static LLVMValueRef
translateFloatBinOp(NodeKind Op, LLVMValueRef ValueE1, LLVMValueRef ValueE2) {
switch (Op) {
case SumOp: return LLVMBuildFAdd(Builder, ValueE1, ValueE2, "");
case SubOp: return LLVMBuildFSub(Builder, ValueE1, ValueE2, "");
case MultOp: return LLVMBuildFMul(Builder, ValueE1, ValueE2, "");
case DivOp: return LLVMBuildFDiv(Builder, ValueE1, ValueE2, "");
case LtOp: return LLVMBuildFCmp(Builder, LLVMRealOLT, ValueE1, ValueE2, "");
case LeOp: return LLVMBuildFCmp(Builder, LLVMRealOLE, ValueE1, ValueE2, "");
case GtOp: return LLVMBuildFCmp(Builder, LLVMRealOGT, ValueE1, ValueE2, "");
case GeOp: return LLVMBuildFCmp(Builder, LLVMRealOGE, ValueE1, ValueE2, "");
case EqOp: return LLVMBuildFCmp(Builder, LLVMRealOEQ, ValueE1, ValueE2, "");
case DiffOp: return LLVMBuildFCmp(Builder, LLVMRealONE, ValueE1, ValueE2, "");
default: return NULL;
}
}
/**
* Special case for converting clamped IEEE-754 floats to unsigned norms.
*
* The mathematical voodoo below may seem excessive but it is actually
* paramount we do it this way for several reasons. First, there is no single
* precision FP to unsigned integer conversion Intel SSE instruction. Second,
* secondly, even if there was, since the FP's mantissa takes only a fraction
* of register bits the typically scale and cast approach would require double
* precision for accurate results, and therefore half the throughput
*
* Although the result values can be scaled to an arbitrary bit width specified
* by dst_width, the actual result type will have the same width.
*
* Ex: src = { float, float, float, float }
* return { i32, i32, i32, i32 } where each value is in [0, 2^dst_width-1].
*/
LLVMValueRef
lp_build_clamped_float_to_unsigned_norm(struct gallivm_state *gallivm,
struct lp_type src_type,
unsigned dst_width,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, src_type);
LLVMValueRef res;
unsigned mantissa;
assert(src_type.floating);
assert(dst_width <= src_type.width);
src_type.sign = FALSE;
mantissa = lp_mantissa(src_type);
if (dst_width <= mantissa) {
/*
* Apply magic coefficients that will make the desired result to appear
* in the lowest significant bits of the mantissa, with correct rounding.
*
* This only works if the destination width fits in the mantissa.
*/
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
ubound = (1ULL << dst_width);
mask = ubound - 1;
scale = (double)mask/ubound;
bias = (double)(1ULL << (mantissa - dst_width));
res = LLVMBuildFMul(builder, src, lp_build_const_vec(gallivm, src_type, scale), "");
res = LLVMBuildFAdd(builder, res, lp_build_const_vec(gallivm, src_type, bias), "");
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
res = LLVMBuildAnd(builder, res,
lp_build_const_int_vec(gallivm, src_type, mask), "");
}
else if (dst_width == (mantissa + 1)) {
/*
* The destination width matches exactly what can be represented in
* floating point (i.e., mantissa + 1 bits). So do a straight
* multiplication followed by casting. No further rounding is necessary.
*/
double scale;
scale = (double)((1ULL << dst_width) - 1);
res = LLVMBuildFMul(builder, src,
lp_build_const_vec(gallivm, src_type, scale), "");
res = LLVMBuildFPToSI(builder, res, int_vec_type, "");
}
else {
/*
* The destination exceeds what can be represented in the floating point.
* So multiply by the largest power two we get away with, and when
* subtract the most significant bit to rescale to normalized values.
*
* The largest power of two factor we can get away is
* (1 << (src_type.width - 1)), because we need to use signed . In theory it
* should be (1 << (src_type.width - 2)), but IEEE 754 rules states
* INT_MIN should be returned in FPToSI, which is the correct result for
* values near 1.0!
