/**
* Converts int16 half-float to float32
* Note this can be performed in 1 instruction if vcvtph2ps exists (f16c/cvt16)
* [llvm.x86.vcvtph2ps / _mm_cvtph_ps]
*
* @param src value to convert
*
*/
LLVMValueRef
lp_build_half_to_float(struct gallivm_state *gallivm,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef src_type = LLVMTypeOf(src);
unsigned src_length = LLVMGetTypeKind(src_type) == LLVMVectorTypeKind ?
LLVMGetVectorSize(src_type) : 1;
struct lp_type f32_type = lp_type_float_vec(32, 32 * src_length);
struct lp_type i32_type = lp_type_int_vec(32, 32 * src_length);
LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
LLVMValueRef h;
if (util_cpu_caps.has_f16c && HAVE_LLVM >= 0x0301 &&
(src_length == 4 || src_length == 8)) {
const char *intrinsic = NULL;
if (src_length == 4) {
src = lp_build_pad_vector(gallivm, src, 8);
intrinsic = "llvm.x86.vcvtph2ps.128";
}
else {
intrinsic = "llvm.x86.vcvtph2ps.256";
}
return lp_build_intrinsic_unary(builder, intrinsic,
lp_build_vec_type(gallivm, f32_type), src);
}
/* Convert int16 vector to int32 vector by zero ext (might generate bad code) */
h = LLVMBuildZExt(builder, src, int_vec_type, "");
return lp_build_smallfloat_to_float(gallivm, f32_type, h, 10, 5, 0, true);
}
/**
* Converts int16 half-float to float32
* Note this can be performed in 1 instruction if vcvtph2ps exists (sse5 i think?)
* [llvm.x86.vcvtph2ps / _mm_cvtph_ps]
*
* @param src_type <vector> type of int16
* @param src value to convert
*
* ref http://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/
*/
LLVMValueRef
lp_build_half_to_float(struct gallivm_state *gallivm,
struct lp_type src_type,
LLVMValueRef src)
{
struct lp_type f32_type = lp_type_float_vec(32, 32 * src_type.length);
struct lp_type i32_type = lp_type_int_vec(32, 32 * src_type.length);
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
LLVMTypeRef float_vec_type = lp_build_vec_type(gallivm, f32_type);
/* Constants */
LLVMValueRef i32_13 = lp_build_const_int_vec(gallivm, i32_type, 13);
LLVMValueRef i32_16 = lp_build_const_int_vec(gallivm, i32_type, 16);
LLVMValueRef i32_mask_nosign = lp_build_const_int_vec(gallivm, i32_type, 0x7fff);
LLVMValueRef i32_was_infnan = lp_build_const_int_vec(gallivm, i32_type, 0x7bff);
LLVMValueRef i32_exp_infnan = lp_build_const_int_vec(gallivm, i32_type, 0xff << 23);
LLVMValueRef f32_magic = LLVMBuildBitCast(builder,
lp_build_const_int_vec(gallivm, i32_type, (254 - 15) << 23),
float_vec_type, "");
/* Convert int16 vector to int32 vector by zero ext */
LLVMValueRef h = LLVMBuildZExt(builder, src, int_vec_type, "");
/* Exponent / mantissa bits */
LLVMValueRef expmant = LLVMBuildAnd(builder, i32_mask_nosign, h, "");
LLVMValueRef shifted = LLVMBuildBitCast(builder, LLVMBuildShl(builder, expmant, i32_13, ""), float_vec_type, "");
/* Exponent adjust */
LLVMValueRef scaled = LLVMBuildBitCast(builder, LLVMBuildFMul(builder, shifted, f32_magic, ""), int_vec_type, "");
/* Make sure Inf/NaN survive */
LLVMValueRef b_wasinfnan = lp_build_compare(gallivm, i32_type, PIPE_FUNC_GREATER, expmant, i32_was_infnan);
LLVMValueRef infnanexp = LLVMBuildAnd(builder, b_wasinfnan, i32_exp_infnan, "");
/* Sign bit */
LLVMValueRef justsign = LLVMBuildXor(builder, h, expmant, "");
LLVMValueRef sign = LLVMBuildShl(builder, justsign, i32_16, "");
/* Combine result */
LLVMValueRef sign_inf = LLVMBuildOr(builder, sign, infnanexp, "");
LLVMValueRef final = LLVMBuildOr(builder, scaled, sign_inf, "");
/* Cast from int32 vector to float32 vector */
return LLVMBuildBitCast(builder, final, float_vec_type, "");
}
/**
* Generate code for performing depth and/or stencil tests.
* We operate on a vector of values (typically a 2x2 quad).
