void
lp_build_alpha_test(struct gallivm_state *gallivm,
unsigned func,
struct lp_type type,
const struct util_format_description *cbuf_format_desc,
struct lp_build_mask_context *mask,
LLVMValueRef alpha,
LLVMValueRef ref,
boolean do_branch)
{
struct lp_build_context bld;
LLVMValueRef test;
lp_build_context_init(&bld, gallivm, type);
/*
* Alpha testing needs to be done in the color buffer precision.
*
* TODO: Ideally, instead of duplicating the color conversion code, we would do
* alpha testing after converting the output colors, but that's not very
* convenient, because it needs to be done before depth testing. Hopefully
* LLVM will detect and remove the duplicate expression.
*
* FIXME: This should be generalized to formats other than rgba8 variants.
*/
if (type.floating &&
util_format_is_rgba8_variant(cbuf_format_desc)) {
const unsigned dst_width = 8;
alpha = lp_build_clamp(&bld, alpha, bld.zero, bld.one);
ref = lp_build_clamp(&bld, ref, bld.zero, bld.one);
alpha = lp_build_clamped_float_to_unsigned_norm(gallivm, type, dst_width, alpha);
ref = lp_build_clamped_float_to_unsigned_norm(gallivm, type, dst_width, ref);
type.floating = 0;
lp_build_context_init(&bld, gallivm, type);
}
test = lp_build_cmp(&bld, func, alpha, ref);
lp_build_name(test, "alpha_mask");
lp_build_mask_update(mask, test);
if (do_branch)
lp_build_mask_check(mask);
}
/**
* Generate color blending and color output.
* \param rt the render target index (to index blend, colormask state)
* \param type the pixel color type
* \param context_ptr pointer to the runtime JIT context
* \param mask execution mask (active fragment/pixel mask)
* \param src colors from the fragment shader
* \param dst_ptr the destination color buffer pointer
*/
static void
generate_blend(const struct pipe_blend_state *blend,
unsigned rt,
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;
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);
const_ptr = lp_jit_context_blend_color(builder, context_ptr);
const_ptr = LLVMBuildBitCast(builder, const_ptr,
LLVMPointerType(vec_type, 0), "");
/* load constant blend color and colors from the dest color buffer */
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]);
}
/* do blend */
lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
/* store results to color buffer */
for(chan = 0; chan < 4; ++chan) {
if(blend->rt[rt].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);
}
/**
* Extract Y, U, V channels from packed YUYV.
* @param packed is a <n x i32> vector with the packed YUYV blocks
* @param i is a <n x i32> vector with the x pixel coordinate (0 or 1)
*/
static void
yuyv_to_yuv_soa(struct gallivm_state *gallivm,
unsigned n,
LLVMValueRef packed,
LLVMValueRef i,
LLVMValueRef *y,
LLVMValueRef *u,
LLVMValueRef *v)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_type type;
LLVMValueRef mask;
memset(&type, 0, sizeof type);
type.width = 32;
type.length = n;
assert(lp_check_value(type, packed));
assert(lp_check_value(type, i));
/*
* y = (yuyv >> 16*i) & 0xff
* u = (yuyv >> 8 ) & 0xff
* v = (yuyv >> 24 ) & 0xff
*/
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
/*
* Avoid shift with per-element count.
* No support on x86, gets translated to roughly 5 instructions
* per element. Didn't measure performance but cuts shader size
* by quite a bit (less difference if cpu has no sse4.1 support).
