/**
* Compute the offset of a pixel.
*
* x, y, y_stride are vectors
*/
LLVMValueRef
lp_build_sample_offset(struct lp_build_context *bld,
const struct util_format_description *format_desc,
LLVMValueRef x,
LLVMValueRef y,
LLVMValueRef y_stride,
LLVMValueRef data_ptr)
{
LLVMValueRef x_stride;
LLVMValueRef offset;
x_stride = lp_build_const_scalar(bld->type, format_desc->block.bits/8);
if(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
LLVMValueRef x_lo, x_hi;
LLVMValueRef y_lo, y_hi;
LLVMValueRef x_stride_lo, x_stride_hi;
LLVMValueRef y_stride_lo, y_stride_hi;
LLVMValueRef x_offset_lo, x_offset_hi;
LLVMValueRef y_offset_lo, y_offset_hi;
LLVMValueRef offset_lo, offset_hi;
x_lo = LLVMBuildAnd(bld->builder, x, bld->one, "");
y_lo = LLVMBuildAnd(bld->builder, y, bld->one, "");
x_hi = LLVMBuildLShr(bld->builder, x, bld->one, "");
y_hi = LLVMBuildLShr(bld->builder, y, bld->one, "");
x_stride_lo = x_stride;
y_stride_lo = lp_build_const_scalar(bld->type, 2*format_desc->block.bits/8);
x_stride_hi = lp_build_const_scalar(bld->type, 4*format_desc->block.bits/8);
y_stride_hi = LLVMBuildShl(bld->builder, y_stride, bld->one, "");
x_offset_lo = lp_build_mul(bld, x_lo, x_stride_lo);
y_offset_lo = lp_build_mul(bld, y_lo, y_stride_lo);
offset_lo = lp_build_add(bld, x_offset_lo, y_offset_lo);
x_offset_hi = lp_build_mul(bld, x_hi, x_stride_hi);
y_offset_hi = lp_build_mul(bld, y_hi, y_stride_hi);
offset_hi = lp_build_add(bld, x_offset_hi, y_offset_hi);
offset = lp_build_add(bld, offset_hi, offset_lo);
}
else {
LLVMValueRef x_offset;
LLVMValueRef y_offset;
x_offset = lp_build_mul(bld, x, x_stride);
y_offset = lp_build_mul(bld, y, y_stride);
offset = lp_build_add(bld, x_offset, y_offset);
}
return offset;
}
/**
* Generate polynomial.
* Ex: coeffs[0] + x * coeffs[1] + x^2 * coeffs[2].
*/
static LLVMValueRef
lp_build_polynomial(struct lp_build_context *bld,
LLVMValueRef x,
const double *coeffs,
unsigned num_coeffs)
{
const struct lp_type type = bld->type;
LLVMValueRef res = NULL;
unsigned i;
/* TODO: optimize the constant case */
if(LLVMIsConstant(x))
debug_printf("%s: inefficient/imprecise constant arithmetic\n",
__FUNCTION__);
for (i = num_coeffs; i--; ) {
LLVMValueRef coeff = lp_build_const_scalar(type, coeffs[i]);
if(res)
res = lp_build_add(bld, coeff, lp_build_mul(bld, x, res));
else
res = coeff;
}
if(res)
return res;
else
return bld->undef;
}
/**
* Compute the offset of a pixel block.
*
* x, y, z, y_stride, z_stride are vectors, and they refer to pixels.
*
* Returns the relative offset and i,j sub-block coordinates
*/
void
lp_build_sample_offset(struct lp_build_context *bld,
const struct util_format_description *format_desc,
LLVMValueRef x,
LLVMValueRef y,
LLVMValueRef z,
LLVMValueRef y_stride,
LLVMValueRef z_stride,
LLVMValueRef *out_offset,
LLVMValueRef *out_i,
LLVMValueRef *out_j)
{
LLVMValueRef x_stride;
LLVMValueRef offset;
x_stride = lp_build_const_vec(bld->gallivm, bld->type,
format_desc->block.bits/8);
lp_build_sample_partial_offset(bld,
format_desc->block.width,
x, x_stride,
&offset, out_i);
if (y && y_stride) {
LLVMValueRef y_offset;
lp_build_sample_partial_offset(bld,
format_desc->block.height,
y, y_stride,
&y_offset, out_j);
offset = lp_build_add(bld, offset, y_offset);
}
else {
*out_j = bld->zero;
}
if (z && z_stride) {
LLVMValueRef z_offset;
LLVMValueRef k;
lp_build_sample_partial_offset(bld,
1, /* pixel blocks are always 2D */
z, z_stride,
&z_offset, &k);
offset = lp_build_add(bld, offset, z_offset);
}
*out_offset = offset;
}
/**
* For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
* two (adjacent) mipmap level indexes. Later, we'll sample from those
* two mipmap levels and interpolate between them.
*/
void
lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
unsigned unit,
LLVMValueRef lod_ipart,
LLVMValueRef *lod_fpart_inout,
LLVMValueRef *level0_out,
LLVMValueRef *level1_out)
{
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_build_context *int_bld = &bld->int_bld;
struct lp_build_context *float_bld = &bld->float_bld;
LLVMValueRef last_level;
LLVMValueRef clamp_min;
LLVMValueRef clamp_max;
*level0_out = lod_ipart;
*level1_out = lp_build_add(int_bld, lod_ipart, int_bld->one);
last_level = bld->dynamic_state->last_level(bld->dynamic_state,
bld->gallivm, unit);
/*
* Clamp both lod_ipart and lod_ipart + 1 to [0, last_level], with the
* minimum number of comparisons, and zeroing lod_fpart in the extreme
* ends in the process.
*/
/* lod_ipart < 0 */
clamp_min = LLVMBuildICmp(builder, LLVMIntSLT,
lod_ipart, int_bld->zero,
"clamp_lod_to_zero");
*level0_out = LLVMBuildSelect(builder, clamp_min,
int_bld->zero, *level0_out, "");
*level1_out = LLVMBuildSelect(builder, clamp_min,
int_bld->zero, *level1_out, "");
*lod_fpart_inout = LLVMBuildSelect(builder, clamp_min,
float_bld->zero, *lod_fpart_inout, "");
/* lod_ipart >= last_level */
clamp_max = LLVMBuildICmp(builder, LLVMIntSGE,
lod_ipart, last_level,
"clamp_lod_to_last");
*level0_out = LLVMBuildSelect(builder, clamp_max,
last_level, *level0_out, "");
*level1_out = LLVMBuildSelect(builder, clamp_max,
last_level, *level1_out, "");
*lod_fpart_inout = LLVMBuildSelect(builder, clamp_max,
float_bld->zero, *lod_fpart_inout, "");
lp_build_name(*level0_out, "sampler%u_miplevel0", unit);
lp_build_name(*level1_out, "sampler%u_miplevel1", unit);
lp_build_name(*lod_fpart_inout, "sampler%u_mipweight", unit);
}
/*
* Bri-linear lod computation
*
* Use a piece-wise linear approximation of log2 such that:
* - round to nearest, for values in the neighborhood of -1, 0, 1, 2, etc.
