/**
* Set the sign of float vector 'a' according to 'sign'.
* If sign==0, return abs(a).
* If sign==1, return -abs(a);
* Other values for sign produce undefined results.
*/
LLVMValueRef
lp_build_set_sign(struct lp_build_context *bld,
LLVMValueRef a, LLVMValueRef sign)
{
const struct lp_type type = bld->type;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMValueRef shift = lp_build_int_const_scalar(type, type.width - 1);
LLVMValueRef mask = lp_build_int_const_scalar(type,
~((unsigned long long) 1 << (type.width - 1)));
LLVMValueRef val, res;
assert(type.floating);
/* val = reinterpret_cast<int>(a) */
val = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
/* val = val & mask */
val = LLVMBuildAnd(bld->builder, val, mask, "");
/* sign = sign << shift */
sign = LLVMBuildShl(bld->builder, sign, shift, "");
/* res = val | sign */
res = LLVMBuildOr(bld->builder, val, sign, "");
/* res = reinterpret_cast<float>(res) */
res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
return res;
}
/**
* Generate abs(a)
*/
LLVMValueRef
lp_build_abs(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
LLVMTypeRef vec_type = lp_build_vec_type(type);
if(!type.sign)
return a;
if(type.floating) {
/* Mask out the sign bit */
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
unsigned long long absMask = ~(1ULL << (type.width - 1));
LLVMValueRef mask = lp_build_int_const_scalar(type, ((unsigned long long) absMask));
a = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
a = LLVMBuildAnd(bld->builder, a, mask, "");
a = LLVMBuildBitCast(bld->builder, a, vec_type, "");
return a;
}
if(type.width*type.length == 128 && util_cpu_caps.has_ssse3) {
switch(type.width) {
case 8:
return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.b.128", vec_type, a);
case 16:
return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.w.128", vec_type, a);
case 32:
return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.d.128", vec_type, a);
}
}
return lp_build_max(bld, a, LLVMBuildNeg(bld->builder, a, ""));
}
/**
* Convert a vector of rgba8 values into 32bit wide SoA vectors.
*
* \param dst_type The desired return type. For pure integer formats
* this should be a 32bit wide int or uint vector type,
* otherwise a float vector type.
*
* \param packed The rgba8 values to pack.
*
* \param rgba The 4 SoA return vectors.
*/
void
lp_build_rgba8_to_fi32_soa(struct gallivm_state *gallivm,
struct lp_type dst_type,
LLVMValueRef packed,
LLVMValueRef *rgba)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef mask = lp_build_const_int_vec(gallivm, dst_type, 0xff);
unsigned chan;
/* XXX technically shouldn't use that for uint dst_type */
packed = LLVMBuildBitCast(builder, packed,
lp_build_int_vec_type(gallivm, dst_type), "");
/* Decode the input vector components */
for (chan = 0; chan < 4; ++chan) {
unsigned start = chan*8;
unsigned stop = start + 8;
LLVMValueRef input;
input = packed;
if (start)
input = LLVMBuildLShr(builder, input,
lp_build_const_int_vec(gallivm, dst_type, start), "");
if (stop < 32)
input = LLVMBuildAnd(builder, input, mask, "");
if (dst_type.floating)
input = lp_build_unsigned_norm_to_float(gallivm, 8, dst_type, input);
rgba[chan] = input;
}
}
/**
* 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.
*/
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 = LLVMBuildMul(builder, src, lp_build_const_scalar(src_type, scale), "");
res = LLVMBuildAdd(builder, res, lp_build_const_scalar(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_int_const_scalar(src_type, shift), "");
/* TODO: Fill in the empty lower bits for additional precision? */
#if 0
{
LLVMValueRef msb;
msb = LLVMBuildLShr(builder, res, lp_build_int_const_scalar(src_type, dst_width - 1), "");
msb = LLVMBuildShl(builder, msb, lp_build_int_const_scalar(src_type, shift), "");
msb = LLVMBuildSub(builder, msb, lp_build_int_const_scalar(src_type, 1), "");
res = LLVMBuildOr(builder, res, msb, "");
}
#elif 0
while(shift > 0) {
res = LLVMBuildOr(builder, res, LLVMBuildLShr(builder, res, lp_build_int_const_scalar(src_type, n), ""), "");
shift -= n;
n *= 2;
}
#endif
}
else
res = LLVMBuildAnd(builder, res, lp_build_int_const_scalar(src_type, mask), "");
return res;
}
void
lp_build_exp2_approx(struct lp_build_context *bld,
LLVMValueRef x,
LLVMValueRef *p_exp2_int_part,
LLVMValueRef *p_frac_part,
LLVMValueRef *p_exp2)
{
const struct lp_type type = bld->type;
LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef ipart = NULL;
LLVMValueRef fpart = NULL;
LLVMValueRef expipart = NULL;
LLVMValueRef expfpart = NULL;
LLVMValueRef res = NULL;
if(p_exp2_int_part || p_frac_part || p_exp2) {
/* TODO: optimize the constant case */
if(LLVMIsConstant(x))
debug_printf("%s: inefficient/imprecise constant arithmetic\n",
__FUNCTION__);
assert(type.floating && type.width == 32);
x = lp_build_min(bld, x, lp_build_const_scalar(type, 129.0));
x = lp_build_max(bld, x, lp_build_const_scalar(type, -126.99999));
/* ipart = int(x - 0.5) */
ipart = LLVMBuildSub(bld->builder, x, lp_build_const_scalar(type, 0.5f), "");
ipart = LLVMBuildFPToSI(bld->builder, ipart, int_vec_type, "");
/* fpart = x - ipart */
fpart = LLVMBuildSIToFP(bld->builder, ipart, vec_type, "");
fpart = LLVMBuildSub(bld->builder, x, fpart, "");
}
if(p_exp2_int_part || p_exp2) {
/* expipart = (float) (1 << ipart) */
expipart = LLVMBuildAdd(bld->builder, ipart, lp_build_int_const_scalar(type, 127), "");
expipart = LLVMBuildShl(bld->builder, expipart, lp_build_int_const_scalar(type, 23), "");
expipart = LLVMBuildBitCast(bld->builder, expipart, vec_type, "");
}
if(p_exp2) {
expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial,
Elements(lp_build_exp2_polynomial));
res = LLVMBuildMul(bld->builder, expipart, expfpart, "");
}
if(p_exp2_int_part)
*p_exp2_int_part = expipart;
if(p_frac_part)
*p_frac_part = fpart;
if(p_exp2)
*p_exp2 = res;
}
/**
* Generate color blending and color output.
