/**
* Gather one element from scatter positions in memory.
* Nearly the same as above, however the individual elements
* may be vectors themselves, and fetches may be float type.
* Can also do pad vector instead of ZExt.
*
* @sa lp_build_gather()
*/
static LLVMValueRef
lp_build_gather_elem_vec(struct gallivm_state *gallivm,
unsigned length,
unsigned src_width,
LLVMTypeRef src_type,
struct lp_type dst_type,
boolean aligned,
LLVMValueRef base_ptr,
LLVMValueRef offsets,
unsigned i,
boolean vector_justify)
{
LLVMValueRef ptr, res;
LLVMTypeRef src_ptr_type = LLVMPointerType(src_type, 0);
assert(LLVMTypeOf(base_ptr) == LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0));
ptr = lp_build_gather_elem_ptr(gallivm, length, base_ptr, offsets, i);
ptr = LLVMBuildBitCast(gallivm->builder, ptr, src_ptr_type, "");
res = LLVMBuildLoad(gallivm->builder, ptr, "");
/* XXX
* On some archs we probably really want to avoid having to deal
* with alignments lower than 4 bytes (if fetch size is a power of
* two >= 32). On x86 it doesn't matter, however.
* We should be able to guarantee full alignment for any kind of texture
* fetch (except ARB_texture_buffer_range, oops), but not vertex fetch
* (there's PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY and friends
* but I don't think that's quite what we wanted).
* For ARB_texture_buffer_range, PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT
* looks like a good fit, but it seems this cap bit (and OpenGL) aren't
* enforcing what we want (which is what d3d10 does, the offset needs to
* be aligned to element size, but GL has bytes regardless of element
* size which would only leave us with minimum alignment restriction of 16
* which doesn't make much sense if the type isn't 4x32bit). Due to
* translation of offsets to first_elem in sampler_views it actually seems
* gallium could not do anything else except 16 no matter what...
*/
if (!aligned) {
LLVMSetAlignment(res, 1);
} else if (!util_is_power_of_two(src_width)) {
/*
* Full alignment is impossible, assume the caller really meant
* the individual elements were aligned (e.g. 3x32bit format).
* And yes the generated code may otherwise crash, llvm will
* really assume 128bit alignment with a 96bit fetch (I suppose
* that makes sense as it can just assume the upper 32bit to be
* whatever).
* Maybe the caller should be able to explicitly set this, but
* this should cover all the 3-channel formats.
*/
if (((src_width / 24) * 24 == src_width) &&
util_is_power_of_two(src_width / 24)) {
LLVMSetAlignment(res, src_width / 24);
} else {
LLVMSetAlignment(res, 1);
}
}
assert(src_width <= dst_type.width * dst_type.length);
if (src_width < dst_type.width * dst_type.length) {
if (dst_type.length > 1) {
res = lp_build_pad_vector(gallivm, res, dst_type.length);
/*
* vector_justify hopefully a non-issue since we only deal
* with src_width >= 32 here?
*/
} else {
LLVMTypeRef dst_elem_type = lp_build_vec_type(gallivm, dst_type);
/*
* Only valid if src_ptr_type is int type...
*/
res = LLVMBuildZExt(gallivm->builder, res, dst_elem_type, "");
#ifdef PIPE_ARCH_BIG_ENDIAN
if (vector_justify) {
res = LLVMBuildShl(gallivm->builder, res,
LLVMConstInt(dst_elem_type,
dst_type.width - src_width, 0), "");
}
if (src_width == 48) {
/* Load 3x16 bit vector.
* The sequence of loads on big-endian hardware proceeds as follows.
* 16-bit fields are denoted by X, Y, Z, and 0. In memory, the sequence
* of three fields appears in the order X, Y, Z.
*
* Load 32-bit word: 0.0.X.Y
* Load 16-bit halfword: 0.0.0.Z
* Rotate left: 0.X.Y.0
* Bitwise OR: 0.X.Y.Z
*
* The order in which we need the fields in the result is 0.Z.Y.X,
* the same as on little-endian; permute 16-bit fields accordingly
* within 64-bit register:
*/
LLVMValueRef shuffles[4] = {
lp_build_const_int32(gallivm, 2),
lp_build_const_int32(gallivm, 1),
lp_build_const_int32(gallivm, 0),
lp_build_const_int32(gallivm, 3),
};
res = LLVMBuildBitCast(gallivm->builder, res,
lp_build_vec_type(gallivm, lp_type_uint_vec(16, 4*16)), "");
res = LLVMBuildShuffleVector(gallivm->builder, res, res, LLVMConstVector(shuffles, 4), "");
res = LLVMBuildBitCast(gallivm->builder, res, dst_elem_type, "");
}
#endif
}
}
return res;
}
/**
* Swizzle one channel into other channels.
*/
LLVMValueRef
lp_build_swizzle_scalar_aos(struct lp_build_context *bld,
LLVMValueRef a,
unsigned channel,
unsigned num_channels)
{
LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
const unsigned n = type.length;
unsigned i, j;
if(a == bld->undef || a == bld->zero || a == bld->one || num_channels == 1)
return a;
assert(num_channels == 2 || num_channels == 4);
/* XXX: SSE3 has PSHUFB which should be better than bitmasks, but forcing
* using shuffles here actually causes worst results. More investigation is
* needed. */
if (LLVMIsConstant(a) ||
type.width >= 16) {
/*
* Shuffle.
*/
LLVMTypeRef elem_type = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
for(j = 0; j < n; j += num_channels)
for(i = 0; i < num_channels; ++i)
shuffles[j + i] = LLVMConstInt(elem_type, j + channel, 0);
return LLVMBuildShuffleVector(builder, a, bld->undef, LLVMConstVector(shuffles, n), "");
}
else if (num_channels == 2) {
/*
* Bit mask and shifts
*
* XY XY .... XY <= input
* 0Y 0Y .... 0Y
* YY YY .... YY
* YY YY .... YY <= output
*/
struct lp_type type2;
LLVMValueRef tmp = NULL;
int shift;
a = LLVMBuildAnd(builder, a,
lp_build_const_mask_aos(bld->gallivm,
type, 1 << channel, num_channels), "");
type2 = type;
type2.floating = FALSE;
type2.width *= 2;
type2.length /= 2;
a = LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, type2), "");
/*
* Vector element 0 is always channel X.
*
* 76 54 32 10 (array numbering)
* Little endian reg in: YX YX YX YX
* Little endian reg out: YY YY YY YY if shift right (shift == -1)
* XX XX XX XX if shift left (shift == 1)
*
* 01 23 45 67 (array numbering)
* Big endian reg in: XY XY XY XY
* Big endian reg out: YY YY YY YY if shift left (shift == 1)
* XX XX XX XX if shift right (shift == -1)
*
*/
#ifdef PIPE_ARCH_LITTLE_ENDIAN
shift = channel == 0 ? 1 : -1;
#else
shift = channel == 0 ? -1 : 1;
#endif
if (shift > 0) {
tmp = LLVMBuildShl(builder, a, lp_build_const_int_vec(bld->gallivm, type2, shift * type.width), "");
} else if (shift < 0) {
tmp = LLVMBuildLShr(builder, a, lp_build_const_int_vec(bld->gallivm, type2, -shift * type.width), "");
}
assert(tmp);
if (tmp) {
a = LLVMBuildOr(builder, a, tmp, "");
}
return LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, type), "");
}
else {
/*
* Bit mask and recursive shifts
*
* Little-endian registers:
*
* 7654 3210
* WZYX WZYX .... WZYX <= input
* 00Y0 00Y0 .... 00Y0 <= mask
* 00YY 00YY .... 00YY <= shift right 1 (shift amount -1)
* YYYY YYYY .... YYYY <= shift left 2 (shift amount 2)
*
* Big-endian registers:
*
* 0123 4567
* XYZW XYZW .... XYZW <= input
* 0Y00 0Y00 .... 0Y00 <= mask
* YY00 YY00 .... YY00 <= shift left 1 (shift amount 1)
* YYYY YYYY .... YYYY <= shift right 2 (shift amount -2)
*
* shifts[] gives little-endian shift amounts; we need to negate for big-endian.