*/
if (util_cpu_caps.has_sse2 && n == 4) {
LLVMValueRef sel, tmp;
struct lp_build_context bld32;
lp_build_context_init(&bld32, gallivm, type);
tmp = LLVMBuildLShr(builder, packed, lp_build_const_int_vec(gallivm, type, 16), "");
sel = lp_build_compare(gallivm, type, PIPE_FUNC_EQUAL, i, lp_build_const_int_vec(gallivm, type, 0));
*y = lp_build_select(&bld32, sel, packed, tmp);
} else
#endif
{
LLVMValueRef shift;
shift = LLVMBuildMul(builder, i, lp_build_const_int_vec(gallivm, type, 16), "");
*y = LLVMBuildLShr(builder, packed, shift, "");
}
*u = LLVMBuildLShr(builder, packed, lp_build_const_int_vec(gallivm, type, 8), "");
*v = LLVMBuildLShr(builder, packed, lp_build_const_int_vec(gallivm, type, 24), "");
mask = lp_build_const_int_vec(gallivm, type, 0xff);
*y = LLVMBuildAnd(builder, *y, mask, "y");
*u = LLVMBuildAnd(builder, *u, mask, "u");
*v = LLVMBuildAnd(builder, *v, mask, "v");
}
/**
* 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);
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->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);
}
/**
* @brief lp_build_fetch_rgba_aos_array
*
* \param format_desc describes format of the image we're fetching from
* \param dst_type output type
* \param base_ptr address of the pixel block (or the texel if uncompressed)
* \param offset ptr offset
*/
LLVMValueRef
lp_build_fetch_rgba_aos_array(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type dst_type,
LLVMValueRef base_ptr,
LLVMValueRef offset)
{
struct lp_build_context bld;
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef src_elem_type, src_vec_type;
LLVMValueRef ptr, res = NULL;
struct lp_type src_type;
memset(&src_type, 0, sizeof src_type);
src_type.floating = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT;
src_type.fixed = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FIXED;
src_type.sign = format_desc->channel[0].type != UTIL_FORMAT_TYPE_UNSIGNED;
src_type.norm = format_desc->channel[0].normalized;
src_type.width = format_desc->channel[0].size;
src_type.length = format_desc->nr_channels;
assert(src_type.length <= dst_type.length);
src_elem_type = lp_build_elem_type(gallivm, src_type);
src_vec_type = lp_build_vec_type(gallivm, src_type);
/* Read whole vector from memory, unaligned */
if (!res) {
ptr = LLVMBuildGEP(builder, base_ptr, &offset, 1, "");
ptr = LLVMBuildPointerCast(builder, ptr, LLVMPointerType(src_vec_type, 0), "");
res = LLVMBuildLoad(builder, ptr, "");
lp_set_load_alignment(res, src_type.width / 8);
}
/* Truncate doubles to float */
if (src_type.floating && src_type.width == 64) {
src_type.width = 32;
src_vec_type = lp_build_vec_type(gallivm, src_type);
res = LLVMBuildFPTrunc(builder, res, src_vec_type, "");
}
/* Expand to correct length */
if (src_type.length < dst_type.length) {
res = lp_build_pad_vector(gallivm, res, src_type, dst_type.length);
src_type.length = dst_type.length;
}
/* Convert to correct format */
lp_build_conv(gallivm, src_type, dst_type, &res, 1, &res, 1);
/* Swizzle it */
lp_build_context_init(&bld, gallivm, dst_type);
return lp_build_format_swizzle_aos(format_desc, &bld, res);
}
/**
* Convert srgb int values to linear float values.
* Several possibilities how to do this, e.g.
* - table
* - doing the pow() with int-to-float and float-to-int tricks
* (http://stackoverflow.com/questions/6475373/optimizations-for-pow-with-const-non-integer-exponent)
* - just using standard polynomial approximation
* (3rd order polynomial is required for crappy but just sufficient accuracy)
*
* @param src integer (vector) value(s) to convert
* (chan_bits bit values unpacked to 32 bit already).
*/
LLVMValueRef
lp_build_srgb_to_linear(struct gallivm_state *gallivm,
struct lp_type src_type,
unsigned chan_bits,
LLVMValueRef src)
{
struct lp_type f32_type = lp_type_float_vec(32, src_type.length * 32);
struct lp_build_context f32_bld;
LLVMValueRef srcf, part_lin, part_pow, is_linear, lin_const, lin_thresh;
double coeffs[4] = {0.0023f,
0.0030f / 255.0f,
0.6935f / (255.0f * 255.0f),
0.3012f / (255.0f * 255.0f * 255.0f)
};
assert(src_type.width == 32);
/* Technically this would work with more bits too but would be inaccurate. */
assert(chan_bits <= 8);
lp_build_context_init(&f32_bld, gallivm, f32_type);
/*
* using polynomial: (src * (src * (src * 0.3012 + 0.6935) + 0.0030) + 0.0023)
* ( poly = 0.3012*x^3 + 0.6935*x^2 + 0.0030*x + 0.0023)
* (found with octave polyfit and some magic as I couldn't get the error
* function right). Using the above mentioned error function, the values stay
* within +-0.35, except for the lowest values - hence tweaking linear segment
* to cover the first 16 instead of the first 11 values (the error stays
* just about acceptable there too).