*/
static void
generate_blend(const struct pipe_blend_state *blend,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef context_ptr,
LLVMValueRef mask,
LLVMValueRef *src,
LLVMValueRef dst_ptr)
{
struct lp_build_context bld;
struct lp_build_flow_context *flow;
struct lp_build_mask_context mask_ctx;
LLVMTypeRef vec_type;
LLVMTypeRef int_vec_type;
LLVMValueRef const_ptr;
LLVMValueRef con[4];
LLVMValueRef dst[4];
LLVMValueRef res[4];
unsigned chan;
lp_build_context_init(&bld, builder, type);
flow = lp_build_flow_create(builder);
/* we'll use this mask context to skip blending if all pixels are dead */
lp_build_mask_begin(&mask_ctx, flow, type, mask);
vec_type = lp_build_vec_type(type);
int_vec_type = lp_build_int_vec_type(type);
const_ptr = lp_jit_context_blend_color(builder, context_ptr);
const_ptr = LLVMBuildBitCast(builder, const_ptr,
LLVMPointerType(vec_type, 0), "");
for(chan = 0; chan < 4; ++chan) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
lp_build_name(con[chan], "con.%c", "rgba"[chan]);
lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
}
lp_build_blend_soa(builder, blend, type, src, dst, con, res);
for(chan = 0; chan < 4; ++chan) {
if(blend->rt[0].colormask & (1 << chan)) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
lp_build_name(res[chan], "res.%c", "rgba"[chan]);
res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
}
}
lp_build_mask_end(&mask_ctx);
lp_build_flow_destroy(flow);
}
/**
* Convert to integer, through whichever rounding method that's fastest,
* typically truncating toward zero.
*/
LLVMValueRef
lp_build_itrunc(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
assert(type.floating);
assert(lp_check_value(type, a));
return LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
}
/**
* Inverse of lp_build_clamped_float_to_unsigned_norm above.
* Ex: src = { i32, i32, i32, i32 } with values in range [0, 2^src_width-1]
* return {float, float, float, float} with values in range [0, 1].
*/
LLVMValueRef
lp_build_unsigned_norm_to_float(LLVMBuilderRef builder,
unsigned src_width,
struct lp_type dst_type,
LLVMValueRef src)
{
LLVMTypeRef vec_type = lp_build_vec_type(dst_type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(dst_type);
LLVMValueRef bias_;
LLVMValueRef res;
unsigned mantissa;
unsigned n;
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
assert(dst_type.floating);
mantissa = lp_mantissa(dst_type);
n = MIN2(mantissa, src_width);
ubound = ((unsigned long long)1 << n);
mask = ubound - 1;
scale = (double)ubound/mask;
bias = (double)((unsigned long long)1 << (mantissa - n));
res = src;
if(src_width > mantissa) {
int shift = src_width - mantissa;
res = LLVMBuildLShr(builder, res, lp_build_const_int_vec(dst_type, shift), "");
}
bias_ = lp_build_const_vec(dst_type, bias);
res = LLVMBuildOr(builder,
res,
LLVMBuildBitCast(builder, bias_, int_vec_type, ""), "");
res = LLVMBuildBitCast(builder, res, vec_type, "");
res = LLVMBuildFSub(builder, res, bias_, "");
res = LLVMBuildFMul(builder, res, lp_build_const_vec(dst_type, scale), "");
return res;
}
/**
* 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);
}
/**
* Convert float[] to int[] with floor().
*/
LLVMValueRef
lp_build_ifloor(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef res;
assert(type.floating);
assert(lp_check_value(type, a));
if(util_cpu_caps.has_sse4_1) {
res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
}
else {
/* Take the sign bit and add it to 1 constant */
LLVMTypeRef vec_type = lp_build_vec_type(type);
unsigned mantissa = lp_mantissa(type);
LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
LLVMValueRef sign;
LLVMValueRef offset;
/* sign = a < 0 ? ~0 : 0 */
sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
sign = LLVMBuildAnd(bld->builder, sign, mask, "");
sign = LLVMBuildAShr(bld->builder, sign, lp_build_int_const_scalar(type, type.width - 1), "");
lp_build_name(sign, "floor.sign");
/* offset = -0.99999(9)f */
offset = lp_build_const_scalar(type, -(double)(((unsigned long long)1 << mantissa) - 1)/((unsigned long long)1 << mantissa));
offset = LLVMConstBitCast(offset, int_vec_type);
/* offset = a < 0 ? -0.99999(9)f : 0.0f */
offset = LLVMBuildAnd(bld->builder, offset, sign, "");
offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "");
lp_build_name(offset, "floor.offset");
res = LLVMBuildAdd(bld->builder, a, offset, "");
lp_build_name(res, "floor.res");
}
res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
lp_build_name(res, "floor");
return res;
}
/**
* Begin a section of code which is predicated on a mask.
* \param mask the mask context, initialized here
* \param flow the flow context
* \param type the type of the mask
* \param value storage for the mask
*/
void
lp_build_mask_begin(struct lp_build_mask_context *mask,
struct gallivm_state *gallivm,
struct lp_type type,
LLVMValueRef value)
{
memset(mask, 0, sizeof *mask);
mask->reg_type = LLVMIntTypeInContext(gallivm->context, type.width * type.length);
mask->var = lp_build_alloca(gallivm,
lp_build_int_vec_type(gallivm, type),
"execution_mask");
LLVMBuildStore(gallivm->builder, value, mask->var);
lp_build_flow_skip_begin(&mask->skip, gallivm);
}
LLVMValueRef
lp_build_trunc(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
assert(type.floating);
assert(lp_check_value(type, a));
if(util_cpu_caps.has_sse4_1)
return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_TRUNCATE);
else {
LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef res;
res = LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
return res;
}
}
/**
* Return (mask & a) | (~mask & b);
*/
LLVMValueRef
lp_build_select_bitwise(struct lp_build_context *bld,
LLVMValueRef mask,
LLVMValueRef a,
LLVMValueRef b)
{
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_type type = bld->type;
LLVMValueRef res;
assert(lp_check_value(type, a));
assert(lp_check_value(type, b));
if (a == b) {
return a;
}
if(type.floating) {
LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->gallivm, type);
a = LLVMBuildBitCast(builder, a, int_vec_type, "");
b = LLVMBuildBitCast(builder, b, int_vec_type, "");
}
a = LLVMBuildAnd(builder, a, mask, "");
/* This often gets translated to PANDN, but sometimes the NOT is
* pre-computed and stored in another constant. The best strategy depends
* on available registers, so it is not a big deal -- hopefully LLVM does
* the right decision attending the rest of the program.