*/
struct lp_type type4;
const int shifts[4][2] = {
{ 1, 2},
{-1, 2},
{ 1, -2},
{-1, -2}
};
unsigned i;
a = LLVMBuildAnd(builder, a,
lp_build_const_mask_aos(bld->gallivm,
type, 1 << channel, 4), "");
/*
* Build a type where each element is an integer that cover the four
* channels.
*/
type4 = type;
type4.floating = FALSE;
type4.width *= 4;
type4.length /= 4;
a = LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, type4), "");
for(i = 0; i < 2; ++i) {
LLVMValueRef tmp = NULL;
int shift = shifts[channel][i];
/* See endianness diagram above */
#ifdef PIPE_ARCH_BIG_ENDIAN
shift = -shift;
#endif
if(shift > 0)
tmp = LLVMBuildShl(builder, a, lp_build_const_int_vec(bld->gallivm, type4, shift*type.width), "");
if(shift < 0)
tmp = LLVMBuildLShr(builder, a, lp_build_const_int_vec(bld->gallivm, type4, -shift*type.width), "");
assert(tmp);
if(tmp)
a = LLVMBuildOr(builder, a, tmp, "");
}
return LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, type), "");
}
}
/**
* Unpack a single pixel into its XYZW components.
*
* @param desc the pixel format for the packed pixel value
* @param packed integer pixel in a format such as PIPE_FORMAT_B8G8R8A8_UNORM
*
* @return XYZW in a float[4] or ubyte[4] or ushort[4] vector.
*/
static INLINE LLVMValueRef
lp_build_unpack_arith_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *desc,
LLVMValueRef packed)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef shifted, casted, scaled, masked;
LLVMValueRef shifts[4];
LLVMValueRef masks[4];
LLVMValueRef scales[4];
boolean normalized;
boolean needs_uitofp;
unsigned i;
/* TODO: Support more formats */
assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
assert(desc->block.width == 1);
assert(desc->block.height == 1);
assert(desc->block.bits <= 32);
/* Do the intermediate integer computations with 32bit integers since it
* matches floating point size */
assert (LLVMTypeOf(packed) == LLVMInt32TypeInContext(gallivm->context));
/* Broadcast the packed value to all four channels
* before: packed = BGRA
* after: packed = {BGRA, BGRA, BGRA, BGRA}
*/
packed = LLVMBuildInsertElement(builder,
LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
packed,
LLVMConstNull(LLVMInt32TypeInContext(gallivm->context)),
"");
packed = LLVMBuildShuffleVector(builder,
packed,
LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
LLVMConstNull(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
"");
/* Initialize vector constants */
normalized = FALSE;
needs_uitofp = FALSE;
/* Loop over 4 color components */
for (i = 0; i < 4; ++i) {
unsigned bits = desc->channel[i].size;
unsigned shift = desc->channel[i].shift;
if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
masks[i] = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
scales[i] = LLVMConstNull(LLVMFloatTypeInContext(gallivm->context));
}
else {
unsigned long long mask = (1ULL << bits) - 1;
assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
if (bits == 32) {
needs_uitofp = TRUE;
}
shifts[i] = lp_build_const_int32(gallivm, shift);
masks[i] = lp_build_const_int32(gallivm, mask);
if (desc->channel[i].normalized) {
scales[i] = lp_build_const_float(gallivm, 1.0 / mask);
normalized = TRUE;
}
else
scales[i] = lp_build_const_float(gallivm, 1.0);
}
}
/* Ex: convert packed = {XYZW, XYZW, XYZW, XYZW}
* into masked = {X, Y, Z, W}
*/
shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), "");
masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), "");
if (!needs_uitofp) {
/* UIToFP can't be expressed in SSE2 */
casted = LLVMBuildSIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
} else {
casted = LLVMBuildUIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
}
/* At this point 'casted' may be a vector of floats such as
* {255.0, 255.0, 255.0, 255.0}. Next, if the pixel values are normalized
* we'll scale this to {1.0, 1.0, 1.0, 1.0}.
*/
if (normalized)
scaled = LLVMBuildFMul(builder, casted, LLVMConstVector(scales, 4), "");
else
scaled = casted;
return scaled;
}
/**
* Pack a single pixel.
*
* @param rgba 4 float vector with the unpacked components.
*
* XXX: This is mostly for reference and testing -- operating a single pixel at
* a time is rarely if ever needed.
*/
LLVMValueRef
lp_build_pack_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *desc,
LLVMValueRef rgba)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef type;
LLVMValueRef packed = NULL;
LLVMValueRef swizzles[4];
LLVMValueRef shifted, casted, scaled, unswizzled;
LLVMValueRef shifts[4];
LLVMValueRef scales[4];
boolean normalized;
unsigned i, j;
assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
assert(desc->block.width == 1);
assert(desc->block.height == 1);
type = LLVMIntTypeInContext(gallivm->context, desc->block.bits);
/* Unswizzle the color components into the source vector. */
for (i = 0; i < 4; ++i) {
for (j = 0; j < 4; ++j) {
if (desc->swizzle[j] == i)
break;
}
if (j < 4)
swizzles[i] = lp_build_const_int32(gallivm, j);
else
swizzles[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
}
unswizzled = LLVMBuildShuffleVector(builder, rgba,
LLVMGetUndef(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4)),
LLVMConstVector(swizzles, 4), "");
normalized = FALSE;
for (i = 0; i < 4; ++i) {
unsigned bits = desc->channel[i].size;
unsigned shift = desc->channel[i].shift;
if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
scales[i] = LLVMGetUndef(LLVMFloatTypeInContext(gallivm->context));
}
else {
unsigned mask = (1 << bits) - 1;
assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
assert(bits < 32);
shifts[i] = lp_build_const_int32(gallivm, shift);
if (desc->channel[i].normalized) {
scales[i] = lp_build_const_float(gallivm, mask);
normalized = TRUE;
}
else
scales[i] = lp_build_const_float(gallivm, 1.0);
}
}
if (normalized)
scaled = LLVMBuildFMul(builder, unswizzled, LLVMConstVector(scales, 4), "");
else
scaled = unswizzled;
casted = LLVMBuildFPToSI(builder, scaled, LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), "");
shifted = LLVMBuildShl(builder, casted, LLVMConstVector(shifts, 4), "");
/* Bitwise or all components */
for (i = 0; i < 4; ++i) {
if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
LLVMValueRef component = LLVMBuildExtractElement(builder, shifted,
lp_build_const_int32(gallivm, i), "");
if (packed)
packed = LLVMBuildOr(builder, packed, component, "");
else
packed = component;
}
}
if (!packed)
packed = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
if (desc->block.bits < 32)
packed = LLVMBuildTrunc(builder, packed, type, "");
return packed;
}
LLVMValueRef
lp_build_swizzle_aos(struct lp_build_context *bld,
LLVMValueRef a,
const unsigned char swizzles[4])
{
LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
const unsigned n = type.length;
unsigned i, j;
if (swizzles[0] == PIPE_SWIZZLE_RED &&
swizzles[1] == PIPE_SWIZZLE_GREEN &&
swizzles[2] == PIPE_SWIZZLE_BLUE &&
swizzles[3] == PIPE_SWIZZLE_ALPHA) {
return a;
}
if (swizzles[0] == swizzles[1] &&
swizzles[1] == swizzles[2] &&
swizzles[2] == swizzles[3]) {
switch (swizzles[0]) {
case PIPE_SWIZZLE_RED:
case PIPE_SWIZZLE_GREEN:
case PIPE_SWIZZLE_BLUE:
case PIPE_SWIZZLE_ALPHA:
return lp_build_swizzle_scalar_aos(bld, a, swizzles[0]);
case PIPE_SWIZZLE_ZERO:
return bld->zero;
case PIPE_SWIZZLE_ONE:
return bld->one;
case LP_BLD_SWIZZLE_DONTCARE:
return bld->undef;
default:
assert(0);
return bld->undef;
}
}
if (type.width >= 16) {
/*
* Shuffle.