* - linear approximation for values in the neighborhood of 0.5, 1.5., etc,
* with the steepness specified in 'factor'
* - exact result for 0.5, 1.5, etc.
*
*
* 1.0 - /----*
* /
* /
* /
* 0.5 - *
* /
* /
* /
* 0.0 - *----/
*
* | |
* 2^0 2^1
*
* This is a technique also commonly used in hardware:
* - http://ixbtlabs.com/articles2/gffx/nv40-rx800-3.html
*
* TODO: For correctness, this should only be applied when texture is known to
* have regular mipmaps, i.e., mipmaps derived from the base level.
*
* TODO: This could be done in fixed point, where applicable.
*/
static void
lp_build_brilinear_lod(struct lp_build_context *bld,
LLVMValueRef lod,
double factor,
LLVMValueRef *out_lod_ipart,
LLVMValueRef *out_lod_fpart)
{
LLVMValueRef lod_fpart;
double pre_offset = (factor - 0.5)/factor - 0.5;
double post_offset = 1 - factor;
if (0) {
lp_build_printf(bld->gallivm, "lod = %f\n", lod);
}
lod = lp_build_add(bld, lod,
lp_build_const_vec(bld->gallivm, bld->type, pre_offset));
lp_build_ifloor_fract(bld, lod, out_lod_ipart, &lod_fpart);
lod_fpart = lp_build_mul(bld, lod_fpart,
lp_build_const_vec(bld->gallivm, bld->type, factor));
lod_fpart = lp_build_add(bld, lod_fpart,
lp_build_const_vec(bld->gallivm, bld->type, post_offset));
/*
* It's not necessary to clamp lod_fpart since:
* - the above expression will never produce numbers greater than one.
* - the mip filtering branch is only taken if lod_fpart is positive
*/
*out_lod_fpart = lod_fpart;
if (0) {
lp_build_printf(bld->gallivm, "lod_ipart = %i\n", *out_lod_ipart);
lp_build_printf(bld->gallivm, "lod_fpart = %f\n\n", *out_lod_fpart);
}
}
/**
* Small vector x scale multiplication optimization.
*/
LLVMValueRef
lp_build_mul_imm(struct lp_build_context *bld,
LLVMValueRef a,
int b)
{
LLVMValueRef factor;
if(b == 0)
return bld->zero;
if(b == 1)
return a;
if(b == -1)
return LLVMBuildNeg(bld->builder, a, "");
if(b == 2 && bld->type.floating)
return lp_build_add(bld, a, a);
if(util_is_pot(b)) {
unsigned shift = ffs(b) - 1;
if(bld->type.floating) {
#if 0
/*
* Power of two multiplication by directly manipulating the mantissa.
*
* XXX: This might not be always faster, it will introduce a small error
* for multiplication by zero, and it will produce wrong results
* for Inf and NaN.
*/
unsigned mantissa = lp_mantissa(bld->type);
factor = lp_build_int_const_scalar(bld->type, (unsigned long long)shift << mantissa);
a = LLVMBuildBitCast(bld->builder, a, lp_build_int_vec_type(bld->type), "");
a = LLVMBuildAdd(bld->builder, a, factor, "");
a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(bld->type), "");
return a;
#endif
}
else {
factor = lp_build_const_scalar(bld->type, shift);
return LLVMBuildShl(bld->builder, a, factor, "");
}
}
factor = lp_build_const_scalar(bld->type, (double)b);
return lp_build_mul(bld, a, factor);
}
/*
* Combined log2 and brilinear lod computation.
*
* It's in all identical to calling lp_build_fast_log2() and
* lp_build_brilinear_lod() above, but by combining we can compute the integer
* and fractional part independently.
*/
static void
lp_build_brilinear_rho(struct lp_build_context *bld,
LLVMValueRef rho,
double factor,
LLVMValueRef *out_lod_ipart,
LLVMValueRef *out_lod_fpart)
{
LLVMValueRef lod_ipart;
LLVMValueRef lod_fpart;
const double pre_factor = (2*factor - 0.5)/(M_SQRT2*factor);
const double post_offset = 1 - 2*factor;
assert(bld->type.floating);
assert(lp_check_value(bld->type, rho));
/*
* The pre factor will make the intersections with the exact powers of two
* happen precisely where we want then to be, which means that the integer
* part will not need any post adjustments.
*/
rho = lp_build_mul(bld, rho,
lp_build_const_vec(bld->gallivm, bld->type, pre_factor));
/* ipart = ifloor(log2(rho)) */
lod_ipart = lp_build_extract_exponent(bld, rho, 0);
/* fpart = rho / 2**ipart */
lod_fpart = lp_build_extract_mantissa(bld, rho);
lod_fpart = lp_build_mul(bld, lod_fpart,
lp_build_const_vec(bld->gallivm, bld->type, factor));
lod_fpart = lp_build_add(bld, lod_fpart,
lp_build_const_vec(bld->gallivm, bld->type, post_offset));
/*
* Like lp_build_brilinear_lod, it's not necessary to clamp lod_fpart since:
* - the above expression will never produce numbers greater than one.
* - the mip filtering branch is only taken if lod_fpart is positive
*/
*out_lod_ipart = lod_ipart;
*out_lod_fpart = lod_fpart;
}
/**
* For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
* mipmap level index.
* Note: this is all scalar code.
