/**
* Generate clamp(a, min, max)
* Do checks for special cases.
*/
LLVMValueRef
lp_build_clamp(struct lp_build_context *bld,
LLVMValueRef a,
LLVMValueRef min,
LLVMValueRef max)
{
a = lp_build_min(bld, a, max);
a = lp_build_max(bld, a, min);
return a;
}
/**
* Build LLVM code for texture coord wrapping, for nearest filtering,
* for scaled integer texcoords.
* \param block_length is the length of the pixel block along the
* coordinate axis
* \param coord 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 out_offset byte offset for the wrapped coordinate
* \param out_i resulting sub-block pixel coordinate for coord0
*/
static void
lp_build_sample_wrap_nearest_int(struct lp_build_sample_context *bld,
unsigned block_length,
LLVMValueRef coord,
LLVMValueRef length,
LLVMValueRef stride,
boolean is_pot,
unsigned wrap_mode,
LLVMValueRef *out_offset,
LLVMValueRef *out_i)
{
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef length_minus_one;
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)
coord = LLVMBuildAnd(builder, coord, 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);
coord = LLVMBuildAdd(builder, coord, bias, "");
coord = LLVMBuildURem(builder, coord, length, "");
}
break;
case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
coord = lp_build_min(int_coord_bld, coord, length_minus_one);
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);
}
lp_build_sample_partial_offset(int_coord_bld, block_length, coord, stride,
out_offset, out_i);
}
/**
* Codegen equivalent for u_minify().
* Return max(1, base_size >> level);
*/
LLVMValueRef
lp_build_minify(struct lp_build_context *bld,
LLVMValueRef base_size,
LLVMValueRef level)
{
LLVMBuilderRef builder = bld->gallivm->builder;
assert(lp_check_value(bld->type, base_size));
assert(lp_check_value(bld->type, level));
if (level == bld->zero) {
/* if we're using mipmap level zero, no minification is needed */
return base_size;
}
else {
LLVMValueRef size =
LLVMBuildLShr(builder, base_size, level, "minify");
assert(bld->type.sign);
size = lp_build_max(bld, size, bld->one);
return size;
}
}
/**
* @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;
}
}
/**
* 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]));
}
}
/**
* Generate code to compute texture level of detail (lambda).
* \param ddx partial derivatives of (s, t, r, q) with respect to X
* \param ddy partial derivatives of (s, t, r, q) with respect to Y
* \param lod_bias optional float vector with the shader lod bias
* \param explicit_lod optional float vector with the explicit lod
* \param width scalar int texture width
* \param height scalar int texture height
* \param depth scalar int texture depth
*
* XXX: The resulting lod is scalar, so ignore all but the first element of
* derivatives, lod_bias, etc that are passed by the shader.
*/
void
lp_build_lod_selector(struct lp_build_sample_context *bld,
unsigned unit,
const LLVMValueRef ddx[4],
const LLVMValueRef ddy[4],
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
unsigned mip_filter,
LLVMValueRef *out_lod_ipart,
LLVMValueRef *out_lod_fpart)
{
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_build_context *float_bld = &bld->float_bld;
LLVMValueRef lod;
*out_lod_ipart = bld->int_bld.zero;
*out_lod_fpart = bld->float_bld.zero;
if (bld->static_state->min_max_lod_equal) {
/* User is forcing sampling from a particular mipmap level.
* This is hit during mipmap generation.
*/
LLVMValueRef min_lod =
bld->dynamic_state->min_lod(bld->dynamic_state, bld->gallivm, unit);
lod = min_lod;
}
else {
LLVMValueRef sampler_lod_bias =
bld->dynamic_state->lod_bias(bld->dynamic_state, bld->gallivm, unit);
LLVMValueRef index0 = lp_build_const_int32(bld->gallivm, 0);
if (explicit_lod) {
lod = LLVMBuildExtractElement(builder, explicit_lod,
index0, "");
}
else {
LLVMValueRef rho;
rho = lp_build_rho(bld, unit, ddx, ddy);
/*
* Compute lod = log2(rho)
*/
if (!lod_bias &&
!bld->static_state->lod_bias_non_zero &&
!bld->static_state->apply_max_lod &&
!bld->static_state->apply_min_lod) {
/*
* Special case when there are no post-log2 adjustments, which
* saves instructions but keeping the integer and fractional lod
* computations separate from the start.
