/**
* Convert linear float soa values to packed srgb AoS values.
* This only handles packed formats which are 4x8bit in size
* (rgba and rgbx plus swizzles), and 16bit 565-style formats
* with no alpha. (In the latter case the return values won't be
* fully packed, it will look like r5g6b5x16r5g6b5x16...)
*
* @param src float SoA (vector) values to convert.
*/
LLVMValueRef
lp_build_float_to_srgb_packed(struct gallivm_state *gallivm,
const struct util_format_description *dst_fmt,
struct lp_type src_type,
LLVMValueRef *src)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned chan;
struct lp_build_context f32_bld;
struct lp_type int32_type = lp_int_type(src_type);
LLVMValueRef tmpsrgb[4], alpha, dst;
lp_build_context_init(&f32_bld, gallivm, src_type);
/* rgb is subject to linear->srgb conversion, alpha is not */
for (chan = 0; chan < 3; chan++) {
unsigned chan_bits = dst_fmt->channel[dst_fmt->swizzle[chan]].size;
tmpsrgb[chan] = lp_build_linear_to_srgb(gallivm, src_type, chan_bits, src[chan]);
}
/*
* can't use lp_build_conv since we want to keep values as 32bit
* here so we can interleave with rgb to go from SoA->AoS.
*/
alpha = lp_build_clamp_zero_one_nanzero(&f32_bld, src[3]);
alpha = lp_build_mul(&f32_bld, alpha,
lp_build_const_vec(gallivm, src_type, 255.0f));
tmpsrgb[3] = lp_build_iround(&f32_bld, alpha);
dst = lp_build_zero(gallivm, int32_type);
for (chan = 0; chan < dst_fmt->nr_channels; chan++) {
if (dst_fmt->swizzle[chan] <= PIPE_SWIZZLE_W) {
unsigned ls;
LLVMValueRef shifted, shift_val;
ls = dst_fmt->channel[dst_fmt->swizzle[chan]].shift;
shift_val = lp_build_const_int_vec(gallivm, int32_type, ls);
shifted = LLVMBuildShl(builder, tmpsrgb[chan], shift_val, "");
dst = LLVMBuildOr(builder, dst, shifted, "");
}
}
return dst;
}
void
lp_build_tgsi_aos(struct gallivm_state *gallivm,
const struct tgsi_token *tokens,
struct lp_type type,
const unsigned char swizzles[4],
LLVMValueRef consts_ptr,
const LLVMValueRef *inputs,
LLVMValueRef *outputs,
struct lp_build_sampler_aos *sampler,
const struct tgsi_shader_info *info)
{
struct lp_build_tgsi_aos_context bld;
struct tgsi_parse_context parse;
uint num_immediates = 0;
unsigned chan;
int pc = 0;
/* Setup build context */
memset(&bld, 0, sizeof bld);
lp_build_context_init(&bld.bld_base.base, gallivm, type);
lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
lp_build_context_init(&bld.int_bld, gallivm, lp_int_type(type));
for (chan = 0; chan < 4; ++chan) {
bld.swizzles[chan] = swizzles[chan];
bld.inv_swizzles[swizzles[chan]] = chan;
}
bld.inputs = inputs;
bld.outputs = outputs;
bld.consts_ptr = consts_ptr;
bld.sampler = sampler;
bld.indirect_files = info->indirect_files;
bld.bld_base.emit_swizzle = swizzle_aos;
bld.bld_base.info = info;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
/* Set opcode actions */
lp_set_default_actions_cpu(&bld.bld_base);
if (!lp_bld_tgsi_list_init(&bld.bld_base)) {
return;
}
tgsi_parse_init(&parse, tokens);
while (!tgsi_parse_end_of_tokens(&parse)) {
tgsi_parse_token(&parse);
switch(parse.FullToken.Token.Type) {
case TGSI_TOKEN_TYPE_DECLARATION:
/* Inputs already interpolated */
lp_emit_declaration_aos(&bld, &parse.FullToken.FullDeclaration);
break;
case TGSI_TOKEN_TYPE_INSTRUCTION:
/* save expanded instruction */
lp_bld_tgsi_add_instruction(&bld.bld_base,
&parse.FullToken.FullInstruction);
break;
case TGSI_TOKEN_TYPE_IMMEDIATE:
/* simply copy the immediate values into the next immediates[] slot */
{
const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
float imm[4];
assert(size <= 4);
assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
for (chan = 0; chan < 4; ++chan) {
imm[chan] = 0.0f;
}
for (chan = 0; chan < size; ++chan) {
unsigned swizzle = bld.swizzles[chan];
imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
}
bld.immediates[num_immediates] =
lp_build_const_aos(gallivm, type,
imm[0], imm[1], imm[2], imm[3],
NULL);
num_immediates++;
}
break;
case TGSI_TOKEN_TYPE_PROPERTY:
break;
default:
assert(0);
}
}
while (pc != -1) {
struct tgsi_full_instruction *instr = bld.bld_base.instructions + pc;
const struct tgsi_opcode_info *opcode_info =
tgsi_get_opcode_info(instr->Instruction.Opcode);
if (!lp_emit_instruction_aos(&bld, instr, opcode_info, &pc))
_debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
opcode_info->mnemonic);
}
if (0) {
LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
LLVMValueRef function = LLVMGetBasicBlockParent(block);
debug_printf("11111111111111111111111111111 \n");
tgsi_dump(tokens, 0);
lp_debug_dump_value(function);
debug_printf("2222222222222222222222222222 \n");
}
tgsi_parse_free(&parse);
FREE(bld.bld_base.instructions);
if (0) {
LLVMModuleRef module = LLVMGetGlobalParent(
LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
LLVMDumpModule(module);
}
}
/**
* Generate code for performing depth and/or stencil tests.