*
* \param depth the depth test state
* \param stencil the front/back stencil state
* \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 stencil_refs the front/back stencil ref values (scalar)
* \param z_src the incoming depth/stencil values (a 2x2 quad)
* \param zs_dst_ptr pointer to depth/stencil values in framebuffer
* \param facing contains float value indicating front/back facing polygon
*/
void
lp_build_depth_stencil_test(LLVMBuilderRef builder,
const struct pipe_depth_state *depth,
const struct pipe_stencil_state stencil[2],
struct lp_type type,
const struct util_format_description *format_desc,
struct lp_build_mask_context *mask,
LLVMValueRef stencil_refs[2],
LLVMValueRef z_src,
LLVMValueRef zs_dst_ptr,
LLVMValueRef face,
LLVMValueRef counter)
{
struct lp_build_context bld;
struct lp_build_context sbld;
struct lp_type s_type;
LLVMValueRef zs_dst, z_dst = NULL;
LLVMValueRef stencil_vals = NULL;
LLVMValueRef z_bitmask = NULL, stencil_shift = NULL;
LLVMValueRef z_pass = NULL, s_pass_mask = NULL;
LLVMValueRef orig_mask = mask->value;
/* Sanity checking */
{
const unsigned z_swizzle = format_desc->swizzle[0];
const unsigned s_swizzle = format_desc->swizzle[1];
assert(z_swizzle != UTIL_FORMAT_SWIZZLE_NONE ||
s_swizzle != UTIL_FORMAT_SWIZZLE_NONE);
assert(depth->enabled || stencil[0].enabled);
assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS);
assert(format_desc->block.width == 1);
assert(format_desc->block.height == 1);
if (stencil[0].enabled) {
assert(format_desc->format == PIPE_FORMAT_Z24_UNORM_S8_USCALED ||
format_desc->format == PIPE_FORMAT_S8_USCALED_Z24_UNORM);
}
assert(z_swizzle < 4);
assert(format_desc->block.bits == type.width);
if (type.floating) {
assert(z_swizzle == 0);
assert(format_desc->channel[z_swizzle].type ==
UTIL_FORMAT_TYPE_FLOAT);
assert(format_desc->channel[z_swizzle].size ==
format_desc->block.bits);
}
else {
assert(format_desc->channel[z_swizzle].type ==
UTIL_FORMAT_TYPE_UNSIGNED);
assert(format_desc->channel[z_swizzle].normalized);
assert(!type.fixed);
assert(!type.sign);
assert(type.norm);
}
}
/* Setup build context for Z vals */
lp_build_context_init(&bld, builder, type);
/* Setup build context for stencil vals */
s_type = lp_type_int_vec(type.width);
lp_build_context_init(&sbld, builder, s_type);
/* Load current z/stencil value from z/stencil buffer */
zs_dst = LLVMBuildLoad(builder, zs_dst_ptr, "");
lp_build_name(zs_dst, "zsbufval");
/* Compute and apply the Z/stencil bitmasks and shifts.
*/
{
unsigned z_shift, z_mask;
unsigned s_shift, s_mask;
if (get_z_shift_and_mask(format_desc, &z_shift, &z_mask)) {
if (z_shift) {
LLVMValueRef shift = lp_build_const_int_vec(type, z_shift);
z_src = LLVMBuildLShr(builder, z_src, shift, "");
}
if (z_mask != 0xffffffff) {
LLVMValueRef mask = lp_build_const_int_vec(type, z_mask);
z_src = LLVMBuildAnd(builder, z_src, mask, "");
z_dst = LLVMBuildAnd(builder, zs_dst, mask, "");
z_bitmask = mask; /* used below */
}
else {
z_dst = zs_dst;
}
lp_build_name(z_dst, "zsbuf.z");
}
if (get_s_shift_and_mask(format_desc, &s_shift, &s_mask)) {
if (s_shift) {
LLVMValueRef shift = lp_build_const_int_vec(type, s_shift);
stencil_vals = LLVMBuildLShr(builder, zs_dst, shift, "");
stencil_shift = shift; /* used below */
}
else {
stencil_vals = zs_dst;
}
if (s_mask != 0xffffffff) {
LLVMValueRef mask = lp_build_const_int_vec(type, s_mask);
stencil_vals = LLVMBuildAnd(builder, stencil_vals, mask, "");
}
lp_build_name(stencil_vals, "stencil");
}
}
if (stencil[0].enabled) {
/* convert scalar stencil refs into vectors */
stencil_refs[0] = lp_build_broadcast_scalar(&bld, stencil_refs[0]);
stencil_refs[1] = lp_build_broadcast_scalar(&bld, stencil_refs[1]);
s_pass_mask = lp_build_stencil_test(&sbld, stencil,
stencil_refs, stencil_vals, face);
/* apply stencil-fail operator */
{
LLVMValueRef s_fail_mask = lp_build_andc(&bld, orig_mask, s_pass_mask);
stencil_vals = lp_build_stencil_op(&sbld, stencil, S_FAIL_OP,
stencil_refs, stencil_vals,
s_fail_mask, face);
}
}
if (depth->enabled) {
/* compare src Z to dst Z, returning 'pass' mask */
z_pass = lp_build_cmp(&bld, depth->func, z_src, z_dst);
if (!stencil[0].enabled) {
/* We can potentially skip all remaining operations here, but only
* if stencil is disabled because we still need to update the stencil
* buffer values. Don't need to update Z buffer values.