*/
b = LLVMBuildAnd(builder, b, LLVMBuildNot(builder, mask, ""), "");
res = LLVMBuildOr(builder, a, b, "");
if(type.floating) {
LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
res = LLVMBuildBitCast(builder, res, vec_type, "");
}
return res;
}
LLVMValueRef
lp_build_sgn(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMValueRef cond;
LLVMValueRef res;
/* Handle non-zero case */
if(!type.sign) {
/* if not zero then sign must be positive */
res = bld->one;
}
else if(type.floating) {
/* Take the sign bit and add it to 1 constant */
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
LLVMValueRef sign;
LLVMValueRef one;
sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
sign = LLVMBuildAnd(bld->builder, sign, mask, "");
one = LLVMConstBitCast(bld->one, int_vec_type);
res = LLVMBuildOr(bld->builder, sign, one, "");
res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
}
else
{
LLVMValueRef minus_one = lp_build_const_scalar(type, -1.0);
cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, bld->zero);
res = lp_build_select(bld, cond, bld->one, minus_one);
}
/* Handle zero */
cond = lp_build_cmp(bld, PIPE_FUNC_EQUAL, a, bld->zero);
res = lp_build_select(bld, cond, bld->zero, bld->one);
return res;
}
LLVMValueRef
lp_build_iceil(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef res;
assert(type.floating);
assert(lp_check_value(type, a));
if(util_cpu_caps.has_sse4_1) {
res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL);
}
else {
assert(0);
res = bld->undef;
}
res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
return res;
}
LLVMValueRef
lp_build_iround(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef res;
assert(type.floating);
assert(lp_check_value(type, a));
if(util_cpu_caps.has_sse4_1) {
res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST);
}
else {
LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
LLVMValueRef sign;
LLVMValueRef half;
/* get sign bit */
sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
sign = LLVMBuildAnd(bld->builder, sign, mask, "");
/* sign * 0.5 */
half = lp_build_const_scalar(type, 0.5);
half = LLVMBuildBitCast(bld->builder, half, int_vec_type, "");
half = LLVMBuildOr(bld->builder, sign, half, "");
half = LLVMBuildBitCast(bld->builder, half, vec_type, "");
res = LLVMBuildAdd(bld->builder, a, half, "");
}
res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
return res;
}
/**
* Store depth/stencil values.
* Incoming values are swizzled (typically n 2x2 quads), stored linear.
* If there's a mask it will do select/store otherwise just store.
*
* \param type the data type of the fragment depth/stencil values
* \param format_desc description of the depth/stencil surface
* \param mask the alive/dead pixel mask for the quad (vector)
* \param z_fb z values read from fb (with padding)
* \param s_fb s values read from fb (with padding)
* \param loop_counter the current loop iteration
* \param depth_ptr pointer to the depth/stencil values of this 4x4 block
* \param depth_stride stride of the depth/stencil buffer
* \param z_value the depth values to store (with padding)
* \param s_value the stencil values to store (with padding)
*/
void
lp_build_depth_stencil_write_swizzled(struct gallivm_state *gallivm,
struct lp_type z_src_type,
const struct util_format_description *format_desc,
struct lp_build_mask_context *mask,
LLVMValueRef z_fb,
LLVMValueRef s_fb,
LLVMValueRef loop_counter,
LLVMValueRef depth_ptr,
LLVMValueRef depth_stride,
LLVMValueRef z_value,
LLVMValueRef s_value)
{
struct lp_build_context z_bld;
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 4];
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef mask_value = NULL;
LLVMValueRef zs_dst1, zs_dst2;
LLVMValueRef zs_dst_ptr1, zs_dst_ptr2;
LLVMValueRef depth_offset1, depth_offset2;
LLVMTypeRef load_ptr_type;
unsigned depth_bytes = format_desc->block.bits / 8;
struct lp_type zs_type = lp_depth_type(format_desc, z_src_type.length);
struct lp_type z_type = zs_type;
struct lp_type zs_load_type = zs_type;
zs_load_type.length = zs_load_type.length / 2;
load_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0);
z_type.width = z_src_type.width;
lp_build_context_init(&z_bld, gallivm, z_type);
/*
* This is far from ideal, at least for late depth write we should do this
* outside the fs loop to avoid all the swizzle stuff.
*/
if (z_src_type.length == 4) {
LLVMValueRef looplsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
LLVMValueRef loopmsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 2), "");
LLVMValueRef offset2 = LLVMBuildMul(builder, loopmsb,
depth_stride, "");
depth_offset1 = LLVMBuildMul(builder, looplsb,
lp_build_const_int32(gallivm, depth_bytes * 2), "");
depth_offset1 = LLVMBuildAdd(builder, depth_offset1, offset2, "");
}
else {
unsigned i;
LLVMValueRef loopx2 = LLVMBuildShl(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
assert(z_src_type.length == 8);
depth_offset1 = LLVMBuildMul(builder, loopx2, depth_stride, "");
/*
* We load 2x4 values, and need to swizzle them (order
* 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
*/
for (i = 0; i < 8; i++) {
shuffles[i] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
}
}
depth_offset2 = LLVMBuildAdd(builder, depth_offset1, depth_stride, "");
zs_dst_ptr1 = LLVMBuildGEP(builder, depth_ptr, &depth_offset1, 1, "");
zs_dst_ptr1 = LLVMBuildBitCast(builder, zs_dst_ptr1, load_ptr_type, "");
zs_dst_ptr2 = LLVMBuildGEP(builder, depth_ptr, &depth_offset2, 1, "");
zs_dst_ptr2 = LLVMBuildBitCast(builder, zs_dst_ptr2, load_ptr_type, "");
if (format_desc->block.bits > 32) {
s_value = LLVMBuildBitCast(builder, s_value, z_bld.vec_type, "");
}
if (mask) {
mask_value = lp_build_mask_value(mask);
z_value = lp_build_select(&z_bld, mask_value, z_value, z_fb);
if (format_desc->block.bits > 32) {
s_fb = LLVMBuildBitCast(builder, s_fb, z_bld.vec_type, "");
s_value = lp_build_select(&z_bld, mask_value, s_value, s_fb);
}
}
if (zs_type.width < z_src_type.width) {
/* Truncate ZS values (e.g., when writing to Z16_UNORM) */
z_value = LLVMBuildTrunc(builder, z_value,
lp_build_int_vec_type(gallivm, zs_type), "");
}
if (format_desc->block.bits <= 32) {
if (z_src_type.length == 4) {
zs_dst1 = lp_build_extract_range(gallivm, z_value, 0, 2);
zs_dst2 = lp_build_extract_range(gallivm, z_value, 2, 2);
}
else {
assert(z_src_type.length == 8);
zs_dst1 = LLVMBuildShuffleVector(builder, z_value, z_value,
LLVMConstVector(&shuffles[0],
zs_load_type.length), "");
zs_dst2 = LLVMBuildShuffleVector(builder, z_value, z_value,
LLVMConstVector(&shuffles[4],
zs_load_type.length), "");
}
}
else {
if (z_src_type.length == 4) {
zs_dst1 = lp_build_interleave2(gallivm, z_type,
z_value, s_value, 0);
zs_dst2 = lp_build_interleave2(gallivm, z_type,
z_value, s_value, 1);
}
else {
unsigned i;
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 2];
assert(z_src_type.length == 8);
for (i = 0; i < 8; i++) {
shuffles[i*2] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
shuffles[i*2+1] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2 +
z_src_type.length);
}
zs_dst1 = LLVMBuildShuffleVector(builder, z_value, s_value,
LLVMConstVector(&shuffles[0],
z_src_type.length), "");
zs_dst2 = LLVMBuildShuffleVector(builder, z_value, s_value,
LLVMConstVector(&shuffles[8],
z_src_type.length), "");
}
zs_dst1 = LLVMBuildBitCast(builder, zs_dst1,
lp_build_vec_type(gallivm, zs_load_type), "");
zs_dst2 = LLVMBuildBitCast(builder, zs_dst2,
lp_build_vec_type(gallivm, zs_load_type), "");
}
LLVMBuildStore(builder, zs_dst1, zs_dst_ptr1);
LLVMBuildStore(builder, zs_dst2, zs_dst_ptr2);
}
/**
* Load depth/stencil values.