* \param lod scalar float texture level of detail
* \param level_out returns integer
*/
void
lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
unsigned unit,
LLVMValueRef lod_ipart,
LLVMValueRef *level_out)
{
struct lp_build_context *int_bld = &bld->int_bld;
LLVMValueRef first_level, last_level, level;
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
bld->gallivm, unit);
last_level = bld->dynamic_state->last_level(bld->dynamic_state,
bld->gallivm, unit);
/* convert float lod to integer */
level = lp_build_add(int_bld, lod_ipart, first_level);
/* clamp level to legal range of levels */
*level_out = lp_build_clamp(int_bld, level, first_level, last_level);
}
static LLVMValueRef fetch_constant(
struct lp_build_tgsi_context * bld_base,
const struct tgsi_full_src_register *reg,
enum tgsi_opcode_type type,
unsigned swizzle)
{
struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
struct lp_build_context * base = &bld_base->base;
const struct tgsi_ind_register *ireg = ®->Indirect;
unsigned idx;
LLVMValueRef args[2];
LLVMValueRef addr;
LLVMValueRef result;
if (swizzle == LP_CHAN_ALL) {
unsigned chan;
LLVMValueRef values[4];
for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
values[chan] = fetch_constant(bld_base, reg, type, chan);
return lp_build_gather_values(bld_base->base.gallivm, values, 4);
}
idx = reg->Register.Index * 4 + swizzle;
if (!reg->Register.Indirect)
return bitcast(bld_base, type, si_shader_ctx->constants[idx]);
args[0] = si_shader_ctx->const_resource;
args[1] = lp_build_const_int32(base->gallivm, idx * 4);
addr = si_shader_ctx->radeon_bld.soa.addr[ireg->Index][ireg->Swizzle];
addr = LLVMBuildLoad(base->gallivm->builder, addr, "load addr reg");
addr = lp_build_mul_imm(&bld_base->uint_bld, addr, 16);
args[1] = lp_build_add(&bld_base->uint_bld, addr, args[1]);
result = build_intrinsic(base->gallivm->builder, "llvm.SI.load.const", base->elem_type,
args, 2, LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
return bitcast(bld_base, type, result);
}
/**
* @sa http://www.opengl.org/sdk/docs/man/xhtml/glBlendEquationSeparate.xml
*/
LLVMValueRef
lp_build_blend_func(struct lp_build_context *bld,
unsigned func,
LLVMValueRef term1,
LLVMValueRef term2)
{
switch (func) {
case PIPE_BLEND_ADD:
return lp_build_add(bld, term1, term2);
case PIPE_BLEND_SUBTRACT:
return lp_build_sub(bld, term1, term2);
case PIPE_BLEND_REVERSE_SUBTRACT:
return lp_build_sub(bld, term2, term1);
case PIPE_BLEND_MIN:
return lp_build_min(bld, term1, term2);
case PIPE_BLEND_MAX:
return lp_build_max(bld, term1, term2);
default:
assert(0);
return bld->zero;
}
}
/**
* Linear interpolation.
*
* This also works for integer values with a few caveats.
*
* @sa http://www.stereopsis.com/doubleblend.html
*/
LLVMValueRef
lp_build_lerp(struct lp_build_context *bld,
LLVMValueRef x,
LLVMValueRef v0,
LLVMValueRef v1)
{
LLVMValueRef delta;
LLVMValueRef res;
delta = lp_build_sub(bld, v1, v0);
res = lp_build_mul(bld, x, delta);
res = lp_build_add(bld, v0, res);
if(bld->type.fixed)
/* XXX: This step is necessary for lerping 8bit colors stored on 16bits,
* but it will be wrong for other uses. Basically we need a more
* powerful lp_type, capable of further distinguishing the values
* interpretation from the value storage. */
res = LLVMBuildAnd(bld->builder, res, lp_build_int_const_scalar(bld->type, (1 << bld->type.width/2) - 1), "");
return res;
}
/**
* Helper used by lp_build_cube_lookup()
* \param sign scalar +1 or -1
* \param coord float vector
* \param ima float vector
*/
static LLVMValueRef
lp_build_cube_coord(struct lp_build_context *coord_bld,
LLVMValueRef sign, int negate_coord,
LLVMValueRef coord, LLVMValueRef ima)
{
/* return negate(coord) * ima * sign + 0.5; */
LLVMValueRef half = lp_build_const_vec(coord_bld->gallivm, coord_bld->type, 0.5);
LLVMValueRef res;
assert(negate_coord == +1 || negate_coord == -1);
if (negate_coord == -1) {
coord = lp_build_negate(coord_bld, coord);
}
res = lp_build_mul(coord_bld, coord, ima);
if (sign) {
sign = lp_build_broadcast_scalar(coord_bld, sign);
res = lp_build_mul(coord_bld, res, sign);
}
res = lp_build_add(coord_bld, res, half);
return res;
}
/**
* Build LLVM code for texture coord wrapping, for linear filtering,
* for scaled integer texcoords.
* \param block_length is the length of the pixel block along the
* coordinate axis
* \param coord0 the incoming texcoord (s,t,r or q) scaled to the texture size
* \param length the texture size along one dimension
* \param stride pixel stride along the coordinate axis (in bytes)
* \param is_pot if TRUE, length is a power of two
* \param wrap_mode one of PIPE_TEX_WRAP_x
* \param offset0 resulting relative offset for coord0
* \param offset1 resulting relative offset for coord0 + 1
* \param i0 resulting sub-block pixel coordinate for coord0
* \param i1 resulting sub-block pixel coordinate for coord0 + 1
*/
static void
lp_build_sample_wrap_linear_int(struct lp_build_sample_context *bld,
unsigned block_length,
LLVMValueRef coord0,
LLVMValueRef length,
LLVMValueRef stride,
boolean is_pot,
unsigned wrap_mode,
LLVMValueRef *offset0,
LLVMValueRef *offset1,
LLVMValueRef *i0,
LLVMValueRef *i1)
{
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef length_minus_one;
LLVMValueRef lmask, umask, mask;
if (block_length != 1) {
/*
* If the pixel block covers more than one pixel then there is no easy
* way to calculate offset1 relative to offset0. Instead, compute them
* independently.
*/
LLVMValueRef coord1;
lp_build_sample_wrap_nearest_int(bld,
block_length,
coord0,
length,
stride,
is_pot,
wrap_mode,
offset0, i0);
coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
lp_build_sample_wrap_nearest_int(bld,
block_length,
coord1,
length,
stride,
is_pot,
wrap_mode,
offset1, i1);
return;
}
/*
* Scalar pixels -- try to compute offset0 and offset1 with a single stride
* multiplication.