*/
LLVMValueRef undef = LLVMGetUndef(lp_build_elem_type(bld->gallivm, type));
LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
LLVMValueRef aux[LP_MAX_VECTOR_LENGTH];
memset(aux, 0, sizeof aux);
for(j = 0; j < n; j += 4) {
for(i = 0; i < 4; ++i) {
unsigned shuffle;
switch (swizzles[i]) {
default:
assert(0);
/* fall through */
case PIPE_SWIZZLE_RED:
case PIPE_SWIZZLE_GREEN:
case PIPE_SWIZZLE_BLUE:
case PIPE_SWIZZLE_ALPHA:
shuffle = j + swizzles[i];
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
break;
case PIPE_SWIZZLE_ZERO:
shuffle = type.length + 0;
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
if (!aux[0]) {
aux[0] = lp_build_const_elem(bld->gallivm, type, 0.0);
}
break;
case PIPE_SWIZZLE_ONE:
shuffle = type.length + 1;
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
if (!aux[1]) {
aux[1] = lp_build_const_elem(bld->gallivm, type, 1.0);
}
break;
case LP_BLD_SWIZZLE_DONTCARE:
shuffles[j + i] = LLVMGetUndef(i32t);
break;
}
}
}
for (i = 0; i < n; ++i) {
if (!aux[i]) {
aux[i] = undef;
}
}
return LLVMBuildShuffleVector(builder, a,
LLVMConstVector(aux, n),
LLVMConstVector(shuffles, n), "");
} else {
/*
* Bit mask and shifts.
*
* For example, this will convert BGRA to RGBA by doing
*
* rgba = (bgra & 0x00ff0000) >> 16
* | (bgra & 0xff00ff00)
* | (bgra & 0x000000ff) << 16
*
* This is necessary not only for faster cause, but because X86 backend
* will refuse shuffles of <4 x i8> vectors
*/
LLVMValueRef res;
struct lp_type type4;
unsigned cond = 0;
unsigned chan;
int shift;
/*
* Start with a mixture of 1 and 0.
*/
for (chan = 0; chan < 4; ++chan) {
if (swizzles[chan] == PIPE_SWIZZLE_ONE) {
cond |= 1 << chan;
}
}
res = lp_build_select_aos(bld, cond, bld->one, bld->zero);
/*
* Build a type where each element is an integer that cover the four
* channels.
*/
type4 = type;
type4.floating = FALSE;
type4.width *= 4;
type4.length /= 4;
a = LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, type4), "");
res = LLVMBuildBitCast(builder, res, lp_build_vec_type(bld->gallivm, type4), "");
/*
* Mask and shift the channels, trying to group as many channels in the
* same shift as possible
*/
for (shift = -3; shift <= 3; ++shift) {
unsigned long long mask = 0;
assert(type4.width <= sizeof(mask)*8);
for (chan = 0; chan < 4; ++chan) {
/* FIXME: big endian */
if (swizzles[chan] < 4 &&
chan - swizzles[chan] == shift) {
mask |= ((1ULL << type.width) - 1) << (swizzles[chan] * type.width);
}
}
if (mask) {
LLVMValueRef masked;
LLVMValueRef shifted;
if (0)
debug_printf("shift = %i, mask = 0x%08llx\n", shift, mask);
masked = LLVMBuildAnd(builder, a,
lp_build_const_int_vec(bld->gallivm, type4, mask), "");
if (shift > 0) {
shifted = LLVMBuildShl(builder, masked,
lp_build_const_int_vec(bld->gallivm, type4, shift*type.width), "");
} else if (shift < 0) {
shifted = LLVMBuildLShr(builder, masked,
lp_build_const_int_vec(bld->gallivm, type4, -shift*type.width), "");
} else {
shifted = masked;
}
res = LLVMBuildOr(builder, res, shifted, "");
}
}
return LLVMBuildBitCast(builder, res,
lp_build_vec_type(bld->gallivm, type), "");
}
}
static void llvm_emit_tex(
const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
struct gallivm_state * gallivm = bld_base->base.gallivm;
LLVMValueRef args[7];
unsigned c, sampler_src;
struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
if (emit_data->inst->Texture.Texture == TGSI_TEXTURE_BUFFER) {
switch (emit_data->inst->Instruction.Opcode) {
case TGSI_OPCODE_TXQ: {
struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
ctx->uses_tex_buffers = true;
bool isEgPlus = (ctx->chip_class >= EVERGREEN);
LLVMValueRef offset = lp_build_const_int32(bld_base->base.gallivm,
isEgPlus ? 0 : 1);
LLVMValueRef cvecval = llvm_load_const_buffer(bld_base, offset,
LLVM_R600_BUFFER_INFO_CONST_BUFFER);
if (!isEgPlus) {
LLVMValueRef maskval[4] = {
lp_build_const_int32(gallivm, 1),
lp_build_const_int32(gallivm, 2),
lp_build_const_int32(gallivm, 3),
lp_build_const_int32(gallivm, 0),
};
LLVMValueRef mask = LLVMConstVector(maskval, 4);
cvecval = LLVMBuildShuffleVector(gallivm->builder, cvecval, cvecval,
mask, "");
}
emit_data->output[0] = cvecval;
return;
}
case TGSI_OPCODE_TXF: {
args[0] = LLVMBuildExtractElement(gallivm->builder, emit_data->args[0], lp_build_const_int32(gallivm, 0), "");
args[1] = lp_build_const_int32(gallivm, R600_MAX_CONST_BUFFERS);
emit_data->output[0] = lp_build_intrinsic(gallivm->builder,
"llvm.R600.load.texbuf",
emit_data->dst_type, args, 2, LLVMReadNoneAttribute);
if (ctx->chip_class >= EVERGREEN)
return;
ctx->uses_tex_buffers = true;
LLVMDumpValue(emit_data->output[0]);
emit_data->output[0] = LLVMBuildBitCast(gallivm->builder,
emit_data->output[0], LLVMVectorType(bld_base->base.int_elem_type, 4),
"");
LLVMValueRef Mask = llvm_load_const_buffer(bld_base,
lp_build_const_int32(gallivm, 0),
LLVM_R600_BUFFER_INFO_CONST_BUFFER);
Mask = LLVMBuildBitCast(gallivm->builder, Mask,
LLVMVectorType(bld_base->base.int_elem_type, 4), "");
emit_data->output[0] = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_AND,
emit_data->output[0],
Mask);
LLVMValueRef WComponent = LLVMBuildExtractElement(gallivm->builder,
emit_data->output[0], lp_build_const_int32(gallivm, 3), "");
Mask = llvm_load_const_buffer(bld_base, lp_build_const_int32(gallivm, 1),
LLVM_R600_BUFFER_INFO_CONST_BUFFER);
Mask = LLVMBuildExtractElement(gallivm->builder, Mask,
lp_build_const_int32(gallivm, 0), "");
Mask = LLVMBuildBitCast(gallivm->builder, Mask,
bld_base->base.int_elem_type, "");
WComponent = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_OR,
WComponent, Mask);
emit_data->output[0] = LLVMBuildInsertElement(gallivm->builder,