* The stored values are linear, swizzle them.
*
* \param type the data type of the fragment depth/stencil values
* \param format_desc description of the depth/stencil surface
* \param loop_counter the current loop iteration
* \param depth_ptr pointer to the depth/stencil values of this 4x4 block
* \param depth_stride stride of the depth/stencil buffer
* \param z_fb contains z values loaded from fb (may include padding)
* \param s_fb contains s values loaded from fb (may include padding)
*/
void
lp_build_depth_stencil_load_swizzled(struct gallivm_state *gallivm,
struct lp_type z_src_type,
const struct util_format_description *format_desc,
LLVMValueRef depth_ptr,
LLVMValueRef depth_stride,
LLVMValueRef *z_fb,
LLVMValueRef *s_fb,
LLVMValueRef loop_counter)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 4];
LLVMValueRef zs_dst1, zs_dst2;
LLVMValueRef zs_dst_ptr;
LLVMValueRef depth_offset1, depth_offset2;
LLVMTypeRef load_ptr_type;
unsigned depth_bytes = format_desc->block.bits / 8;
struct lp_type zs_type = lp_depth_type(format_desc, z_src_type.length);
struct lp_type zs_load_type = zs_type;
zs_load_type.length = zs_load_type.length / 2;
load_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0);
if (z_src_type.length == 4) {
unsigned i;
LLVMValueRef looplsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
LLVMValueRef loopmsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 2), "");
LLVMValueRef offset2 = LLVMBuildMul(builder, loopmsb,
depth_stride, "");
depth_offset1 = LLVMBuildMul(builder, looplsb,
lp_build_const_int32(gallivm, depth_bytes * 2), "");
depth_offset1 = LLVMBuildAdd(builder, depth_offset1, offset2, "");
/* just concatenate the loaded 2x2 values into 4-wide vector */
for (i = 0; i < 4; i++) {
shuffles[i] = lp_build_const_int32(gallivm, i);
}
}
else {
unsigned i;
LLVMValueRef loopx2 = LLVMBuildShl(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
assert(z_src_type.length == 8);
depth_offset1 = LLVMBuildMul(builder, loopx2, depth_stride, "");
/*
* We load 2x4 values, and need to swizzle them (order
* 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
*/
for (i = 0; i < 8; i++) {
shuffles[i] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
}
}
depth_offset2 = LLVMBuildAdd(builder, depth_offset1, depth_stride, "");
/* Load current z/stencil values from z/stencil buffer */
zs_dst_ptr = LLVMBuildGEP(builder, depth_ptr, &depth_offset1, 1, "");
zs_dst_ptr = LLVMBuildBitCast(builder, zs_dst_ptr, load_ptr_type, "");
zs_dst1 = LLVMBuildLoad(builder, zs_dst_ptr, "");
zs_dst_ptr = LLVMBuildGEP(builder, depth_ptr, &depth_offset2, 1, "");
zs_dst_ptr = LLVMBuildBitCast(builder, zs_dst_ptr, load_ptr_type, "");
zs_dst2 = LLVMBuildLoad(builder, zs_dst_ptr, "");
*z_fb = LLVMBuildShuffleVector(builder, zs_dst1, zs_dst2,
LLVMConstVector(shuffles, zs_type.length), "");
*s_fb = *z_fb;
if (format_desc->block.bits < z_src_type.width) {
/* Extend destination ZS values (e.g., when reading from Z16_UNORM) */
*z_fb = LLVMBuildZExt(builder, *z_fb,
lp_build_int_vec_type(gallivm, z_src_type), "");
}
else if (format_desc->block.bits > 32) {
/* rely on llvm to handle too wide vector we have here nicely */
unsigned i;
struct lp_type typex2 = zs_type;
struct lp_type s_type = zs_type;
LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH / 4];
LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH / 4];
LLVMValueRef tmp;
typex2.width = typex2.width / 2;
typex2.length = typex2.length * 2;
s_type.width = s_type.width / 2;
s_type.floating = 0;
tmp = LLVMBuildBitCast(builder, *z_fb,
lp_build_vec_type(gallivm, typex2), "");
for (i = 0; i < zs_type.length; i++) {
shuffles1[i] = lp_build_const_int32(gallivm, i * 2);
shuffles2[i] = lp_build_const_int32(gallivm, i * 2 + 1);
}
*z_fb = LLVMBuildShuffleVector(builder, tmp, tmp,
LLVMConstVector(shuffles1, zs_type.length), "");
*s_fb = LLVMBuildShuffleVector(builder, tmp, tmp,
LLVMConstVector(shuffles2, zs_type.length), "");
*s_fb = LLVMBuildBitCast(builder, *s_fb,
lp_build_vec_type(gallivm, s_type), "");
lp_build_name(*s_fb, "s_dst");
}
lp_build_name(*z_fb, "z_dst");
lp_build_name(*s_fb, "s_dst");
lp_build_name(*z_fb, "z_dst");
}
/**
* See http://www.devmaster.net/forums/showthread.php?p=43580
*/
void
lp_build_log2_approx(struct lp_build_context *bld,
LLVMValueRef x,
LLVMValueRef *p_exp,
LLVMValueRef *p_floor_log2,
LLVMValueRef *p_log2)
{
const struct lp_type type = bld->type;
LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef expmask = lp_build_int_const_scalar(type, 0x7f800000);
LLVMValueRef mantmask = lp_build_int_const_scalar(type, 0x007fffff);
LLVMValueRef one = LLVMConstBitCast(bld->one, int_vec_type);
LLVMValueRef i = NULL;
LLVMValueRef exp = NULL;
LLVMValueRef mant = NULL;
LLVMValueRef logexp = NULL;
LLVMValueRef logmant = NULL;
LLVMValueRef res = NULL;
if(p_exp || p_floor_log2 || p_log2) {
/* TODO: optimize the constant case */
if(LLVMIsConstant(x))
debug_printf("%s: inefficient/imprecise constant arithmetic\n",
__FUNCTION__);
assert(type.floating && type.width == 32);
i = LLVMBuildBitCast(bld->builder, x, int_vec_type, "");
/* exp = (float) exponent(x) */
exp = LLVMBuildAnd(bld->builder, i, expmask, "");
}
if(p_floor_log2 || p_log2) {
logexp = LLVMBuildLShr(bld->builder, exp, lp_build_int_const_scalar(type, 23), "");
logexp = LLVMBuildSub(bld->builder, logexp, lp_build_int_const_scalar(type, 127), "");
logexp = LLVMBuildSIToFP(bld->builder, logexp, vec_type, "");
}
if(p_log2) {
/* mant = (float) mantissa(x) */
mant = LLVMBuildAnd(bld->builder, i, mantmask, "");
mant = LLVMBuildOr(bld->builder, mant, one, "");
mant = LLVMBuildBitCast(bld->builder, mant, vec_type, "");
logmant = lp_build_polynomial(bld, mant, lp_build_log2_polynomial,
Elements(lp_build_log2_polynomial));
/* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
logmant = LLVMBuildMul(bld->builder, logmant, LLVMBuildSub(bld->builder, mant, bld->one, ""), "");
res = LLVMBuildAdd(bld->builder, logmant, logexp, "");
}
if(p_exp)
*p_exp = exp;
if(p_floor_log2)
*p_floor_log2 = logexp;
if(p_log2)
*p_log2 = res;
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
* Note that the function which we generate operates on a block of 16
* pixels at at time. The block contains 2x2 quads. Each quad contains
* 2x2 pixels.
