assert(0);
return LLVMGetUndef(dst_vec_type);
break;
}
if (res) {
res = LLVMBuildBitCast(builder, res, dst_vec_type, "");
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;
}
/**
* Non-interleaved pack and saturate.
*
* Same as lp_build_pack2 but will saturate values so that they fit into the
* destination type.
*/
LLVMValueRef
/**
* 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;
}