* Hence: lin = src > 15 ? poly : src / 12.6
* This function really only makes sense for vectors, should use LUT otherwise.
* All in all (including float conversion) 11 instructions (with sse4.1),
* 6 constants (polynomial could be done with 1 instruction less at the cost
* of slightly worse dependency chain, fma should also help).
*/
/* doing the 1/255 mul as part of the approximation */
srcf = lp_build_int_to_float(&f32_bld, src);
if (chan_bits != 8) {
/* could adjust all the constants instead */
LLVMValueRef rescale_const = lp_build_const_vec(gallivm, f32_type,
255.0f / ((1 << chan_bits) - 1));
srcf = lp_build_mul(&f32_bld, srcf, rescale_const);
}
lin_const = lp_build_const_vec(gallivm, f32_type, 1.0f / (12.6f * 255.0f));
part_lin = lp_build_mul(&f32_bld, srcf, lin_const);
part_pow = lp_build_polynomial(&f32_bld, srcf, coeffs, 4);
lin_thresh = lp_build_const_vec(gallivm, f32_type, 15.0f);
is_linear = lp_build_compare(gallivm, f32_type, PIPE_FUNC_LEQUAL, srcf, lin_thresh);
return lp_build_select(&f32_bld, is_linear, part_lin, part_pow);
}
/**
* 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;
}
/**
* Unpack several pixels in SoA.
*
* It takes a vector of packed pixels:
*
* packed = {P0, P1, P2, P3, ..., Pn}
*
* And will produce four vectors:
*
* red = {R0, R1, R2, R3, ..., Rn}
* green = {G0, G1, G2, G3, ..., Gn}
* blue = {B0, B1, B2, B3, ..., Bn}
* alpha = {A0, A1, A2, A3, ..., An}
*
* It requires that a packed pixel fits into an element of the output
* channels. The common case is when converting pixel with a depth of 32 bit or
* less into floats.
*
* \param format_desc the format of the 'packed' incoming pixel vector
* \param type the desired type for rgba_out (type.length = n, above)
* \param packed the incoming vector of packed pixels
* \param rgba_out returns the SoA R,G,B,A vectors
*/
void
lp_build_unpack_rgba_soa(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
LLVMValueRef packed,
LLVMValueRef rgba_out[4])
{
struct lp_build_context bld;
LLVMValueRef inputs[4];
unsigned chan;
assert(format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
assert(format_desc->block.width == 1);
assert(format_desc->block.height == 1);
assert(format_desc->block.bits <= type.width);
/* FIXME: Support more output types */
assert(type.width == 32);
lp_build_context_init(&bld, gallivm, type);
/* Decode the input vector components */
for (chan = 0; chan < format_desc->nr_channels; ++chan) {
struct util_format_channel_description chan_desc = format_desc->channel[chan];
boolean srgb_chan = FALSE;
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB &&
format_desc->swizzle[3] != chan) {
srgb_chan = TRUE;
}
inputs[chan] = lp_build_extract_soa_chan(&bld,
format_desc->block.bits,
srgb_chan,
chan_desc,
packed);
}
lp_build_format_swizzle_soa(format_desc, &bld, inputs, rgba_out);
}
/*
* Build LLVM function that exercises the unary operator builder.