*
* This means we get (src_type.width - 1) correct bits for values near 0.0,
* and (mantissa + 1) correct bits for values near 1.0. Equally or more
* important, we also get exact results for 0.0 and 1.0.
*/
unsigned n = MIN2(src_type.width - 1, dst_width);
double scale = (double)(1ULL << n);
unsigned lshift = dst_width - n;
unsigned rshift = n;
LLVMValueRef lshifted;
LLVMValueRef rshifted;
res = LLVMBuildFMul(builder, src,
lp_build_const_vec(gallivm, src_type, scale), "");
res = LLVMBuildFPToSI(builder, res, int_vec_type, "");
/*
* Align the most significant bit to its final place.
*
* This will cause 1.0 to overflow to 0, but the later adjustment will
* get it right.
*/
if (lshift) {
lshifted = LLVMBuildShl(builder, res,
lp_build_const_int_vec(gallivm, src_type,
lshift), "");
} else {
lshifted = res;
}
/*
* Align the most significant bit to the right.
*/
rshifted = LLVMBuildLShr(builder, res,
lp_build_const_int_vec(gallivm, src_type, rshift),
"");
/*
* Subtract the MSB to the LSB, therefore re-scaling from
* (1 << dst_width) to ((1 << dst_width) - 1).
*/
res = LLVMBuildSub(builder, lshifted, rshifted, "");
}
return res;
}
/**
* 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;
}
/**
* Initialize the bld->a, dadq fields. This involves fetching
* those values from the arrays which are passed into the JIT function.
*/
static void
coeffs_init(struct lp_build_interp_soa_context *bld,
LLVMValueRef a0_ptr,
LLVMValueRef dadx_ptr,
LLVMValueRef dady_ptr)
{
struct lp_build_context *coeff_bld = &bld->coeff_bld;
struct lp_build_context *setup_bld = &bld->setup_bld;
struct gallivm_state *gallivm = coeff_bld->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef pixoffx, pixoffy;
unsigned attrib;
unsigned chan;
unsigned i;
pixoffx = coeff_bld->undef;
pixoffy = coeff_bld->undef;
for (i = 0; i < coeff_bld->type.length; i++) {
LLVMValueRef nr = lp_build_const_int32(gallivm, i);
LLVMValueRef pixxf = lp_build_const_float(gallivm, quad_offset_x[i]);
LLVMValueRef pixyf = lp_build_const_float(gallivm, quad_offset_y[i]);
pixoffx = LLVMBuildInsertElement(builder, pixoffx, pixxf, nr, "");
pixoffy = LLVMBuildInsertElement(builder, pixoffy, pixyf, nr, "");
}
for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
const unsigned mask = bld->mask[attrib];
const unsigned interp = bld->interp[attrib];
LLVMValueRef index = lp_build_const_int32(gallivm,
attrib * TGSI_NUM_CHANNELS);
LLVMValueRef ptr;
LLVMValueRef dadxaos = setup_bld->zero;
LLVMValueRef dadyaos = setup_bld->zero;
LLVMValueRef a0aos = setup_bld->zero;
/* always fetch all 4 values for performance/simplicity */
switch (interp) {
case LP_INTERP_PERSPECTIVE:
/* fall-through */
case LP_INTERP_LINEAR:
ptr = LLVMBuildGEP(builder, dadx_ptr, &index, 1, "");
ptr = LLVMBuildBitCast(builder, ptr,
LLVMPointerType(setup_bld->vec_type, 0), "");
dadxaos = LLVMBuildLoad(builder, ptr, "");
ptr = LLVMBuildGEP(builder, dady_ptr, &index, 1, "");
ptr = LLVMBuildBitCast(builder, ptr,
LLVMPointerType(setup_bld->vec_type, 0), "");