emit_data->output[0], WComponent, lp_build_const_int32(gallivm, 3), "");
emit_data->output[0] = LLVMBuildBitCast(gallivm->builder,
emit_data->output[0], LLVMVectorType(bld_base->base.elem_type, 4), "");
}
return;
default:
break;
}
}
if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_TEX ||
emit_data->inst->Instruction.Opcode == TGSI_OPCODE_TXP) {
LLVMValueRef Vector[4] = {
LLVMBuildExtractElement(gallivm->builder, emit_data->args[0],
lp_build_const_int32(gallivm, 0), ""),
LLVMBuildExtractElement(gallivm->builder, emit_data->args[0],
lp_build_const_int32(gallivm, 1), ""),
LLVMBuildExtractElement(gallivm->builder, emit_data->args[0],
lp_build_const_int32(gallivm, 2), ""),
LLVMBuildExtractElement(gallivm->builder, emit_data->args[0],
lp_build_const_int32(gallivm, 3), ""),
};
switch (emit_data->inst->Texture.Texture) {
case TGSI_TEXTURE_2D:
case TGSI_TEXTURE_RECT:
Vector[2] = Vector[3] = LLVMGetUndef(bld_base->base.elem_type);
break;
case TGSI_TEXTURE_1D:
Vector[1] = Vector[2] = Vector[3] = LLVMGetUndef(bld_base->base.elem_type);
break;
default:
break;
}
args[0] = lp_build_gather_values(gallivm, Vector, 4);
} else {
args[0] = emit_data->args[0];
}
assert(emit_data->arg_count + 2 <= Elements(args));
for (c = 1; c < emit_data->arg_count; ++c)
args[c] = emit_data->args[c];
if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_TXF) {
args[1] = LLVMBuildShl(gallivm->builder, args[1], lp_build_const_int32(gallivm, 1), "");
args[2] = LLVMBuildShl(gallivm->builder, args[2], lp_build_const_int32(gallivm, 1), "");
args[3] = LLVMBuildShl(gallivm->builder, args[3], lp_build_const_int32(gallivm, 1), "");
}
sampler_src = emit_data->inst->Instruction.NumSrcRegs-1;
args[c++] = lp_build_const_int32(gallivm,
emit_data->inst->Src[sampler_src].Register.Index + R600_MAX_CONST_BUFFERS);
args[c++] = lp_build_const_int32(gallivm,
emit_data->inst->Src[sampler_src].Register.Index);
args[c++] = lp_build_const_int32(gallivm,
emit_data->inst->Texture.Texture);
if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_TXF &&
(emit_data->inst->Texture.Texture == TGSI_TEXTURE_2D_MSAA ||
emit_data->inst->Texture.Texture == TGSI_TEXTURE_2D_ARRAY_MSAA)) {
switch (emit_data->inst->Texture.Texture) {
case TGSI_TEXTURE_2D_MSAA:
args[6] = lp_build_const_int32(gallivm, TGSI_TEXTURE_2D);
break;
case TGSI_TEXTURE_2D_ARRAY_MSAA:
args[6] = lp_build_const_int32(gallivm, TGSI_TEXTURE_2D_ARRAY);
break;
default:
break;
}
if (ctx->has_compressed_msaa_texturing) {
LLVMValueRef ldptr_args[10] = {
args[0], // Coord
args[1], // Offset X
args[2], // Offset Y
args[3], // Offset Z
args[4],
args[5],
lp_build_const_int32(gallivm, 1),
lp_build_const_int32(gallivm, 1),
lp_build_const_int32(gallivm, 1),
lp_build_const_int32(gallivm, 1)
};
LLVMValueRef ptr = lp_build_intrinsic(gallivm->builder,
"llvm.R600.ldptr",
emit_data->dst_type, ldptr_args, 10, LLVMReadNoneAttribute);
LLVMValueRef Tmp = LLVMBuildExtractElement(gallivm->builder, args[0],
lp_build_const_int32(gallivm, 3), "");
Tmp = LLVMBuildMul(gallivm->builder, Tmp,
lp_build_const_int32(gallivm, 4), "");
LLVMValueRef ResX = LLVMBuildExtractElement(gallivm->builder, ptr,
lp_build_const_int32(gallivm, 0), "");
ResX = LLVMBuildBitCast(gallivm->builder, ResX,
bld_base->base.int_elem_type, "");
Tmp = LLVMBuildLShr(gallivm->builder, ResX, Tmp, "");
Tmp = LLVMBuildAnd(gallivm->builder, Tmp,
lp_build_const_int32(gallivm, 0xF), "");
args[0] = LLVMBuildInsertElement(gallivm->builder, args[0], Tmp,
lp_build_const_int32(gallivm, 3), "");
args[c++] = lp_build_const_int32(gallivm,
emit_data->inst->Texture.Texture);
}
}
emit_data->output[0] = lp_build_intrinsic(gallivm->builder,
action->intr_name,
emit_data->dst_type, args, c, LLVMReadNoneAttribute);
if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_TXQ &&
((emit_data->inst->Texture.Texture == TGSI_TEXTURE_CUBE_ARRAY ||
emit_data->inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE_ARRAY)))
if (emit_data->inst->Dst[0].Register.WriteMask & 4) {
LLVMValueRef offset = lp_build_const_int32(bld_base->base.gallivm, 0);
LLVMValueRef ZLayer = LLVMBuildExtractElement(gallivm->builder,
llvm_load_const_buffer(bld_base, offset, LLVM_R600_BUFFER_INFO_CONST_BUFFER),
lp_build_const_int32(gallivm, 0), "");
emit_data->output[0] = LLVMBuildInsertElement(gallivm->builder, emit_data->output[0], ZLayer, lp_build_const_int32(gallivm, 2), "");
struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
ctx->has_txq_cube_array_z_comp = true;
}
}
static LLVMValueRef
emit_fetch_constant(
struct lp_build_tgsi_context * bld_base,
const struct tgsi_full_src_register * reg,
enum tgsi_opcode_type stype,
unsigned swizzle)
{
struct lp_build_tgsi_aos_context * bld = lp_aos_context(bld_base);
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
struct lp_type type = bld_base->base.type;
LLVMValueRef res;
unsigned chan;
assert(!reg->Register.Indirect);
/*
* Get the constants components
*/
res = bld->bld_base.base.undef;
for (chan = 0; chan < 4; ++chan) {
LLVMValueRef index;
LLVMValueRef scalar_ptr;
LLVMValueRef scalar;
LLVMValueRef swizzle;
index = lp_build_const_int32(bld->bld_base.base.gallivm,
reg->Register.Index * 4 + chan);
scalar_ptr = LLVMBuildGEP(builder, bld->consts_ptr, &index, 1, "");
scalar = LLVMBuildLoad(builder, scalar_ptr, "");
lp_build_name(scalar, "const[%u].%c", reg->Register.Index, "xyzw"[chan]);
/*
* NOTE: constants array is always assumed to be RGBA
*/
swizzle = lp_build_const_int32(bld->bld_base.base.gallivm,
bld->swizzles[chan]);
res = LLVMBuildInsertElement(builder, res, scalar, swizzle, "");
}
/*
* Broadcast the first quaternion to all others.
*
* XXX: could be factored into a reusable function.