* We operate on a vector of values (typically n 2x2 quads).
*
* \param depth the depth test state
* \param stencil the front/back stencil state
* \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 stencil_refs the front/back stencil ref values (scalar)
* \param z_src the incoming depth/stencil values (n 2x2 quad values, float32)
* \param zs_dst the depth/stencil values in framebuffer
* \param face contains boolean value indicating front/back facing polygon
*/
void
lp_build_depth_stencil_test(struct gallivm_state *gallivm,
const struct pipe_depth_state *depth,
const struct pipe_stencil_state stencil[2],
struct lp_type z_src_type,
const struct util_format_description *format_desc,
struct lp_build_mask_context *mask,
LLVMValueRef stencil_refs[2],
LLVMValueRef z_src,
LLVMValueRef z_fb,
LLVMValueRef s_fb,
LLVMValueRef face,
LLVMValueRef *z_value,
LLVMValueRef *s_value,
boolean do_branch)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_type z_type;
struct lp_build_context z_bld;
struct lp_build_context s_bld;
struct lp_type s_type;
unsigned z_shift = 0, z_width = 0, z_mask = 0;
LLVMValueRef z_dst = NULL;
LLVMValueRef stencil_vals = NULL;
LLVMValueRef z_bitmask = NULL, stencil_shift = NULL;
LLVMValueRef z_pass = NULL, s_pass_mask = NULL;
LLVMValueRef orig_mask = lp_build_mask_value(mask);
LLVMValueRef front_facing = NULL;
boolean have_z, have_s;
/*
* Depths are expected to be between 0 and 1, even if they are stored in
* floats. Setting these bits here will ensure that the lp_build_conv() call
* below won't try to unnecessarily clamp the incoming values.
*/
if(z_src_type.floating) {
z_src_type.sign = FALSE;
z_src_type.norm = TRUE;
}
else {
assert(!z_src_type.sign);
assert(z_src_type.norm);
}
/* Pick the type matching the depth-stencil format. */
z_type = lp_depth_type(format_desc, z_src_type.length);
/* Pick the intermediate type for depth operations. */
z_type.width = z_src_type.width;
assert(z_type.length == z_src_type.length);
/* FIXME: for non-float depth/stencil might generate better code
* if we'd always split it up to use 128bit operations.
* For stencil we'd almost certainly want to pack to 8xi16 values,
* for z just run twice.
*/
/* Sanity checking */
{
const unsigned z_swizzle = format_desc->swizzle[0];
const unsigned s_swizzle = format_desc->swizzle[1];
assert(z_swizzle != UTIL_FORMAT_SWIZZLE_NONE ||
s_swizzle != UTIL_FORMAT_SWIZZLE_NONE);
assert(depth->enabled || stencil[0].enabled);
assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS);
assert(format_desc->block.width == 1);
assert(format_desc->block.height == 1);
if (stencil[0].enabled) {
assert(s_swizzle < 4);
assert(format_desc->channel[s_swizzle].type == UTIL_FORMAT_TYPE_UNSIGNED);
assert(format_desc->channel[s_swizzle].pure_integer);
assert(!format_desc->channel[s_swizzle].normalized);
assert(format_desc->channel[s_swizzle].size == 8);
}
if (depth->enabled) {
assert(z_swizzle < 4);
if (z_type.floating) {
assert(z_swizzle == 0);
assert(format_desc->channel[z_swizzle].type ==
UTIL_FORMAT_TYPE_FLOAT);
assert(format_desc->channel[z_swizzle].size == 32);
}
else {
assert(format_desc->channel[z_swizzle].type ==
UTIL_FORMAT_TYPE_UNSIGNED);
assert(format_desc->channel[z_swizzle].normalized);
assert(!z_type.fixed);
}
}
}
/* Setup build context for Z vals */
lp_build_context_init(&z_bld, gallivm, z_type);
/* Setup build context for stencil vals */
s_type = lp_int_type(z_type);
lp_build_context_init(&s_bld, gallivm, s_type);
/* Compute and apply the Z/stencil bitmasks and shifts.
*/
{
unsigned s_shift, s_mask;
z_dst = z_fb;
stencil_vals = s_fb;
have_z = get_z_shift_and_mask(format_desc, &z_shift, &z_width, &z_mask);
have_s = get_s_shift_and_mask(format_desc, &s_shift, &s_mask);
if (have_z) {
if (z_mask != 0xffffffff) {
z_bitmask = lp_build_const_int_vec(gallivm, z_type, z_mask);
}
/*
* Align the framebuffer Z 's LSB to the right.