dadyaos = LLVMBuildLoad(builder, ptr, "");
attrib_name(dadxaos, attrib, 0, ".dadxaos");
attrib_name(dadyaos, attrib, 0, ".dadyaos");
/* fall-through */
case LP_INTERP_CONSTANT:
case LP_INTERP_FACING:
ptr = LLVMBuildGEP(builder, a0_ptr, &index, 1, "");
ptr = LLVMBuildBitCast(builder, ptr,
LLVMPointerType(setup_bld->vec_type, 0), "");
a0aos = LLVMBuildLoad(builder, ptr, "");
attrib_name(a0aos, attrib, 0, ".a0aos");
break;
case LP_INTERP_POSITION:
/* Nothing to do as the position coeffs are already setup in slot 0 */
continue;
default:
assert(0);
break;
}
/*
* a = a0 + (x * dadx + y * dady)
* a0aos is the attrib value at top left corner of stamp
*/
if (interp != LP_INTERP_CONSTANT &&
interp != LP_INTERP_FACING) {
LLVMValueRef x = lp_build_broadcast_scalar(setup_bld, bld->x);
LLVMValueRef y = lp_build_broadcast_scalar(setup_bld, bld->y);
a0aos = lp_build_fmuladd(builder, x, dadxaos, a0aos);
a0aos = lp_build_fmuladd(builder, y, dadyaos, a0aos);
}
/*
* dadq = {0, dadx, dady, dadx + dady}
* for two quads (side by side) this is:
* {0, dadx, dady, dadx+dady, 2*dadx, 2*dadx+dady, 3*dadx+dady}
*/
for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
/* this generates a CRAPLOAD of shuffles... */
if (mask & (1 << chan)) {
LLVMValueRef dadx, dady;
LLVMValueRef dadq, dadq2;
LLVMValueRef a;
LLVMValueRef chan_index = lp_build_const_int32(gallivm, chan);
if (attrib == 0 && chan == 0) {
a = bld->x;
if (bld->pos_offset) {
a = LLVMBuildFAdd(builder, a, lp_build_const_float(gallivm, bld->pos_offset), "");
}
a = lp_build_broadcast_scalar(coeff_bld, a);
dadx = coeff_bld->one;
dady = coeff_bld->zero;
}
else if (attrib == 0 && chan == 1) {
a = bld->y;
if (bld->pos_offset) {
a = LLVMBuildFAdd(builder, a, lp_build_const_float(gallivm, bld->pos_offset), "");
}
a = lp_build_broadcast_scalar(coeff_bld, a);
dady = coeff_bld->one;
dadx = coeff_bld->zero;
}
else {
dadx = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, dadxaos, chan_index);
dady = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, dadyaos, chan_index);
/*
* a = {a, a, a, a}
*/
a = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, a0aos, chan_index);
}
dadx = LLVMBuildFMul(builder, dadx, pixoffx, "");
dady = LLVMBuildFMul(builder, dady, pixoffy, "");
dadq = LLVMBuildFAdd(builder, dadx, dady, "");
/*
* Compute the attrib values on the upper-left corner of each
* group of quads.
* Note that if we process 2 quads at once this doesn't
* really exactly to what we want.
* We need to access elem 0 and 2 respectively later if we process
* 2 quads at once.
*/
if (interp != LP_INTERP_CONSTANT &&
interp != LP_INTERP_FACING) {
dadq2 = LLVMBuildFAdd(builder, dadq, dadq, "");
a = LLVMBuildFAdd(builder, a, dadq2, "");
}
#if PERSPECTIVE_DIVIDE_PER_QUAD
/*
* a *= 1 / w
*/
/*
* XXX since we're only going to access elements 0,2 out of 8
* if we have 8-wide vectors we should do the division only 4-wide.
* a is really a 2-elements in a 4-wide vector disguised as 8-wide
* in this case.