*/
lp_build_mask_update(mask, z_pass);
}
if (depth->writemask) {
LLVMValueRef zselectmask = mask->value;
/* mask off bits that failed Z test */
zselectmask = LLVMBuildAnd(builder, zselectmask, z_pass, "");
/* mask off bits that failed stencil test */
if (s_pass_mask) {
zselectmask = LLVMBuildAnd(builder, zselectmask, s_pass_mask, "");
}
/* if combined Z/stencil format, mask off the stencil bits */
if (z_bitmask) {
zselectmask = LLVMBuildAnd(builder, zselectmask, z_bitmask, "");
}
/* Mix the old and new Z buffer values.
* z_dst[i] = (zselectmask[i] & z_src[i]) | (~zselectmask[i] & z_dst[i])
*/
z_dst = lp_build_select_bitwise(&bld, zselectmask, z_src, z_dst);
}
if (stencil[0].enabled) {
/* update stencil buffer values according to z pass/fail result */
LLVMValueRef z_fail_mask, z_pass_mask;
/* apply Z-fail operator */
z_fail_mask = lp_build_andc(&bld, orig_mask, z_pass);
stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_FAIL_OP,
stencil_refs, stencil_vals,
z_fail_mask, face);
/* apply Z-pass operator */
z_pass_mask = LLVMBuildAnd(bld.builder, orig_mask, z_pass, "");
stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_PASS_OP,
stencil_refs, stencil_vals,
z_pass_mask, face);
}
}
else {
/* No depth test: apply Z-pass operator to stencil buffer values which
* passed the stencil test.
*/
s_pass_mask = LLVMBuildAnd(bld.builder, orig_mask, s_pass_mask, "");
stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_PASS_OP,
stencil_refs, stencil_vals,
s_pass_mask, face);
}
/* The Z bits are already in the right place but we may need to shift the
* stencil bits before ORing Z with Stencil to make the final pixel value.
*/
if (stencil_vals && stencil_shift)
stencil_vals = LLVMBuildShl(bld.builder, stencil_vals,
stencil_shift, "");
/* Finally, merge/store the z/stencil values */
if ((depth->enabled && depth->writemask) ||
(stencil[0].enabled && stencil[0].writemask)) {
if (z_dst && stencil_vals)
zs_dst = LLVMBuildOr(bld.builder, z_dst, stencil_vals, "");
else if (z_dst)
zs_dst = z_dst;
else
zs_dst = stencil_vals;
LLVMBuildStore(builder, zs_dst, zs_dst_ptr);
}
if (s_pass_mask)
lp_build_mask_update(mask, s_pass_mask);
if (depth->enabled && stencil[0].enabled)
lp_build_mask_update(mask, z_pass);
if (counter)
lp_build_occlusion_count(builder, type, mask->value, counter);
}
/**
* Gather elements from scatter positions in memory into a single vector.
* Use for fetching texels from a texture.
* For SSE, typical values are length=4, src_width=32, dst_width=32.
*
* When src_width < dst_width, the return value can be justified in
* one of two ways:
* "integer justification" is used when the caller treats the destination
* as a packed integer bitmask, as described by the channels' "shift" and
* "width" fields;
* "vector justification" is used when the caller casts the destination
* to a vector and needs channel X to be in vector element 0.
*
* @param length length of the offsets
* @param src_width src element width in bits
* @param dst_type result element type (src will be expanded to fit,
* but truncation is not allowed)
* (this may be a vector, must be pot sized)
* @param aligned whether the data is guaranteed to be aligned (to src_width)
* @param base_ptr base pointer, needs to be a i8 pointer type.
* @param offsets vector with offsets
* @param vector_justify select vector rather than integer justification
*/
LLVMValueRef
lp_build_gather(struct gallivm_state *gallivm,
unsigned length,
unsigned src_width,
struct lp_type dst_type,
boolean aligned,
LLVMValueRef base_ptr,
LLVMValueRef offsets,
boolean vector_justify)
{
LLVMValueRef res;
boolean need_expansion = src_width < dst_type.width * dst_type.length;
boolean vec_fetch;
struct lp_type fetch_type, fetch_dst_type;
LLVMTypeRef src_type;
assert(src_width <= dst_type.width * dst_type.length);
/*
* This is quite a mess...
* Figure out if the fetch should be done as:
* a) scalar or vector
* b) float or int
*
* As an example, for a 96bit fetch expanded into 4x32bit, it is better
* to use (3x32bit) vector type (then pad the vector). Otherwise, the
* zext will cause extra instructions.