*/
if (type.length > 4) {
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
unsigned i;
for (chan = 0; chan < 4; ++chan) {
shuffles[chan] = lp_build_const_int32(bld->bld_base.base.gallivm, chan);
}
for (i = 4; i < type.length; ++i) {
shuffles[i] = shuffles[i % 4];
}
res = LLVMBuildShuffleVector(builder,
res, bld->bld_base.base.undef,
LLVMConstVector(shuffles, type.length),
"");
}
return res;
}
/* This function returns an i1 1 if the value is not equal to 0 and
an i1 0 if the value is equal to 0. */
struct cl2llvm_val_t *cl2llvm_to_bool_ne_0(struct cl2llvm_val_t *value)
{
LLVMValueRef const_zero;
LLVMValueRef zero_vec[16];
LLVMTypeRef switch_type;
int i;
int veclength;
struct cl2llvm_val_t *bool_val = cl2llvm_val_create_w_init(value->val, cl2llvmTypeWrapGetSign(value->type));
/* if value is i1 no conversion necessary */
if (LLVMTypeOf(value->val) == LLVMInt1Type())
return bool_val;
/* If value is a vector create a vector of constant zeros, else
create a scalar 0. */
if (LLVMGetTypeKind(cl2llvmTypeWrapGetLlvmType(value->type)) == LLVMVectorTypeKind)
{
switch_type = LLVMGetElementType(cl2llvmTypeWrapGetLlvmType(value->type));
veclength = LLVMGetVectorSize(cl2llvmTypeWrapGetLlvmType(value->type));
switch (LLVMGetTypeKind(LLVMGetElementType(cl2llvmTypeWrapGetLlvmType(value->type))))
{
case LLVMIntegerTypeKind:
/* Create zero vector */
for (i = 0; i < veclength; i++)
zero_vec[i] = LLVMConstInt(switch_type, 0, 0);
break;
case LLVMFloatTypeKind:
case LLVMDoubleTypeKind:
case LLVMHalfTypeKind:
/* Create zero vector */
for (i = 0; i < veclength; i++)
zero_vec[i] = LLVMConstReal(switch_type, 0);
break;
default:
cl2llvm_yyerror("unreachable code reached");
}
const_zero = LLVMConstVector(zero_vec, veclength);
}
else if (LLVMGetTypeKind(cl2llvmTypeWrapGetLlvmType(value->type)) == LLVMIntegerTypeKind)
{
const_zero = LLVMConstInt(cl2llvmTypeWrapGetLlvmType(value->type), 0, 0);
switch_type = cl2llvmTypeWrapGetLlvmType(value->type);
}
else if (LLVMGetTypeKind(cl2llvmTypeWrapGetLlvmType(value->type)) == LLVMFloatTypeKind
|| LLVMGetTypeKind(cl2llvmTypeWrapGetLlvmType(value->type)) == LLVMDoubleTypeKind
|| LLVMGetTypeKind(cl2llvmTypeWrapGetLlvmType(value->type)) == LLVMHalfTypeKind)
{
const_zero = LLVMConstReal(cl2llvmTypeWrapGetLlvmType(value->type), 0);
switch_type = cl2llvmTypeWrapGetLlvmType(value->type);
}
/* Create comparison */
snprintf(temp_var_name, sizeof temp_var_name,
"tmp_%d", temp_var_count++);
switch (LLVMGetTypeKind(switch_type))
{
case LLVMFloatTypeKind:
case LLVMDoubleTypeKind:
case LLVMHalfTypeKind:
bool_val->val = LLVMBuildFCmp(cl2llvm_builder, LLVMRealONE,
value->val, const_zero, temp_var_name);
break;
case LLVMIntegerTypeKind:
bool_val->val = LLVMBuildICmp(cl2llvm_builder, LLVMIntNE,
value->val, const_zero, temp_var_name);
break;
default:
cl2llvm_yyerror("unreachable code reached");
break;
}
cl2llvmTypeWrapSetLlvmType(bool_val->type, LLVMInt1Type());
cl2llvmTypeWrapSetSign(bool_val->type, 0);
return bool_val;
}
/**
* 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 the texture/mipmap using given image filter and mip filter.
* data0_ptr and data1_ptr point to the two mipmap levels to sample
* from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
* If we're using nearest miplevel sampling the '1' values will be null/unused.
*/
static void
lp_build_sample_mipmap(struct lp_build_sample_context *bld,
unsigned img_filter,
unsigned mip_filter,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
LLVMValueRef ilevel0,
LLVMValueRef ilevel1,
LLVMValueRef lod_fpart,
LLVMValueRef colors_lo_var,
LLVMValueRef colors_hi_var)
{
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef size0;
LLVMValueRef size1;
LLVMValueRef row_stride0_vec;
LLVMValueRef row_stride1_vec;
LLVMValueRef img_stride0_vec;
LLVMValueRef img_stride1_vec;
LLVMValueRef data_ptr0;
LLVMValueRef data_ptr1;
LLVMValueRef colors0_lo, colors0_hi;
LLVMValueRef colors1_lo, colors1_hi;
/* sample the first mipmap level */
lp_build_mipmap_level_sizes(bld, ilevel0,
&size0,
&row_stride0_vec, &img_stride0_vec);
data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
lp_build_sample_image_nearest(bld,
size0,
row_stride0_vec, img_stride0_vec,
data_ptr0, s, t, r,
&colors0_lo, &colors0_hi);
}
else {
assert(img_filter == PIPE_TEX_FILTER_LINEAR);
lp_build_sample_image_linear(bld,
size0,
row_stride0_vec, img_stride0_vec,
data_ptr0, s, t, r,
&colors0_lo, &colors0_hi);
}
/* Store the first level's colors in the output variables */
LLVMBuildStore(builder, colors0_lo, colors_lo_var);
LLVMBuildStore(builder, colors0_hi, colors_hi_var);
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
LLVMValueRef h16_scale = lp_build_const_float(bld->gallivm, 256.0);
LLVMTypeRef i32_type = LLVMIntTypeInContext(bld->gallivm->context, 32);
struct lp_build_if_state if_ctx;
LLVMValueRef need_lerp;
lod_fpart = LLVMBuildFMul(builder, lod_fpart, h16_scale, "");
lod_fpart = LLVMBuildFPToSI(builder, lod_fpart, i32_type, "lod_fpart.fixed16");
/* need_lerp = lod_fpart > 0 */
need_lerp = LLVMBuildICmp(builder, LLVMIntSGT,
lod_fpart, LLVMConstNull(i32_type),
"need_lerp");
lp_build_if(&if_ctx, bld->gallivm, need_lerp);
{
struct lp_build_context h16_bld;
lp_build_context_init(&h16_bld, bld->gallivm, lp_type_ufixed(16));
/* sample the second mipmap level */
lp_build_mipmap_level_sizes(bld, ilevel1,
&size1,
&row_stride1_vec, &img_stride1_vec);
data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
lp_build_sample_image_nearest(bld,
size1,
row_stride1_vec, img_stride1_vec,
data_ptr1, s, t, r,
&colors1_lo, &colors1_hi);
}
else {
lp_build_sample_image_linear(bld,
size1,
row_stride1_vec, img_stride1_vec,
data_ptr1, s, t, r,
&colors1_lo, &colors1_hi);
}
/* interpolate samples from the two mipmap levels */
lod_fpart = LLVMBuildTrunc(builder, lod_fpart, h16_bld.elem_type, "");
lod_fpart = lp_build_broadcast_scalar(&h16_bld, lod_fpart);
#if HAVE_LLVM == 0x208
/* This is a work-around for a bug in LLVM 2.8.
* Evidently, something goes wrong in the construction of the
* lod_fpart short[8] vector. Adding this no-effect shuffle seems
* to force the vector to be properly constructed.