*/
if (interp == LP_INTERP_PERSPECTIVE) {
LLVMValueRef w = bld->a[0][3];
assert(attrib != 0);
assert(bld->mask[0] & TGSI_WRITEMASK_W);
if (!bld->oow) {
bld->oow = lp_build_rcp(coeff_bld, w);
lp_build_name(bld->oow, "oow");
}
a = lp_build_mul(coeff_bld, a, bld->oow);
}
#endif
attrib_name(a, attrib, chan, ".a");
attrib_name(dadq, attrib, chan, ".dadq");
bld->a[attrib][chan] = lp_build_alloca(gallivm,
LLVMTypeOf(a), "");
LLVMBuildStore(builder, a, bld->a[attrib][chan]);
bld->dadq[attrib][chan] = dadq;
}
}
}
}
/**
* Interpolate the shader input attribute values.
* This is called for each (group of) quad(s).
*/
static void
attribs_update_simple(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;
struct lp_build_context *setup_bld = &bld->setup_bld;
LLVMValueRef oow = NULL;
unsigned attrib;
LLVMValueRef pixoffx;
LLVMValueRef pixoffy;
LLVMValueRef ptr;
/* could do this with code-generated passed in pixel offsets too */
assert(loop_iter);
ptr = LLVMBuildGEP(builder, bld->xoffset_store, &loop_iter, 1, "");
pixoffx = LLVMBuildLoad(builder, ptr, "");
ptr = LLVMBuildGEP(builder, bld->yoffset_store, &loop_iter, 1, "");
pixoffy = LLVMBuildLoad(builder, ptr, "");
pixoffx = LLVMBuildFAdd(builder, pixoffx,
lp_build_broadcast_scalar(coeff_bld, bld->x), "");
pixoffy = LLVMBuildFAdd(builder, pixoffy,
lp_build_broadcast_scalar(coeff_bld, bld->y), "");
for (attrib = start; attrib < end; attrib++) {
const unsigned mask = bld->mask[attrib];
const unsigned interp = bld->interp[attrib];
unsigned chan;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
if (mask & (1 << chan)) {
LLVMValueRef index;
LLVMValueRef dadx = coeff_bld->zero;
LLVMValueRef dady = coeff_bld->zero;
LLVMValueRef a = coeff_bld->zero;
index = lp_build_const_int32(gallivm, chan);
switch (interp) {
case LP_INTERP_PERSPECTIVE:
/* fall-through */
case LP_INTERP_LINEAR:
if (attrib == 0 && chan == 0) {
dadx = coeff_bld->one;
if (bld->pos_offset) {
a = lp_build_const_vec(gallivm, coeff_bld->type, bld->pos_offset);
}
}
else if (attrib == 0 && chan == 1) {
dady = coeff_bld->one;
if (bld->pos_offset) {
a = lp_build_const_vec(gallivm, coeff_bld->type, bld->pos_offset);
}
}
else {
dadx = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, bld->dadxaos[attrib],
index);
dady = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, bld->dadyaos[attrib],
index);
a = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, bld->a0aos[attrib],
index);
}
/*
* a = a0 + (x * dadx + y * dady)
*/
a = lp_build_fmuladd(builder, dadx, pixoffx, a);
a = lp_build_fmuladd(builder, dady, pixoffy, a);
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);
}
break;
case LP_INTERP_CONSTANT:
case LP_INTERP_FACING:
a = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, bld->a0aos[attrib],
index);
break;
case LP_INTERP_POSITION:
assert(attrib > 0);
a = bld->attribs[0][chan];
break;
default:
assert(0);
break;
}
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);
}
bld->attribs[attrib][chan] = a;
}
}
}
}
/* Perform view culling and small primitive elimination and return true
* if the primitive is accepted and initially_accepted == true. */
static LLVMValueRef cull_bbox(struct ac_llvm_context *ctx,
LLVMValueRef pos[3][4],
LLVMValueRef initially_accepted,
struct ac_position_w_info *w,
LLVMValueRef vp_scale[2],
LLVMValueRef vp_translate[2],
LLVMValueRef small_prim_precision,
bool cull_view_xy,
bool cull_view_near_z,
bool cull_view_far_z,
bool cull_small_prims,
bool use_halfz_clip_space)
{
LLVMBuilderRef builder = ctx->builder;
if (!cull_view_xy && !cull_view_near_z && !cull_view_far_z && !cull_small_prims)
return ctx->i1true;
/* Skip the culling if the primitive has already been rejected or
* if any W is negative. The bounding box culling doesn't work when
* W is negative.