*/
*i0 = int_coord_bld->zero;
*i1 = int_coord_bld->zero;
length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
switch(wrap_mode) {
case PIPE_TEX_WRAP_REPEAT:
if (is_pot) {
coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, "");
}
else {
/* Add a bias to the texcoord to handle negative coords */
LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024);
coord0 = LLVMBuildAdd(builder, coord0, bias, "");
coord0 = LLVMBuildURem(builder, coord0, length, "");
}
mask = lp_build_compare(bld->gallivm, int_coord_bld->type,
PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
*offset0 = lp_build_mul(int_coord_bld, coord0, stride);
*offset1 = LLVMBuildAnd(builder,
lp_build_add(int_coord_bld, *offset0, stride),
mask, "");
break;
case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
lmask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
PIPE_FUNC_GEQUAL, coord0, int_coord_bld->zero);
umask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
PIPE_FUNC_LESS, coord0, length_minus_one);
coord0 = lp_build_select(int_coord_bld, lmask, coord0, int_coord_bld->zero);
coord0 = lp_build_select(int_coord_bld, umask, coord0, length_minus_one);
mask = LLVMBuildAnd(builder, lmask, umask, "");
*offset0 = lp_build_mul(int_coord_bld, coord0, stride);
*offset1 = lp_build_add(int_coord_bld,
*offset0,
LLVMBuildAnd(builder, stride, mask, ""));
break;
case PIPE_TEX_WRAP_CLAMP:
case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
case PIPE_TEX_WRAP_MIRROR_REPEAT:
case PIPE_TEX_WRAP_MIRROR_CLAMP:
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
default:
assert(0);
*offset0 = int_coord_bld->zero;
*offset1 = int_coord_bld->zero;
break;
}
}
/**
* Convert linear float values to srgb int values.
* Several possibilities how to do this, e.g.
* - use table (based on exponent/highest order mantissa bits) and do
* linear interpolation (https://gist.github.com/rygorous/2203834)
* - Chebyshev polynomial
* - Approximation using reciprocals
* - using int-to-float and float-to-int tricks for pow()
* (http://stackoverflow.com/questions/6475373/optimizations-for-pow-with-const-non-integer-exponent)
*
* @param src float (vector) value(s) to convert.
*/
static LLVMValueRef
lp_build_linear_to_srgb(struct gallivm_state *gallivm,
struct lp_type src_type,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_build_context f32_bld;
LLVMValueRef lin_thresh, lin, lin_const, is_linear, tmp, pow_final;
lp_build_context_init(&f32_bld, gallivm, src_type);
src = lp_build_clamp(&f32_bld, src, f32_bld.zero, f32_bld.one);
if (0) {
/*
* using int-to-float and float-to-int trick for pow().
* This is much more accurate than necessary thanks to the correction,
* but it most certainly makes no sense without rsqrt available.
* Bonus points if you understand how this works...
* All in all (including min/max clamp, conversion) 19 instructions.
*/
float exp_f = 2.0f / 3.0f;
/* some compilers can't do exp2f, so this is exp2f(127.0f/exp_f - 127.0f) */
float exp2f_c = 1.30438178253e+19f;
float coeff_f = 0.62996f;
LLVMValueRef pow_approx, coeff, x2, exponent, pow_1, pow_2;
struct lp_type int_type = lp_int_type(src_type);
/*
* First calculate approx x^8/12
*/
exponent = lp_build_const_vec(gallivm, src_type, exp_f);
coeff = lp_build_const_vec(gallivm, src_type,
exp2f_c * powf(coeff_f, 1.0f / exp_f));
/* premultiply src */
tmp = lp_build_mul(&f32_bld, coeff, src);
/* "log2" */
tmp = LLVMBuildBitCast(builder, tmp, lp_build_vec_type(gallivm, int_type), "");
tmp = lp_build_int_to_float(&f32_bld, tmp);
/* multiply for pow */
tmp = lp_build_mul(&f32_bld, tmp, exponent);
/* "exp2" */
pow_approx = lp_build_itrunc(&f32_bld, tmp);
pow_approx = LLVMBuildBitCast(builder, pow_approx,
lp_build_vec_type(gallivm, src_type), "");
/*
* Since that pow was inaccurate (like 3 bits, though each sqrt step would
* give another bit), compensate the error (which is why we chose another
* exponent in the first place).
*/
/* x * x^(8/12) = x^(20/12) */
pow_1 = lp_build_mul(&f32_bld, pow_approx, src);
/* x * x * x^(-4/12) = x^(20/12) */
/* Should avoid using rsqrt if it's not available, but
* using x * x^(4/12) * x^(4/12) instead will change error weight */
tmp = lp_build_fast_rsqrt(&f32_bld, pow_approx);
x2 = lp_build_mul(&f32_bld, src, src);
pow_2 = lp_build_mul(&f32_bld, x2, tmp);
/* average the values so the errors cancel out, compensate bias,
* we also squeeze the 1.055 mul of the srgb conversion plus the 255.0 mul
* for conversion to int in here */
tmp = lp_build_add(&f32_bld, pow_1, pow_2);
coeff = lp_build_const_vec(gallivm, src_type,
1.0f / (3.0f * coeff_f) * 0.999852f *
powf(1.055f * 255.0f, 4.0f));
pow_final = lp_build_mul(&f32_bld, tmp, coeff);
/* x^(5/12) = rsqrt(rsqrt(x^20/12)) */
if (lp_build_fast_rsqrt_available(src_type)) {
pow_final = lp_build_fast_rsqrt(&f32_bld,
lp_build_fast_rsqrt(&f32_bld, pow_final));
}
else {
pow_final = lp_build_sqrt(&f32_bld, lp_build_sqrt(&f32_bld, pow_final));
}
pow_final = lp_build_add(&f32_bld, pow_final,
lp_build_const_vec(gallivm, src_type, -0.055f * 255.0f));
}
else {
/*
* using "rational polynomial" approximation here.
* Essentially y = a*x^0.375 + b*x^0.5 + c, with also
* factoring in the 255.0 mul and the scaling mul.
* (a is closer to actual value so has higher weight than b.)
* Note: the constants are magic values. They were found empirically,
* possibly could be improved but good enough (be VERY careful with
* error metric if you'd want to tweak them, they also MUST fit with
* the crappy polynomial above for srgb->linear since it is required
* that each srgb value maps back to the same value).
* This function has an error of max +-0.17 (and we'd only require +-0.6),
* for the approximated srgb->linear values the error is naturally larger
* (+-0.42) but still accurate enough (required +-0.5 essentially).
* All in all (including min/max clamp, conversion) 15 instructions.
* FMA would help (minus 2 instructions).
