/**
* Build a manual selection sequence for cube face sc/tc coordinates and
* major axis vector (multiplied by 2 for consistency) for the given
* vec3 \p coords, for the face implied by \p selcoords.
*
* For the major axis, we always adjust the sign to be in the direction of
* selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
* the selcoords major axis.
*/
static void build_cube_select(LLVMBuilderRef builder,
const struct cube_selection_coords *selcoords,
const LLVMValueRef *coords,
LLVMValueRef *out_st,
LLVMValueRef *out_ma)
{
LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
LLVMValueRef is_ma_positive;
LLVMValueRef sgn_ma;
LLVMValueRef is_ma_z, is_not_ma_z;
LLVMValueRef is_ma_y;
LLVMValueRef is_ma_x;
LLVMValueRef sgn;
LLVMValueRef tmp;
is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
selcoords->ma, LLVMConstReal(f32, 0.0), "");
sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
/* Select sc */
tmp = LLVMBuildSelect(builder, is_ma_z, coords[2], coords[0], "");
sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
LLVMBuildSelect(builder, is_ma_x, sgn_ma,
LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
/* Select tc */
tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
sgn = LLVMBuildSelect(builder, is_ma_y, LLVMBuildFNeg(builder, sgn_ma, ""),
LLVMConstReal(f32, -1.0), "");
out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
/* Select ma */
tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
sgn = LLVMBuildSelect(builder, is_ma_positive,
LLVMConstReal(f32, 2.0), LLVMConstReal(f32, -2.0), "");
*out_ma = LLVMBuildFMul(builder, tmp, sgn, "");
}
/* Perform front/back face culling and return true if the primitive is accepted. */
static LLVMValueRef ac_cull_face(struct ac_llvm_context *ctx,
LLVMValueRef pos[3][4],
struct ac_position_w_info *w,
bool cull_front,
bool cull_back,
bool cull_zero_area)
{
LLVMBuilderRef builder = ctx->builder;
if (cull_front && cull_back)
return ctx->i1false;
if (!cull_front && !cull_back && !cull_zero_area)
return ctx->i1true;
/* Front/back face culling. Also if the determinant == 0, the triangle
* area is 0.
*/
LLVMValueRef det_t0 = LLVMBuildFSub(builder, pos[2][0], pos[0][0], "");
LLVMValueRef det_t1 = LLVMBuildFSub(builder, pos[1][1], pos[0][1], "");
LLVMValueRef det_t2 = LLVMBuildFSub(builder, pos[0][0], pos[1][0], "");
LLVMValueRef det_t3 = LLVMBuildFSub(builder, pos[0][1], pos[2][1], "");
LLVMValueRef det_p0 = LLVMBuildFMul(builder, det_t0, det_t1, "");
LLVMValueRef det_p1 = LLVMBuildFMul(builder, det_t2, det_t3, "");
LLVMValueRef det = LLVMBuildFSub(builder, det_p0, det_p1, "");
/* Negative W negates the determinant. */
det = LLVMBuildSelect(builder, w->w_reflection,
LLVMBuildFNeg(builder, det, ""),
det, "");
LLVMValueRef accepted = NULL;
if (cull_front) {
LLVMRealPredicate cond = cull_zero_area ? LLVMRealOGT : LLVMRealOGE;
accepted = LLVMBuildFCmp(builder, cond, det, ctx->f32_0, "");
} else if (cull_back) {
LLVMRealPredicate cond = cull_zero_area ? LLVMRealOLT : LLVMRealOLE;
accepted = LLVMBuildFCmp(builder, cond, det, ctx->f32_0, "");
} else if (cull_zero_area) {
accepted = LLVMBuildFCmp(builder, LLVMRealONE, det, ctx->f32_0, "");
}
return accepted;
}
/**
* Inverse of lp_build_clamped_float_to_unsigned_norm above.
* Ex: src = { i32, i32, i32, i32 } with values in range [0, 2^src_width-1]
* return {float, float, float, float} with values in range [0, 1].
*/
LLVMValueRef
lp_build_unsigned_norm_to_float(LLVMBuilderRef builder,
unsigned src_width,
struct lp_type dst_type,
LLVMValueRef src)
{
LLVMTypeRef vec_type = lp_build_vec_type(dst_type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(dst_type);
LLVMValueRef bias_;
LLVMValueRef res;
unsigned mantissa;
unsigned n;
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
assert(dst_type.floating);
mantissa = lp_mantissa(dst_type);
n = MIN2(mantissa, src_width);
ubound = ((unsigned long long)1 << n);
mask = ubound - 1;
scale = (double)ubound/mask;
bias = (double)((unsigned long long)1 << (mantissa - n));
res = src;
if(src_width > mantissa) {
int shift = src_width - mantissa;
res = LLVMBuildLShr(builder, res, lp_build_const_int_vec(dst_type, shift), "");
}
bias_ = lp_build_const_vec(dst_type, bias);
res = LLVMBuildOr(builder,
res,
LLVMBuildBitCast(builder, bias_, int_vec_type, ""), "");
res = LLVMBuildBitCast(builder, res, vec_type, "");
res = LLVMBuildFSub(builder, res, bias_, "");
res = LLVMBuildFMul(builder, res, lp_build_const_vec(dst_type, scale), "");
return res;
}
static LLVMValueRef
translateFloatBinOp(NodeKind Op, LLVMValueRef ValueE1, LLVMValueRef ValueE2) {
switch (Op) {
case SumOp: return LLVMBuildFAdd(Builder, ValueE1, ValueE2, "");
case SubOp: return LLVMBuildFSub(Builder, ValueE1, ValueE2, "");
case MultOp: return LLVMBuildFMul(Builder, ValueE1, ValueE2, "");
case DivOp: return LLVMBuildFDiv(Builder, ValueE1, ValueE2, "");
case LtOp: return LLVMBuildFCmp(Builder, LLVMRealOLT, ValueE1, ValueE2, "");
case LeOp: return LLVMBuildFCmp(Builder, LLVMRealOLE, ValueE1, ValueE2, "");
case GtOp: return LLVMBuildFCmp(Builder, LLVMRealOGT, ValueE1, ValueE2, "");
case GeOp: return LLVMBuildFCmp(Builder, LLVMRealOGE, ValueE1, ValueE2, "");
case EqOp: return LLVMBuildFCmp(Builder, LLVMRealOEQ, ValueE1, ValueE2, "");
case DiffOp: return LLVMBuildFCmp(Builder, LLVMRealONE, ValueE1, ValueE2, "");
default: return NULL;
}
}
/**
* Converts int16 half-float to float32
* Note this can be performed in 1 instruction if vcvtph2ps exists (sse5 i think?)