*/
LLVMValueRef cond = LLVMBuildAnd(builder, initially_accepted,
w->all_w_positive, "");
LLVMValueRef accepted_var = ac_build_alloca_undef(ctx, ctx->i1, "");
LLVMBuildStore(builder, initially_accepted, accepted_var);
ac_build_ifcc(ctx, cond, 10000000 /* does this matter? */);
{
LLVMValueRef bbox_min[3], bbox_max[3];
LLVMValueRef accepted = initially_accepted;
/* Compute the primitive bounding box for easy culling. */
for (unsigned chan = 0; chan < 3; chan++) {
bbox_min[chan] = ac_build_fmin(ctx, pos[0][chan], pos[1][chan]);
bbox_min[chan] = ac_build_fmin(ctx, bbox_min[chan], pos[2][chan]);
bbox_max[chan] = ac_build_fmax(ctx, pos[0][chan], pos[1][chan]);
bbox_max[chan] = ac_build_fmax(ctx, bbox_max[chan], pos[2][chan]);
}
/* View culling. */
if (cull_view_xy || cull_view_near_z || cull_view_far_z) {
for (unsigned chan = 0; chan < 3; chan++) {
LLVMValueRef visible;
if ((cull_view_xy && chan <= 1) ||
(cull_view_near_z && chan == 2)) {
float t = chan == 2 && use_halfz_clip_space ? 0 : -1;
visible = LLVMBuildFCmp(builder, LLVMRealOGE, bbox_max[chan],
LLVMConstReal(ctx->f32, t), "");
accepted = LLVMBuildAnd(builder, accepted, visible, "");
}
if ((cull_view_xy && chan <= 1) ||
(cull_view_far_z && chan == 2)) {
visible = LLVMBuildFCmp(builder, LLVMRealOLE, bbox_min[chan],
ctx->f32_1, "");
accepted = LLVMBuildAnd(builder, accepted, visible, "");
}
}
}
/* Small primitive elimination. */
if (cull_small_prims) {
/* Assuming a sample position at (0.5, 0.5), if we round
* the bounding box min/max extents and the results of
* the rounding are equal in either the X or Y direction,
* the bounding box does not intersect the sample.
*
* See these GDC slides for pictures:
* https://frostbite-wp-prd.s3.amazonaws.com/wp-content/uploads/2016/03/29204330/GDC_2016_Compute.pdf
*/
LLVMValueRef min, max, not_equal[2], visible;
for (unsigned chan = 0; chan < 2; chan++) {
/* Convert the position to screen-space coordinates. */
min = ac_build_fmad(ctx, bbox_min[chan],
vp_scale[chan], vp_translate[chan]);
max = ac_build_fmad(ctx, bbox_max[chan],
vp_scale[chan], vp_translate[chan]);
/* Scale the bounding box according to the precision of
* the rasterizer and the number of MSAA samples. */
min = LLVMBuildFSub(builder, min, small_prim_precision, "");
max = LLVMBuildFAdd(builder, max, small_prim_precision, "");
/* Determine if the bbox intersects the sample point.
* It also works for MSAA, but vp_scale, vp_translate,
* and small_prim_precision are computed differently.