* However, the same isn't true for 3x16bit (the codegen for that is
* completely worthless on x86 simd, and for 3x8bit is is way worse
* still, don't try that... (To get really good code out of llvm for
* these cases, the only way is to decompose the fetches manually
* into 1x32bit/1x16bit, or 1x16/1x8bit respectively, although the latter
* case requires sse41, otherwise simple scalar zext is way better.
* But probably not important enough, so don't bother.)
* Also, we try to honor the floating bit of destination (but isn't
* possible if caller asks for instance for 2x32bit dst_type with
* 48bit fetch - the idea would be to use 3x16bit fetch, pad and
* cast to 2x32f type, so the fetch is always int and on top of that
* we avoid the vec pad and use scalar zext due the above mentioned
* issue).
* Note this is optimized for x86 sse2 and up backend. Could be tweaked
* for other archs if necessary...
*/
if (((src_width % 32) == 0) && ((src_width % dst_type.width) == 0) &&
(dst_type.length > 1)) {
/* use vector fetch (if dst_type is vector) */
vec_fetch = TRUE;
if (dst_type.floating) {
fetch_type = lp_type_float_vec(dst_type.width, src_width);
} else {
fetch_type = lp_type_int_vec(dst_type.width, src_width);
}
/* intentionally not using lp_build_vec_type here */
src_type = LLVMVectorType(lp_build_elem_type(gallivm, fetch_type),
fetch_type.length);
fetch_dst_type = fetch_type;
fetch_dst_type.length = dst_type.length;
} else {
/* use scalar fetch */
vec_fetch = FALSE;
if (dst_type.floating && ((src_width == 32) || (src_width == 64))) {
fetch_type = lp_type_float(src_width);
} else {
fetch_type = lp_type_int(src_width);
}
src_type = lp_build_vec_type(gallivm, fetch_type);
fetch_dst_type = fetch_type;
fetch_dst_type.width = dst_type.width * dst_type.length;
}
if (length == 1) {
/* Scalar */
res = lp_build_gather_elem_vec(gallivm, length,
src_width, src_type, fetch_dst_type,
aligned, base_ptr, offsets, 0,
vector_justify);
return LLVMBuildBitCast(gallivm->builder, res,
lp_build_vec_type(gallivm, dst_type), "");
/*
* Excluding expansion from these paths because if you need it for
* 32bit/64bit fetches you're doing it wrong (this is gather, not
* conversion) and it would be awkward for floats.
*/
} else if (util_cpu_caps.has_avx2 && !need_expansion &&
src_width == 32 && (length == 4 || length == 8)) {
return lp_build_gather_avx2(gallivm, length, src_width, dst_type,
base_ptr, offsets);
/*
* This looks bad on paper wrt throughtput/latency on Haswell.
* Even on Broadwell it doesn't look stellar.
* Albeit no measurements were done (but tested to work).
* Should definitely enable on Skylake.
* (In general, should be more of a win if the fetch is 256bit wide -
* this is true for the 32bit case above too.)
*/
} else if (0 && util_cpu_caps.has_avx2 && !need_expansion &&
src_width == 64 && (length == 2 || length == 4)) {
return lp_build_gather_avx2(gallivm, length, src_width, dst_type,
base_ptr, offsets);
} else {
/* Vector */
LLVMValueRef elems[LP_MAX_VECTOR_WIDTH / 8];
unsigned i;
boolean vec_zext = FALSE;
struct lp_type res_type, gather_res_type;
LLVMTypeRef res_t, gather_res_t;
res_type = fetch_dst_type;
res_type.length *= length;
gather_res_type = res_type;
if (src_width == 16 && dst_type.width == 32 && dst_type.length == 1) {
/*
* Note that llvm is never able to optimize zext/insert combos
* directly (i.e. zero the simd reg, then place the elements into
* the appropriate place directly). (I think this has to do with
* scalar/vector transition.) And scalar 16->32bit zext simd loads
* aren't possible (instead loading to scalar reg first).
* No idea about other archs...
* We could do this manually, but instead we just use a vector
* zext, which is simple enough (and, in fact, llvm might optimize
* this away).
* (We're not trying that with other bit widths as that might not be
* easier, in particular with 8 bit values at least with only sse2.)