* Tested with mesa-demos/src/tests/mipmap_limits.c (press t, f).
*/
{
LLVMValueRef shuffles[8], shuffle;
int i;
assert(h16_bld.type.length <= Elements(shuffles));
for (i = 0; i < h16_bld.type.length; i++)
shuffles[i] = lp_build_const_int32(bld->gallivm, 2 * (i & 1));
shuffle = LLVMConstVector(shuffles, h16_bld.type.length);
lod_fpart = LLVMBuildShuffleVector(builder,
lod_fpart, lod_fpart,
shuffle, "");
}
#endif
colors0_lo = lp_build_lerp(&h16_bld, lod_fpart,
colors0_lo, colors1_lo);
colors0_hi = lp_build_lerp(&h16_bld, lod_fpart,
colors0_hi, colors1_hi);
LLVMBuildStore(builder, colors0_lo, colors_lo_var);
LLVMBuildStore(builder, colors0_hi, colors_hi_var);
}
lp_build_endif(&if_ctx);
}
}
/**
* 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);
}
}
/**
* 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);
}
}
/**
* Expand the relevent bits of mask_input to a 4-dword mask for the
* four pixels in a 2x2 quad. This will set the four elements of the
* quad mask vector to 0 or ~0.
*
* \param quad which quad of the quad group to test, in [0,3]
* \param mask_input bitwise mask for the whole 4x4 stamp
*/
static LLVMValueRef
generate_quad_mask(LLVMBuilderRef builder,
struct lp_type fs_type,
unsigned quad,
LLVMValueRef mask_input) /* int32 */
{
struct lp_type mask_type;
LLVMTypeRef i32t = LLVMInt32Type();
LLVMValueRef bits[4];
LLVMValueRef mask;
int shift;
/*
* XXX: We'll need a different path for 16 x u8
*/
assert(fs_type.width == 32);
assert(fs_type.length == 4);
mask_type = lp_int_type(fs_type);
/*
* mask_input >>= (quad * 4)
*/
switch (quad) {
case 0:
shift = 0;
break;
case 1:
shift = 2;
break;
case 2:
shift = 8;
break;
case 3:
shift = 10;
break;
default:
assert(0);
shift = 0;
}
mask_input = LLVMBuildLShr(builder,
mask_input,
LLVMConstInt(i32t, shift, 0),
"");
/*
* mask = { mask_input & (1 << i), for i in [0,3] }
*/
mask = lp_build_broadcast(builder, lp_build_vec_type(mask_type), mask_input);
bits[0] = LLVMConstInt(i32t, 1 << 0, 0);
bits[1] = LLVMConstInt(i32t, 1 << 1, 0);
bits[2] = LLVMConstInt(i32t, 1 << 4, 0);
bits[3] = LLVMConstInt(i32t, 1 << 5, 0);
mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), "");
/*
* mask = mask != 0 ? ~0 : 0
*/
mask = lp_build_compare(builder,
mask_type, PIPE_FUNC_NOTEQUAL,
mask,
lp_build_const_int_vec(mask_type, 0));
return mask;
}
LLVMValueRef
lp_build_swizzle_aos(struct lp_build_context *bld,
LLVMValueRef a,
const unsigned char swizzles[4])
{
LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
const unsigned n = type.length;
unsigned i, j;
if (swizzles[0] == PIPE_SWIZZLE_X &&
swizzles[1] == PIPE_SWIZZLE_Y &&
swizzles[2] == PIPE_SWIZZLE_Z &&
swizzles[3] == PIPE_SWIZZLE_W) {
return a;
}
if (swizzles[0] == swizzles[1] &&
swizzles[1] == swizzles[2] &&
swizzles[2] == swizzles[3]) {
switch (swizzles[0]) {
case PIPE_SWIZZLE_X:
case PIPE_SWIZZLE_Y:
case PIPE_SWIZZLE_Z:
case PIPE_SWIZZLE_W:
return lp_build_swizzle_scalar_aos(bld, a, swizzles[0], 4);
case PIPE_SWIZZLE_0:
return bld->zero;
case PIPE_SWIZZLE_1:
return bld->one;
case LP_BLD_SWIZZLE_DONTCARE:
return bld->undef;
default:
assert(0);
return bld->undef;
}
}
if (LLVMIsConstant(a) ||
type.width >= 16) {
/*
* Shuffle.
*/
LLVMValueRef undef = LLVMGetUndef(lp_build_elem_type(bld->gallivm, type));
LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
LLVMValueRef aux[LP_MAX_VECTOR_LENGTH];
memset(aux, 0, sizeof aux);
for(j = 0; j < n; j += 4) {
for(i = 0; i < 4; ++i) {
unsigned shuffle;
switch (swizzles[i]) {
default:
assert(0);
/* fall through */
case PIPE_SWIZZLE_X:
case PIPE_SWIZZLE_Y:
case PIPE_SWIZZLE_Z:
case PIPE_SWIZZLE_W:
shuffle = j + swizzles[i];
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
break;
case PIPE_SWIZZLE_0:
shuffle = type.length + 0;
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
if (!aux[0]) {
aux[0] = lp_build_const_elem(bld->gallivm, type, 0.0);
}
break;
case PIPE_SWIZZLE_1:
shuffle = type.length + 1;
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
if (!aux[1]) {
aux[1] = lp_build_const_elem(bld->gallivm, type, 1.0);
}
break;
case LP_BLD_SWIZZLE_DONTCARE:
shuffles[j + i] = LLVMGetUndef(i32t);
break;
}
}
}
for (i = 0; i < n; ++i) {
if (!aux[i]) {
aux[i] = undef;
}
}
return LLVMBuildShuffleVector(builder, a,
LLVMConstVector(aux, n),
LLVMConstVector(shuffles, n), "");
} else {
/*
* Bit mask and shifts.
*
* For example, this will convert BGRA to RGBA by doing
*
* Little endian:
* rgba = (bgra & 0x00ff0000) >> 16
* | (bgra & 0xff00ff00)
* | (bgra & 0x000000ff) << 16
*
* Big endian:A
* rgba = (bgra & 0x0000ff00) << 16
* | (bgra & 0x00ff00ff)
* | (bgra & 0xff000000) >> 16
*
* This is necessary not only for faster cause, but because X86 backend
* will refuse shuffles of <4 x i8> vectors
*/
LLVMValueRef res;
struct lp_type type4;
unsigned cond = 0;
unsigned chan;
int shift;
/*
* Start with a mixture of 1 and 0.
*/
for (chan = 0; chan < 4; ++chan) {
if (swizzles[chan] == PIPE_SWIZZLE_1) {
cond |= 1 << chan;
}
}
res = lp_build_select_aos(bld, cond, bld->one, bld->zero, 4);
/*
* Build a type where each element is an integer that cover the four
* channels.
*/
type4 = type;
type4.floating = FALSE;
type4.width *= 4;
type4.length /= 4;
a = LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, type4), "");
res = LLVMBuildBitCast(builder, res, lp_build_vec_type(bld->gallivm, type4), "");
/*
* Mask and shift the channels, trying to group as many channels in the
* same shift as possible. The shift amount is positive for shifts left
* and negative for shifts right.
*/
for (shift = -3; shift <= 3; ++shift) {
uint64_t mask = 0;
assert(type4.width <= sizeof(mask)*8);
/*
* Vector element numbers follow the XYZW order, so 0 is always X, etc.
* After widening 4 times we have:
*
* 3210
* Little-endian register layout: WZYX
*
* 0123
* Big-endian register layout: XYZW
*
* For little-endian, higher-numbered channels are obtained by a shift right
* (negative shift amount) and lower-numbered channels by a shift left
* (positive shift amount). The opposite is true for big-endian.