*/
if (z_shift) {
LLVMValueRef shift = lp_build_const_int_vec(gallivm, z_type, z_shift);
z_dst = LLVMBuildLShr(builder, z_dst, shift, "z_dst");
} else if (z_bitmask) {
z_dst = LLVMBuildAnd(builder, z_dst, z_bitmask, "z_dst");
} else {
lp_build_name(z_dst, "z_dst");
}
}
if (have_s) {
if (s_shift) {
LLVMValueRef shift = lp_build_const_int_vec(gallivm, s_type, s_shift);
stencil_vals = LLVMBuildLShr(builder, stencil_vals, shift, "");
stencil_shift = shift; /* used below */
}
if (s_mask != 0xffffffff) {
LLVMValueRef mask = lp_build_const_int_vec(gallivm, s_type, s_mask);
stencil_vals = LLVMBuildAnd(builder, stencil_vals, mask, "");
}
lp_build_name(stencil_vals, "s_dst");
}
}
if (stencil[0].enabled) {
if (face) {
LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
/* front_facing = face != 0 ? ~0 : 0 */
front_facing = LLVMBuildICmp(builder, LLVMIntNE, face, zero, "");
front_facing = LLVMBuildSExt(builder, front_facing,
LLVMIntTypeInContext(gallivm->context,
s_bld.type.length*s_bld.type.width),
"");
front_facing = LLVMBuildBitCast(builder, front_facing,
s_bld.int_vec_type, "");
}
/* convert scalar stencil refs into vectors */
stencil_refs[0] = lp_build_broadcast_scalar(&s_bld, stencil_refs[0]);
stencil_refs[1] = lp_build_broadcast_scalar(&s_bld, stencil_refs[1]);
s_pass_mask = lp_build_stencil_test(&s_bld, stencil,
stencil_refs, stencil_vals,
front_facing);
/* apply stencil-fail operator */
{
LLVMValueRef s_fail_mask = lp_build_andnot(&s_bld, orig_mask, s_pass_mask);
stencil_vals = lp_build_stencil_op(&s_bld, stencil, S_FAIL_OP,
stencil_refs, stencil_vals,
s_fail_mask, front_facing);
}
}
if (depth->enabled) {
/*
* Convert fragment Z to the desired type, aligning the LSB to the right.
*/
assert(z_type.width == z_src_type.width);
assert(z_type.length == z_src_type.length);
assert(lp_check_value(z_src_type, z_src));
if (z_src_type.floating) {
/*
* Convert from floating point values
*/
if (!z_type.floating) {
z_src = lp_build_clamped_float_to_unsigned_norm(gallivm,
z_src_type,
z_width,
z_src);
}
} else {
/*
* Convert from unsigned normalized values.
*/
assert(!z_src_type.sign);
assert(!z_src_type.fixed);
assert(z_src_type.norm);
assert(!z_type.floating);
if (z_src_type.width > z_width) {
LLVMValueRef shift = lp_build_const_int_vec(gallivm, z_src_type,
z_src_type.width - z_width);
z_src = LLVMBuildLShr(builder, z_src, shift, "");
}
}
assert(lp_check_value(z_type, z_src));
lp_build_name(z_src, "z_src");
/* compare src Z to dst Z, returning 'pass' mask */
z_pass = lp_build_cmp(&z_bld, depth->func, z_src, z_dst);
if (!stencil[0].enabled) {
/* We can potentially skip all remaining operations here, but only
* if stencil is disabled because we still need to update the stencil
* buffer values. Don't need to update Z buffer values.
*/
lp_build_mask_update(mask, z_pass);
if (do_branch) {
lp_build_mask_check(mask);
do_branch = FALSE;
}
}
if (depth->writemask) {
LLVMValueRef zselectmask;
/* mask off bits that failed Z test */
zselectmask = LLVMBuildAnd(builder, orig_mask, z_pass, "");
/* mask off bits that failed stencil test */
if (s_pass_mask) {
zselectmask = LLVMBuildAnd(builder, zselectmask, s_pass_mask, "");
}
/* Mix the old and new Z buffer values.
* z_dst[i] = zselectmask[i] ? z_src[i] : z_dst[i]
*/
z_dst = lp_build_select(&z_bld, zselectmask, z_src, z_dst);
}
if (stencil[0].enabled) {
/* update stencil buffer values according to z pass/fail result */
LLVMValueRef z_fail_mask, z_pass_mask;
/* apply Z-fail operator */
z_fail_mask = lp_build_andnot(&s_bld, orig_mask, z_pass);
stencil_vals = lp_build_stencil_op(&s_bld, stencil, Z_FAIL_OP,
stencil_refs, stencil_vals,
z_fail_mask, front_facing);
/* apply Z-pass operator */
z_pass_mask = LLVMBuildAnd(builder, orig_mask, z_pass, "");
stencil_vals = lp_build_stencil_op(&s_bld, stencil, Z_PASS_OP,
stencil_refs, stencil_vals,
z_pass_mask, front_facing);
}
}
else {
/* No depth test: apply Z-pass operator to stencil buffer values which
* passed the stencil test.