*/
static LLVMValueRef
build_unary_test_func(struct gallivm_state *gallivm,
const struct unary_test_t *test)
{
struct lp_type type = lp_type_float_vec(32, lp_native_vector_width);
LLVMContextRef context = gallivm->context;
LLVMModuleRef module = gallivm->module;
LLVMTypeRef vf32t = lp_build_vec_type(gallivm, type);
LLVMTypeRef args[2] = { LLVMPointerType(vf32t, 0), LLVMPointerType(vf32t, 0) };
LLVMValueRef func = LLVMAddFunction(module, test->name,
LLVMFunctionType(LLVMVoidTypeInContext(context),
args, Elements(args), 0));
LLVMValueRef arg0 = LLVMGetParam(func, 0);
LLVMValueRef arg1 = LLVMGetParam(func, 1);
LLVMBuilderRef builder = gallivm->builder;
LLVMBasicBlockRef block = LLVMAppendBasicBlockInContext(context, func, "entry");
LLVMValueRef ret;
struct lp_build_context bld;
lp_build_context_init(&bld, gallivm, type);
LLVMSetFunctionCallConv(func, LLVMCCallConv);
LLVMPositionBuilderAtEnd(builder, block);
arg1 = LLVMBuildLoad(builder, arg1, "");
ret = test->builder(&bld, arg1);
LLVMBuildStore(builder, ret, arg0);
LLVMBuildRetVoid(builder);
gallivm_verify_function(gallivm, func);
return func;
}
/**
* Non-interleaved pack and saturate.
*
* Same as lp_build_pack2 but will saturate values so that they fit into the
* destination type.
*/
LLVMValueRef
lp_build_packs2(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
LLVMValueRef lo,
LLVMValueRef hi)
{
boolean clamp;
assert(!src_type.floating);
assert(!dst_type.floating);
assert(src_type.sign == dst_type.sign);
assert(src_type.width == dst_type.width * 2);
assert(src_type.length * 2 == dst_type.length);
clamp = TRUE;
/* All X86 SSE non-interleaved pack instructions take signed inputs and
* saturate them, so no need to clamp for those cases. */
if(util_cpu_caps.has_sse2 &&
src_type.width * src_type.length >= 128 &&
src_type.sign &&
(src_type.width == 32 || src_type.width == 16))
clamp = FALSE;
if(clamp) {
struct lp_build_context bld;
unsigned dst_bits = dst_type.sign ? dst_type.width - 1 : dst_type.width;
LLVMValueRef dst_max = lp_build_const_int_vec(gallivm, src_type, ((unsigned long long)1 << dst_bits) - 1);
lp_build_context_init(&bld, gallivm, src_type);
lo = lp_build_min(&bld, lo, dst_max);
hi = lp_build_min(&bld, hi, dst_max);
/* FIXME: What about lower bound? */
}
return lp_build_pack2(gallivm, src_type, dst_type, lo, hi);
}
void
lp_build_alpha_to_coverage(struct gallivm_state *gallivm,
struct lp_type type,
struct lp_build_mask_context *mask,
LLVMValueRef alpha,
boolean do_branch)
{
struct lp_build_context bld;
LLVMValueRef test;
LLVMValueRef alpha_ref_value;
lp_build_context_init(&bld, gallivm, type);
alpha_ref_value = lp_build_const_vec(gallivm, type, 0.5);
test = lp_build_cmp(&bld, PIPE_FUNC_GREATER, alpha, alpha_ref_value);
lp_build_name(test, "alpha_to_coverage");
lp_build_mask_update(mask, test);
if (do_branch)
lp_build_mask_check(mask);
}
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);
}
}
/**
* Initialize fragment shader input attribute info.
*/
void
lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
struct gallivm_state *gallivm,
unsigned num_inputs,
const struct lp_shader_input *inputs,
boolean pixel_center_integer,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef a0_ptr,
LLVMValueRef dadx_ptr,
LLVMValueRef dady_ptr,
LLVMValueRef x0,
LLVMValueRef y0)
{
struct lp_type coeff_type;
struct lp_type setup_type;
unsigned attrib;
unsigned chan;
memset(bld, 0, sizeof *bld);
memset(&coeff_type, 0, sizeof coeff_type);
coeff_type.floating = TRUE;
coeff_type.sign = TRUE;
coeff_type.width = 32;
coeff_type.length = type.length;
memset(&setup_type, 0, sizeof setup_type);
setup_type.floating = TRUE;
setup_type.sign = TRUE;
setup_type.width = 32;
setup_type.length = TGSI_NUM_CHANNELS;
/* XXX: we don't support interpolating into any other types */
assert(memcmp(&coeff_type, &type, sizeof coeff_type) == 0);
lp_build_context_init(&bld->coeff_bld, gallivm, coeff_type);
lp_build_context_init(&bld->setup_bld, gallivm, setup_type);
/* For convenience */
bld->pos = bld->attribs[0];
bld->inputs = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) bld->attribs[1];
/* Position */
bld->mask[0] = TGSI_WRITEMASK_XYZW;
bld->interp[0] = LP_INTERP_LINEAR;
/* Inputs */
for (attrib = 0; attrib < num_inputs; ++attrib) {
bld->mask[1 + attrib] = inputs[attrib].usage_mask;
bld->interp[1 + attrib] = inputs[attrib].interp;
}
bld->num_attribs = 1 + num_inputs;
/* Ensure all masked out input channels have a valid value */
for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
bld->attribs[attrib][chan] = bld->coeff_bld.undef;
}
}
if (pixel_center_integer) {
bld->pos_offset = 0.0;
} else {
bld->pos_offset = 0.5;
}
pos_init(bld, x0, y0);
/*
* Simple method (single step interpolation) may be slower if vector length
* is just 4, but the results are different (generally less accurate) with
* the other method, so always use more accurate version.