*/
for (chan = 0; chan < 4; ++chan) {
if (swizzles[chan] < 4) {
/* We need to move channel swizzles[chan] into channel chan */
#ifdef PIPE_ARCH_LITTLE_ENDIAN
if (swizzles[chan] - chan == -shift) {
mask |= ((1ULL << type.width) - 1) << (swizzles[chan] * type.width);
}
#else
if (swizzles[chan] - chan == shift) {
mask |= ((1ULL << type.width) - 1) << (type4.width - type.width) >> (swizzles[chan] * type.width);
}
#endif
}
}
if (mask) {
LLVMValueRef masked;
LLVMValueRef shifted;
if (0)
debug_printf("shift = %i, mask = %" PRIx64 "\n", shift, mask);
masked = LLVMBuildAnd(builder, a,
lp_build_const_int_vec(bld->gallivm, type4, mask), "");
if (shift > 0) {
shifted = LLVMBuildShl(builder, masked,
lp_build_const_int_vec(bld->gallivm, type4, shift*type.width), "");
} else if (shift < 0) {
shifted = LLVMBuildLShr(builder, masked,
lp_build_const_int_vec(bld->gallivm, type4, -shift*type.width), "");
} else {
shifted = masked;
}
res = LLVMBuildOr(builder, res, shifted, "");
}
}
return LLVMBuildBitCast(builder, res,
lp_build_vec_type(bld->gallivm, type), "");
}
/**
* Swizzle one channel into all other three channels.
*/
LLVMValueRef
lp_build_swizzle_scalar_aos(struct lp_build_context *bld,
LLVMValueRef a,
unsigned channel)
{
LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
const unsigned n = type.length;
unsigned i, j;
if(a == bld->undef || a == bld->zero || a == bld->one)
return a;
/* XXX: SSE3 has PSHUFB which should be better than bitmasks, but forcing
* using shuffles here actually causes worst results. More investigation is
* needed. */
if (type.width >= 16) {
/*
* Shuffle.
*/
LLVMTypeRef elem_type = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
for(j = 0; j < n; j += 4)
for(i = 0; i < 4; ++i)
shuffles[j + i] = LLVMConstInt(elem_type, j + channel, 0);
return LLVMBuildShuffleVector(builder, a, bld->undef, LLVMConstVector(shuffles, n), "");
}
else {
/*
* Bit mask and recursive shifts
*
* XYZW XYZW .... XYZW <= input
* 0Y00 0Y00 .... 0Y00
* YY00 YY00 .... YY00
* YYYY YYYY .... YYYY <= output
*/
struct lp_type type4;
const char shifts[4][2] = {
{ 1, 2},
{-1, 2},
{ 1, -2},
{-1, -2}
};
unsigned i;
a = LLVMBuildAnd(builder, a,
lp_build_const_mask_aos(bld->gallivm,
type, 1 << channel), "");
/*
* Build a type where each element is an integer that cover the four
* channels.
*/
type4 = type;
type4.floating = FALSE;
type4.width *= 4;
type4.length /= 4;
a = LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, type4), "");
for(i = 0; i < 2; ++i) {
LLVMValueRef tmp = NULL;
int shift = shifts[channel][i];
#ifdef PIPE_ARCH_LITTLE_ENDIAN
shift = -shift;
#endif
if(shift > 0)
tmp = LLVMBuildLShr(builder, a, lp_build_const_int_vec(bld->gallivm, type4, shift*type.width), "");
if(shift < 0)
tmp = LLVMBuildShl(builder, a, lp_build_const_int_vec(bld->gallivm, type4, -shift*type.width), "");
assert(tmp);
if(tmp)
a = LLVMBuildOr(builder, a, tmp, "");
}
return LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, type), "");
}
}
/**
* Perform the occlusion test and increase the counter.
* Test the depth mask. Add the number of channel which has none zero mask
* into the occlusion counter. e.g. maskvalue is {-1, -1, -1, -1}.
* The counter will add 4.
*
* \param type holds element type of the mask vector.
* \param maskvalue is the depth test mask.
* \param counter is a pointer of the uint32 counter.
*/
void
lp_build_occlusion_count(struct gallivm_state *gallivm,
struct lp_type type,
LLVMValueRef maskvalue,
LLVMValueRef counter)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMContextRef context = gallivm->context;
LLVMValueRef countmask = lp_build_const_int_vec(gallivm, type, 1);
LLVMValueRef count, newcount;
assert(type.length <= 16);
assert(type.floating);
if(util_cpu_caps.has_sse && type.length == 4) {
const char *movmskintr = "llvm.x86.sse.movmsk.ps";
const char *popcntintr = "llvm.ctpop.i32";
LLVMValueRef bits = LLVMBuildBitCast(builder, maskvalue,
lp_build_vec_type(gallivm, type), "");
bits = lp_build_intrinsic_unary(builder, movmskintr,
LLVMInt32TypeInContext(context), bits);
count = lp_build_intrinsic_unary(builder, popcntintr,
LLVMInt32TypeInContext(context), bits);
}
else if(util_cpu_caps.has_avx && type.length == 8) {
const char *movmskintr = "llvm.x86.avx.movmsk.ps.256";
const char *popcntintr = "llvm.ctpop.i32";
LLVMValueRef bits = LLVMBuildBitCast(builder, maskvalue,
lp_build_vec_type(gallivm, type), "");
bits = lp_build_intrinsic_unary(builder, movmskintr,
LLVMInt32TypeInContext(context), bits);
count = lp_build_intrinsic_unary(builder, popcntintr,
LLVMInt32TypeInContext(context), bits);
}
else {
unsigned i;
LLVMValueRef countv = LLVMBuildAnd(builder, maskvalue, countmask, "countv");
LLVMTypeRef counttype = LLVMIntTypeInContext(context, type.length * 8);
LLVMTypeRef i8vntype = LLVMVectorType(LLVMInt8TypeInContext(context), type.length * 4);
LLVMValueRef shufflev, countd;
LLVMValueRef shuffles[16];
const char *popcntintr = NULL;
countv = LLVMBuildBitCast(builder, countv, i8vntype, "");
for (i = 0; i < type.length; i++) {
shuffles[i] = lp_build_const_int32(gallivm, 4*i);
}
shufflev = LLVMConstVector(shuffles, type.length);
countd = LLVMBuildShuffleVector(builder, countv, LLVMGetUndef(i8vntype), shufflev, "");
countd = LLVMBuildBitCast(builder, countd, counttype, "countd");
/*
* XXX FIXME
* this is bad on cpus without popcount (on x86 supported by intel
* nehalem, amd barcelona, and up - not tied to sse42).
* Would be much faster to just sum the 4 elements of the vector with
* some horizontal add (shuffle/add/shuffle/add after the initial and).