*/
LLVMValueRef x05, x0375, a_const, b_const, c_const, tmp2;
if (lp_build_fast_rsqrt_available(src_type)) {
tmp = lp_build_fast_rsqrt(&f32_bld, src);
x05 = lp_build_mul(&f32_bld, src, tmp);
}
else {
/*
* I don't really expect this to be practical without rsqrt
* but there's no reason for triple punishment so at least
* save the otherwise resulting division and unnecessary mul...
*/
x05 = lp_build_sqrt(&f32_bld, src);
}
tmp = lp_build_mul(&f32_bld, x05, src);
if (lp_build_fast_rsqrt_available(src_type)) {
x0375 = lp_build_fast_rsqrt(&f32_bld, lp_build_fast_rsqrt(&f32_bld, tmp));
}
else {
x0375 = lp_build_sqrt(&f32_bld, lp_build_sqrt(&f32_bld, tmp));
}
a_const = lp_build_const_vec(gallivm, src_type, 0.675f * 1.0622 * 255.0f);
b_const = lp_build_const_vec(gallivm, src_type, 0.325f * 1.0622 * 255.0f);
c_const = lp_build_const_vec(gallivm, src_type, -0.0620f * 255.0f);
tmp = lp_build_mul(&f32_bld, a_const, x0375);
tmp2 = lp_build_mul(&f32_bld, b_const, x05);
tmp2 = lp_build_add(&f32_bld, tmp2, c_const);
pow_final = lp_build_add(&f32_bld, tmp, tmp2);
}
/* linear part is easy */
lin_const = lp_build_const_vec(gallivm, src_type, 12.92f * 255.0f);
lin = lp_build_mul(&f32_bld, src, lin_const);
lin_thresh = lp_build_const_vec(gallivm, src_type, 0.0031308f);
is_linear = lp_build_compare(gallivm, src_type, PIPE_FUNC_LEQUAL, src, lin_thresh);
tmp = lp_build_select(&f32_bld, is_linear, lin, pow_final);
f32_bld.type.sign = 0;
return lp_build_iround(&f32_bld, tmp);
}
/**
* Increment the shader input attribute values.
* This is called when we move from one quad to the next.
*/
static void
attribs_update(struct lp_build_interp_soa_context *bld,
struct gallivm_state *gallivm,
LLVMValueRef loop_iter,
int start,
int end)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_build_context *coeff_bld = &bld->coeff_bld;
LLVMValueRef oow = NULL;
unsigned attrib;
unsigned chan;
for(attrib = start; attrib < end; ++attrib) {
const unsigned mask = bld->mask[attrib];
const unsigned interp = bld->interp[attrib];
for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
if(mask & (1 << chan)) {
LLVMValueRef a;
if (interp == LP_INTERP_CONSTANT ||
interp == LP_INTERP_FACING) {
a = LLVMBuildLoad(builder, bld->a[attrib][chan], "");
}
else if (interp == LP_INTERP_POSITION) {
assert(attrib > 0);
a = bld->attribs[0][chan];
}
else {
LLVMValueRef dadq;
a = bld->a[attrib][chan];
/*
* Broadcast the attribute value for this quad into all elements
*/
{
/* stored as vector load as float */
LLVMTypeRef ptr_type = LLVMPointerType(LLVMFloatTypeInContext(
gallivm->context), 0);
LLVMValueRef ptr;
a = LLVMBuildBitCast(builder, a, ptr_type, "");
ptr = LLVMBuildGEP(builder, a, &loop_iter, 1, "");
a = LLVMBuildLoad(builder, ptr, "");
a = lp_build_broadcast_scalar(&bld->coeff_bld, a);
}
/*
* Get the derivatives.
*/
dadq = bld->dadq[attrib][chan];
#if PERSPECTIVE_DIVIDE_PER_QUAD
if (interp == LP_INTERP_PERSPECTIVE) {
LLVMValueRef dwdq = bld->dadq[0][3];
if (oow == NULL) {
assert(bld->oow);
oow = LLVMBuildShuffleVector(coeff_bld->builder,
bld->oow, coeff_bld->undef,
shuffle, "");
}
dadq = lp_build_sub(coeff_bld,
dadq,
lp_build_mul(coeff_bld, a, dwdq));
dadq = lp_build_mul(coeff_bld, dadq, oow);
}
#endif
/*
* Add the derivatives
*/
a = lp_build_add(coeff_bld, a, dadq);
#if !PERSPECTIVE_DIVIDE_PER_QUAD
if (interp == LP_INTERP_PERSPECTIVE) {
if (oow == NULL) {
LLVMValueRef w = bld->attribs[0][3];
assert(attrib != 0);
assert(bld->mask[0] & TGSI_WRITEMASK_W);
oow = lp_build_rcp(coeff_bld, w);
}
a = lp_build_mul(coeff_bld, a, oow);
}
#endif
if (attrib == 0 && chan == 2) {
/* FIXME: Depth values can exceed 1.0, due to the fact that
* setup interpolation coefficients refer to (0,0) which causes
* precision loss. So we must clamp to 1.0 here to avoid artifacts
*/
a = lp_build_min(coeff_bld, a, coeff_bld->one);
}
attrib_name(a, attrib, chan, "");
}
bld->attribs[attrib][chan] = a;
}
}
}
}
/**
* Sample a single texture image with nearest sampling.
* If sampling a cube texture, r = cube face in [0,5].
* Return filtered color as two vectors of 16-bit fixed point values.