*/
if (1) {
bld->simple_interp = TRUE;
{
/* XXX this should use a global static table */
unsigned i;
unsigned num_loops = 16 / type.length;
LLVMValueRef pixoffx, pixoffy, index;
LLVMValueRef ptr;
bld->xoffset_store = lp_build_array_alloca(gallivm,
lp_build_vec_type(gallivm, type),
lp_build_const_int32(gallivm, num_loops),
"");
bld->yoffset_store = lp_build_array_alloca(gallivm,
lp_build_vec_type(gallivm, type),
lp_build_const_int32(gallivm, num_loops),
"");
for (i = 0; i < num_loops; i++) {
index = lp_build_const_int32(gallivm, i);
calc_offsets(&bld->coeff_bld, i*type.length/4, &pixoffx, &pixoffy);
ptr = LLVMBuildGEP(builder, bld->xoffset_store, &index, 1, "");
LLVMBuildStore(builder, pixoffx, ptr);
ptr = LLVMBuildGEP(builder, bld->yoffset_store, &index, 1, "");
LLVMBuildStore(builder, pixoffy, ptr);
}
}
coeffs_init_simple(bld, a0_ptr, dadx_ptr, dady_ptr);
}
else {
bld->simple_interp = FALSE;
coeffs_init(bld, a0_ptr, dadx_ptr, dady_ptr);
}
}
/**
* Generate blend code in SOA mode.
* \param src src/fragment color
* \param dst dst/framebuffer color
* \param con constant blend color
* \param res the result/output
*/
void
lp_build_blend_soa(LLVMBuilderRef builder,
const struct pipe_blend_state *blend,
struct lp_type type,
LLVMValueRef src[4],
LLVMValueRef dst[4],
LLVMValueRef con[4],
LLVMValueRef res[4])
{
struct lp_build_blend_soa_context bld;
unsigned i, j, k;
/* Setup build context */
memset(&bld, 0, sizeof bld);
lp_build_context_init(&bld.base, builder, type);
for (i = 0; i < 4; ++i) {
bld.src[i] = src[i];
bld.dst[i] = dst[i];
bld.con[i] = con[i];
}
for (i = 0; i < 4; ++i) {
if (blend->colormask & (1 << i)) {
if (blend->logicop_enable) {
if(!type.floating) {
res[i] = lp_build_logicop(builder, blend->logicop_func, src[i], dst[i]);
}
else
res[i] = dst[i];
}
else if (blend->blend_enable) {
unsigned src_factor = i < 3 ? blend->rgb_src_factor : blend->alpha_src_factor;
unsigned dst_factor = i < 3 ? blend->rgb_dst_factor : blend->alpha_dst_factor;
unsigned func = i < 3 ? blend->rgb_func : blend->alpha_func;
boolean func_commutative = lp_build_blend_func_commutative(func);
/* It makes no sense to blend unless values are normalized */
assert(type.norm);
/*
* Compute src/dst factors.