* [llvm.x86.vcvtph2ps / _mm_cvtph_ps]
*
* @param src_type <vector> type of int16
* @param src value to convert
*
* ref http://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/
*/
LLVMValueRef
lp_build_half_to_float(struct gallivm_state *gallivm,
struct lp_type src_type,
LLVMValueRef src)
{
struct lp_type f32_type = lp_type_float_vec(32, 32 * src_type.length);
struct lp_type i32_type = lp_type_int_vec(32, 32 * src_type.length);
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
LLVMTypeRef float_vec_type = lp_build_vec_type(gallivm, f32_type);
/* Constants */
LLVMValueRef i32_13 = lp_build_const_int_vec(gallivm, i32_type, 13);
LLVMValueRef i32_16 = lp_build_const_int_vec(gallivm, i32_type, 16);
LLVMValueRef i32_mask_nosign = lp_build_const_int_vec(gallivm, i32_type, 0x7fff);
LLVMValueRef i32_was_infnan = lp_build_const_int_vec(gallivm, i32_type, 0x7bff);
LLVMValueRef i32_exp_infnan = lp_build_const_int_vec(gallivm, i32_type, 0xff << 23);
LLVMValueRef f32_magic = LLVMBuildBitCast(builder,
lp_build_const_int_vec(gallivm, i32_type, (254 - 15) << 23),
float_vec_type, "");
/* Convert int16 vector to int32 vector by zero ext */
LLVMValueRef h = LLVMBuildZExt(builder, src, int_vec_type, "");
/* Exponent / mantissa bits */
LLVMValueRef expmant = LLVMBuildAnd(builder, i32_mask_nosign, h, "");
LLVMValueRef shifted = LLVMBuildBitCast(builder, LLVMBuildShl(builder, expmant, i32_13, ""), float_vec_type, "");
/* Exponent adjust */
LLVMValueRef scaled = LLVMBuildBitCast(builder, LLVMBuildFMul(builder, shifted, f32_magic, ""), int_vec_type, "");
/* Make sure Inf/NaN survive */
LLVMValueRef b_wasinfnan = lp_build_compare(gallivm, i32_type, PIPE_FUNC_GREATER, expmant, i32_was_infnan);
LLVMValueRef infnanexp = LLVMBuildAnd(builder, b_wasinfnan, i32_exp_infnan, "");
/* Sign bit */
LLVMValueRef justsign = LLVMBuildXor(builder, h, expmant, "");
LLVMValueRef sign = LLVMBuildShl(builder, justsign, i32_16, "");
/* Combine result */
LLVMValueRef sign_inf = LLVMBuildOr(builder, sign, infnanexp, "");
LLVMValueRef final = LLVMBuildOr(builder, scaled, sign_inf, "");
/* Cast from int32 vector to float32 vector */
return LLVMBuildBitCast(builder, final, float_vec_type, "");
}
/**
* Unpack several pixels in SoA.
*
* It takes a vector of packed pixels:
*
* packed = {P0, P1, P2, P3, ..., Pn}
*
* And will produce four vectors:
*
* red = {R0, R1, R2, R3, ..., Rn}
* green = {G0, G1, G2, G3, ..., Gn}
* blue = {B0, B1, B2, B3, ..., Bn}
* alpha = {A0, A1, A2, A3, ..., An}
*
* It requires that a packed pixel fits into an element of the output
* channels. The common case is when converting pixel with a depth of 32 bit or
* less into floats.
*
* \param format_desc the format of the 'packed' incoming pixel vector
* \param type the desired type for rgba_out (type.length = n, above)
* \param packed the incoming vector of packed pixels
* \param rgba_out returns the SoA R,G,B,A vectors
*/
void
lp_build_unpack_rgba_soa(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
LLVMValueRef packed,
LLVMValueRef rgba_out[4])
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_build_context bld;
LLVMValueRef inputs[4];
unsigned chan;
assert(format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
assert(format_desc->block.width == 1);
assert(format_desc->block.height == 1);
assert(format_desc->block.bits <= type.width);
/* FIXME: Support more output types */
assert(type.width == 32);
lp_build_context_init(&bld, gallivm, type);
/* Decode the input vector components */
for (chan = 0; chan < format_desc->nr_channels; ++chan) {
const unsigned width = format_desc->channel[chan].size;
const unsigned start = format_desc->channel[chan].shift;
const unsigned stop = start + width;
LLVMValueRef input;
input = packed;
switch(format_desc->channel[chan].type) {
case UTIL_FORMAT_TYPE_VOID:
input = lp_build_undef(gallivm, type);
break;
case UTIL_FORMAT_TYPE_UNSIGNED:
/*
* Align the LSB
*/
if (start) {
input = LLVMBuildLShr(builder, input, lp_build_const_int_vec(gallivm, type, start), "");
}
/*
* Zero the MSBs
*/
if (stop < format_desc->block.bits) {
unsigned mask = ((unsigned long long)1 << width) - 1;
input = LLVMBuildAnd(builder, input, lp_build_const_int_vec(gallivm, type, mask), "");
}
/*
* Type conversion
*/
if (type.floating) {
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
assert(width == 8);
if (format_desc->swizzle[3] == chan) {
input = lp_build_unsigned_norm_to_float(gallivm, width, type, input);
}
else {
struct lp_type conv_type = lp_uint_type(type);
input = lp_build_srgb_to_linear(gallivm, conv_type, input);
}
}
else {
if(format_desc->channel[chan].normalized)
input = lp_build_unsigned_norm_to_float(gallivm, width, type, input);
else
input = LLVMBuildSIToFP(builder, input,
lp_build_vec_type(gallivm, type), "");
}
}
else if (format_desc->channel[chan].pure_integer) {
/* Nothing to do */
} else {
/* FIXME */
assert(0);
}
break;
case UTIL_FORMAT_TYPE_SIGNED:
/*
* Align the sign bit first.