*/
assert(vec_fetch == FALSE);
gather_res_type.width /= 2;
fetch_dst_type = fetch_type;
src_type = lp_build_vec_type(gallivm, fetch_type);
vec_zext = TRUE;
}
res_t = lp_build_vec_type(gallivm, res_type);
gather_res_t = lp_build_vec_type(gallivm, gather_res_type);
res = LLVMGetUndef(gather_res_t);
for (i = 0; i < length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
elems[i] = lp_build_gather_elem_vec(gallivm, length,
src_width, src_type, fetch_dst_type,
aligned, base_ptr, offsets, i,
vector_justify);
if (!vec_fetch) {
res = LLVMBuildInsertElement(gallivm->builder, res, elems[i], index, "");
}
}
if (vec_zext) {
res = LLVMBuildZExt(gallivm->builder, res, res_t, "");
if (vector_justify) {
#ifdef PIPE_ARCH_BIG_ENDIAN
unsigned sv = dst_type.width - src_width;
res = LLVMBuildShl(gallivm->builder, res,
lp_build_const_int_vec(gallivm, res_type, sv), "");
#endif
}
}
if (vec_fetch) {
/*
* Do bitcast now otherwise llvm might get some funny ideas wrt
* float/int types...
*/
for (i = 0; i < length; i++) {
elems[i] = LLVMBuildBitCast(gallivm->builder, elems[i],
lp_build_vec_type(gallivm, dst_type), "");
}
res = lp_build_concat(gallivm, elems, dst_type, length);
} else {
struct lp_type really_final_type = dst_type;
assert(res_type.length * res_type.width ==
dst_type.length * dst_type.width * length);
really_final_type.length *= length;
res = LLVMBuildBitCast(gallivm->builder, res,
lp_build_vec_type(gallivm, really_final_type), "");
}
}
return res;
}
/**
* 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;
}
/**
* 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;
}
/**
* Sample a single texture image with nearest sampling.
* If sampling a cube texture, r = cube face in [0,5].
* Return filtered color as two vectors of 16-bit fixed point values.
*/
static void
lp_build_sample_image_nearest(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;
LLVMValueRef width_vec, height_vec, depth_vec;
LLVMValueRef s_ipart, t_ipart = NULL, r_ipart = NULL;
LLVMValueRef x_stride;
LLVMValueRef x_offset, offset;
LLVMValueRef x_subcoord, y_subcoord, z_subcoord;
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, "");
/* 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, "");
/* get pixel, row, image strides */
x_stride = lp_build_const_vec(bld->gallivm,
bld->int_coord_bld.type,
bld->format_desc->block.bits/8);
/* Do texcoord wrapping, compute texel offset */
lp_build_sample_wrap_nearest_int(bld,
bld->format_desc->block.width,
s_ipart, width_vec, x_stride,
bld->static_state->pot_width,
bld->static_state->wrap_s,
&x_offset, &x_subcoord);
offset = x_offset;
if (dims >= 2) {
LLVMValueRef y_offset;
lp_build_sample_wrap_nearest_int(bld,
bld->format_desc->block.height,
t_ipart, height_vec, row_stride_vec,
bld->static_state->pot_height,
bld->static_state->wrap_t,
&y_offset, &y_subcoord);
offset = lp_build_add(&bld->int_coord_bld, offset, y_offset);
if (dims >= 3) {
LLVMValueRef z_offset;
lp_build_sample_wrap_nearest_int(bld,
1, /* block length (depth) */
r_ipart, depth_vec, img_stride_vec,
bld->static_state->pot_height,
bld->static_state->wrap_r,
&z_offset, &z_subcoord);
offset = lp_build_add(&bld->int_coord_bld, offset, z_offset);
}
else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
LLVMValueRef z_offset;
/* The r coord is the cube face in [0,5] */
z_offset = lp_build_mul(&bld->int_coord_bld, r, img_stride_vec);
offset = lp_build_add(&bld->int_coord_bld, offset, z_offset);
}
}
/*
* 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.
*/
{
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);
rgba8 = LLVMBuildBitCast(builder, rgba8, u8n_vec_type, "");
}
else {
rgba8 = lp_build_fetch_rgba_aos(bld->gallivm,
bld->format_desc,
u8n.type,
data_ptr, offset,
x_subcoord,
y_subcoord);
}
/* Expand one 4*rgba8 to two 2*rgba16 */
lp_build_unpack2(bld->gallivm, u8n.type, h16.type,
rgba8,
colors_lo, colors_hi);
}
}
/**
* Fetch a texels from a texture, returning them in SoA layout.
*
* \param type the desired return type for 'rgba'. The vector length
* is the number of texels to fetch
* \param aligned if the offset is guaranteed to be aligned to element width
*
* \param base_ptr points to the base of the texture mip tree.
* \param offset offset to start of the texture image block. For non-
* compressed formats, this simply is an offset to the texel.
* For compressed formats, it is an offset to the start of the
* compressed data block.
*
* \param i, j the sub-block pixel coordinates. For non-compressed formats
* these will always be (0,0). For compressed formats, i will
* be in [0, block_width-1] and j will be in [0, block_height-1].