*/
bld.factor[0][0][i] = src[i];
bld.factor[0][1][i] = lp_build_blend_soa_factor(&bld, src_factor, i);
bld.factor[1][0][i] = dst[i];
bld.factor[1][1][i] = lp_build_blend_soa_factor(&bld, dst_factor, i);
/*
* Compute src/dst terms
*/
for(k = 0; k < 2; ++k) {
/* See if this multiplication has been previously computed */
for(j = 0; j < i; ++j) {
if((bld.factor[k][0][j] == bld.factor[k][0][i] &&
bld.factor[k][1][j] == bld.factor[k][1][i]) ||
(bld.factor[k][0][j] == bld.factor[k][1][i] &&
bld.factor[k][1][j] == bld.factor[k][0][i]))
break;
}
if(j < i)
bld.term[k][i] = bld.term[k][j];
else
bld.term[k][i] = lp_build_mul(&bld.base, bld.factor[k][0][i], bld.factor[k][1][i]);
}
/*
* Combine terms
*/
/* See if this function has been previously applied */
for(j = 0; j < i; ++j) {
unsigned prev_func = j < 3 ? blend->rgb_func : blend->alpha_func;
unsigned func_reverse = lp_build_blend_func_reverse(func, prev_func);
if((!func_reverse &&
bld.term[0][j] == bld.term[0][i] &&
bld.term[1][j] == bld.term[1][i]) ||
((func_commutative || func_reverse) &&
bld.term[0][j] == bld.term[1][i] &&
bld.term[1][j] == bld.term[0][i]))
break;
}
if(j < i)
res[i] = res[j];
else
res[i] = lp_build_blend_func(&bld.base, func, bld.term[0][i], bld.term[1][i]);
}
else {
res[i] = src[i];
}
}
else {
res[i] = dst[i];
}
}
}
/**
* @brief lp_build_fetch_rgba_aos_array
*
* \param format_desc describes format of the image we're fetching from
* \param dst_type output type
* \param base_ptr address of the pixel block (or the texel if uncompressed)
* \param offset ptr offset
*/
LLVMValueRef
lp_build_fetch_rgba_aos_array(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type dst_type,
LLVMValueRef base_ptr,
LLVMValueRef offset)
{
struct lp_build_context bld;
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef src_vec_type;
LLVMValueRef ptr, res = NULL;
struct lp_type src_type;
boolean pure_integer = format_desc->channel[0].pure_integer;
struct lp_type tmp_type;
lp_type_from_format_desc(&src_type, format_desc);
assert(src_type.length <= dst_type.length);
src_vec_type = lp_build_vec_type(gallivm, src_type);
/* Read whole vector from memory, unaligned */
ptr = LLVMBuildGEP(builder, base_ptr, &offset, 1, "");
ptr = LLVMBuildPointerCast(builder, ptr, LLVMPointerType(src_vec_type, 0), "");
res = LLVMBuildLoad(builder, ptr, "");
LLVMSetAlignment(res, src_type.width / 8);
/* Truncate doubles to float */
if (src_type.floating && src_type.width == 64) {
src_type.width = 32;
src_vec_type = lp_build_vec_type(gallivm, src_type);
res = LLVMBuildFPTrunc(builder, res, src_vec_type, "");
}
/* Expand to correct length */
if (src_type.length < dst_type.length) {
res = lp_build_pad_vector(gallivm, res, dst_type.length);
src_type.length = dst_type.length;
}
tmp_type = dst_type;
if (pure_integer) {
/* some callers expect (fake) floats other real ints. */
tmp_type.floating = 0;
tmp_type.sign = src_type.sign;
}
/* Convert to correct format */
lp_build_conv(gallivm, src_type, tmp_type, &res, 1, &res, 1);
/* Swizzle it */
lp_build_context_init(&bld, gallivm, tmp_type);
res = lp_build_format_swizzle_aos(format_desc, &bld, res);
/* Bitcast to floats (for pure integers) when requested */
if (pure_integer && dst_type.floating) {
res = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, dst_type), "");
}
return res;
}
/**
* Generate blend code in SOA mode.