*/
static void
lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
LLVMValueRef int_size,
LLVMValueRef row_stride_vec,
LLVMValueRef img_stride_vec,
LLVMValueRef data_ptr,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
LLVMValueRef *colors_lo,
LLVMValueRef *colors_hi)
{
const unsigned dims = bld->dims;
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_build_context i32, h16, u8n;
LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
LLVMValueRef i32_c8;
LLVMValueRef width_vec, height_vec, depth_vec;
LLVMValueRef s_ipart, t_ipart = NULL, r_ipart = NULL;
LLVMValueRef x_stride;
LLVMValueRef x_offset, offset;
LLVMValueRef x_subcoord, y_subcoord, z_subcoord;
lp_build_context_init(&i32, bld->gallivm, lp_type_int_vec(32));
lp_build_context_init(&h16, bld->gallivm, lp_type_ufixed(16));
lp_build_context_init(&u8n, bld->gallivm, lp_type_unorm(8));
i32_vec_type = lp_build_vec_type(bld->gallivm, i32.type);
h16_vec_type = lp_build_vec_type(bld->gallivm, h16.type);
u8n_vec_type = lp_build_vec_type(bld->gallivm, u8n.type);
lp_build_extract_image_sizes(bld,
bld->int_size_type,
bld->int_coord_type,
int_size,
&width_vec,
&height_vec,
&depth_vec);
if (bld->static_state->normalized_coords) {
LLVMValueRef scaled_size;
LLVMValueRef flt_size;
/* scale size by 256 (8 fractional bits) */
scaled_size = lp_build_shl_imm(&bld->int_size_bld, int_size, 8);
flt_size = lp_build_int_to_float(&bld->float_size_bld, scaled_size);
lp_build_unnormalized_coords(bld, flt_size, &s, &t, &r);
}
else {
/* scale coords by 256 (8 fractional bits) */
s = lp_build_mul_imm(&bld->coord_bld, s, 256);
if (dims >= 2)
t = lp_build_mul_imm(&bld->coord_bld, t, 256);
if (dims >= 3)
r = lp_build_mul_imm(&bld->coord_bld, r, 256);
}
/* convert float to int */
s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
if (dims >= 2)
t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
if (dims >= 3)
r = LLVMBuildFPToSI(builder, r, i32_vec_type, "");
/* compute floor (shift right 8) */
i32_c8 = lp_build_const_int_vec(bld->gallivm, i32.type, 8);
s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
if (dims >= 2)
t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
if (dims >= 3)
r_ipart = LLVMBuildAShr(builder, r, i32_c8, "");
/* get pixel, row, image strides */
x_stride = lp_build_const_vec(bld->gallivm,
bld->int_coord_bld.type,
bld->format_desc->block.bits/8);
/* Do texcoord wrapping, compute texel offset */
lp_build_sample_wrap_nearest_int(bld,
bld->format_desc->block.width,
s_ipart, width_vec, x_stride,
bld->static_state->pot_width,
bld->static_state->wrap_s,
&x_offset, &x_subcoord);
offset = x_offset;
if (dims >= 2) {
LLVMValueRef y_offset;
lp_build_sample_wrap_nearest_int(bld,
bld->format_desc->block.height,
t_ipart, height_vec, row_stride_vec,
bld->static_state->pot_height,
bld->static_state->wrap_t,
&y_offset, &y_subcoord);
offset = lp_build_add(&bld->int_coord_bld, offset, y_offset);
if (dims >= 3) {
LLVMValueRef z_offset;
lp_build_sample_wrap_nearest_int(bld,
1, /* block length (depth) */
r_ipart, depth_vec, img_stride_vec,
bld->static_state->pot_height,
bld->static_state->wrap_r,
&z_offset, &z_subcoord);
offset = lp_build_add(&bld->int_coord_bld, offset, z_offset);
}
else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
LLVMValueRef z_offset;
/* The r coord is the cube face in [0,5] */
z_offset = lp_build_mul(&bld->int_coord_bld, r, img_stride_vec);
offset = lp_build_add(&bld->int_coord_bld, offset, z_offset);
}
}
/*
* Fetch the pixels as 4 x 32bit (rgba order might differ):
*
* rgba0 rgba1 rgba2 rgba3
*
* bit cast them into 16 x u8
*
* r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
*
* unpack them into two 8 x i16:
*
* r0 g0 b0 a0 r1 g1 b1 a1
* r2 g2 b2 a2 r3 g3 b3 a3
*
* The higher 8 bits of the resulting elements will be zero.
*/
{
LLVMValueRef rgba8;
if (util_format_is_rgba8_variant(bld->format_desc)) {
/*
* Given the format is a rgba8, just read the pixels as is,
* without any swizzling. Swizzling will be done later.
*/
rgba8 = lp_build_gather(bld->gallivm,
bld->texel_type.length,
bld->format_desc->block.bits,
bld->texel_type.width,
data_ptr, offset);
rgba8 = LLVMBuildBitCast(builder, rgba8, u8n_vec_type, "");
}
else {
rgba8 = lp_build_fetch_rgba_aos(bld->gallivm,
bld->format_desc,
u8n.type,
data_ptr, offset,
x_subcoord,
y_subcoord);
}
/* Expand one 4*rgba8 to two 2*rgba16 */
lp_build_unpack2(bld->gallivm, u8n.type, h16.type,
rgba8,
colors_lo, colors_hi);
}
}
/**
* Emit LLVM for one TGSI instruction.
* \param return TRUE for success, FALSE otherwise
*/
boolean
lp_emit_instruction_aos(
struct lp_build_tgsi_aos_context *bld,
const struct tgsi_full_instruction *inst,
const struct tgsi_opcode_info *info,
int *pc)
{
LLVMValueRef src0, src1, src2;
LLVMValueRef tmp0, tmp1;
LLVMValueRef dst0 = NULL;
/*
* Stores and write masks are handled in a general fashion after the long
* instruction opcode switch statement.
*
* Although not stricitly necessary, we avoid generating instructions for
* channels which won't be stored, in cases where's that easy. For some
* complex instructions, like texture sampling, it is more convenient to
* assume a full writemask and then let LLVM optimization passes eliminate
* redundant code.