*/
min = ac_build_round(ctx, min);
max = ac_build_round(ctx, max);
not_equal[chan] = LLVMBuildFCmp(builder, LLVMRealONE, min, max, "");
}
visible = LLVMBuildAnd(builder, not_equal[0], not_equal[1], "");
accepted = LLVMBuildAnd(builder, accepted, visible, "");
}
LLVMBuildStore(builder, accepted, accepted_var);
}
ac_build_endif(ctx, 10000000);
return LLVMBuildLoad(builder, accepted_var, "");
}
void si_prepare_cube_coords(struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data,
LLVMValueRef *coords_arg,
LLVMValueRef *derivs_arg)
{
unsigned target = emit_data->inst->Texture.Texture;
unsigned opcode = emit_data->inst->Instruction.Opcode;
struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef coords[4];
unsigned i;
si_llvm_cube_to_2d_coords(bld_base, coords_arg, coords);
if (opcode == TGSI_OPCODE_TXD && derivs_arg) {
LLVMValueRef derivs[4];
int axis;
/* Convert cube derivatives to 2D derivatives. */
for (axis = 0; axis < 2; axis++) {
LLVMValueRef shifted_cube_coords[4], shifted_coords[4];
/* Shift the cube coordinates by the derivatives to get
* the cube coordinates of the "neighboring pixel".
*/
for (i = 0; i < 3; i++)
shifted_cube_coords[i] =
LLVMBuildFAdd(builder, coords_arg[i],
derivs_arg[axis*3+i], "");
shifted_cube_coords[3] = LLVMGetUndef(bld_base->base.elem_type);
/* Project the shifted cube coordinates onto the face. */
si_llvm_cube_to_2d_coords(bld_base, shifted_cube_coords,
shifted_coords);
/* Subtract both sets of 2D coordinates to get 2D derivatives.
* This won't work if the shifted coordinates ended up
* in a different face.
*/
for (i = 0; i < 2; i++)
derivs[axis * 2 + i] =
LLVMBuildFSub(builder, shifted_coords[i],
coords[i], "");
}
memcpy(derivs_arg, derivs, sizeof(derivs));
}
if (target == TGSI_TEXTURE_CUBE_ARRAY ||
target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
/* for cube arrays coord.z = coord.w(array_index) * 8 + face */
/* coords_arg.w component - array_index for cube arrays */
coords[2] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
coords_arg[3], lp_build_const_float(gallivm, 8.0), coords[2]);
}
/* Preserve compare/lod/bias. Put it in coords.w. */
if (opcode == TGSI_OPCODE_TEX2 ||
opcode == TGSI_OPCODE_TXB2 ||
opcode == TGSI_OPCODE_TXL2) {
coords[3] = coords_arg[4];
} else if (opcode == TGSI_OPCODE_TXB ||
opcode == TGSI_OPCODE_TXL ||
target == TGSI_TEXTURE_SHADOWCUBE) {
coords[3] = coords_arg[3];
}
memcpy(coords_arg, coords, sizeof(coords));
}
/**
* Special case for converting clamped IEEE-754 floats to unsigned norms.
*
* The mathematical voodoo below may seem excessive but it is actually
* paramount we do it this way for several reasons. First, there is no single
* precision FP to unsigned integer conversion Intel SSE instruction. Second,
* secondly, even if there was, since the FP's mantissa takes only a fraction
* of register bits the typically scale and cast approach would require double
* precision for accurate results, and therefore half the throughput
*
* Although the result values can be scaled to an arbitrary bit width specified
* by dst_width, the actual result type will have the same width.
*
* Ex: src = { float, float, float, float }
* return { i32, i32, i32, i32 } where each value is in [0, 2^dst_width-1].
*/
LLVMValueRef
lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
struct lp_type src_type,
unsigned dst_width,
LLVMValueRef src)
{
LLVMTypeRef int_vec_type = lp_build_int_vec_type(src_type);
LLVMValueRef res;
unsigned mantissa;
unsigned n;
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
assert(src_type.floating);
mantissa = lp_mantissa(src_type);
/* We cannot carry more bits than the mantissa */
n = MIN2(mantissa, dst_width);
/* This magic coefficients will make the desired result to appear in the
* lowest significant bits of the mantissa.