*/
if (stop < type.width) {
unsigned bits = type.width - stop;
LLVMValueRef bits_val = lp_build_const_int_vec(gallivm, type, bits);
input = LLVMBuildShl(builder, input, bits_val, "");
}
/*
* Align the LSB (with an arithmetic shift to preserve the sign)
*/
if (format_desc->channel[chan].size < type.width) {
unsigned bits = type.width - format_desc->channel[chan].size;
LLVMValueRef bits_val = lp_build_const_int_vec(gallivm, type, bits);
input = LLVMBuildAShr(builder, input, bits_val, "");
}
/*
* Type conversion
*/
if (type.floating) {
input = LLVMBuildSIToFP(builder, input, lp_build_vec_type(gallivm, type), "");
if (format_desc->channel[chan].normalized) {
double scale = 1.0 / ((1 << (format_desc->channel[chan].size - 1)) - 1);
LLVMValueRef scale_val = lp_build_const_vec(gallivm, type, scale);
input = LLVMBuildFMul(builder, input, scale_val, "");
/* the formula above will produce value below -1.0 for most negative
* value but everything seems happy with that hence disable for now */
if (0)
input = lp_build_max(&bld, input,
lp_build_const_vec(gallivm, type, -1.0f));
}
}
else if (format_desc->channel[chan].pure_integer) {
/* Nothing to do */
} else {
/* FIXME */
assert(0);
}
break;
case UTIL_FORMAT_TYPE_FLOAT:
if (type.floating) {
assert(start == 0);
assert(stop == 32);
assert(type.width == 32);
input = LLVMBuildBitCast(builder, input, lp_build_vec_type(gallivm, type), "");
}
else {
/* FIXME */
assert(0);
input = lp_build_undef(gallivm, type);
}
break;
case UTIL_FORMAT_TYPE_FIXED:
if (type.floating) {
double scale = 1.0 / ((1 << (format_desc->channel[chan].size/2)) - 1);
LLVMValueRef scale_val = lp_build_const_vec(gallivm, type, scale);
input = LLVMBuildSIToFP(builder, input, lp_build_vec_type(gallivm, type), "");
input = LLVMBuildFMul(builder, input, scale_val, "");
}
else {
/* FIXME */
assert(0);
input = lp_build_undef(gallivm, type);
}
break;
default:
assert(0);
input = lp_build_undef(gallivm, type);
break;
}
inputs[chan] = input;
}
lp_build_format_swizzle_soa(format_desc, &bld, inputs, rgba_out);
}
/**
* Generic type conversion.
*
* TODO: Take a precision argument, or even better, add a new precision member
* to the lp_type union.
*/
void
lp_build_conv(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
const LLVMValueRef *src, unsigned num_srcs,
LLVMValueRef *dst, unsigned num_dsts)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_type tmp_type;
LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
unsigned num_tmps;
unsigned i;
/* We must not loose or gain channels. Only precision */
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
assert(num_srcs <= LP_MAX_VECTOR_LENGTH);
assert(num_dsts <= LP_MAX_VECTOR_LENGTH);
tmp_type = src_type;
for(i = 0; i < num_srcs; ++i) {
assert(lp_check_value(src_type, src[i]));
tmp[i] = src[i];
}
num_tmps = num_srcs;
/* Special case 4x4f --> 1x16ub
*/
if (src_type.floating == 1 &&
src_type.fixed == 0 &&
src_type.sign == 1 &&
src_type.norm == 0 &&
src_type.width == 32 &&
src_type.length == 4 &&
dst_type.floating == 0 &&
dst_type.fixed == 0 &&
dst_type.sign == 0 &&
dst_type.norm == 1 &&
dst_type.width == 8 &&
dst_type.length == 16 &&
4 * num_dsts == num_srcs &&
util_cpu_caps.has_sse2)
{
struct lp_build_context bld;
struct lp_type int16_type = dst_type;
struct lp_type int32_type = dst_type;
LLVMValueRef const_255f;
unsigned i, j;
lp_build_context_init(&bld, gallivm, src_type);
int16_type.width *= 2;
int16_type.length /= 2;
int16_type.sign = 1;
int32_type.width *= 4;
int32_type.length /= 4;
int32_type.sign = 1;
const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);
for (i = 0; i < num_dsts; ++i, src += 4) {
LLVMValueRef lo, hi;
for (j = 0; j < 4; ++j) {
tmp[j] = LLVMBuildFMul(builder, src[j], const_255f, "");
tmp[j] = lp_build_iround(&bld, tmp[j]);
}
/* relying on clamping behavior of sse2 intrinsics here */
lo = lp_build_pack2(gallivm, int32_type, int16_type, tmp[0], tmp[1]);
hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
}
return;
}
/* Special case 2x8f --> 1x16ub
*/
else if (src_type.floating == 1 &&
src_type.fixed == 0 &&
src_type.sign == 1 &&
src_type.norm == 0 &&
src_type.width == 32 &&
src_type.length == 8 &&
dst_type.floating == 0 &&
dst_type.fixed == 0 &&
dst_type.sign == 0 &&
dst_type.norm == 1 &&
dst_type.width == 8 &&
dst_type.length == 16 &&
2 * num_dsts == num_srcs &&
util_cpu_caps.has_avx) {
struct lp_build_context bld;
struct lp_type int16_type = dst_type;
struct lp_type int32_type = dst_type;
LLVMValueRef const_255f;
unsigned i;
lp_build_context_init(&bld, gallivm, src_type);
int16_type.width *= 2;
int16_type.length /= 2;
int16_type.sign = 1;
int32_type.width *= 4;
int32_type.length /= 4;
int32_type.sign = 1;
const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);
for (i = 0; i < num_dsts; ++i, src += 2) {
LLVMValueRef lo, hi, a, b;
a = LLVMBuildFMul(builder, src[0], const_255f, "");
b = LLVMBuildFMul(builder, src[1], const_255f, "");
a = lp_build_iround(&bld, a);
b = lp_build_iround(&bld, b);
tmp[0] = lp_build_extract_range(gallivm, a, 0, 4);
tmp[1] = lp_build_extract_range(gallivm, a, 4, 4);
tmp[2] = lp_build_extract_range(gallivm, b, 0, 4);
tmp[3] = lp_build_extract_range(gallivm, b, 4, 4);
/* relying on clamping behavior of sse2 intrinsics here */
lo = lp_build_pack2(gallivm, int32_type, int16_type, tmp[0], tmp[1]);
hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
}
return;
}
/* Pre convert half-floats to floats
*/
else if (src_type.floating && src_type.width == 16)
{
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_half_to_float(gallivm, src_type, tmp[i]);
tmp_type.width = 32;
}
/*
* Clamp if necessary
*/
if(memcmp(&src_type, &dst_type, sizeof src_type) != 0) {
struct lp_build_context bld;
double src_min = lp_const_min(src_type);
double dst_min = lp_const_min(dst_type);
double src_max = lp_const_max(src_type);
double dst_max = lp_const_max(dst_type);
LLVMValueRef thres;
lp_build_context_init(&bld, gallivm, tmp_type);
if(src_min < dst_min) {
if(dst_min == 0.0)
thres = bld.zero;
else