*/
static void
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
struct lp_fragment_shader_variant *variant,
unsigned do_tri_test)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
const struct lp_fragment_shader_variant_key *key = &variant->key;
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_vec_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
LLVMTypeRef arg_types[14];
LLVMTypeRef func_type;
LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
LLVMValueRef y0;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef blend_in_color[NUM_CHANNELS];
LLVMValueRef function;
unsigned num_fs;
unsigned i;
unsigned chan;
unsigned cbuf;
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 elements per vector */
num_fs = 4; /* number of quads per block */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_vec_type = lp_build_vec_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
blend_int_vec_type = lp_build_int_vec_type(blend_type);
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[6] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
arg_types[8] = LLVMInt32Type(); /* c0 */
arg_types[9] = LLVMInt32Type(); /* c1 */
arg_types[10] = LLVMInt32Type(); /* c2 */
/* Note: the step arrays are built as int32[16] but we interpret
* them here as int32_vec4[4].
*/
arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
function = LLVMAddFunction(screen->module, "shader", func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
variant->function[do_tri_test] = function;
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
*/
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(function, 0);
x = LLVMGetParam(function, 1);
y = LLVMGetParam(function, 2);
a0_ptr = LLVMGetParam(function, 3);
dadx_ptr = LLVMGetParam(function, 4);
dady_ptr = LLVMGetParam(function, 5);
color_ptr_ptr = LLVMGetParam(function, 6);
depth_ptr = LLVMGetParam(function, 7);
c0 = LLVMGetParam(function, 8);
c1 = LLVMGetParam(function, 9);
c2 = LLVMGetParam(function, 10);
step0_ptr = LLVMGetParam(function, 11);
step1_ptr = LLVMGetParam(function, 12);
step2_ptr = LLVMGetParam(function, 13);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr");
lp_build_name(depth_ptr, "depth");
lp_build_name(c0, "c0");
lp_build_name(c1, "c1");
lp_build_name(c2, "c2");
lp_build_name(step0_ptr, "step0");
lp_build_name(step1_ptr, "step1");
lp_build_name(step2_ptr, "step2");
/*
* Function body
*/
block = LLVMAppendBasicBlock(function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
generate_pos0(builder, x, y, &x0, &y0);
lp_build_interp_soa_init(&interp,
shader->base.tokens,
key->flatshade,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x0, y0);
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
/* loop over quads in the block */
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
int cbuf;
if(i != 0)
lp_build_interp_soa_update(&interp, i);
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i], /* output */
out_color,
depth_ptr_i,
do_tri_test,
c0, c1, c2,
step0_ptr, step1_ptr, step2_ptr);
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
}
sampler->destroy(sampler);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
LLVMValueRef color_ptr;
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[cbuf][chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
}
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
color_ptr = LLVMBuildLoad(builder,
LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
"");
lp_build_name(color_ptr, "color_ptr%d", cbuf);
/*
* Blending.
*/
generate_blend(&key->blend,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
}
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/* Verify the LLVM IR. If invalid, dump and abort */
#ifdef DEBUG
if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
if (1)
LLVMDumpValue(function);
abort();
}
#endif
/* Apply optimizations to LLVM IR */
if (1)
LLVMRunFunctionPassManager(screen->pass, function);
if (LP_DEBUG & DEBUG_JIT) {
/* Print the LLVM IR to stderr */
LLVMDumpValue(function);
debug_printf("\n");
}
/*
* Translate the LLVM IR into machine code.
*/
variant->jit_function[do_tri_test] = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, function);
if (LP_DEBUG & DEBUG_ASM)
lp_disassemble(variant->jit_function[do_tri_test]);
}
/**
* Generate the fragment shader, depth/stencil test, and alpha tests.
* \param i which quad in the tile, in range [0,3]
* \param do_tri_test if 1, do triangle edge in/out testing
*/
static void
generate_fs(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
const struct lp_fragment_shader_variant_key *key,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef context_ptr,
unsigned i,
const struct lp_build_interp_soa_context *interp,
struct lp_build_sampler_soa *sampler,
LLVMValueRef *pmask,
LLVMValueRef (*color)[4],
LLVMValueRef depth_ptr,
unsigned do_tri_test,
LLVMValueRef c0,
LLVMValueRef c1,
LLVMValueRef c2,
LLVMValueRef step0_ptr,
LLVMValueRef step1_ptr,
LLVMValueRef step2_ptr)
{
const struct tgsi_token *tokens = shader->base.tokens;
LLVMTypeRef elem_type;
LLVMTypeRef vec_type;
LLVMTypeRef int_vec_type;
LLVMValueRef consts_ptr;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
LLVMValueRef z = interp->pos[2];
struct lp_build_flow_context *flow;
struct lp_build_mask_context mask;
boolean early_depth_test;
unsigned attrib;
unsigned chan;
unsigned cbuf;
assert(i < 4);
elem_type = lp_build_elem_type(type);
vec_type = lp_build_vec_type(type);
int_vec_type = lp_build_int_vec_type(type);
consts_ptr = lp_jit_context_constants(builder, context_ptr);
flow = lp_build_flow_create(builder);
memset(outputs, 0, sizeof outputs);
lp_build_flow_scope_begin(flow);
/* Declare the color and z variables */
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
color[cbuf][chan] = LLVMGetUndef(vec_type);
lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
}
}
lp_build_flow_scope_declare(flow, &z);
/* do triangle edge testing */
if (do_tri_test) {
generate_tri_edge_mask(builder, i, pmask,
c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
}
else {
*pmask = build_int32_vec_const(~0);
}
/* 'mask' will control execution based on quad's pixel alive/killed state */
lp_build_mask_begin(&mask, flow, type, *pmask);
if (key->scissor) {
LLVMValueRef smask =
generate_scissor_test(builder, context_ptr, interp, type);
lp_build_mask_update(&mask, smask);
}
early_depth_test =
key->depth.enabled &&
!key->alpha.enabled &&
!shader->info.uses_kill &&
!shader->info.writes_z;
if(early_depth_test)
generate_depth(builder, key,
type, &mask,
z, depth_ptr);
lp_build_tgsi_soa(builder, tokens, type, &mask,
consts_ptr, interp->pos, interp->inputs,
outputs, sampler);
for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
if(outputs[attrib][chan]) {
LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
switch (shader->info.output_semantic_name[attrib]) {
case TGSI_SEMANTIC_COLOR:
{
