param loop_counter the current loop iteration
* \param depth_ptr pointer to the depth/stencil values of this 4x4 block
* \param depth_stride stride of the depth/stencil buffer
* \param z_value the depth values to store (with padding)
* \param s_value the stencil values to store (with padding)
*/
void
lp_build_depth_stencil_write_swizzled(struct gallivm_state *gallivm,
struct lp_type z_src_type,
const struct util_format_description *format_desc,
struct lp_build_mask_context *mask,
LLVMValueRef z_fb,
LLVMValueRef s_fb,
LLVMValueRef loop_counter,
LLVMValueRef depth_ptr,
LLVMValueRef depth_stride,
LLVMValueRef z_value,
LLVMValueRef s_value)
{
struct lp_build_context z_bld;
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 4];
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef mask_value = NULL;
LLVMValueRef zs_dst1, zs_dst2;
LLVMValueRef zs_dst_ptr1, zs_dst_ptr2;
LLVMValueRef depth_offset1, depth_offset2;
LLVMTypeRef load_ptr_type;
unsigned depth_bytes = format_desc->block.bits / 8;
struct lp_type zs_type = lp_depth_type(format_desc, z_src_type.length);
struct lp_type z_type = zs_type;
struct lp_type zs_load_type = zs_type;
zs_load_type.length = zs_load_type.length / 2;
load_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0);
z_type.width = z_src_type.width;
lp_build_context_init(&z_bld, gallivm, z_type);
/*
* This is far from ideal, at least for late depth write we should do this
* outside the fs loop to avoid all the swizzle stuff.
*/
if (z_src_type.length == 4) {
LLVMValueRef looplsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
LLVMValueRef loopmsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 2), "");
LLVMValueRef offset2 = LLVMBuildMul(builder, loopmsb,
depth_stride, "");
depth_offset1 = LLVMBuildMul(builder, looplsb,
lp_build_const_int32(gallivm, depth_bytes * 2), "");
depth_offset1 = LLVMBuildAdd(builder, depth_offset1, offset2, "");
}
else {
unsigned i;
LLVMValueRef loopx2 = LLVMBuildShl(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
assert(z_src_type.length == 8);
depth_offset1 = LLVMBuildMul(builder, loopx2, depth_stride, "");
/*
* We load 2x4 values, and need to swizzle them (order
* 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
*/
for (i = 0; i < 8; i++) {
shuffles[i] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
}
}
depth_offset2 = LLVMBuildAdd(builder, depth_offset1, depth_stride, "");
zs_dst_ptr1 = LLVMBuildGEP(builder, depth_ptr, &depth_offset1, 1, "");
zs_dst_ptr1 = LLVMBuildBitCast(builder, zs_dst_ptr1, load_ptr_type, "");
zs_dst_ptr2 = LLVMBuildGEP(builder, depth_ptr, &depth_offset2, 1, "");
zs_dst_ptr2 = LLVMBuildBitCast(builder, zs_dst_ptr2, load_ptr_type, "");
if (format_desc->block.bits > 32) {
s_value = LLVMBuildBitCast(builder, s_value, z_bld.vec_type, "");
}
if (mask) {
mask_value = lp_build_mask_value(mask);
z_value = lp_build_select(&z_bld, mask_value, z_value, z_fb);
if (format_desc->block.bits > 32) {
s_fb = LLVMBuildBitCast(builder, s_fb, z_bld.vec_type, "");
s_value = lp_build_select(&z_bld, mask_value, s_value, s_fb);
}
}
if (zs_type.width < z_src_type.width) {
/* Truncate ZS values (e.g., when writing to Z16_UNORM) */