* \param rt render target index (to index the blend / colormask state)
* \param src src/fragment color
* \param dst dst/framebuffer color
* \param con constant blend color
* \param res the result/output
*/
void
lp_build_blend_soa(struct gallivm_state *gallivm,
const struct pipe_blend_state *blend,
struct lp_type type,
unsigned rt,
LLVMValueRef src[4],
LLVMValueRef dst[4],
LLVMValueRef con[4],
LLVMValueRef res[4])
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_build_blend_soa_context bld;
unsigned i, j, k;
assert(rt < PIPE_MAX_COLOR_BUFS);
/* Setup build context */
memset(&bld, 0, sizeof bld);
lp_build_context_init(&bld.base, gallivm, type);
for (i = 0; i < 4; ++i) {
bld.src[i] = src[i];
bld.dst[i] = dst[i];
bld.con[i] = con[i];
}
for (i = 0; i < 4; ++i) {
/* only compute blending for the color channels enabled for writing */
if (blend->rt[rt].colormask & (1 << i)) {
if (blend->logicop_enable) {
if(!type.floating) {
res[i] = lp_build_logicop(builder, blend->logicop_func, src[i], dst[i]);
}
else
res[i] = dst[i];
}
else if (blend->rt[rt].blend_enable) {
unsigned src_factor = i < 3 ? blend->rt[rt].rgb_src_factor : blend->rt[rt].alpha_src_factor;
unsigned dst_factor = i < 3 ? blend->rt[rt].rgb_dst_factor : blend->rt[rt].alpha_dst_factor;
unsigned func = i < 3 ? blend->rt[rt].rgb_func : blend->rt[rt].alpha_func;
boolean func_commutative = lp_build_blend_func_commutative(func);
/*
* Compute src/dst factors.
*/
bld.factor[0][0][i] = src[i];
bld.factor[0][1][i] = lp_build_blend_soa_factor(&bld, src_factor, i);
bld.factor[1][0][i] = dst[i];
bld.factor[1][1][i] = lp_build_blend_soa_factor(&bld, dst_factor, i);
/*
* Check if lp_build_blend can perform any optimisations
*/
res[i] = lp_build_blend(&bld.base,
func,
src_factor,
dst_factor,
bld.factor[0][0][i],
bld.factor[1][0][i],
bld.factor[0][1][i],
bld.factor[1][1][i],
true,
true);
if (res[i]) {
continue;
}
/*
* Compute src/dst terms
*/
for(k = 0; k < 2; ++k) {
/* See if this multiplication has been previously computed */
for(j = 0; j < i; ++j) {
if((bld.factor[k][0][j] == bld.factor[k][0][i] &&
bld.factor[k][1][j] == bld.factor[k][1][i]) ||
(bld.factor[k][0][j] == bld.factor[k][1][i] &&
bld.factor[k][1][j] == bld.factor[k][0][i]))
break;
}
if(j < i && bld.term[k][j])
bld.term[k][i] = bld.term[k][j];
else
bld.term[k][i] = lp_build_mul(&bld.base, bld.factor[k][0][i], bld.factor[k][1][i]);
if (src_factor == PIPE_BLENDFACTOR_ZERO &&
(dst_factor == PIPE_BLENDFACTOR_DST_ALPHA ||
dst_factor == PIPE_BLENDFACTOR_INV_DST_ALPHA)) {
/* XXX special case these combos to work around an apparent
* bug in LLVM.
* This hack disables the check for multiplication by zero
* in lp_bld_mul(). When we optimize away the
* multiplication, something goes wrong during code
* generation and we segfault at runtime.