*/
(*pc)++;
assert(info->num_dst <= 1);
if (info->num_dst) {
dst0 = bld->bld_base.base.undef;
}
switch (inst->Instruction.Opcode) {
case TGSI_OPCODE_ARL:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_floor(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_MOV:
dst0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
break;
case TGSI_OPCODE_LIT:
return FALSE;
case TGSI_OPCODE_RCP:
/* TGSI_OPCODE_RECIP */
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_rcp(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_RSQ:
/* TGSI_OPCODE_RECIPSQRT */
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = lp_build_emit_llvm_unary(&bld->bld_base, TGSI_OPCODE_ABS, src0);
dst0 = lp_build_rsqrt(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_EXP:
return FALSE;
case TGSI_OPCODE_LOG:
return FALSE;
case TGSI_OPCODE_MUL:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
dst0 = lp_build_mul(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_ADD:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
dst0 = lp_build_add(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_DP3:
/* TGSI_OPCODE_DOT3 */
return FALSE;
case TGSI_OPCODE_DP4:
/* TGSI_OPCODE_DOT4 */
return FALSE;
case TGSI_OPCODE_DST:
return FALSE;
case TGSI_OPCODE_MIN:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
dst0 = lp_build_max(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_MAX:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
dst0 = lp_build_max(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_SLT:
/* TGSI_OPCODE_SETLT */
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_SGE:
/* TGSI_OPCODE_SETGE */
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_MAD:
/* TGSI_OPCODE_MADD */
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
tmp0 = lp_build_mul(&bld->bld_base.base, src0, src1);
dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
break;
case TGSI_OPCODE_SUB:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
dst0 = lp_build_sub(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_LRP:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
tmp0 = lp_build_sub(&bld->bld_base.base, src1, src2);
tmp0 = lp_build_mul(&bld->bld_base.base, src0, tmp0);
dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
break;
case TGSI_OPCODE_CND:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
tmp1 = lp_build_const_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, 0.5);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src2, tmp1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, src0, src1);
break;
case TGSI_OPCODE_DP2A:
return FALSE;
case TGSI_OPCODE_FRC:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = lp_build_floor(&bld->bld_base.base, src0);
dst0 = lp_build_sub(&bld->bld_base.base, src0, tmp0);
break;
case TGSI_OPCODE_CLAMP:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
tmp0 = lp_build_max(&bld->bld_base.base, src0, src1);
dst0 = lp_build_min(&bld->bld_base.base, tmp0, src2);
break;
case TGSI_OPCODE_FLR:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_floor(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_ROUND:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_round(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_EX2:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = lp_build_swizzle_scalar_aos(&bld->bld_base.base, src0, TGSI_SWIZZLE_X, TGSI_NUM_CHANNELS);
dst0 = lp_build_exp2(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_LG2:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_log2(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_POW:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
dst0 = lp_build_pow(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_XPD:
return FALSE;
case TGSI_OPCODE_RCC:
/* deprecated? */
assert(0);
return FALSE;
case TGSI_OPCODE_DPH:
return FALSE;
case TGSI_OPCODE_COS:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_cos(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_DDX:
return FALSE;
case TGSI_OPCODE_DDY:
return FALSE;
case TGSI_OPCODE_KILP:
/* predicated kill */
return FALSE;
case TGSI_OPCODE_KIL:
/* conditional kill */
return FALSE;
case TGSI_OPCODE_PK2H:
return FALSE;
break;
case TGSI_OPCODE_PK2US:
return FALSE;
break;
case TGSI_OPCODE_PK4B:
return FALSE;
break;
case TGSI_OPCODE_PK4UB:
return FALSE;
case TGSI_OPCODE_RFL:
return FALSE;
case TGSI_OPCODE_SEQ:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_EQUAL, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_SFL:
dst0 = bld->bld_base.base.zero;
break;
case TGSI_OPCODE_SGT:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_SIN:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_sin(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_SLE:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LEQUAL, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_SNE:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_NOTEQUAL, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_STR:
dst0 = bld->bld_base.base.one;
break;
case TGSI_OPCODE_TEX:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
break;
case TGSI_OPCODE_TXD:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
break;
case TGSI_OPCODE_UP2H:
/* deprecated */
assert (0);
return FALSE;
break;
case TGSI_OPCODE_UP2US:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_UP4B:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_UP4UB:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_X2D:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_ARA:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_ARR:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_round(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_BRA:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_CAL:
return FALSE;
case TGSI_OPCODE_RET:
return FALSE;
case TGSI_OPCODE_END:
*pc = -1;
break;
case TGSI_OPCODE_SSG:
/* TGSI_OPCODE_SGN */
tmp0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_sgn(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_CMP:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, bld->bld_base.base.zero);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, src1, src2);
break;
case TGSI_OPCODE_SCS:
return FALSE;
case TGSI_OPCODE_TXB:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
break;
case TGSI_OPCODE_NRM:
/* fall-through */
case TGSI_OPCODE_NRM4:
return FALSE;
case TGSI_OPCODE_DIV:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_DP2:
return FALSE;
case TGSI_OPCODE_TXL:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
break;
case TGSI_OPCODE_TXP:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
break;
case TGSI_OPCODE_BRK:
return FALSE;
case TGSI_OPCODE_IF:
return FALSE;
case TGSI_OPCODE_BGNLOOP:
return FALSE;
case TGSI_OPCODE_BGNSUB:
return FALSE;
case TGSI_OPCODE_ELSE:
return FALSE;
case TGSI_OPCODE_ENDIF:
return FALSE;
case TGSI_OPCODE_ENDLOOP:
return FALSE;
case TGSI_OPCODE_ENDSUB:
return FALSE;
case TGSI_OPCODE_PUSHA:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_POPA:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_CEIL:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_ceil(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_I2F:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_NOT:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_TRUNC:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_trunc(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_SHL:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_ISHR:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_AND:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_OR:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_MOD:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_XOR:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_SAD:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_TXF:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_TXQ:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_CONT:
return FALSE;
case TGSI_OPCODE_EMIT:
return FALSE;
break;
case TGSI_OPCODE_ENDPRIM:
return FALSE;
break;
case TGSI_OPCODE_NOP:
break;
default:
return FALSE;
}
if (info->num_dst) {
lp_emit_store_aos(bld, inst, 0, dst0);
}
return TRUE;
}
/**
* 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;
}
/**
* Apply the stencil operator (add/sub/keep/etc) to the given vector
* of stencil values.
* \return new stencil values vector
*/
static LLVMValueRef
lp_build_stencil_op_single(struct lp_build_context *bld,
const struct pipe_stencil_state *stencil,
enum stencil_op op,
LLVMValueRef stencilRef,
LLVMValueRef stencilVals)
{
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_type type = bld->type;
LLVMValueRef res;
LLVMValueRef max = lp_build_const_int_vec(bld->gallivm, type, 0xff);
unsigned stencil_op;
assert(type.sign);
switch (op) {
case S_FAIL_OP:
stencil_op = stencil->fail_op;
break;
case Z_FAIL_OP:
stencil_op = stencil->zfail_op;
break;
case Z_PASS_OP:
stencil_op = stencil->zpass_op;
break;
default:
assert(0 && "Invalid stencil_op mode");
stencil_op = PIPE_STENCIL_OP_KEEP;
}
switch (stencil_op) {
case PIPE_STENCIL_OP_KEEP:
res = stencilVals;
/* we can return early for this case */
return res;
case PIPE_STENCIL_OP_ZERO:
res = bld->zero;
break;
case PIPE_STENCIL_OP_REPLACE:
res = stencilRef;
break;
case PIPE_STENCIL_OP_INCR:
res = lp_build_add(bld, stencilVals, bld->one);
res = lp_build_min(bld, res, max);
break;
case PIPE_STENCIL_OP_DECR:
res = lp_build_sub(bld, stencilVals, bld->one);
res = lp_build_max(bld, res, bld->zero);
break;
case PIPE_STENCIL_OP_INCR_WRAP:
res = lp_build_add(bld, stencilVals, bld->one);
res = LLVMBuildAnd(builder, res, max, "");
break;
case PIPE_STENCIL_OP_DECR_WRAP:
res = lp_build_sub(bld, stencilVals, bld->one);
res = LLVMBuildAnd(builder, res, max, "");
break;
case PIPE_STENCIL_OP_INVERT:
res = LLVMBuildNot(builder, stencilVals, "");
res = LLVMBuildAnd(builder, res, max, "");
break;
default:
assert(0 && "bad stencil op mode");
res = bld->undef;
}
return res;
}
/**
* Apply the stencil operator (add/sub/keep/etc) to the given vector
* of stencil values.