unsigned cbuf = shader->info.output_semantic_index[attrib];
lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
/* Alpha test */
/* XXX: should the alpha reference value be passed separately? */
/* XXX: should only test the final assignment to alpha */
if(cbuf == 0 && chan == 3) {
LLVMValueRef alpha = out;
LLVMValueRef alpha_ref_value;
alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
lp_build_alpha_test(builder, &key->alpha, type,
&mask, alpha, alpha_ref_value);
}
color[cbuf][chan] = out;
break;
}
case TGSI_SEMANTIC_POSITION:
if(chan == 2)
z = out;
break;
}
}
}
}
if(!early_depth_test)
generate_depth(builder, key,
type, &mask,
z, depth_ptr);
lp_build_mask_end(&mask);
lp_build_flow_scope_end(flow);
lp_build_flow_destroy(flow);
*pmask = mask.value;
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
*/
static struct lp_fragment_shader_variant *
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
const struct lp_fragment_shader_variant_key *key)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
struct lp_fragment_shader_variant *variant;
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_vec_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
LLVMTypeRef arg_types[9];
LLVMTypeRef func_type;
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef mask_ptr;
LLVMValueRef color_ptr;
LLVMValueRef depth_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
LLVMValueRef y0;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef blend_in_color[NUM_CHANNELS];
unsigned num_fs;
unsigned i;
unsigned chan;
#ifdef DEBUG
tgsi_dump(shader->base.tokens, 0);
if(key->depth.enabled) {
debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format));
debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
debug_printf("depth.writemask = %u\n", key->depth.writemask);
}
if(key->alpha.enabled) {
debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
}
if(key->blend.logicop_enable) {
debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
}
else if(key->blend.blend_enable) {
debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key->blend.rgb_func, TRUE));
debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
debug_printf("alpha_func = %s\n", debug_dump_blend_func (key->blend.alpha_func, TRUE));
debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
}
debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
if(key->sampler[i].format) {
debug_printf("sampler[%u] = \n", i);
debug_printf(" .format = %s\n",
pf_name(key->sampler[i].format));
debug_printf(" .target = %s\n",
debug_dump_tex_target(key->sampler[i].target, TRUE));
debug_printf(" .pot = %u %u %u\n",
key->sampler[i].pot_width,
key->sampler[i].pot_height,
key->sampler[i].pot_depth);
debug_printf(" .wrap = %s %s %s\n",
debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
debug_printf(" .min_img_filter = %s\n",
debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
debug_printf(" .min_mip_filter = %s\n",
debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
debug_printf(" .mag_img_filter = %s\n",
debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
if(key->sampler[i].compare_mode)
debug_printf(" .compare_mode = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE));
debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
debug_printf(" .prefilter = %u\n", key->sampler[i].prefilter);
}
}
#endif
variant = CALLOC_STRUCT(lp_fragment_shader_variant);
if(!variant)
return NULL;
variant->shader = shader;
memcpy(&variant->key, key, sizeof *key);
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 element per vector */
num_fs = 4;
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_vec_type = lp_build_vec_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
blend_int_vec_type = lp_build_int_vec_type(blend_type);
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
arg_types[7] = LLVMPointerType(blend_vec_type, 0); /* color */
arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
variant->function = LLVMAddFunction(screen->module, "shader", func_type);
LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(variant->function, 0);
x = LLVMGetParam(variant->function, 1);
y = LLVMGetParam(variant->function, 2);
a0_ptr = LLVMGetParam(variant->function, 3);
dadx_ptr = LLVMGetParam(variant->function, 4);
dady_ptr = LLVMGetParam(variant->function, 5);
mask_ptr = LLVMGetParam(variant->function, 6);
color_ptr = LLVMGetParam(variant->function, 7);
depth_ptr = LLVMGetParam(variant->function, 8);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(mask_ptr, "mask");
lp_build_name(color_ptr, "color");
lp_build_name(depth_ptr, "depth");
/*
* Function body
*/
block = LLVMAppendBasicBlock(variant->function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
generate_pos0(builder, x, y, &x0, &y0);
lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x0, y0, 2, 0);
#if 0
/* C texture sampling */
sampler = lp_c_sampler_soa_create(context_ptr);
#else
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
#endif
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
if(i != 0)
lp_build_interp_soa_update(&interp);
fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i],
out_color,
depth_ptr_i);
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[chan][i] = out_color[chan];
}
sampler->destroy(sampler);
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
}
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
/*
* Blending.
*/
generate_blend(&key->blend,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/*
* Translate the LLVM IR into machine code.
*/
if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
LLVMDumpValue(variant->function);
abort();
}
LLVMRunFunctionPassManager(screen->pass, variant->function);
#ifdef DEBUG
LLVMDumpValue(variant->function);
debug_printf("\n");
#endif
variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
#ifdef DEBUG
lp_disassemble(variant->jit_function);
#endif
variant->next = shader->variants;
shader->variants = variant;
return variant;
}
/**
* 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;
}
/**
* Build code to compare two values 'a' and 'b' of 'type' using the given func.
* \param func one of PIPE_FUNC_x
* The result values will be 0 for false or ~0 for true.
*/
LLVMValueRef
lp_build_compare(struct gallivm_state *gallivm,
const struct lp_type type,
unsigned func,
LLVMValueRef a,
LLVMValueRef b)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, type);
LLVMValueRef zeros = LLVMConstNull(int_vec_type);
LLVMValueRef ones = LLVMConstAllOnes(int_vec_type);
LLVMValueRef cond;
LLVMValueRef res;
assert(func >= PIPE_FUNC_NEVER);
assert(func <= PIPE_FUNC_ALWAYS);
assert(lp_check_value(type, a));
assert(lp_check_value(type, b));
if(func == PIPE_FUNC_NEVER)
return zeros;
if(func == PIPE_FUNC_ALWAYS)
return ones;
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
/*
* There are no unsigned integer comparison instructions in SSE.