* \param cache optional value pointing to a lp_build_format_cache structure
*/
void
lp_build_fetch_rgba_soa(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
boolean aligned,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j,
LLVMValueRef cache,
LLVMValueRef rgba_out[4])
{
LLVMBuilderRef builder = gallivm->builder;
enum pipe_format format = format_desc->format;
struct lp_type fetch_type;
if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
format_desc->block.bits <= type.width &&
(format_desc->channel[0].type != UTIL_FORMAT_TYPE_FLOAT ||
format_desc->channel[0].size == 32 ||
format_desc->channel[0].size == 16))
{
/*
* The packed pixel fits into an element of the destination format. Put
* the packed pixels into a vector and extract each component for all
* vector elements in parallel.
*/
LLVMValueRef packed;
/*
* gather the texels from the texture
* Ex: packed = {XYZW, XYZW, XYZW, XYZW}
*/
assert(format_desc->block.bits <= type.width);
fetch_type = lp_type_uint(type.width);
packed = lp_build_gather(gallivm,
type.length,
format_desc->block.bits,
fetch_type,
aligned,
base_ptr, offset, FALSE);
/*
* convert texels to float rgba
*/
lp_build_unpack_rgba_soa(gallivm,
format_desc,
type,
packed, rgba_out);
return;
}
if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
format_desc->block.bits > type.width &&
((format_desc->block.bits <= type.width * type.length &&
format_desc->channel[0].size <= type.width) ||
(format_desc->channel[0].size == 64 &&
format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT &&
type.floating)))
{
/*
* Similar to above, but the packed pixel is larger than what fits
* into an element of the destination format. The packed pixels will be
* shuffled into SoA vectors appropriately, and then the extraction will
* be done in parallel as much as possible.
* Good for 16xn (n > 2) and 32xn (n > 1) formats, care is taken so
* the gathered vectors can be shuffled easily (even with avx).
* 64xn float -> 32xn float is handled too but it's a bit special as
* it does the conversion pre-shuffle.
*/
LLVMValueRef packed[4], dst[4], output[4], shuffles[LP_MAX_VECTOR_WIDTH/32];
struct lp_type fetch_type, gather_type = type;
unsigned num_gather, fetch_width, i, j;
struct lp_build_context bld;
boolean fp64 = format_desc->channel[0].size == 64;
lp_build_context_init(&bld, gallivm, type);
assert(type.width == 32);
assert(format_desc->block.bits > type.width);
/*
* First, figure out fetch order.
*/
fetch_width = util_next_power_of_two(format_desc->block.bits);
num_gather = fetch_width / type.width;
/*
* fp64 are treated like fp32 except we fetch twice wide values
* (as we shuffle after trunc). The shuffles for that work out
* mostly fine (slightly suboptimal for 4-wide, perfect for AVX)
* albeit we miss the potential opportunity for hw gather (as it
* only handles native size).
*/
num_gather = fetch_width / type.width;
gather_type.width *= num_gather;
if (fp64) {
num_gather /= 2;
}
gather_type.length /= num_gather;
for (i = 0; i < num_gather; i++) {
LLVMValueRef offsetr, shuf_vec;
if(num_gather == 4) {
for (j = 0; j < gather_type.length; j++) {
unsigned idx = i + 4*j;
shuffles[j] = lp_build_const_int32(gallivm, idx);
}
shuf_vec = LLVMConstVector(shuffles, gather_type.length);
offsetr = LLVMBuildShuffleVector(builder, offset, offset, shuf_vec, "");
}
else if (num_gather == 2) {
assert(num_gather == 2);
for (j = 0; j < gather_type.length; j++) {
unsigned idx = i*2 + (j%2) + (j/2)*4;
shuffles[j] = lp_build_const_int32(gallivm, idx);
}
shuf_vec = LLVMConstVector(shuffles, gather_type.length);
offsetr = LLVMBuildShuffleVector(builder, offset, offset, shuf_vec, "");
}
else {
assert(num_gather == 1);
offsetr = offset;
}
if (gather_type.length == 1) {
LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
offsetr = LLVMBuildExtractElement(builder, offsetr, zero, "");
}
/*
* Determine whether to use float or int loads. This is mostly
* to outsmart the (stupid) llvm int/float shuffle logic, we
* don't really care much if the data is floats or ints...
* But llvm will refuse to use single float shuffle with int data
* and instead use 3 int shuffles instead, the code looks atrocious.
* (Note bitcasts often won't help, as llvm is too smart to be
* fooled by that.)
* Nobody cares about simd float<->int domain transition penalties,
* which usually don't even exist for shuffles anyway.
* With 4x32bit (and 3x32bit) fetch, we use float vec (the data is
* going into transpose, which is unpacks, so doesn't really matter
* much).
* With 2x32bit or 4x16bit fetch, we use float vec, since those
* go into the weird channel separation shuffle. With floats,
* this is (with 128bit vectors):
* - 2 movq, 2 movhpd, 2 shufps
* With ints it would be:
* - 4 movq, 2 punpcklqdq, 4 pshufd, 2 blendw
* I've seen texture functions increase in code size by 15% just due
* to that (there's lots of such fetches in them...)