*/
switch (type.length) {
case 4:
popcntintr = "llvm.ctpop.i32";
break;
case 8:
popcntintr = "llvm.ctpop.i64";
break;
case 16:
popcntintr = "llvm.ctpop.i128";
break;
default:
assert(0);
}
count = lp_build_intrinsic_unary(builder, popcntintr, counttype, countd);
if (type.length > 4) {
count = LLVMBuildTrunc(builder, count, LLVMIntTypeInContext(context, 32), "");
}
}
newcount = LLVMBuildLoad(builder, counter, "origcount");
newcount = LLVMBuildAdd(builder, newcount, count, "newcount");
LLVMBuildStore(builder, newcount, counter);
}
struct cl2llvm_val_t *llvm_type_cast(struct cl2llvm_val_t * original_val,
struct cl2llvmTypeWrap *totype_w_sign)
{
struct cl2llvm_val_t *llvm_val = cl2llvm_val_create();
int i;
struct cl2llvmTypeWrap *elem_type;
struct cl2llvm_val_t *cast_original_val;
LLVMValueRef index;
LLVMValueRef vector_addr;
LLVMValueRef vector;
LLVMValueRef const_elems[16];
LLVMTypeRef fromtype = cl2llvmTypeWrapGetLlvmType(original_val->type);
LLVMTypeRef totype = cl2llvmTypeWrapGetLlvmType(totype_w_sign);
int fromsign = cl2llvmTypeWrapGetSign(original_val->type);
int tosign = cl2llvmTypeWrapGetSign(totype_w_sign);
/*By default the return value is the same as the original_val*/
llvm_val->val = original_val->val;
cl2llvmTypeWrapSetLlvmType(llvm_val->type, cl2llvmTypeWrapGetLlvmType(original_val->type));
cl2llvmTypeWrapSetSign(llvm_val->type, cl2llvmTypeWrapGetSign(original_val->type));
snprintf(temp_var_name, sizeof temp_var_name,
"tmp_%d", temp_var_count++);
/* Check that fromtype is not a vector, unless both types are identical. */
if (LLVMGetTypeKind(fromtype) == LLVMVectorTypeKind)
{
if ((LLVMGetVectorSize(fromtype) != LLVMGetVectorSize(totype)
|| LLVMGetElementType(fromtype)
!= LLVMGetElementType(totype))
|| fromsign != tosign)
{
if (LLVMGetTypeKind(totype) == LLVMVectorTypeKind)
cl2llvm_yyerror("Casts between vector types are forbidden");
cl2llvm_yyerror("A vector may not be cast to any other type.");
}
}
/* If totype is a vector, create a vector whose components are equal to
original_val */
if (LLVMGetTypeKind(totype) == LLVMVectorTypeKind
&& LLVMGetTypeKind(fromtype) != LLVMVectorTypeKind)
{
/*Go to entry block and declare vector*/
LLVMPositionBuilder(cl2llvm_builder, cl2llvm_current_function->entry_block,
cl2llvm_current_function->branch_instr);
snprintf(temp_var_name, sizeof temp_var_name,
"tmp_%d", temp_var_count++);
vector_addr = LLVMBuildAlloca(cl2llvm_builder,
totype, temp_var_name);
LLVMPositionBuilderAtEnd(cl2llvm_builder, current_basic_block);
/* Load vector */
snprintf(temp_var_name, sizeof temp_var_name,
"tmp_%d", temp_var_count++);
vector = LLVMBuildLoad(cl2llvm_builder, vector_addr, temp_var_name);
/* Create object to represent element type of totype */
elem_type = cl2llvmTypeWrapCreate(LLVMGetElementType(totype), tosign);
/* If original_val is constant create a constant vector */
if (LLVMIsConstant(original_val->val))
{
cast_original_val = llvm_type_cast(original_val, elem_type);
for (i = 0; i < LLVMGetVectorSize(totype); i++)
const_elems[i] = cast_original_val->val;
vector = LLVMConstVector(const_elems,
LLVMGetVectorSize(totype));
llvm_val->val = vector;
cl2llvm_val_free(cast_original_val);
}
/* If original value is not constant insert elements */
else
{
for (i = 0; i < LLVMGetVectorSize(totype); i++)
{
index = LLVMConstInt(LLVMInt32Type(), i, 0);
cast_original_val = llvm_type_cast(original_val, elem_type);
snprintf(temp_var_name, sizeof temp_var_name,
"tmp_%d", temp_var_count++);
vector = LLVMBuildInsertElement(cl2llvm_builder,
vector, cast_original_val->val, index, temp_var_name);
cl2llvm_val_free(cast_original_val);
}
}
cl2llvmTypeWrapFree(elem_type);
llvm_val->val = vector;
}
if (fromtype == LLVMInt64Type())
{
if (totype == LLVMDoubleType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMInt64Type())
{
if (tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
temp_var_count--;
}
else if (totype == LLVMInt32Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt32Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt16Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt16Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt8Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt8Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
}
else if (fromtype == LLVMInt32Type())
{
if (totype == LLVMDoubleType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMInt64Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
if (tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt32Type())
{
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
temp_var_count--;
}
else if (totype == LLVMInt16Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt16Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt8Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt8Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
}
else if (fromtype == LLVMInt16Type())
{
if (totype == LLVMDoubleType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMInt64Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
if (tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt32Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt16Type())
{
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
temp_var_count--;
}
else if (totype == LLVMInt8Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt8Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
}
else if (fromtype == LLVMInt8Type())
{
if (totype == LLVMDoubleType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMInt64Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
if (tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt32Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt16Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt16Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt16Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt8Type())
{
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
temp_var_count--;
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
}
else if (fromtype == LLVMInt1Type())
{
if (totype == LLVMDoubleType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMInt64Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
if (tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt32Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt16Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt16Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt16Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt8Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt8Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt8Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt1Type())
{
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
temp_var_count--;
}
}
/*We now know that from type must be a floating point.*/
/*Floating point to signed integer conversions*/
else if (tosign && LLVMGetTypeKind(totype) == 8)
{
if (totype == LLVMInt64Type())
{
llvm_val->val = LLVMBuildFPToSI(cl2llvm_builder,
original_val->val, LLVMInt64Type(), temp_var_name);
}
else if (totype == LLVMInt32Type())
{
llvm_val->val = LLVMBuildFPToSI(cl2llvm_builder,
original_val->val, LLVMInt32Type(), temp_var_name);
}
else if (totype == LLVMInt16Type())
{
llvm_val->val = LLVMBuildFPToSI(cl2llvm_builder,
original_val->val, LLVMInt16Type(), temp_var_name);
}
else if (totype == LLVMInt8Type())
{
llvm_val->val = LLVMBuildFPToSI(cl2llvm_builder,
original_val->val, LLVMInt8Type(), temp_var_name);
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildFPToSI(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
/*Floating point to unsigned integer conversions*/
else if (!tosign)
{
if (totype == LLVMInt64Type())
{
llvm_val->val = LLVMBuildFPToUI(cl2llvm_builder,
original_val->val, LLVMInt64Type(), temp_var_name);
}
else if (totype == LLVMInt32Type())
{
llvm_val->val = LLVMBuildFPToUI(cl2llvm_builder,
original_val->val, LLVMInt32Type(), temp_var_name);
}
else if (totype == LLVMInt16Type())
{
llvm_val->val = LLVMBuildFPToUI(cl2llvm_builder,
original_val->val, LLVMInt16Type(), temp_var_name);
}
else if (totype == LLVMInt8Type())
{
llvm_val->val = LLVMBuildFPToUI(cl2llvm_builder,
original_val->val, LLVMInt8Type(), temp_var_name);
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildFPToUI(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMDoubleType())
{
llvm_val->val = LLVMBuildFPExt(cl2llvm_builder,
original_val->val, LLVMDoubleType(), temp_var_name);
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromtype == LLVMDoubleType())
{
llvm_val->val = LLVMBuildFPTrunc(cl2llvm_builder,
original_val->val, LLVMFloatType(), temp_var_name);
}
else if (fromtype == LLVMHalfType())
{
llvm_val->val = LLVMBuildFPExt(cl2llvm_builder,
original_val->val, LLVMFloatType(), temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
llvm_val->val = LLVMBuildFPTrunc(cl2llvm_builder,
original_val->val, LLVMHalfType(), temp_var_name);
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
cl2llvmTypeWrapSetLlvmType(llvm_val->type, totype);
cl2llvmTypeWrapSetSign(llvm_val->type, tosign);
return llvm_val;
}