*/
if (!type.floating && !type.sign &&
type.width * type.length == 128 &&
util_cpu_caps.has_sse2 &&
(func == PIPE_FUNC_LESS ||
func == PIPE_FUNC_LEQUAL ||
func == PIPE_FUNC_GREATER ||
func == PIPE_FUNC_GEQUAL) &&
(gallivm_debug & GALLIVM_DEBUG_PERF)) {
debug_printf("%s: inefficient <%u x i%u> unsigned comparison\n",
__FUNCTION__, type.length, type.width);
}
#endif
#if HAVE_LLVM < 0x0207
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
if(type.width * type.length == 128) {
if(type.floating && util_cpu_caps.has_sse) {
/* float[4] comparison */
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
LLVMValueRef args[3];
unsigned cc;
boolean swap;
swap = FALSE;
switch(func) {
case PIPE_FUNC_EQUAL:
cc = 0;
break;
case PIPE_FUNC_NOTEQUAL:
cc = 4;
break;
case PIPE_FUNC_LESS:
cc = 1;
break;
case PIPE_FUNC_LEQUAL:
cc = 2;
break;
case PIPE_FUNC_GREATER:
cc = 1;
swap = TRUE;
break;
case PIPE_FUNC_GEQUAL:
cc = 2;
swap = TRUE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
if(swap) {
args[0] = b;
args[1] = a;
}
else {
args[0] = a;
args[1] = b;
}
args[2] = LLVMConstInt(LLVMInt8TypeInContext(gallivm->context), cc, 0);
res = lp_build_intrinsic(builder,
"llvm.x86.sse.cmp.ps",
vec_type,
args, 3);
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
return res;
}
else if(util_cpu_caps.has_sse2) {
/* int[4] comparison */
static const struct {
unsigned swap:1;
unsigned eq:1;
unsigned gt:1;
unsigned not:1;
} table[] = {
{0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
{1, 0, 1, 0}, /* PIPE_FUNC_LESS */
{0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
{0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
{0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
{0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
{1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
{0, 0, 0, 0} /* PIPE_FUNC_ALWAYS */
};
const char *pcmpeq;
const char *pcmpgt;
LLVMValueRef args[2];
LLVMValueRef res;
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
switch (type.width) {
case 8:
pcmpeq = "llvm.x86.sse2.pcmpeq.b";
pcmpgt = "llvm.x86.sse2.pcmpgt.b";
break;
case 16:
pcmpeq = "llvm.x86.sse2.pcmpeq.w";
pcmpgt = "llvm.x86.sse2.pcmpgt.w";
break;
case 32:
pcmpeq = "llvm.x86.sse2.pcmpeq.d";
pcmpgt = "llvm.x86.sse2.pcmpgt.d";
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
/* There are no unsigned comparison instructions. So flip the sign bit
* so that the results match.
*/
if (table[func].gt && !type.sign) {
LLVMValueRef msb = lp_build_const_int_vec(gallivm, type, (unsigned long long)1 << (type.width - 1));
a = LLVMBuildXor(builder, a, msb, "");
b = LLVMBuildXor(builder, b, msb, "");
}
if(table[func].swap) {
args[0] = b;
args[1] = a;
}
else {
args[0] = a;
args[1] = b;
}
if(table[func].eq)
res = lp_build_intrinsic(builder, pcmpeq, vec_type, args, 2);
else if (table[func].gt)
res = lp_build_intrinsic(builder, pcmpgt, vec_type, args, 2);
else
res = LLVMConstNull(vec_type);
if(table[func].not)
res = LLVMBuildNot(builder, res, "");
return res;
}
} /* if (type.width * type.length == 128) */
#endif
#endif /* HAVE_LLVM < 0x0207 */
/* XXX: It is not clear if we should use the ordered or unordered operators */
if(type.floating) {
LLVMRealPredicate op;
switch(func) {
case PIPE_FUNC_NEVER:
op = LLVMRealPredicateFalse;
break;
case PIPE_FUNC_ALWAYS:
op = LLVMRealPredicateTrue;
break;
case PIPE_FUNC_EQUAL:
op = LLVMRealUEQ;
break;
case PIPE_FUNC_NOTEQUAL:
op = LLVMRealUNE;
break;
case PIPE_FUNC_LESS:
op = LLVMRealULT;
break;
case PIPE_FUNC_LEQUAL:
op = LLVMRealULE;
break;
case PIPE_FUNC_GREATER:
op = LLVMRealUGT;
break;
case PIPE_FUNC_GEQUAL:
op = LLVMRealUGE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
#if HAVE_LLVM >= 0x0207
cond = LLVMBuildFCmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
#else
if (type.length == 1) {
cond = LLVMBuildFCmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
}
else {
unsigned i;
res = LLVMGetUndef(int_vec_type);
debug_printf("%s: warning: using slow element-wise float"
" vector comparison\n", __FUNCTION__);
for (i = 0; i < type.length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
cond = LLVMBuildFCmp(builder, op,
LLVMBuildExtractElement(builder, a, index, ""),
LLVMBuildExtractElement(builder, b, index, ""),
"");
cond = LLVMBuildSelect(builder, cond,
LLVMConstExtractElement(ones, index),
LLVMConstExtractElement(zeros, index),
"");
res = LLVMBuildInsertElement(builder, res, cond, index, "");
}
}
#endif
}
else {
LLVMIntPredicate op;
switch(func) {
case PIPE_FUNC_EQUAL:
op = LLVMIntEQ;
break;
case PIPE_FUNC_NOTEQUAL:
op = LLVMIntNE;
break;
case PIPE_FUNC_LESS:
op = type.sign ? LLVMIntSLT : LLVMIntULT;
break;
case PIPE_FUNC_LEQUAL:
op = type.sign ? LLVMIntSLE : LLVMIntULE;
break;
case PIPE_FUNC_GREATER:
op = type.sign ? LLVMIntSGT : LLVMIntUGT;
break;
case PIPE_FUNC_GEQUAL:
op = type.sign ? LLVMIntSGE : LLVMIntUGE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
#if HAVE_LLVM >= 0x0207
cond = LLVMBuildICmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
#else
if (type.length == 1) {
cond = LLVMBuildICmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
}
else {
unsigned i;
res = LLVMGetUndef(int_vec_type);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: using slow element-wise int"
" vector comparison\n", __FUNCTION__);
}
for(i = 0; i < type.length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
cond = LLVMBuildICmp(builder, op,
LLVMBuildExtractElement(builder, a, index, ""),
LLVMBuildExtractElement(builder, b, index, ""),
"");
cond = LLVMBuildSelect(builder, cond,
LLVMConstExtractElement(ones, index),
LLVMConstExtractElement(zeros, index),
"");
res = LLVMBuildInsertElement(builder, res, cond, index, "");
}
}
#endif
}
return res;
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
* Note that the function which we generate operates on a block of 16
* pixels at at time. The block contains 2x2 quads. Each quad contains
* 2x2 pixels.
*/
static void
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
struct lp_fragment_shader_variant *variant,
unsigned partial_mask)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
const struct lp_fragment_shader_variant_key *key = &variant->key;
char func_name[256];
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef arg_types[11];
LLVMTypeRef func_type;
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
LLVMValueRef mask_input;
LLVMValueRef counter = NULL;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef function;
LLVMValueRef facing;
unsigned num_fs;
unsigned i;
unsigned chan;
unsigned cbuf;
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 elements per vector */
num_fs = 4; /* number of quads per block */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
shader->no, variant->no, partial_mask ? "partial" : "whole");
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMFloatType(); /* facing */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
arg_types[9] = LLVMInt32Type(); /* mask_input */
arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
function = LLVMAddFunction(screen->module, func_name, func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
variant->function[partial_mask] = function;
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
*/
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(function, 0);
x = LLVMGetParam(function, 1);
y = LLVMGetParam(function, 2);
facing = LLVMGetParam(function, 3);
a0_ptr = LLVMGetParam(function, 4);
dadx_ptr = LLVMGetParam(function, 5);
dady_ptr = LLVMGetParam(function, 6);
color_ptr_ptr = LLVMGetParam(function, 7);
depth_ptr = LLVMGetParam(function, 8);
mask_input = LLVMGetParam(function, 9);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr_ptr");
lp_build_name(depth_ptr, "depth");
lp_build_name(mask_input, "mask_input");
if (key->occlusion_count) {
counter = LLVMGetParam(function, 10);
lp_build_name(counter, "counter");
}
/*
* Function body
*/
block = LLVMAppendBasicBlock(function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
/*
* The shader input interpolation info is not explicitely baked in the
* shader key, but everything it derives from (TGSI, and flatshade) is
* already included in the shader key.