thres = lp_build_const_vec(gallivm, src_type, dst_min);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_max(&bld, tmp[i], thres);
}
if(src_max > dst_max) {
if(dst_max == 1.0)
thres = bld.one;
else
thres = lp_build_const_vec(gallivm, src_type, dst_max);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_min(&bld, tmp[i], thres);
}
}
/*
* Scale to the narrowest range
*/
if(dst_type.floating) {
/* Nothing to do */
}
else if(tmp_type.floating) {
if(!dst_type.fixed && !dst_type.sign && dst_type.norm) {
for(i = 0; i < num_tmps; ++i) {
tmp[i] = lp_build_clamped_float_to_unsigned_norm(gallivm,
tmp_type,
dst_type.width,
tmp[i]);
}
tmp_type.floating = FALSE;
}
else {
double dst_scale = lp_const_scale(dst_type);
LLVMTypeRef tmp_vec_type;
if (dst_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, dst_scale);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = FALSE;
tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
else
tmp[i] = LLVMBuildFPToUI(builder, tmp[i], tmp_vec_type, "");
#else
/* FIXME: there is no SSE counterpart for LLVMBuildFPToUI */
tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
#endif
}
}
}
else {
unsigned src_shift = lp_const_shift(src_type);
unsigned dst_shift = lp_const_shift(dst_type);
unsigned src_offset = lp_const_offset(src_type);
unsigned dst_offset = lp_const_offset(dst_type);
/* Compensate for different offsets */
if (dst_offset > src_offset && src_type.width > dst_type.width) {
for (i = 0; i < num_tmps; ++i) {
LLVMValueRef shifted;
LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, src_shift - 1);
if(src_type.sign)
shifted = LLVMBuildAShr(builder, tmp[i], shift, "");
else
shifted = LLVMBuildLShr(builder, tmp[i], shift, "");
tmp[i] = LLVMBuildSub(builder, tmp[i], shifted, "");
}
}
if(src_shift > dst_shift) {
LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type,
src_shift - dst_shift);
for(i = 0; i < num_tmps; ++i)
if(src_type.sign)
tmp[i] = LLVMBuildAShr(builder, tmp[i], shift, "");
else
tmp[i] = LLVMBuildLShr(builder, tmp[i], shift, "");
}
}
/*
* Truncate or expand bit width
*
* No data conversion should happen here, although the sign bits are
* crucial to avoid bad clamping.
*/
{
struct lp_type new_type;
new_type = tmp_type;
new_type.sign = dst_type.sign;
new_type.width = dst_type.width;
new_type.length = dst_type.length;
lp_build_resize(gallivm, tmp_type, new_type, tmp, num_srcs, tmp, num_dsts);
tmp_type = new_type;
num_tmps = num_dsts;
}
/*
* Scale to the widest range
*/
if(src_type.floating) {
/* Nothing to do */
}
else if(!src_type.floating && dst_type.floating) {
if(!src_type.fixed && !src_type.sign && src_type.norm) {
for(i = 0; i < num_tmps; ++i) {
tmp[i] = lp_build_unsigned_norm_to_float(gallivm,
src_type.width,
dst_type,
tmp[i]);
}
tmp_type.floating = TRUE;
}
else {
double src_scale = lp_const_scale(src_type);
LLVMTypeRef tmp_vec_type;
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = TRUE;
tmp_type.sign = TRUE;
tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
else
tmp[i] = LLVMBuildUIToFP(builder, tmp[i], tmp_vec_type, "");
#else
/* FIXME: there is no SSE counterpart for LLVMBuildUIToFP */
tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
#endif
}
if (src_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, 1.0/src_scale);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
}
}
else {
unsigned src_shift = lp_const_shift(src_type);
unsigned dst_shift = lp_const_shift(dst_type);
unsigned src_offset = lp_const_offset(src_type);
unsigned dst_offset = lp_const_offset(dst_type);
if (src_shift < dst_shift) {
LLVMValueRef pre_shift[LP_MAX_VECTOR_LENGTH];
LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, dst_shift - src_shift);
for (i = 0; i < num_tmps; ++i) {
pre_shift[i] = tmp[i];
tmp[i] = LLVMBuildShl(builder, tmp[i], shift, "");
}
/* Compensate for different offsets */
if (dst_offset > src_offset) {
for (i = 0; i < num_tmps; ++i) {
tmp[i] = LLVMBuildSub(builder, tmp[i], pre_shift[i], "");
}
}
}
}
for(i = 0; i < num_dsts; ++i) {
dst[i] = tmp[i];
assert(lp_check_value(dst_type, dst[i]));
}
}
/**
* Inverse of lp_build_clamped_float_to_unsigned_norm above.
* Ex: src = { i32, i32, i32, i32 } with values in range [0, 2^src_width-1]
* return {float, float, float, float} with values in range [0, 1].
*/
LLVMValueRef
lp_build_unsigned_norm_to_float(struct gallivm_state *gallivm,
unsigned src_width,
struct lp_type dst_type,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, dst_type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, dst_type);
LLVMValueRef bias_;
LLVMValueRef res;
unsigned mantissa;
unsigned n;
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
assert(dst_type.floating);
mantissa = lp_mantissa(dst_type);
if (src_width <= (mantissa + 1)) {
/*
* The source width matches fits what can be represented in floating
* point (i.e., mantissa + 1 bits). So do a straight multiplication
* followed by casting. No further rounding is necessary.
*/
scale = 1.0/(double)((1ULL << src_width) - 1);
res = LLVMBuildSIToFP(builder, src, vec_type, "");
res = LLVMBuildFMul(builder, res,
lp_build_const_vec(gallivm, dst_type, scale), "");
return res;
}
else {
/*
* The source width exceeds what can be represented in floating
* point. So truncate the incoming values.
*/
n = MIN2(mantissa, src_width);
ubound = ((unsigned long long)1 << n);
mask = ubound - 1;
scale = (double)ubound/mask;
bias = (double)((unsigned long long)1 << (mantissa - n));
res = src;
if (src_width > mantissa) {
int shift = src_width - mantissa;
res = LLVMBuildLShr(builder, res,
lp_build_const_int_vec(gallivm, dst_type, shift), "");
}
bias_ = lp_build_const_vec(gallivm, dst_type, bias);
res = LLVMBuildOr(builder,
res,
LLVMBuildBitCast(builder, bias_, int_vec_type, ""), "");
res = LLVMBuildBitCast(builder, res, vec_type, "");
res = LLVMBuildFSub(builder, res, bias_, "");
res = LLVMBuildFMul(builder, res, lp_build_const_vec(gallivm, dst_type, scale), "");
}
return res;
}
/**
* Special case for converting clamped IEEE-754 floats to unsigned norms.