* (We could chose a different gather order to improve this somewhat
* for the int path, but it would basically just drop the blends,
* so the float path with this order really is optimal.)
* Albeit it is tricky sometimes llvm doesn't ignore the float->int
* casts so must avoid them until we're done with the float shuffle...
* 3x16bit formats (the same is also true for 3x8) are pretty bad but
* there's nothing we can do about them (we could overallocate by
* those couple bytes and use unaligned but pot sized load).
* Note that this is very much x86 specific. I don't know if this
* affect other archs at all.
*/
if (num_gather > 1) {
/*
* We always want some float type here (with x86)
* due to shuffles being float ones afterwards (albeit for
* the num_gather == 4 case int should work fine too
* (unless there's some problems with avx but not avx2).
*/
if (format_desc->channel[0].size == 64) {
fetch_type = lp_type_float_vec(64, gather_type.width);
} else {
fetch_type = lp_type_int_vec(32, gather_type.width);
}
}
else {
/* type doesn't matter much */
if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT &&
(format_desc->channel[0].size == 32 ||
format_desc->channel[0].size == 64)) {
fetch_type = lp_type_float(gather_type.width);
} else {
fetch_type = lp_type_uint(gather_type.width);
}
}
/* Now finally gather the values */
packed[i] = lp_build_gather(gallivm, gather_type.length,
format_desc->block.bits,
fetch_type, aligned,
base_ptr, offsetr, FALSE);
if (fp64) {
struct lp_type conv_type = type;
conv_type.width *= 2;
packed[i] = LLVMBuildBitCast(builder, packed[i],
lp_build_vec_type(gallivm, conv_type), "");
packed[i] = LLVMBuildFPTrunc(builder, packed[i], bld.vec_type, "");
}
}
/* shuffle the gathered values to SoA */
if (num_gather == 2) {
for (i = 0; i < num_gather; i++) {
for (j = 0; j < type.length; j++) {
unsigned idx = (j%2)*2 + (j/4)*4 + i;
if ((j/2)%2)
idx += type.length;
shuffles[j] = lp_build_const_int32(gallivm, idx);
}
dst[i] = LLVMBuildShuffleVector(builder, packed[0], packed[1],
LLVMConstVector(shuffles, type.length), "");
}
}
else if (num_gather == 4) {
lp_build_transpose_aos(gallivm, lp_int_type(type), packed, dst);
}
else {
assert(num_gather == 1);
dst[0] = packed[0];
}
/*
* And finally unpack exactly as above, except that
* chan shift is adjusted and the right vector selected.
*/
if (!fp64) {
for (i = 0; i < num_gather; i++) {
dst[i] = LLVMBuildBitCast(builder, dst[i], bld.int_vec_type, "");
}
for (i = 0; i < format_desc->nr_channels; i++) {
struct util_format_channel_description chan_desc = format_desc->channel[i];
unsigned blockbits = type.width;
unsigned vec_nr = chan_desc.shift / type.width;
chan_desc.shift %= type.width;
output[i] = lp_build_extract_soa_chan(&bld,
blockbits,
FALSE,
chan_desc,
dst[vec_nr]);
}
}
else {
for (i = 0; i < format_desc->nr_channels; i++) {
output[i] = dst[i];
}
}
lp_build_format_swizzle_soa(format_desc, &bld, output, rgba_out);
return;
}
if (format == PIPE_FORMAT_R11G11B10_FLOAT ||
format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
/*
* similar conceptually to above but requiring special
* AoS packed -> SoA float conversion code.
*/
LLVMValueRef packed;
struct lp_type fetch_type = lp_type_uint(type.width);
assert(type.floating);
assert(type.width == 32);
packed = lp_build_gather(gallivm, type.length,
format_desc->block.bits,
fetch_type, aligned,
base_ptr, offset, FALSE);
if (format == PIPE_FORMAT_R11G11B10_FLOAT) {
lp_build_r11g11b10_to_float(gallivm, packed, rgba_out);
}
else {
lp_build_rgb9e5_to_float(gallivm, packed, rgba_out);
}
return;
}
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS &&
format_desc->block.bits == 64) {
/*
* special case the format is 64 bits but we only require
* 32bit (or 8bit) from each block.
*/
LLVMValueRef packed;
struct lp_type fetch_type = lp_type_uint(type.width);
if (format == PIPE_FORMAT_X32_S8X24_UINT) {
/*
* for stencil simply fix up offsets - could in fact change
* base_ptr instead even outside the shader.