*/
s_pass_mask = LLVMBuildAnd(builder, orig_mask, s_pass_mask, "");
stencil_vals = lp_build_stencil_op(&s_bld, stencil, Z_PASS_OP,
stencil_refs, stencil_vals,
s_pass_mask, front_facing);
}
/* Put Z and stencil bits in the right place */
if (have_z && z_shift) {
LLVMValueRef shift = lp_build_const_int_vec(gallivm, z_type, z_shift);
z_dst = LLVMBuildShl(builder, z_dst, shift, "");
}
if (stencil_vals && stencil_shift)
stencil_vals = LLVMBuildShl(builder, stencil_vals,
stencil_shift, "");
/* Finally, merge the z/stencil values */
if (format_desc->block.bits <= 32) {
if (have_z && have_s)
*z_value = LLVMBuildOr(builder, z_dst, stencil_vals, "");
else if (have_z)
*z_value = z_dst;
else
*z_value = stencil_vals;
*s_value = *z_value;
}
else {
*z_value = z_dst;
*s_value = stencil_vals;
}
if (s_pass_mask)
lp_build_mask_update(mask, s_pass_mask);
if (depth->enabled && stencil[0].enabled)
lp_build_mask_update(mask, z_pass);
}
/**
* Convert linear float values to srgb int values.
* Several possibilities how to do this, e.g.
* - use table (based on exponent/highest order mantissa bits) and do
* linear interpolation (https://gist.github.com/rygorous/2203834)
* - Chebyshev polynomial
* - Approximation using reciprocals
* - using int-to-float and float-to-int tricks for pow()
* (http://stackoverflow.com/questions/6475373/optimizations-for-pow-with-const-non-integer-exponent)
*
* @param src float (vector) value(s) to convert.
*/
static LLVMValueRef
lp_build_linear_to_srgb(struct gallivm_state *gallivm,
struct lp_type src_type,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_build_context f32_bld;
LLVMValueRef lin_thresh, lin, lin_const, is_linear, tmp, pow_final;
lp_build_context_init(&f32_bld, gallivm, src_type);
src = lp_build_clamp(&f32_bld, src, f32_bld.zero, f32_bld.one);
if (0) {
/*
* using int-to-float and float-to-int trick for pow().
* This is much more accurate than necessary thanks to the correction,
* but it most certainly makes no sense without rsqrt available.
* Bonus points if you understand how this works...
* All in all (including min/max clamp, conversion) 19 instructions.
*/
float exp_f = 2.0f / 3.0f;
/* some compilers can't do exp2f, so this is exp2f(127.0f/exp_f - 127.0f) */
float exp2f_c = 1.30438178253e+19f;
float coeff_f = 0.62996f;
LLVMValueRef pow_approx, coeff, x2, exponent, pow_1, pow_2;
struct lp_type int_type = lp_int_type(src_type);
/*
* First calculate approx x^8/12
*/
exponent = lp_build_const_vec(gallivm, src_type, exp_f);
coeff = lp_build_const_vec(gallivm, src_type,
exp2f_c * powf(coeff_f, 1.0f / exp_f));
/* premultiply src */
tmp = lp_build_mul(&f32_bld, coeff, src);
/* "log2" */
tmp = LLVMBuildBitCast(builder, tmp, lp_build_vec_type(gallivm, int_type), "");
tmp = lp_build_int_to_float(&f32_bld, tmp);
/* multiply for pow */
tmp = lp_build_mul(&f32_bld, tmp, exponent);
/* "exp2" */
pow_approx = lp_build_itrunc(&f32_bld, tmp);
pow_approx = LLVMBuildBitCast(builder, pow_approx,
lp_build_vec_type(gallivm, src_type), "");
/*
* Since that pow was inaccurate (like 3 bits, though each sqrt step would
* give another bit), compensate the error (which is why we chose another
* exponent in the first place).
*/
/* x * x^(8/12) = x^(20/12) */
pow_1 = lp_build_mul(&f32_bld, pow_approx, src);
/* x * x * x^(-4/12) = x^(20/12) */
/* Should avoid using rsqrt if it's not available, but
* using x * x^(4/12) * x^(4/12) instead will change error weight */
tmp = lp_build_fast_rsqrt(&f32_bld, pow_approx);
x2 = lp_build_mul(&f32_bld, src, src);
pow_2 = lp_build_mul(&f32_bld, x2, tmp);
/* average the values so the errors cancel out, compensate bias,
* we also squeeze the 1.055 mul of the srgb conversion plus the 255.0 mul
* for conversion to int in here */
tmp = lp_build_add(&f32_bld, pow_1, pow_2);
coeff = lp_build_const_vec(gallivm, src_type,
1.0f / (3.0f * coeff_f) * 0.999852f *
powf(1.055f * 255.0f, 4.0f));
pow_final = lp_build_mul(&f32_bld, tmp, coeff);
/* x^(5/12) = rsqrt(rsqrt(x^20/12)) */
if (lp_build_fast_rsqrt_available(src_type)) {
pow_final = lp_build_fast_rsqrt(&f32_bld,
lp_build_fast_rsqrt(&f32_bld, pow_final));
}
else {
pow_final = lp_build_sqrt(&f32_bld, lp_build_sqrt(&f32_bld, pow_final));
}
pow_final = lp_build_add(&f32_bld, pow_final,
lp_build_const_vec(gallivm, src_type, -0.055f * 255.0f));
}
else {
/*
* using "rational polynomial" approximation here.