*/
ubound = ((unsigned long long)1 << n);
mask = ubound - 1;
scale = (double)mask/ubound;
bias = (double)((unsigned long long)1 << (mantissa - n));
res = LLVMBuildFMul(builder, src, lp_build_const_vec(src_type, scale), "");
res = LLVMBuildFAdd(builder, res, lp_build_const_vec(src_type, bias), "");
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
if(dst_width > n) {
int shift = dst_width - n;
res = LLVMBuildShl(builder, res, lp_build_const_int_vec(src_type, shift), "");
/* TODO: Fill in the empty lower bits for additional precision? */
/* YES: this fixes progs/trivial/tri-z-eq.c.
* Otherwise vertex Z=1.0 values get converted to something like
* 0xfffffb00 and the test for equality with 0xffffffff fails.
*/
#if 0
{
LLVMValueRef msb;
msb = LLVMBuildLShr(builder, res, lp_build_const_int_vec(src_type, dst_width - 1), "");
msb = LLVMBuildShl(builder, msb, lp_build_const_int_vec(src_type, shift), "");
msb = LLVMBuildSub(builder, msb, lp_build_const_int_vec(src_type, 1), "");
res = LLVMBuildOr(builder, res, msb, "");
}
#elif 0
while(shift > 0) {
res = LLVMBuildOr(builder, res, LLVMBuildLShr(builder, res, lp_build_const_int_vec(src_type, n), ""), "");
shift -= n;
n *= 2;
}
#endif
}
else
res = LLVMBuildAnd(builder, res, lp_build_const_int_vec(src_type, mask), "");
return res;
}
void
ac_prepare_cube_coords(struct ac_llvm_context *ctx,
bool is_deriv, bool is_array,
LLVMValueRef *coords_arg,
LLVMValueRef *derivs_arg)
{
LLVMBuilderRef builder = ctx->builder;
struct cube_selection_coords selcoords;
LLVMValueRef coords[3];
LLVMValueRef invma;
build_cube_intrinsic(ctx, coords_arg, &selcoords);
invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
for (int i = 0; i < 2; ++i)
coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
coords[2] = selcoords.id;
if (is_deriv && derivs_arg) {
LLVMValueRef derivs[4];
int axis;
/* Convert cube derivatives to 2D derivatives. */
for (axis = 0; axis < 2; axis++) {
LLVMValueRef deriv_st[2];
LLVMValueRef deriv_ma;
/* Transform the derivative alongside the texture
* coordinate. Mathematically, the correct formula is
* as follows. Assume we're projecting onto the +Z face
* and denote by dx/dh the derivative of the (original)
* X texture coordinate with respect to horizontal
* window coordinates. The projection onto the +Z face
* plane is:
*
* f(x,z) = x/z
*
* Then df/dh = df/dx * dx/dh + df/dz * dz/dh
* = 1/z * dx/dh - x/z * 1/z * dz/dh.
*
* This motivatives the implementation below.
*
* Whether this actually gives the expected results for
* apps that might feed in derivatives obtained via
* finite differences is anyone's guess. The OpenGL spec
* seems awfully quiet about how textureGrad for cube
* maps should be handled.
*/
build_cube_select(builder, &selcoords, &derivs_arg[axis * 3],
deriv_st, &deriv_ma);
deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
for (int i = 0; i < 2; ++i)
derivs[axis * 2 + i] =
LLVMBuildFSub(builder,
LLVMBuildFMul(builder, deriv_st[i], invma, ""),
LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
}
memcpy(derivs_arg, derivs, sizeof(derivs));
}
/* Shift the texture coordinate. This must be applied after the
* derivative calculation.
*/
for (int i = 0; i < 2; ++i)
coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
if (is_array) {
/* for cube arrays coord.z = coord.w(array_index) * 8 + face */
/* coords_arg.w component - array_index for cube arrays */
LLVMValueRef tmp = LLVMBuildFMul(ctx->builder, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), "");
coords[2] = LLVMBuildFAdd(ctx->builder, tmp, coords[2], "");
}
memcpy(coords_arg, coords, sizeof(coords));
}