*
* The mathematical voodoo below may seem excessive but it is actually
* paramount we do it this way for several reasons. First, there is no single
* precision FP to unsigned integer conversion Intel SSE instruction. Second,
* secondly, even if there was, since the FP's mantissa takes only a fraction
* of register bits the typically scale and cast approach would require double
* precision for accurate results, and therefore half the throughput
*
* Although the result values can be scaled to an arbitrary bit width specified
* by dst_width, the actual result type will have the same width.
*
* Ex: src = { float, float, float, float }
* return { i32, i32, i32, i32 } where each value is in [0, 2^dst_width-1].
*/
LLVMValueRef
lp_build_clamped_float_to_unsigned_norm(struct gallivm_state *gallivm,
struct lp_type src_type,
unsigned dst_width,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, src_type);
LLVMValueRef res;
unsigned mantissa;
assert(src_type.floating);
assert(dst_width <= src_type.width);
src_type.sign = FALSE;
mantissa = lp_mantissa(src_type);
if (dst_width <= mantissa) {
/*
* Apply magic coefficients that will make the desired result to appear
* in the lowest significant bits of the mantissa, with correct rounding.
*
* This only works if the destination width fits in the mantissa.
*/
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
ubound = (1ULL << dst_width);
mask = ubound - 1;
scale = (double)mask/ubound;
bias = (double)(1ULL << (mantissa - dst_width));
res = LLVMBuildFMul(builder, src, lp_build_const_vec(gallivm, src_type, scale), "");
res = LLVMBuildFAdd(builder, res, lp_build_const_vec(gallivm, src_type, bias), "");
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
res = LLVMBuildAnd(builder, res,
lp_build_const_int_vec(gallivm, src_type, mask), "");
}
else if (dst_width == (mantissa + 1)) {
/*
* The destination width matches exactly what can be represented in
* floating point (i.e., mantissa + 1 bits). So do a straight
* multiplication followed by casting. No further rounding is necessary.
*/
double scale;
scale = (double)((1ULL << dst_width) - 1);
res = LLVMBuildFMul(builder, src,
lp_build_const_vec(gallivm, src_type, scale), "");
res = LLVMBuildFPToSI(builder, res, int_vec_type, "");
}
else {
/*
* The destination exceeds what can be represented in the floating point.
* So multiply by the largest power two we get away with, and when
* subtract the most significant bit to rescale to normalized values.
*
* The largest power of two factor we can get away is
* (1 << (src_type.width - 1)), because we need to use signed . In theory it
* should be (1 << (src_type.width - 2)), but IEEE 754 rules states
* INT_MIN should be returned in FPToSI, which is the correct result for
* values near 1.0!
*
* This means we get (src_type.width - 1) correct bits for values near 0.0,
* and (mantissa + 1) correct bits for values near 1.0. Equally or more
* important, we also get exact results for 0.0 and 1.0.
*/
unsigned n = MIN2(src_type.width - 1, dst_width);
double scale = (double)(1ULL << n);
unsigned lshift = dst_width - n;
unsigned rshift = n;
LLVMValueRef lshifted;
LLVMValueRef rshifted;
res = LLVMBuildFMul(builder, src,
lp_build_const_vec(gallivm, src_type, scale), "");
res = LLVMBuildFPToSI(builder, res, int_vec_type, "");
/*
* Align the most significant bit to its final place.
*
* This will cause 1.0 to overflow to 0, but the later adjustment will
* get it right.
*/
if (lshift) {
lshifted = LLVMBuildShl(builder, res,
lp_build_const_int_vec(gallivm, src_type,
lshift), "");
} else {
lshifted = res;
}
/*
* Align the most significant bit to the right.
*/
rshifted = LLVMBuildLShr(builder, res,
lp_build_const_int_vec(gallivm, src_type, rshift),
"");
/*
* Subtract the MSB to the LSB, therefore re-scaling from
* (1 << dst_width) to ((1 << dst_width) - 1).
*/
res = LLVMBuildSub(builder, lshifted, rshifted, "");
}
return res;
}
/**
* Initialize the bld->a, dadq fields. This involves fetching
* those values from the arrays which are passed into the JIT function.
*/
static void
coeffs_init(struct lp_build_interp_soa_context *bld,
LLVMValueRef a0_ptr,
LLVMValueRef dadx_ptr,
LLVMValueRef dady_ptr)
{
struct lp_build_context *coeff_bld = &bld->coeff_bld;
struct lp_build_context *setup_bld = &bld->setup_bld;
struct gallivm_state *gallivm = coeff_bld->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef pixoffx, pixoffy;
unsigned attrib;
unsigned chan;
unsigned i;
pixoffx = coeff_bld->undef;
pixoffy = coeff_bld->undef;
for (i = 0; i < coeff_bld->type.length; i++) {
LLVMValueRef nr = lp_build_const_int32(gallivm, i);
LLVMValueRef pixxf = lp_build_const_float(gallivm, quad_offset_x[i]);
LLVMValueRef pixyf = lp_build_const_float(gallivm, quad_offset_y[i]);
pixoffx = LLVMBuildInsertElement(builder, pixoffx, pixxf, nr, "");
pixoffy = LLVMBuildInsertElement(builder, pixoffy, pixyf, nr, "");
}
for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
const unsigned mask = bld->mask[attrib];
const unsigned interp = bld->interp[attrib];
LLVMValueRef index = lp_build_const_int32(gallivm,
attrib * TGSI_NUM_CHANNELS);
LLVMValueRef ptr;
LLVMValueRef dadxaos = setup_bld->zero;
LLVMValueRef dadyaos = setup_bld->zero;
LLVMValueRef a0aos = setup_bld->zero;
/* always fetch all 4 values for performance/simplicity */
switch (interp) {
case LP_INTERP_PERSPECTIVE:
/* fall-through */
case LP_INTERP_LINEAR:
ptr = LLVMBuildGEP(builder, dadx_ptr, &index, 1, "");
ptr = LLVMBuildBitCast(builder, ptr,
LLVMPointerType(setup_bld->vec_type, 0), "");
dadxaos = LLVMBuildLoad(builder, ptr, "");
ptr = LLVMBuildGEP(builder, dady_ptr, &index, 1, "");
ptr = LLVMBuildBitCast(builder, ptr,