* Essentially y = a*x^0.375 + b*x^0.5 + c, with also
* factoring in the 255.0 mul and the scaling mul.
* (a is closer to actual value so has higher weight than b.)
* Note: the constants are magic values. They were found empirically,
* possibly could be improved but good enough (be VERY careful with
* error metric if you'd want to tweak them, they also MUST fit with
* the crappy polynomial above for srgb->linear since it is required
* that each srgb value maps back to the same value).
* This function has an error of max +-0.17 (and we'd only require +-0.6),
* for the approximated srgb->linear values the error is naturally larger
* (+-0.42) but still accurate enough (required +-0.5 essentially).
* All in all (including min/max clamp, conversion) 15 instructions.
* FMA would help (minus 2 instructions).
*/
LLVMValueRef x05, x0375, a_const, b_const, c_const, tmp2;
if (lp_build_fast_rsqrt_available(src_type)) {
tmp = lp_build_fast_rsqrt(&f32_bld, src);
x05 = lp_build_mul(&f32_bld, src, tmp);
}
else {
/*
* I don't really expect this to be practical without rsqrt
* but there's no reason for triple punishment so at least
* save the otherwise resulting division and unnecessary mul...
*/
x05 = lp_build_sqrt(&f32_bld, src);
}
tmp = lp_build_mul(&f32_bld, x05, src);
if (lp_build_fast_rsqrt_available(src_type)) {
x0375 = lp_build_fast_rsqrt(&f32_bld, lp_build_fast_rsqrt(&f32_bld, tmp));
}
else {
x0375 = lp_build_sqrt(&f32_bld, lp_build_sqrt(&f32_bld, tmp));
}
a_const = lp_build_const_vec(gallivm, src_type, 0.675f * 1.0622 * 255.0f);
b_const = lp_build_const_vec(gallivm, src_type, 0.325f * 1.0622 * 255.0f);
c_const = lp_build_const_vec(gallivm, src_type, -0.0620f * 255.0f);
tmp = lp_build_mul(&f32_bld, a_const, x0375);
tmp2 = lp_build_mul(&f32_bld, b_const, x05);
tmp2 = lp_build_add(&f32_bld, tmp2, c_const);
pow_final = lp_build_add(&f32_bld, tmp, tmp2);
}
/* linear part is easy */
lin_const = lp_build_const_vec(gallivm, src_type, 12.92f * 255.0f);
lin = lp_build_mul(&f32_bld, src, lin_const);
lin_thresh = lp_build_const_vec(gallivm, src_type, 0.0031308f);
is_linear = lp_build_compare(gallivm, src_type, PIPE_FUNC_LEQUAL, src, lin_thresh);
tmp = lp_build_select(&f32_bld, is_linear, lin, pow_final);
f32_bld.type.sign = 0;
return lp_build_iround(&f32_bld, tmp);
}
/**
* Fetch a texels from a texture, returning them in SoA layout.
*
* \param type the desired return type for 'rgba'. The vector length
* is the number of texels to fetch
* \param aligned if the offset is guaranteed to be aligned to element width
*
* \param base_ptr points to the base of the texture mip tree.
* \param offset offset to start of the texture image block. For non-
* compressed formats, this simply is an offset to the texel.
* For compressed formats, it is an offset to the start of the
* compressed data block.
*
* \param i, j the sub-block pixel coordinates. For non-compressed formats
* these will always be (0,0). For compressed formats, i will
* be in [0, block_width-1] and j will be in [0, block_height-1].
* \param cache optional value pointing to a lp_build_format_cache structure
*/
void
lp_build_fetch_rgba_soa(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
boolean aligned,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j,
LLVMValueRef cache,
LLVMValueRef rgba_out[4])
{
LLVMBuilderRef builder = gallivm->builder;
enum pipe_format format = format_desc->format;
struct lp_type fetch_type;
if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
format_desc->block.bits <= type.width &&
(format_desc->channel[0].type != UTIL_FORMAT_TYPE_FLOAT ||
format_desc->channel[0].size == 32 ||
format_desc->channel[0].size == 16))
{
/*
* The packed pixel fits into an element of the destination format. Put
* the packed pixels into a vector and extract each component for all
* vector elements in parallel.
*/
LLVMValueRef packed;
/*
* gather the texels from the texture
* Ex: packed = {XYZW, XYZW, XYZW, XYZW}
*/
assert(format_desc->block.bits <= type.width);
fetch_type = lp_type_uint(type.width);
packed = lp_build_gather(gallivm,
type.length,
format_desc->block.bits,
fetch_type,
aligned,
base_ptr, offset, FALSE);
/*
* convert texels to float rgba
*/
lp_build_unpack_rgba_soa(gallivm,
format_desc,
type,
packed, rgba_out);
return;
}
if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
format_desc->block.bits > type.width &&
((format_desc->block.bits <= type.width * type.length &&
format_desc->channel[0].size <= type.width) ||
(format_desc->channel[0].size == 64 &&
format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT &&
type.floating)))
{
/*
* Similar to above, but the packed pixel is larger than what fits
* into an element of the destination format. The packed pixels will be
* shuffled into SoA vectors appropriately, and then the extraction will
* be done in parallel as much as possible.