*/
lp_build_interp_soa_init(&interp,
lp->num_inputs,
lp->inputs,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x, y);
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
/* loop over quads in the block */
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
if(i != 0)
lp_build_interp_soa_update(&interp, i);
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i], /* output */
out_color,
depth_ptr_i,
facing,
partial_mask,
mask_input,
counter);
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
}
sampler->destroy(sampler);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
LLVMValueRef color_ptr;
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
LLVMValueRef blend_in_color[NUM_CHANNELS];
unsigned rt;
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[cbuf][chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
}
if (partial_mask || !variant->opaque) {
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
} else {
blend_mask = lp_build_const_int_vec(blend_type, ~0);
}
color_ptr = LLVMBuildLoad(builder,
LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
"");
lp_build_name(color_ptr, "color_ptr%d", cbuf);
/* which blend/colormask state to use */
rt = key->blend.independent_blend_enable ? cbuf : 0;
/*
* Blending.
*/
generate_blend(&key->blend,
rt,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
}
#ifdef PIPE_ARCH_X86
/* Avoid corrupting the FPU stack on 32bit OSes. */
lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0);
#endif
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/* Verify the LLVM IR. If invalid, dump and abort */
#ifdef DEBUG
if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
if (1)
lp_debug_dump_value(function);
abort();
}
#endif
/* Apply optimizations to LLVM IR */
LLVMRunFunctionPassManager(screen->pass, function);
if (gallivm_debug & GALLIVM_DEBUG_IR) {
/* Print the LLVM IR to stderr */
lp_debug_dump_value(function);
debug_printf("\n");
}
/*
* Translate the LLVM IR into machine code.
*/
{
void *f = LLVMGetPointerToGlobal(screen->engine, function);
variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
if (gallivm_debug & GALLIVM_DEBUG_ASM) {
lp_disassemble(f);
}
lp_func_delete_body(function);
}
}
/**
* Inverse of lp_build_clamped_float_to_unsigned_norm above.
* Ex: src = { i32, i32, i32, i32 } with values in range [0, 2^src_width-1]
* return {float, float, float, float} with values in range [0, 1].
*/
LLVMValueRef
lp_build_unsigned_norm_to_float(struct gallivm_state *gallivm,
unsigned src_width,
struct lp_type dst_type,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, dst_type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, dst_type);
LLVMValueRef bias_;
LLVMValueRef res;
unsigned mantissa;
unsigned n;
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
assert(dst_type.floating);
mantissa = lp_mantissa(dst_type);
if (src_width <= (mantissa + 1)) {
/*
* The source width matches fits what can be represented in floating
* point (i.e., mantissa + 1 bits). So do a straight multiplication
* followed by casting. No further rounding is necessary.
*/
scale = 1.0/(double)((1ULL << src_width) - 1);
res = LLVMBuildSIToFP(builder, src, vec_type, "");
res = LLVMBuildFMul(builder, res,
lp_build_const_vec(gallivm, dst_type, scale), "");
return res;
}
else {
/*
* The source width exceeds what can be represented in floating
* point. So truncate the incoming values.
*/
n = MIN2(mantissa, src_width);
ubound = ((unsigned long long)1 << n);
mask = ubound - 1;
scale = (double)ubound/mask;
bias = (double)((unsigned long long)1 << (mantissa - n));
res = src;
if (src_width > mantissa) {
int shift = src_width - mantissa;
res = LLVMBuildLShr(builder, res,
lp_build_const_int_vec(gallivm, dst_type, shift), "");
}
bias_ = lp_build_const_vec(gallivm, dst_type, bias);
res = LLVMBuildOr(builder,
res,
LLVMBuildBitCast(builder, bias_, int_vec_type, ""), "");
res = LLVMBuildBitCast(builder, res, vec_type, "");
res = LLVMBuildFSub(builder, res, bias_, "");
res = LLVMBuildFMul(builder, res, lp_build_const_vec(gallivm, dst_type, scale), "");
}
return res;
}
/**
* Generate the fragment shader, depth/stencil test, and alpha tests.
*/
static void
generate_fs(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
const struct lp_fragment_shader_variant_key *key,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef context_ptr,
unsigned i,
const struct lp_build_interp_soa_context *interp,
struct lp_build_sampler_soa *sampler,
LLVMValueRef *pmask,
LLVMValueRef *color,
LLVMValueRef depth_ptr)
{
const struct tgsi_token *tokens = shader->base.tokens;
LLVMTypeRef elem_type;
LLVMTypeRef vec_type;
LLVMTypeRef int_vec_type;
LLVMValueRef consts_ptr;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
LLVMValueRef z = interp->pos[2];
struct lp_build_flow_context *flow;
struct lp_build_mask_context mask;
boolean early_depth_test;
unsigned attrib;
unsigned chan;
elem_type = lp_build_elem_type(type);
vec_type = lp_build_vec_type(type);
int_vec_type = lp_build_int_vec_type(type);
consts_ptr = lp_jit_context_constants(builder, context_ptr);
flow = lp_build_flow_create(builder);
memset(outputs, 0, sizeof outputs);
lp_build_flow_scope_begin(flow);
/* Declare the color and z variables */
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
color[chan] = LLVMGetUndef(vec_type);
lp_build_flow_scope_declare(flow, &color[chan]);
}
lp_build_flow_scope_declare(flow, &z);
lp_build_mask_begin(&mask, flow, type, *pmask);
early_depth_test =
key->depth.enabled &&
!key->alpha.enabled &&
!shader->info.uses_kill &&
!shader->info.writes_z;
if(early_depth_test)
generate_depth(builder, key,
type, &mask,
z, depth_ptr);
lp_build_tgsi_soa(builder, tokens, type, &mask,
consts_ptr, interp->pos, interp->inputs,
outputs, sampler);
for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
if(outputs[attrib][chan]) {
lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
switch (shader->info.output_semantic_name[attrib]) {
case TGSI_SEMANTIC_COLOR:
{
unsigned cbuf = shader->info.output_semantic_index[attrib];
lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
/* Alpha test */
/* XXX: should the alpha reference value be passed separately? */
if(cbuf == 0 && chan == 3) {
LLVMValueRef alpha = outputs[attrib][chan];
LLVMValueRef alpha_ref_value;
alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
lp_build_alpha_test(builder, &key->alpha, type,
&mask, alpha, alpha_ref_value);
}
if(cbuf == 0)
color[chan] = outputs[attrib][chan];
break;
}
case TGSI_SEMANTIC_POSITION:
if(chan == 2)
z = outputs[attrib][chan];
break;
}
}
}
}
if(!early_depth_test)
generate_depth(builder, key,
type, &mask,
z, depth_ptr);
lp_build_mask_end(&mask);
lp_build_flow_scope_end(flow);
lp_build_flow_destroy(flow);
*pmask = mask.value;
}