LLVMPointerType(setup_bld->vec_type, 0), "");
dadyaos = LLVMBuildLoad(builder, ptr, "");
attrib_name(dadxaos, attrib, 0, ".dadxaos");
attrib_name(dadyaos, attrib, 0, ".dadyaos");
/* fall-through */
case LP_INTERP_CONSTANT:
case LP_INTERP_FACING:
ptr = LLVMBuildGEP(builder, a0_ptr, &index, 1, "");
ptr = LLVMBuildBitCast(builder, ptr,
LLVMPointerType(setup_bld->vec_type, 0), "");
a0aos = LLVMBuildLoad(builder, ptr, "");
attrib_name(a0aos, attrib, 0, ".a0aos");
break;
case LP_INTERP_POSITION:
/* Nothing to do as the position coeffs are already setup in slot 0 */
continue;
default:
assert(0);
break;
}
/*
* a = a0 + (x * dadx + y * dady)
* a0aos is the attrib value at top left corner of stamp
*/
if (interp != LP_INTERP_CONSTANT &&
interp != LP_INTERP_FACING) {
LLVMValueRef x = lp_build_broadcast_scalar(setup_bld, bld->x);
LLVMValueRef y = lp_build_broadcast_scalar(setup_bld, bld->y);
a0aos = lp_build_fmuladd(builder, x, dadxaos, a0aos);
a0aos = lp_build_fmuladd(builder, y, dadyaos, a0aos);
}
/*
* dadq = {0, dadx, dady, dadx + dady}
* for two quads (side by side) this is:
* {0, dadx, dady, dadx+dady, 2*dadx, 2*dadx+dady, 3*dadx+dady}
*/
for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
/* this generates a CRAPLOAD of shuffles... */
if (mask & (1 << chan)) {
LLVMValueRef dadx, dady;
LLVMValueRef dadq, dadq2;
LLVMValueRef a;
LLVMValueRef chan_index = lp_build_const_int32(gallivm, chan);
if (attrib == 0 && chan == 0) {
a = bld->x;
if (bld->pos_offset) {
a = LLVMBuildFAdd(builder, a, lp_build_const_float(gallivm, bld->pos_offset), "");
}
a = lp_build_broadcast_scalar(coeff_bld, a);
dadx = coeff_bld->one;
dady = coeff_bld->zero;
}
else if (attrib == 0 && chan == 1) {
a = bld->y;
if (bld->pos_offset) {
a = LLVMBuildFAdd(builder, a, lp_build_const_float(gallivm, bld->pos_offset), "");
}
a = lp_build_broadcast_scalar(coeff_bld, a);
dady = coeff_bld->one;
dadx = coeff_bld->zero;
}
else {
dadx = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, dadxaos, chan_index);
dady = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, dadyaos, chan_index);
/*
* a = {a, a, a, a}
*/
a = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, a0aos, chan_index);
}
dadx = LLVMBuildFMul(builder, dadx, pixoffx, "");
dady = LLVMBuildFMul(builder, dady, pixoffy, "");
dadq = LLVMBuildFAdd(builder, dadx, dady, "");
/*
* Compute the attrib values on the upper-left corner of each
* group of quads.
* Note that if we process 2 quads at once this doesn't
* really exactly to what we want.
* We need to access elem 0 and 2 respectively later if we process
* 2 quads at once.
*/
if (interp != LP_INTERP_CONSTANT &&
interp != LP_INTERP_FACING) {
dadq2 = LLVMBuildFAdd(builder, dadq, dadq, "");
a = LLVMBuildFAdd(builder, a, dadq2, "");
}
#if PERSPECTIVE_DIVIDE_PER_QUAD
/*
* a *= 1 / w
*/
/*
* XXX since we're only going to access elements 0,2 out of 8
* if we have 8-wide vectors we should do the division only 4-wide.
* a is really a 2-elements in a 4-wide vector disguised as 8-wide
* in this case.
*/
if (interp == LP_INTERP_PERSPECTIVE) {
LLVMValueRef w = bld->a[0][3];
assert(attrib != 0);
assert(bld->mask[0] & TGSI_WRITEMASK_W);
if (!bld->oow) {
bld->oow = lp_build_rcp(coeff_bld, w);
lp_build_name(bld->oow, "oow");
}
a = lp_build_mul(coeff_bld, a, bld->oow);
}
#endif
attrib_name(a, attrib, chan, ".a");
attrib_name(dadq, attrib, chan, ".dadq");
bld->a[attrib][chan] = lp_build_alloca(gallivm,
LLVMTypeOf(a), "");
LLVMBuildStore(builder, a, bld->a[attrib][chan]);
bld->dadq[attrib][chan] = dadq;
}
}
}
}
/**
* 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);
}
}
/**
* Special case for converting clamped IEEE-754 floats to unsigned norms.
*
* The mathematical voodoo below may seem excessive but it is actually
* paramount we do it this way for several reasons. First, there is no single
* precision FP to unsigned integer conversion Intel SSE instruction. Second,
* secondly, even if there was, since the FP's mantissa takes only a fraction
* of register bits the typically scale and cast approach would require double
* precision for accurate results, and therefore half the throughput
*
* Although the result values can be scaled to an arbitrary bit width specified
* by dst_width, the actual result type will have the same width.
*
* Ex: src = { float, float, float, float }
* return { i32, i32, i32, i32 } where each value is in [0, 2^dst_width-1].
*/
LLVMValueRef
lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
struct lp_type src_type,
unsigned dst_width,
LLVMValueRef src)
{
LLVMTypeRef int_vec_type = lp_build_int_vec_type(src_type);
LLVMValueRef res;
unsigned mantissa;
unsigned n;
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
assert(src_type.floating);
mantissa = lp_mantissa(src_type);
/* We cannot carry more bits than the mantissa */
n = MIN2(mantissa, dst_width);
/* This magic coefficients will make the desired result to appear in the
* lowest significant bits of the mantissa.
*/
ubound = ((unsigned long long)1 << n);
mask = ubound - 1;
scale = (double)mask/ubound;
bias = (double)((unsigned long long)1 << (mantissa - n));
res = LLVMBuildFMul(builder, src, lp_build_const_vec(src_type, scale), "");
res = LLVMBuildFAdd(builder, res, lp_build_const_vec(src_type, bias), "");
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
if(dst_width > n) {
int shift = dst_width - n;
res = LLVMBuildShl(builder, res, lp_build_const_int_vec(src_type, shift), "");
/* TODO: Fill in the empty lower bits for additional precision? */
/* YES: this fixes progs/trivial/tri-z-eq.c.
* Otherwise vertex Z=1.0 values get converted to something like
* 0xfffffb00 and the test for equality with 0xffffffff fails.