* Good for 16xn (n > 2) and 32xn (n > 1) formats, care is taken so
* the gathered vectors can be shuffled easily (even with avx).
* 64xn float -> 32xn float is handled too but it's a bit special as
* it does the conversion pre-shuffle.
*/
LLVMValueRef packed[4], dst[4], output[4], shuffles[LP_MAX_VECTOR_WIDTH/32];
struct lp_type fetch_type, gather_type = type;
unsigned num_gather, fetch_width, i, j;
struct lp_build_context bld;
boolean fp64 = format_desc->channel[0].size == 64;
lp_build_context_init(&bld, gallivm, type);
assert(type.width == 32);
assert(format_desc->block.bits > type.width);
/*
* First, figure out fetch order.
*/
fetch_width = util_next_power_of_two(format_desc->block.bits);
num_gather = fetch_width / type.width;
/*
* fp64 are treated like fp32 except we fetch twice wide values
* (as we shuffle after trunc). The shuffles for that work out
* mostly fine (slightly suboptimal for 4-wide, perfect for AVX)
* albeit we miss the potential opportunity for hw gather (as it
* only handles native size).
*/
num_gather = fetch_width / type.width;
gather_type.width *= num_gather;
if (fp64) {
num_gather /= 2;
}
gather_type.length /= num_gather;
for (i = 0; i < num_gather; i++) {
LLVMValueRef offsetr, shuf_vec;
if(num_gather == 4) {
for (j = 0; j < gather_type.length; j++) {
unsigned idx = i + 4*j;
shuffles[j] = lp_build_const_int32(gallivm, idx);
}
shuf_vec = LLVMConstVector(shuffles, gather_type.length);
offsetr = LLVMBuildShuffleVector(builder, offset, offset, shuf_vec, "");
}
else if (num_gather == 2) {
assert(num_gather == 2);
for (j = 0; j < gather_type.length; j++) {
unsigned idx = i*2 + (j%2) + (j/2)*4;
shuffles[j] = lp_build_const_int32(gallivm, idx);
}
shuf_vec = LLVMConstVector(shuffles, gather_type.length);
offsetr = LLVMBuildShuffleVector(builder, offset, offset, shuf_vec, "");
}
else {
assert(num_gather == 1);
offsetr = offset;
}
if (gather_type.length == 1) {
LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
offsetr = LLVMBuildExtractElement(builder, offsetr, zero, "");
}
/*
* Determine whether to use float or int loads. This is mostly
* to outsmart the (stupid) llvm int/float shuffle logic, we
* don't really care much if the data is floats or ints...
* But llvm will refuse to use single float shuffle with int data
* and instead use 3 int shuffles instead, the code looks atrocious.
* (Note bitcasts often won't help, as llvm is too smart to be
* fooled by that.)
* Nobody cares about simd float<->int domain transition penalties,
* which usually don't even exist for shuffles anyway.
* With 4x32bit (and 3x32bit) fetch, we use float vec (the data is
* going into transpose, which is unpacks, so doesn't really matter
* much).
* With 2x32bit or 4x16bit fetch, we use float vec, since those
* go into the weird channel separation shuffle. With floats,
* this is (with 128bit vectors):
* - 2 movq, 2 movhpd, 2 shufps
* With ints it would be:
* - 4 movq, 2 punpcklqdq, 4 pshufd, 2 blendw
* I've seen texture functions increase in code size by 15% just due
* to that (there's lots of such fetches in them...)
* (We could chose a different gather order to improve this somewhat
* for the int path, but it would basically just drop the blends,
* so the float path with this order really is optimal.)
* Albeit it is tricky sometimes llvm doesn't ignore the float->int
* casts so must avoid them until we're done with the float shuffle...
* 3x16bit formats (the same is also true for 3x8) are pretty bad but
* there's nothing we can do about them (we could overallocate by
* those couple bytes and use unaligned but pot sized load).
* Note that this is very much x86 specific. I don't know if this
* affect other archs at all.
*/
if (num_gather > 1) {
/*
* We always want some float type here (with x86)
* due to shuffles being float ones afterwards (albeit for
* the num_gather == 4 case int should work fine too
* (unless there's some problems with avx but not avx2).