*/
unsigned mask = (1 << 8) - 1;
LLVMValueRef s_offset = lp_build_const_int_vec(gallivm, type, 4);
offset = LLVMBuildAdd(builder, offset, s_offset, "");
packed = lp_build_gather(gallivm, type.length, 32, fetch_type,
aligned, base_ptr, offset, FALSE);
packed = LLVMBuildAnd(builder, packed,
lp_build_const_int_vec(gallivm, type, mask), "");
}
else {
assert (format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT);
packed = lp_build_gather(gallivm, type.length, 32, fetch_type,
aligned, base_ptr, offset, TRUE);
packed = LLVMBuildBitCast(builder, packed,
lp_build_vec_type(gallivm, type), "");
}
/* for consistency with lp_build_unpack_rgba_soa() return sss1 or zzz1 */
rgba_out[0] = rgba_out[1] = rgba_out[2] = packed;
rgba_out[3] = lp_build_const_vec(gallivm, type, 1.0f);
return;
}
/*
* Try calling lp_build_fetch_rgba_aos for all pixels.
* Should only really hit subsampled, compressed
* (for s3tc srgb too, for rgtc the unorm ones only) by now.
* (This is invalid for plain 8unorm formats because we're lazy with
* the swizzle since some results would arrive swizzled, some not.)
*/
if ((format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN) &&
(util_format_fits_8unorm(format_desc) ||
format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) &&
type.floating && type.width == 32 &&
(type.length == 1 || (type.length % 4 == 0))) {
struct lp_type tmp_type;
struct lp_build_context bld;
LLVMValueRef packed, rgba[4];
const struct util_format_description *flinear_desc;
const struct util_format_description *frgba8_desc;
unsigned chan;
lp_build_context_init(&bld, gallivm, type);
/*
* Make sure the conversion in aos really only does convert to rgba8
* and not anything more (so use linear format, adjust type).
*/
flinear_desc = util_format_description(util_format_linear(format));
memset(&tmp_type, 0, sizeof tmp_type);
tmp_type.width = 8;
tmp_type.length = type.length * 4;
tmp_type.norm = TRUE;
packed = lp_build_fetch_rgba_aos(gallivm, flinear_desc, tmp_type,
aligned, base_ptr, offset, i, j, cache);
packed = LLVMBuildBitCast(builder, packed, bld.int_vec_type, "");
/*
* The values are now packed so they match ordinary (srgb) RGBA8 format,
* hence need to use matching format for unpack.
*/
frgba8_desc = util_format_description(PIPE_FORMAT_R8G8B8A8_UNORM);
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
assert(format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC);
frgba8_desc = util_format_description(PIPE_FORMAT_R8G8B8A8_SRGB);
}
lp_build_unpack_rgba_soa(gallivm,
frgba8_desc,
type,
packed, rgba);
/*
* We converted 4 channels. Make sure llvm can drop unneeded ones
* (luckily the rgba order is fixed, only LA needs special case).
*/
for (chan = 0; chan < 4; chan++) {
enum pipe_swizzle swizzle = format_desc->swizzle[chan];
if (chan == 3 && util_format_is_luminance_alpha(format)) {
swizzle = PIPE_SWIZZLE_W;
}
rgba_out[chan] = lp_build_swizzle_soa_channel(&bld, rgba, swizzle);
}
return;
}
/*
* Fallback to calling lp_build_fetch_rgba_aos for each pixel.
*
* This is not the most efficient way of fetching pixels, as we
* miss some opportunities to do vectorization, but this is
* convenient for formats or scenarios for which there was no
* opportunity or incentive to optimize.
*
* We do NOT want to end up here, this typically is quite terrible,
* in particular if the formats have less than 4 channels.
*
* Right now, this should only be hit for:
* - RGTC snorm formats
* (those miss fast fetch functions hence they are terrible anyway)
*/
{
unsigned k;
struct lp_type tmp_type;
LLVMValueRef aos_fetch[LP_MAX_VECTOR_WIDTH / 32];
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: AoS fetch fallback for %s\n",
__FUNCTION__, format_desc->short_name);
}
tmp_type = type;
tmp_type.length = 4;
/*
* Note that vector transpose can be worse compared to insert/extract
* for aos->soa conversion (for formats with 1 or 2 channels). However,
* we should try to avoid getting here for just about all formats, so
* don't bother.
*/
/* loop over number of pixels */
for(k = 0; k < type.length; ++k) {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
LLVMValueRef offset_elem;
LLVMValueRef i_elem, j_elem;
offset_elem = LLVMBuildExtractElement(builder, offset,
index, "");
i_elem = LLVMBuildExtractElement(builder, i, index, "");
j_elem = LLVMBuildExtractElement(builder, j, index, "");
/* Get a single float[4]={R,G,B,A} pixel */
aos_fetch[k] = lp_build_fetch_rgba_aos(gallivm, format_desc, tmp_type,
aligned, base_ptr, offset_elem,
i_elem, j_elem, cache);
}
convert_to_soa(gallivm, aos_fetch, rgba_out, type);
}
}