*/
#if 0
{
LLVMValueRef msb;
msb = LLVMBuildLShr(builder, res, lp_build_const_int_vec(src_type, dst_width - 1), "");
msb = LLVMBuildShl(builder, msb, lp_build_const_int_vec(src_type, shift), "");
msb = LLVMBuildSub(builder, msb, lp_build_const_int_vec(src_type, 1), "");
res = LLVMBuildOr(builder, res, msb, "");
}
#elif 0
while(shift > 0) {
res = LLVMBuildOr(builder, res, LLVMBuildLShr(builder, res, lp_build_const_int_vec(src_type, n), ""), "");
shift -= n;
n *= 2;
}
#endif
}
else
res = LLVMBuildAnd(builder, res, lp_build_const_int_vec(src_type, mask), "");
return res;
}
/**
* Generic type conversion.
*
* TODO: Take a precision argument, or even better, add a new precision member
* to the lp_type union.
*/
void
lp_build_conv(LLVMBuilderRef builder,
struct lp_type src_type,
struct lp_type dst_type,
const LLVMValueRef *src, unsigned num_srcs,
LLVMValueRef *dst, unsigned num_dsts)
{
struct lp_type tmp_type;
LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
unsigned num_tmps;
unsigned i;
/* We must not loose or gain channels. Only precision */
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
assert(num_srcs <= LP_MAX_VECTOR_LENGTH);
assert(num_dsts <= LP_MAX_VECTOR_LENGTH);
tmp_type = src_type;
for(i = 0; i < num_srcs; ++i) {
assert(lp_check_value(src_type, src[i]));
tmp[i] = src[i];
}
num_tmps = num_srcs;
/*
* Clamp if necessary
*/
if(memcmp(&src_type, &dst_type, sizeof src_type) != 0) {
struct lp_build_context bld;
double src_min = lp_const_min(src_type);
double dst_min = lp_const_min(dst_type);
double src_max = lp_const_max(src_type);
double dst_max = lp_const_max(dst_type);
LLVMValueRef thres;
lp_build_context_init(&bld, builder, tmp_type);
if(src_min < dst_min) {
if(dst_min == 0.0)
thres = bld.zero;
else
thres = lp_build_const_vec(src_type, dst_min);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_max(&bld, tmp[i], thres);
}
if(src_max > dst_max) {
if(dst_max == 1.0)
thres = bld.one;
else
thres = lp_build_const_vec(src_type, dst_max);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_min(&bld, tmp[i], thres);
}
}
/*
* Scale to the narrowest range
*/
if(dst_type.floating) {
/* Nothing to do */
}
else if(tmp_type.floating) {
if(!dst_type.fixed && !dst_type.sign && dst_type.norm) {
for(i = 0; i < num_tmps; ++i) {
tmp[i] = lp_build_clamped_float_to_unsigned_norm(builder,
tmp_type,
dst_type.width,
tmp[i]);
}
tmp_type.floating = FALSE;
}
else {
double dst_scale = lp_const_scale(dst_type);
LLVMTypeRef tmp_vec_type;
if (dst_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(tmp_type, dst_scale);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = FALSE;
tmp_vec_type = lp_build_vec_type(tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
else
tmp[i] = LLVMBuildFPToUI(builder, tmp[i], tmp_vec_type, "");
#else
/* FIXME: there is no SSE counterpart for LLVMBuildFPToUI */
tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
#endif
}
}
}
else {
unsigned src_shift = lp_const_shift(src_type);
unsigned dst_shift = lp_const_shift(dst_type);
/* FIXME: compensate different offsets too */
if(src_shift > dst_shift) {
LLVMValueRef shift = lp_build_const_int_vec(tmp_type, src_shift - dst_shift);
for(i = 0; i < num_tmps; ++i)
if(src_type.sign)
tmp[i] = LLVMBuildAShr(builder, tmp[i], shift, "");
else
tmp[i] = LLVMBuildLShr(builder, tmp[i], shift, "");
}
}
/*
* Truncate or expand bit width
*
* No data conversion should happen here, although the sign bits are
* crucial to avoid bad clamping.
*/
{
struct lp_type new_type;
new_type = tmp_type;
new_type.sign = dst_type.sign;
new_type.width = dst_type.width;
new_type.length = dst_type.length;
lp_build_resize(builder, tmp_type, new_type, tmp, num_srcs, tmp, num_dsts);
tmp_type = new_type;
num_tmps = num_dsts;
}
/*
* Scale to the widest range
*/
if(src_type.floating) {
/* Nothing to do */
}
else if(!src_type.floating && dst_type.floating) {
if(!src_type.fixed && !src_type.sign && src_type.norm) {
for(i = 0; i < num_tmps; ++i) {
tmp[i] = lp_build_unsigned_norm_to_float(builder,
src_type.width,
dst_type,
tmp[i]);
}
tmp_type.floating = TRUE;
}
else {
double src_scale = lp_const_scale(src_type);
LLVMTypeRef tmp_vec_type;
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = TRUE;
tmp_type.sign = TRUE;
tmp_vec_type = lp_build_vec_type(tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
else
tmp[i] = LLVMBuildUIToFP(builder, tmp[i], tmp_vec_type, "");
#else
/* FIXME: there is no SSE counterpart for LLVMBuildUIToFP */
tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
#endif
}
if (src_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(tmp_type, 1.0/src_scale);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
}
}
else {
unsigned src_shift = lp_const_shift(src_type);
unsigned dst_shift = lp_const_shift(dst_type);
/* FIXME: compensate different offsets too */
if(src_shift < dst_shift) {
LLVMValueRef shift = lp_build_const_int_vec(tmp_type, dst_shift - src_shift);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildShl(builder, tmp[i], shift, "");
}
}
for(i = 0; i < num_dsts; ++i) {
dst[i] = tmp[i];
assert(lp_check_value(dst_type, dst[i]));
}
}
void
ac_prepare_cube_coords(struct ac_llvm_context *ctx,
bool is_deriv, bool is_array,
LLVMValueRef *coords_arg,
LLVMValueRef *derivs_arg)
{
LLVMBuilderRef builder = ctx->builder;
struct cube_selection_coords selcoords;
LLVMValueRef coords[3];
LLVMValueRef invma;
build_cube_intrinsic(ctx, coords_arg, &selcoords);
invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
for (int i = 0; i < 2; ++i)
coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
coords[2] = selcoords.id;
if (is_deriv && derivs_arg) {
LLVMValueRef derivs[4];
int axis;
/* Convert cube derivatives to 2D derivatives. */
for (axis = 0; axis < 2; axis++) {
LLVMValueRef deriv_st[2];
LLVMValueRef deriv_ma;
/* Transform the derivative alongside the texture
* coordinate. Mathematically, the correct formula is
* as follows. Assume we're projecting onto the +Z face
* and denote by dx/dh the derivative of the (original)
* X texture coordinate with respect to horizontal
* window coordinates. The projection onto the +Z face
* plane is:
*
* f(x,z) = x/z
*
* Then df/dh = df/dx * dx/dh + df/dz * dz/dh
* = 1/z * dx/dh - x/z * 1/z * dz/dh.