z_value = LLVMBuildTrunc(builder, z_value,
lp_build_int_vec_type(gallivm, zs_type), "");
}
if (format_desc->block.bits <= 32) {
if (z_src_type.length == 4) {
zs_dst1 = lp_build_extract_range(gallivm, z_value, 0, 2);
zs_dst2 = lp_build_extract_range(gallivm, z_value, 2, 2);
}
else {
assert(z_src_type.length == 8);
zs_dst1 = LLVMBuildShuffleVector(builder, z_value, z_value,
LLVMConstVector(&shuffles[0],
zs_load_type.length), "");
zs_dst2 = LLVMBuildShuffleVector(builder, z_value, z_value,
LLVMConstVector(&shuffles[4],
zs_load_type.length), "");
}
}
else {
if (z_src_type.length == 4) {
zs_dst1 = lp_build_interleave2(gallivm, z_type,
z_value, s_value, 0);
zs_dst2 = lp_build_interleave2(gallivm, z_type,
z_value, s_value, 1);
}
else {
unsigned i;
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 2];
assert(z_src_type.length == 8);
for (i = 0; i < 8; i++) {
shuffles[i*2] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
shuffles[i*2+1] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2 +
z_src_type.length);
}
zs_dst1 = LLVMBuildShuffleVector(builder, z_value, s_value,
LLVMConstVector(&shuffles[0],
z_src_type.length), "");
zs_dst2 = LLVMBuildShuffleVector(builder, z_value, s_value,
LLVMConstVector(&shuffles[8],
z_src_type.length), "");
}
zs_dst1 = LLVMBuildBitCast(builder, zs_dst1,
lp_build_vec_type(gallivm, zs_load_type), "");
zs_dst2 = LLVMBuildBitCast(builder, zs_dst2,
lp_build_vec_type(gallivm, zs_load_type), "");
}
LLVMBuildStore(builder, zs_dst1, zs_dst_ptr1);
LLVMBuildStore(builder, zs_dst2, zs_dst_ptr2);
}
/**
* Converts float32 to int16 half-float
* Note this can be performed in 1 instruction if vcvtps2ph exists (f16c/cvt16)
* [llvm.x86.vcvtps2ph / _mm_cvtps_ph]
*
* @param src value to convert
*
* Convert float32 to half floats, preserving Infs and NaNs,
* with rounding towards zero (trunc).
*/
LLVMValueRef
lp_build_float_to_half(struct gallivm_state *gallivm,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef f32_vec_type = LLVMTypeOf(src);
unsigned length = LLVMGetTypeKind(f32_vec_type) == LLVMVectorTypeKind
? LLVMGetVectorSize(f32_vec_type) : 1;
struct lp_type i32_type = lp_type_int_vec(32, 32 * length);
struct lp_type i16_type = lp_type_int_vec(16, 16 * length);
LLVMValueRef result;
if (util_cpu_caps.has_f16c && HAVE_LLVM >= 0x0301 &&
(length == 4 || length == 8)) {
struct lp_type i168_type = lp_type_int_vec(16, 16 * 8);
unsigned mode = 3; /* same as LP_BUILD_ROUND_TRUNCATE */
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
const char *intrinsic = NULL;
if (length == 4) {
intrinsic = "llvm.x86.vcvtps2ph.128";
}
else {
intrinsic = "llvm.x86.vcvtps2ph.256";
}
result = lp_build_intrinsic_binary(builder, intrinsic,
lp_build_vec_type(gallivm, i168_type),
src, LLVMConstInt(i32t, mode, 0));
if (length == 4) {
result = lp_build_extract_range(gallivm, result, 0, 4);
}
}
else {
result = lp_build_float_to_smallfloat(gallivm, i32_type, src, 10, 5, 0, true);
/* Convert int32 vector to int16 vector by trunc (might generate bad code) */
result = LLVMBuildTrunc(builder, result, lp_build_vec_type(gallivm, i16_type), "");
}
/*
* Debugging code.
