/**
* 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 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].
*/
void
lp_build_fetch_rgba_soa(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j,
LLVMValueRef rgba_out[4])
{
LLVMBuilderRef builder = gallivm->builder;
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))
{
/*
* 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);
packed = lp_build_gather(gallivm,
type.length,
format_desc->block.bits,
type.width,
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->format == PIPE_FORMAT_R11G11B10_FLOAT ||
format_desc->format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
/*
* similar conceptually to above but requiring special
* AoS packed -> SoA float conversion code.
*/
LLVMValueRef packed;
assert(type.floating);
assert(type.width == 32);
packed = lp_build_gather(gallivm, type.length,
format_desc->block.bits,
type.width, base_ptr, offset,
FALSE);
if (format_desc->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;
if (format_desc->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, type.width, base_ptr, offset, FALSE);
packed = LLVMBuildAnd(builder, packed,
lp_build_const_int_vec(gallivm, type, mask), "");
}
else {
assert (format_desc->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT);
packed = lp_build_gather(gallivm, type.length,
32, type.width, 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.
*/
if (util_format_fits_8unorm(format_desc) &&
type.floating && type.width == 32 &&
(type.length == 1 || (type.length % 4 == 0))) {
struct lp_type tmp_type;
LLVMValueRef tmp;
memset(&tmp_type, 0, sizeof tmp_type);
tmp_type.width = 8;
tmp_type.length = type.length * 4;
tmp_type.norm = TRUE;
tmp = lp_build_fetch_rgba_aos(gallivm, format_desc, tmp_type,
base_ptr, offset, i, j);
lp_build_rgba8_to_fi32_soa(gallivm,
type,
tmp,
rgba_out);
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.
*/
{
unsigned k, chan;
struct lp_type tmp_type;
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: scalar unpacking of %s\n",
__FUNCTION__, format_desc->short_name);
}
tmp_type = type;
tmp_type.length = 4;
for (chan = 0; chan < 4; ++chan) {
rgba_out[chan] = lp_build_undef(gallivm, type);
}
/* 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;
LLVMValueRef tmp;
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 */
tmp = lp_build_fetch_rgba_aos(gallivm, format_desc, tmp_type,
base_ptr, offset_elem,
i_elem, j_elem);
/*
* Insert the AoS tmp value channels into the SoA result vectors at
* position = 'index'.
*/
for (chan = 0; chan < 4; ++chan) {
LLVMValueRef chan_val = lp_build_const_int32(gallivm, chan),
tmp_chan = LLVMBuildExtractElement(builder, tmp, chan_val, "");
rgba_out[chan] = LLVMBuildInsertElement(builder, rgba_out[chan],
tmp_chan, index, "");
}
}
}
}
/**
* 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);
}
}