*
* This motivatives the implementation below.
*
* Whether this actually gives the expected results for
* apps that might feed in derivatives obtained via
* finite differences is anyone's guess. The OpenGL spec
* seems awfully quiet about how textureGrad for cube
* maps should be handled.
*/
build_cube_select(builder, &selcoords, &derivs_arg[axis * 3],
deriv_st, &deriv_ma);
deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
for (int i = 0; i < 2; ++i)
derivs[axis * 2 + i] =
LLVMBuildFSub(builder,
LLVMBuildFMul(builder, deriv_st[i], invma, ""),
LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
}
memcpy(derivs_arg, derivs, sizeof(derivs));
}
/* Shift the texture coordinate. This must be applied after the
* derivative calculation.
*/
for (int i = 0; i < 2; ++i)
coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
if (is_array) {
/* for cube arrays coord.z = coord.w(array_index) * 8 + face */
/* coords_arg.w component - array_index for cube arrays */
LLVMValueRef tmp = LLVMBuildFMul(ctx->builder, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), "");
coords[2] = LLVMBuildFAdd(ctx->builder, tmp, coords[2], "");
}
memcpy(coords_arg, coords, sizeof(coords));
}
/**
* Unpack a single pixel into its RGBA 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 RGBA 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 shift;
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;
shift = 0;
/* Loop over 4 color components */
for (i = 0; i < 4; ++i) {
unsigned bits = desc->channel[i].size;
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);
}
shift += bits;
}
/* Ex: convert packed = {BGRA, BGRA, BGRA, BGRA}
* into masked = {B, G, R, A}
*/
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 shift;
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;
shift = 0;
for (i = 0; i < 4; ++i) {
unsigned bits = desc->channel[i].size;
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);
}
shift += bits;
}
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;
}
static LLVMValueRef
lp_build_extract_soa_chan(struct lp_build_context *bld,
unsigned blockbits,
boolean srgb_chan,
struct util_format_channel_description chan_desc,
LLVMValueRef packed)
{
struct gallivm_state *gallivm = bld->gallivm;
LLVMBuilderRef builder = gallivm->builder;
struct lp_type type = bld->type;
LLVMValueRef input = packed;
const unsigned width = chan_desc.size;
const unsigned start = chan_desc.shift;
const unsigned stop = start + width;
/* Decode the input vector component */
switch(chan_desc.type) {
case UTIL_FORMAT_TYPE_VOID:
input = bld->undef;
break;
case UTIL_FORMAT_TYPE_UNSIGNED:
/*
* Align the LSB
*/
if (start) {
input = LLVMBuildLShr(builder, input,
lp_build_const_int_vec(gallivm, type, start), "");
}
/*
* Zero the MSBs
*/
if (stop < blockbits) {
unsigned mask = ((unsigned long long)1 << width) - 1;
input = LLVMBuildAnd(builder, input,
lp_build_const_int_vec(gallivm, type, mask), "");
}
/*
* Type conversion
*/
if (type.floating) {
if (srgb_chan) {
struct lp_type conv_type = lp_uint_type(type);
input = lp_build_srgb_to_linear(gallivm, conv_type, width, input);
}
else {
if(chan_desc.normalized)
input = lp_build_unsigned_norm_to_float(gallivm, width, type, input);
else
input = LLVMBuildSIToFP(builder, input, bld->vec_type, "");
}
}
else if (chan_desc.pure_integer) {
/* Nothing to do */
} else {
/* FIXME */
assert(0);
}
break;
case UTIL_FORMAT_TYPE_SIGNED:
/*
* Align the sign bit first.
*/
if (stop < type.width) {
unsigned bits = type.width - stop;
LLVMValueRef bits_val = lp_build_const_int_vec(gallivm, type, bits);
input = LLVMBuildShl(builder, input, bits_val, "");
}
/*
* Align the LSB (with an arithmetic shift to preserve the sign)
*/
if (chan_desc.size < type.width) {
unsigned bits = type.width - chan_desc.size;
LLVMValueRef bits_val = lp_build_const_int_vec(gallivm, type, bits);
input = LLVMBuildAShr(builder, input, bits_val, "");
}
/*
* Type conversion
*/
if (type.floating) {
input = LLVMBuildSIToFP(builder, input, bld->vec_type, "");
if (chan_desc.normalized) {
double scale = 1.0 / ((1 << (chan_desc.size - 1)) - 1);
LLVMValueRef scale_val = lp_build_const_vec(gallivm, type, scale);
input = LLVMBuildFMul(builder, input, scale_val, "");
/*
* The formula above will produce value below -1.0 for most negative
* value but everything seems happy with that hence disable for now.
*/
if (0)
input = lp_build_max(bld, input,
lp_build_const_vec(gallivm, type, -1.0f));
}
}
else if (chan_desc.pure_integer) {
/* Nothing to do */
} else {
/* FIXME */
assert(0);
}
break;
case UTIL_FORMAT_TYPE_FLOAT:
if (type.floating) {
if (chan_desc.size == 16) {
struct lp_type f16i_type = type;
f16i_type.width /= 2;
f16i_type.floating = 0;
if (start) {
input = LLVMBuildLShr(builder, input,
lp_build_const_int_vec(gallivm, type, start), "");
}
input = LLVMBuildTrunc(builder, input,
lp_build_vec_type(gallivm, f16i_type), "");
input = lp_build_half_to_float(gallivm, input);
} else {
assert(start == 0);
assert(stop == 32);
assert(type.width == 32);
}
input = LLVMBuildBitCast(builder, input, bld->vec_type, "");
}
else {
/* FIXME */
assert(0);
input = bld->undef;
}
break;
case UTIL_FORMAT_TYPE_FIXED:
if (type.floating) {
double scale = 1.0 / ((1 << (chan_desc.size/2)) - 1);
LLVMValueRef scale_val = lp_build_const_vec(gallivm, type, scale);
input = LLVMBuildSIToFP(builder, input, bld->vec_type, "");
input = LLVMBuildFMul(builder, input, scale_val, "");
}
else {
/* FIXME */
assert(0);
input = bld->undef;
}
break;
default:
assert(0);
input = bld->undef;
break;
}
return input;
}