*/
if (format_desc->channel[0].size == 64) {
fetch_type = lp_type_float_vec(64, gather_type.width);
} else {
fetch_type = lp_type_int_vec(32, gather_type.width);
}
}
else {
/* type doesn't matter much */
if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT &&
(format_desc->channel[0].size == 32 ||
format_desc->channel[0].size == 64)) {
fetch_type = lp_type_float(gather_type.width);
} else {
fetch_type = lp_type_uint(gather_type.width);
}
}
/* Now finally gather the values */
packed[i] = lp_build_gather(gallivm, gather_type.length,
format_desc->block.bits,
fetch_type, aligned,
base_ptr, offsetr, FALSE);
if (fp64) {
struct lp_type conv_type = type;
conv_type.width *= 2;
packed[i] = LLVMBuildBitCast(builder, packed[i],
lp_build_vec_type(gallivm, conv_type), "");
packed[i] = LLVMBuildFPTrunc(builder, packed[i], bld.vec_type, "");
}
}
/* shuffle the gathered values to SoA */
if (num_gather == 2) {
for (i = 0; i < num_gather; i++) {
for (j = 0; j < type.length; j++) {
unsigned idx = (j%2)*2 + (j/4)*4 + i;
if ((j/2)%2)
idx += type.length;
shuffles[j] = lp_build_const_int32(gallivm, idx);
}
dst[i] = LLVMBuildShuffleVector(builder, packed[0], packed[1],
LLVMConstVector(shuffles, type.length), "");
}
}
else if (num_gather == 4) {
lp_build_transpose_aos(gallivm, lp_int_type(type), packed, dst);
}
else {
assert(num_gather == 1);
dst[0] = packed[0];
}
/*
* And finally unpack exactly as above, except that
* chan shift is adjusted and the right vector selected.
*/
if (!fp64) {
for (i = 0; i < num_gather; i++) {
dst[i] = LLVMBuildBitCast(builder, dst[i], bld.int_vec_type, "");
}
for (i = 0; i < format_desc->nr_channels; i++) {
struct util_format_channel_description chan_desc = format_desc->channel[i];
unsigned blockbits = type.width;
unsigned vec_nr = chan_desc.shift / type.width;
chan_desc.shift %= type.width;
output[i] = lp_build_extract_soa_chan(&bld,
blockbits,
FALSE,
chan_desc,
dst[vec_nr]);
}
}
else {
for (i = 0; i < format_desc->nr_channels; i++) {
output[i] = dst[i];
}
}
lp_build_format_swizzle_soa(format_desc, &bld, output, rgba_out);
return;
}
if (format == PIPE_FORMAT_R11G11B10_FLOAT ||
format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
/*
* similar conceptually to above but requiring special
* AoS packed -> SoA float conversion code.
*/
LLVMValueRef packed;
struct lp_type fetch_type = lp_type_uint(type.width);
assert(type.floating);
assert(type.width == 32);
packed = lp_build_gather(gallivm, type.length,
format_desc->block.bits,
fetch_type, aligned,
base_ptr, offset, FALSE);
if (format == PIPE_FORMAT_R11G11B10_FLOAT) {
lp_build_r11g11b10_to_float(gallivm, packed, rgba_out);
}
else {
lp_build_rgb9e5_to_float(gallivm, packed, rgba_out);
}
return;
}
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS &&
format_desc->block.bits == 64) {
/*
* special case the format is 64 bits but we only require
* 32bit (or 8bit) from each block.
*/
LLVMValueRef packed;
struct lp_type fetch_type = lp_type_uint(type.width);
if (format == PIPE_FORMAT_X32_S8X24_UINT) {
/*
* for stencil simply fix up offsets - could in fact change
* base_ptr instead even outside the shader.
*/
unsigned mask = (1 << 8) - 1;
LLVMValueRef s_offset = lp_build_const_int_vec(gallivm, type, 4);
offset = LLVMBuildAdd(builder, offset, s_offset, "");
packed = lp_build_gather(gallivm, type.length, 32, fetch_type,
aligned, base_ptr, offset, FALSE);
packed = LLVMBuildAnd(builder, packed,
lp_build_const_int_vec(gallivm, type, mask), "");
}
else {
assert (format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT);
packed = lp_build_gather(gallivm, type.length, 32, fetch_type,
aligned, base_ptr, offset, TRUE);
packed = LLVMBuildBitCast(builder, packed,
lp_build_vec_type(gallivm, type), "");
}
/* for consistency with lp_build_unpack_rgba_soa() return sss1 or zzz1 */
rgba_out[0] = rgba_out[1] = rgba_out[2] = packed;
rgba_out[3] = lp_build_const_vec(gallivm, type, 1.0f);
return;
}
/*
* Try calling lp_build_fetch_rgba_aos for all pixels.
* Should only really hit subsampled, compressed
* (for s3tc srgb too, for rgtc the unorm ones only) by now.
* (This is invalid for plain 8unorm formats because we're lazy with
* the swizzle since some results would arrive swizzled, some not.)
*/
if ((format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN) &&
(util_format_fits_8unorm(format_desc) ||
format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) &&
type.floating && type.width == 32 &&
(type.length == 1 || (type.length % 4 == 0))) {
struct lp_type tmp_type;
struct lp_build_context bld;
LLVMValueRef packed, rgba[4];
const struct util_format_description *flinear_desc;
const struct util_format_description *frgba8_desc;
unsigned chan;
lp_build_context_init(&bld, gallivm, type);
/*
* Make sure the conversion in aos really only does convert to rgba8
* and not anything more (so use linear format, adjust type).