* \return new stencil values vector
*/
static LLVMValueRef
lp_build_stencil_op_single(struct lp_build_context *bld,
const struct pipe_stencil_state *stencil,
enum stencil_op op,
LLVMValueRef stencilRef,
LLVMValueRef stencilVals,
LLVMValueRef mask)
{
const unsigned stencilMax = 255; /* XXX fix */
struct lp_type type = bld->type;
LLVMValueRef res;
LLVMValueRef max = lp_build_const_int_vec(type, stencilMax);
unsigned stencil_op;
assert(type.sign);
switch (op) {
case S_FAIL_OP:
stencil_op = stencil->fail_op;
break;
case Z_FAIL_OP:
stencil_op = stencil->zfail_op;
break;
case Z_PASS_OP:
stencil_op = stencil->zpass_op;
break;
default:
assert(0 && "Invalid stencil_op mode");
stencil_op = PIPE_STENCIL_OP_KEEP;
}
switch (stencil_op) {
case PIPE_STENCIL_OP_KEEP:
res = stencilVals;
/* we can return early for this case */
return res;
case PIPE_STENCIL_OP_ZERO:
res = bld->zero;
break;
case PIPE_STENCIL_OP_REPLACE:
res = stencilRef;
break;
case PIPE_STENCIL_OP_INCR:
res = lp_build_add(bld, stencilVals, bld->one);
res = lp_build_min(bld, res, max);
break;
case PIPE_STENCIL_OP_DECR:
res = lp_build_sub(bld, stencilVals, bld->one);
res = lp_build_max(bld, res, bld->zero);
break;
case PIPE_STENCIL_OP_INCR_WRAP:
res = lp_build_add(bld, stencilVals, bld->one);
res = LLVMBuildAnd(bld->builder, res, max, "");
break;
case PIPE_STENCIL_OP_DECR_WRAP:
res = lp_build_sub(bld, stencilVals, bld->one);
res = LLVMBuildAnd(bld->builder, res, max, "");
break;
case PIPE_STENCIL_OP_INVERT:
res = LLVMBuildNot(bld->builder, stencilVals, "");
res = LLVMBuildAnd(bld->builder, res, max, "");
break;
default:
assert(0 && "bad stencil op mode");
res = NULL;
}
if (stencil->writemask != stencilMax) {
/* mask &= stencil->writemask */
LLVMValueRef writemask = lp_build_const_int_vec(type, stencil->writemask);
mask = LLVMBuildAnd(bld->builder, mask, writemask, "");
/* res = (res & mask) | (stencilVals & ~mask) */
res = lp_build_select_bitwise(bld, writemask, res, stencilVals);
}
else {
/* res = mask ? res : stencilVals */
res = lp_build_select(bld, mask, res, stencilVals);
}
return res;
}
/**
* Performs optimisations and blending independent of SoA/AoS
*
* @param func the blend function
* @param factor_src PIPE_BLENDFACTOR_xxx
* @param factor_dst PIPE_BLENDFACTOR_xxx
* @param src source rgba
* @param dst dest rgba
* @param src_factor src factor computed value
* @param dst_factor dst factor computed value
* @param not_alpha_dependent same factors accross all channels of src/dst
*
* not_alpha_dependent should be:
* SoA: always true as it is only one channel at a time
* AoS: rgb_src_factor == alpha_src_factor && rgb_dst_factor == alpha_dst_factor
*
* Note that pretty much every possible optimisation can only be done on non-unorm targets
* due to unorm values not going above 1.0 meaning factorisation can change results.
* e.g. (0.9 * 0.9) + (0.9 * 0.9) != 0.9 * (0.9 + 0.9) as result of + is always <= 1.
*/
LLVMValueRef
lp_build_blend(struct lp_build_context *bld,
unsigned func,
unsigned factor_src,
unsigned factor_dst,
LLVMValueRef src,
LLVMValueRef dst,
LLVMValueRef src_factor,
LLVMValueRef dst_factor,
boolean not_alpha_dependent,
boolean optimise_only)
{
LLVMValueRef result, src_term, dst_term;
/* If we are not alpha dependent we can mess with the src/dst factors */
if (not_alpha_dependent) {
if (lp_build_blend_factor_complementary(factor_src, factor_dst)) {
if (func == PIPE_BLEND_ADD) {
if (factor_src < factor_dst) {
return lp_build_lerp(bld, src_factor, dst, src, 0);
} else {
return lp_build_lerp(bld, dst_factor, src, dst, 0);
}
} else if(bld->type.floating && func == PIPE_BLEND_SUBTRACT) {
result = lp_build_add(bld, src, dst);
if (factor_src < factor_dst) {
result = lp_build_mul(bld, result, src_factor);
return lp_build_sub(bld, result, dst);
} else {
result = lp_build_mul(bld, result, dst_factor);
return lp_build_sub(bld, src, result);
}
} else if(bld->type.floating && func == PIPE_BLEND_REVERSE_SUBTRACT) {
result = lp_build_add(bld, src, dst);
if (factor_src < factor_dst) {
result = lp_build_mul(bld, result, src_factor);
return lp_build_sub(bld, dst, result);
} else {
result = lp_build_mul(bld, result, dst_factor);
return lp_build_sub(bld, result, src);
}
}
}
if (bld->type.floating && factor_src == factor_dst) {
if (func == PIPE_BLEND_ADD ||
func == PIPE_BLEND_SUBTRACT ||
func == PIPE_BLEND_REVERSE_SUBTRACT) {
LLVMValueRef result;
result = lp_build_blend_func(bld, func, src, dst);
return lp_build_mul(bld, result, src_factor);
}
}
}
if (optimise_only)
return NULL;
src_term = lp_build_mul(bld, src, src_factor);
dst_term = lp_build_mul(bld, dst, dst_factor);
return lp_build_blend_func(bld, func, src_term, dst_term);
}