*/
if (mip_filter == PIPE_TEX_MIPFILTER_NONE ||
mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
*out_lod_ipart = lp_build_ilog2(float_bld, rho);
*out_lod_fpart = bld->float_bld.zero;
return;
}
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR &&
!(gallivm_debug & GALLIVM_DEBUG_NO_BRILINEAR)) {
lp_build_brilinear_rho(float_bld, rho, BRILINEAR_FACTOR,
out_lod_ipart, out_lod_fpart);
return;
}
}
if (0) {
lod = lp_build_log2(float_bld, rho);
}
else {
lod = lp_build_fast_log2(float_bld, rho);
}
/* add shader lod bias */
if (lod_bias) {
lod_bias = LLVMBuildExtractElement(builder, lod_bias,
index0, "");
lod = LLVMBuildFAdd(builder, lod, lod_bias, "shader_lod_bias");
}
}
/* add sampler lod bias */
if (bld->static_state->lod_bias_non_zero)
lod = LLVMBuildFAdd(builder, lod, sampler_lod_bias, "sampler_lod_bias");
/* clamp lod */
if (bld->static_state->apply_max_lod) {
LLVMValueRef max_lod =
bld->dynamic_state->max_lod(bld->dynamic_state, bld->gallivm, unit);
lod = lp_build_min(float_bld, lod, max_lod);
}
if (bld->static_state->apply_min_lod) {
LLVMValueRef min_lod =
bld->dynamic_state->min_lod(bld->dynamic_state, bld->gallivm, unit);
lod = lp_build_max(float_bld, lod, min_lod);
}
}
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
if (!(gallivm_debug & GALLIVM_DEBUG_NO_BRILINEAR)) {
lp_build_brilinear_lod(float_bld, lod, BRILINEAR_FACTOR,
out_lod_ipart, out_lod_fpart);
}
else {
lp_build_ifloor_fract(float_bld, lod, out_lod_ipart, out_lod_fpart);
}
lp_build_name(*out_lod_fpart, "lod_fpart");
}
else {
*out_lod_ipart = lp_build_iround(float_bld, lod);
}
lp_build_name(*out_lod_ipart, "lod_ipart");
return;
}
/**
* Generate code to compute coordinate gradient (rho).
* \param ddx partial derivatives of (s, t, r, q) with respect to X
* \param ddy partial derivatives of (s, t, r, q) with respect to Y
*
* XXX: The resulting rho is scalar, so we ignore all but the first element of
* derivatives that are passed by the shader.
*/
static LLVMValueRef
lp_build_rho(struct lp_build_sample_context *bld,
unsigned unit,
const LLVMValueRef ddx[4],
const LLVMValueRef ddy[4])
{
struct lp_build_context *int_size_bld = &bld->int_size_bld;
struct lp_build_context *float_size_bld = &bld->float_size_bld;
struct lp_build_context *float_bld = &bld->float_bld;
const unsigned dims = bld->dims;
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
LLVMValueRef index1 = LLVMConstInt(i32t, 1, 0);
LLVMValueRef index2 = LLVMConstInt(i32t, 2, 0);
LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy;
LLVMValueRef rho_x, rho_y;
LLVMValueRef rho_vec;
LLVMValueRef int_size, float_size;
LLVMValueRef rho;
LLVMValueRef first_level, first_level_vec;
dsdx = ddx[0];
dsdy = ddy[0];
if (dims <= 1) {
rho_x = dsdx;
rho_y = dsdy;
}
else {
rho_x = float_size_bld->undef;
rho_y = float_size_bld->undef;
rho_x = LLVMBuildInsertElement(builder, rho_x, dsdx, index0, "");
rho_y = LLVMBuildInsertElement(builder, rho_y, dsdy, index0, "");
dtdx = ddx[1];
dtdy = ddy[1];
rho_x = LLVMBuildInsertElement(builder, rho_x, dtdx, index1, "");
rho_y = LLVMBuildInsertElement(builder, rho_y, dtdy, index1, "");
if (dims >= 3) {
drdx = ddx[2];
drdy = ddy[2];
rho_x = LLVMBuildInsertElement(builder, rho_x, drdx, index2, "");
rho_y = LLVMBuildInsertElement(builder, rho_y, drdy, index2, "");
}
}
rho_x = lp_build_abs(float_size_bld, rho_x);
rho_y = lp_build_abs(float_size_bld, rho_y);
rho_vec = lp_build_max(float_size_bld, rho_x, rho_y);
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
bld->gallivm, unit);
first_level_vec = lp_build_broadcast_scalar(&bld->int_size_bld, first_level);
int_size = lp_build_minify(int_size_bld, bld->int_size, first_level_vec);
float_size = lp_build_int_to_float(float_size_bld, int_size);
rho_vec = lp_build_mul(float_size_bld, rho_vec, float_size);
if (dims <= 1) {
rho = rho_vec;
}
else {
if (dims >= 2) {
LLVMValueRef rho_s, rho_t, rho_r;
rho_s = LLVMBuildExtractElement(builder, rho_vec, index0, "");
rho_t = LLVMBuildExtractElement(builder, rho_vec, index1, "");
rho = lp_build_max(float_bld, rho_s, rho_t);
if (dims >= 3) {
rho_r = LLVMBuildExtractElement(builder, rho_vec, index2, "");
rho = lp_build_max(float_bld, rho, rho_r);
}
}
}
return rho;
}
/**
* 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;
}
/**
* Register store.