*/
flinear_desc = util_format_description(util_format_linear(format));
memset(&tmp_type, 0, sizeof tmp_type);
tmp_type.width = 8;
tmp_type.length = type.length * 4;
tmp_type.norm = TRUE;
packed = lp_build_fetch_rgba_aos(gallivm, flinear_desc, tmp_type,
aligned, base_ptr, offset, i, j, cache);
packed = LLVMBuildBitCast(builder, packed, bld.int_vec_type, "");
/*
* The values are now packed so they match ordinary (srgb) RGBA8 format,
* hence need to use matching format for unpack.
*/
frgba8_desc = util_format_description(PIPE_FORMAT_R8G8B8A8_UNORM);
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
assert(format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC);
frgba8_desc = util_format_description(PIPE_FORMAT_R8G8B8A8_SRGB);
}
lp_build_unpack_rgba_soa(gallivm,
frgba8_desc,
type,
packed, rgba);
/*
* We converted 4 channels. Make sure llvm can drop unneeded ones
* (luckily the rgba order is fixed, only LA needs special case).
*/
for (chan = 0; chan < 4; chan++) {
enum pipe_swizzle swizzle = format_desc->swizzle[chan];
if (chan == 3 && util_format_is_luminance_alpha(format)) {
swizzle = PIPE_SWIZZLE_W;
}
rgba_out[chan] = lp_build_swizzle_soa_channel(&bld, rgba, swizzle);
}
return;
}
/*
* Fallback to calling lp_build_fetch_rgba_aos for each pixel.
*
* This is not the most efficient way of fetching pixels, as we
* miss some opportunities to do vectorization, but this is
* convenient for formats or scenarios for which there was no
* opportunity or incentive to optimize.
*
* We do NOT want to end up here, this typically is quite terrible,
* in particular if the formats have less than 4 channels.
*
* Right now, this should only be hit for:
* - RGTC snorm formats
* (those miss fast fetch functions hence they are terrible anyway)
*/
{
unsigned k;
struct lp_type tmp_type;
LLVMValueRef aos_fetch[LP_MAX_VECTOR_WIDTH / 32];
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: AoS fetch fallback for %s\n",
__FUNCTION__, format_desc->short_name);
}
tmp_type = type;
tmp_type.length = 4;
/*
* Note that vector transpose can be worse compared to insert/extract
* for aos->soa conversion (for formats with 1 or 2 channels). However,
* we should try to avoid getting here for just about all formats, so
* don't bother.
*/
/* loop over number of pixels */
for(k = 0; k < type.length; ++k) {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
LLVMValueRef offset_elem;
LLVMValueRef i_elem, j_elem;
offset_elem = LLVMBuildExtractElement(builder, offset,
index, "");
i_elem = LLVMBuildExtractElement(builder, i, index, "");
j_elem = LLVMBuildExtractElement(builder, j, index, "");
/* Get a single float[4]={R,G,B,A} pixel */
aos_fetch[k] = lp_build_fetch_rgba_aos(gallivm, format_desc, tmp_type,
aligned, base_ptr, offset_elem,
i_elem, j_elem, cache);
}
convert_to_soa(gallivm, aos_fetch, rgba_out, type);
}
}
/**
* Expand the relevent bits of mask_input to a 4-dword mask for the
* four pixels in a 2x2 quad. This will set the four elements of the
* quad mask vector to 0 or ~0.
*
* \param quad which quad of the quad group to test, in [0,3]
* \param mask_input bitwise mask for the whole 4x4 stamp
*/
static LLVMValueRef
generate_quad_mask(LLVMBuilderRef builder,
struct lp_type fs_type,
unsigned quad,
LLVMValueRef mask_input) /* int32 */
{
struct lp_type mask_type;
LLVMTypeRef i32t = LLVMInt32Type();
LLVMValueRef bits[4];
LLVMValueRef mask;
int shift;
/*
* XXX: We'll need a different path for 16 x u8
*/
assert(fs_type.width == 32);
assert(fs_type.length == 4);
mask_type = lp_int_type(fs_type);
/*
* mask_input >>= (quad * 4)
*/
switch (quad) {
case 0:
shift = 0;
break;
case 1:
shift = 2;
break;
case 2:
shift = 8;
break;
case 3:
shift = 10;
break;
default:
assert(0);
shift = 0;
}
mask_input = LLVMBuildLShr(builder,
mask_input,
LLVMConstInt(i32t, shift, 0),
"");
/*
* mask = { mask_input & (1 << i), for i in [0,3] }
*/
mask = lp_build_broadcast(builder, lp_build_vec_type(mask_type), mask_input);
bits[0] = LLVMConstInt(i32t, 1 << 0, 0);
bits[1] = LLVMConstInt(i32t, 1 << 1, 0);
bits[2] = LLVMConstInt(i32t, 1 << 4, 0);
bits[3] = LLVMConstInt(i32t, 1 << 5, 0);
mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), "");
/*
* mask = mask != 0 ? ~0 : 0
*/
mask = lp_build_compare(builder,
mask_type, PIPE_FUNC_NOTEQUAL,
mask,
lp_build_const_int_vec(mask_type, 0));
return mask;
}