*/
if (0) {
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef i16t = LLVMInt16TypeInContext(gallivm->context);
LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
LLVMValueRef ref_result = LLVMGetUndef(LLVMVectorType(i16t, length));
unsigned i;
LLVMTypeRef func_type = LLVMFunctionType(i16t, &f32t, 1, 0);
LLVMValueRef func = lp_build_const_int_pointer(gallivm, func_to_pointer((func_pointer)util_float_to_half));
func = LLVMBuildBitCast(builder, func, LLVMPointerType(func_type, 0), "util_float_to_half");
for (i = 0; i < length; ++i) {
LLVMValueRef index = LLVMConstInt(i32t, i, 0);
LLVMValueRef f32 = LLVMBuildExtractElement(builder, src, index, "");
#if 0
/* XXX: not really supported by backends */
LLVMValueRef f16 = lp_build_intrinsic_unary(builder, "llvm.convert.to.fp16", i16t, f32);
#else
LLVMValueRef f16 = LLVMBuildCall(builder, func, &f32, 1, "");
#endif
ref_result = LLVMBuildInsertElement(builder, ref_result, f16, index, "");
}
lp_build_print_value(gallivm, "src = ", src);
lp_build_print_value(gallivm, "llvm = ", result);
lp_build_print_value(gallivm, "util = ", ref_result);
lp_build_printf(gallivm, "\n");
}
return result;
}
/**
* Call intrinsic with arguments adapted to intrinsic vector length.
*
* Split vectors which are too large for the hw, or expand them if they
* are too small, so a caller calling a function which might use intrinsics
* doesn't need to do splitting/expansion on its own.
* This only supports intrinsics where src and dst types match.
*/
LLVMValueRef
lp_build_intrinsic_binary_anylength(struct gallivm_state *gallivm,
const char *name,
struct lp_type src_type,
unsigned intr_size,
LLVMValueRef a,
LLVMValueRef b)
{
unsigned i;
struct lp_type intrin_type = src_type;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef i32undef = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
LLVMValueRef anative, bnative;
unsigned intrin_length = intr_size / src_type.width;
intrin_type.length = intrin_length;
if (intrin_length > src_type.length) {
LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
LLVMValueRef constvec, tmp;
for (i = 0; i < src_type.length; i++) {
elems[i] = lp_build_const_int32(gallivm, i);
}
for (; i < intrin_length; i++) {
elems[i] = i32undef;
}
if (src_type.length == 1) {
LLVMTypeRef elem_type = lp_build_elem_type(gallivm, intrin_type);
a = LLVMBuildBitCast(builder, a, LLVMVectorType(elem_type, 1), "");
b = LLVMBuildBitCast(builder, b, LLVMVectorType(elem_type, 1), "");
}
constvec = LLVMConstVector(elems, intrin_length);
anative = LLVMBuildShuffleVector(builder, a, a, constvec, "");
bnative = LLVMBuildShuffleVector(builder, b, b, constvec, "");
tmp = lp_build_intrinsic_binary(builder, name,
lp_build_vec_type(gallivm, intrin_type),
anative, bnative);
if (src_type.length > 1) {
constvec = LLVMConstVector(elems, src_type.length);
return LLVMBuildShuffleVector(builder, tmp, tmp, constvec, "");
}
else {
return LLVMBuildExtractElement(builder, tmp, elems[0], "");
}
}
else if (intrin_length < src_type.length) {
unsigned num_vec = src_type.length / intrin_length;
LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
/* don't support arbitrary size here as this is so yuck */
if (src_type.length % intrin_length) {
/* FIXME: This is something which should be supported
* but there doesn't seem to be any need for it currently
* so crash and burn.
*/
debug_printf("%s: should handle arbitrary vector size\n",
__FUNCTION__);
assert(0);
return NULL;
}
for (i = 0; i < num_vec; i++) {
anative = lp_build_extract_range(gallivm, a, i*intrin_length,
intrin_length);
bnative = lp_build_extract_range(gallivm, b, i*intrin_length,
intrin_length);
tmp[i] = lp_build_intrinsic_binary(builder, name,
lp_build_vec_type(gallivm, intrin_type),
anative, bnative);
}
return lp_build_concat(gallivm, tmp, intrin_type, num_vec);
}
else {
return lp_build_intrinsic_binary(builder, name,
lp_build_vec_type(gallivm, src_type),
a, b);
}
}