*/
void
lp_emit_store_aos(
struct lp_build_tgsi_aos_context *bld,
const struct tgsi_full_instruction *inst,
unsigned index,
LLVMValueRef value)
{
LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
const struct tgsi_full_dst_register *reg = &inst->Dst[index];
LLVMValueRef mask = NULL;
LLVMValueRef ptr;
/*
* Saturate the value
*/
switch (inst->Instruction.Saturate) {
case TGSI_SAT_NONE:
break;
case TGSI_SAT_ZERO_ONE:
value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
break;
case TGSI_SAT_MINUS_PLUS_ONE:
value = lp_build_max(&bld->bld_base.base, value, lp_build_const_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, -1.0));
value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
break;
default:
assert(0);
}
/*
* Translate the register file
*/
assert(!reg->Register.Indirect);
switch (reg->Register.File) {
case TGSI_FILE_OUTPUT:
ptr = bld->outputs[reg->Register.Index];
break;
case TGSI_FILE_TEMPORARY:
ptr = bld->temps[reg->Register.Index];
break;
case TGSI_FILE_ADDRESS:
ptr = bld->addr[reg->Indirect.Index];
break;
case TGSI_FILE_PREDICATE:
ptr = bld->preds[reg->Register.Index];
break;
default:
assert(0);
return;
}
if (!ptr)
return;
/*
* Predicate
*/
if (inst->Instruction.Predicate) {
LLVMValueRef pred;
assert(inst->Predicate.Index < LP_MAX_TGSI_PREDS);
pred = LLVMBuildLoad(builder,
bld->preds[inst->Predicate.Index], "");
/*
* Convert the value to an integer mask.
*/
pred = lp_build_compare(bld->bld_base.base.gallivm,
bld->bld_base.base.type,
PIPE_FUNC_NOTEQUAL,
pred,
bld->bld_base.base.zero);
if (inst->Predicate.Negate) {
pred = LLVMBuildNot(builder, pred, "");
}
pred = bld->bld_base.emit_swizzle(&bld->bld_base, pred,
inst->Predicate.SwizzleX,
inst->Predicate.SwizzleY,
inst->Predicate.SwizzleZ,
inst->Predicate.SwizzleW);
if (mask) {
mask = LLVMBuildAnd(builder, mask, pred, "");
} else {
mask = pred;
}
}
/*
* Writemask
*/
if (reg->Register.WriteMask != TGSI_WRITEMASK_XYZW) {
LLVMValueRef writemask;
writemask = lp_build_const_mask_aos_swizzled(bld->bld_base.base.gallivm,
bld->bld_base.base.type,
reg->Register.WriteMask,
TGSI_NUM_CHANNELS,
bld->swizzles);
if (mask) {
mask = LLVMBuildAnd(builder, mask, writemask, "");
} else {
mask = writemask;
}
}
if (mask) {
LLVMValueRef orig_value;
orig_value = LLVMBuildLoad(builder, ptr, "");
value = lp_build_select(&bld->bld_base.base,
mask, value, orig_value);
}
LLVMBuildStore(builder, value, ptr);
}
/**
* 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;
}
/**
* 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]));
}
}
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;
}
/**
* Register store.
*/
void
lp_emit_store_aos(
struct lp_build_tgsi_aos_context *bld,
const struct tgsi_full_instruction *inst,
unsigned index,
LLVMValueRef value)
{
LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
const struct tgsi_full_dst_register *reg = &inst->Dst[index];
LLVMValueRef mask = NULL;
LLVMValueRef ptr;
/*
* Saturate the value
*/
if (inst->Instruction.Saturate) {
value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
}
/*
* Translate the register file
*/
assert(!reg->Register.Indirect);
switch (reg->Register.File) {
case TGSI_FILE_OUTPUT:
ptr = bld->outputs[reg->Register.Index];
break;
case TGSI_FILE_TEMPORARY:
ptr = bld->temps[reg->Register.Index];
break;
case TGSI_FILE_ADDRESS:
ptr = bld->addr[reg->Indirect.Index];
break;
default:
assert(0);
return;
}
if (!ptr)
return;
/*
* Writemask
*/
if (reg->Register.WriteMask != TGSI_WRITEMASK_XYZW) {
LLVMValueRef writemask;
writemask = lp_build_const_mask_aos_swizzled(bld->bld_base.base.gallivm,
bld->bld_base.base.type,
reg->Register.WriteMask,
TGSI_NUM_CHANNELS,
bld->swizzles);
if (mask) {
mask = LLVMBuildAnd(builder, mask, writemask, "");
} else {
mask = writemask;
}
}
if (mask) {
LLVMValueRef orig_value;
orig_value = LLVMBuildLoad(builder, ptr, "");
value = lp_build_select(&bld->bld_base.base,
mask, value, orig_value);
}
LLVMBuildStore(builder, value, ptr);
}