*/
LLVMValueRef zeroSave = bld.base.zero;
bld.base.zero = NULL;
bld.term[k][i] = lp_build_mul(&bld.base, bld.factor[k][0][i],
bld.factor[k][1][i]);
bld.base.zero = zeroSave;
}
}
/*
* Combine terms
*/
/* See if this function has been previously applied */
for(j = 0; j < i; ++j) {
unsigned prev_func = j < 3 ? blend->rt[rt].rgb_func : blend->rt[rt].alpha_func;
unsigned func_reverse = lp_build_blend_func_reverse(func, prev_func);
if((!func_reverse &&
bld.term[0][j] == bld.term[0][i] &&
bld.term[1][j] == bld.term[1][i]) ||
((func_commutative || func_reverse) &&
bld.term[0][j] == bld.term[1][i] &&
bld.term[1][j] == bld.term[0][i]))
break;
}
if(j < i)
res[i] = res[j];
else
res[i] = lp_build_blend_func(&bld.base, func, bld.term[0][i], bld.term[1][i]);
}
else {
res[i] = src[i];
}
}
else {
res[i] = dst[i];
}
}
}
/**
* Performs blending of src and dst pixels
*
* @param blend the blend state of the shader variant
* @param cbuf_format format of the colour buffer
* @param type data type of the pixel vector
* @param rt render target index
* @param src blend src
* @param src_alpha blend src alpha (if not included in src)
* @param src1 second blend src (for dual source blend)
* @param src1_alpha second blend src alpha (if not included in src1)
* @param dst blend dst
* @param mask optional mask to apply to the blending result
* @param const_ const blend color
* @param const_alpha const blend color alpha (if not included in const_)
* @param swizzle swizzle values for RGBA
*
* @return the result of blending src and dst
*/
LLVMValueRef
lp_build_blend_aos(struct gallivm_state *gallivm,
const struct pipe_blend_state *blend,
enum pipe_format cbuf_format,
struct lp_type type,
unsigned rt,
LLVMValueRef src,
LLVMValueRef src_alpha,
LLVMValueRef src1,
LLVMValueRef src1_alpha,
LLVMValueRef dst,
LLVMValueRef mask,
LLVMValueRef const_,
LLVMValueRef const_alpha,
const unsigned char swizzle[4],
int nr_channels)
{
const struct pipe_rt_blend_state * state = &blend->rt[rt];
const struct util_format_description * desc;
struct lp_build_blend_aos_context bld;
LLVMValueRef src_factor, dst_factor;
LLVMValueRef result;
unsigned alpha_swizzle = UTIL_FORMAT_SWIZZLE_NONE;
unsigned i;
desc = util_format_description(cbuf_format);
/* Setup build context */
memset(&bld, 0, sizeof bld);
lp_build_context_init(&bld.base, gallivm, type);
bld.src = src;
bld.src1 = src1;
bld.dst = dst;
bld.const_ = const_;
bld.src_alpha = src_alpha;
bld.src1_alpha = src1_alpha;
bld.const_alpha = const_alpha;
/* Find the alpha channel if not provided seperately */
if (!src_alpha) {
for (i = 0; i < 4; ++i) {
if (swizzle[i] == 3) {
alpha_swizzle = i;
}
}
}
if (blend->logicop_enable) {
if(!type.floating) {
result = lp_build_logicop(gallivm->builder, blend->logicop_func, src, dst);
}
else {
result = src;
}
} else if (!state->blend_enable) {
result = src;
} else {
boolean rgb_alpha_same = (state->rgb_src_factor == state->rgb_dst_factor && state->alpha_src_factor == state->alpha_dst_factor) || nr_channels == 1;
src_factor = lp_build_blend_factor(&bld, state->rgb_src_factor,
state->alpha_src_factor,
alpha_swizzle,
nr_channels);
dst_factor = lp_build_blend_factor(&bld, state->rgb_dst_factor,
state->alpha_dst_factor,
alpha_swizzle,
nr_channels);
result = lp_build_blend(&bld.base,
state->rgb_func,
state->rgb_src_factor,
state->rgb_dst_factor,
src,
dst,
src_factor,
dst_factor,
rgb_alpha_same,
false);
if(state->rgb_func != state->alpha_func && nr_channels > 1 && alpha_swizzle != UTIL_FORMAT_SWIZZLE_NONE) {
LLVMValueRef alpha;
alpha = lp_build_blend(&bld.base,
state->alpha_func,
state->alpha_src_factor,
state->alpha_dst_factor,
src,
dst,
src_factor,
dst_factor,
rgb_alpha_same,
false);
result = lp_build_blend_swizzle(&bld,
result,
alpha,
LP_BUILD_BLEND_SWIZZLE_RGBA,
alpha_swizzle,
nr_channels);
}
}
/* Check if color mask is necessary */
if (!util_format_colormask_full(desc, state->colormask)) {
LLVMValueRef color_mask;
color_mask = lp_build_const_mask_aos_swizzled(gallivm, bld.base.type, state->colormask, nr_channels, swizzle);
lp_build_name(color_mask, "color_mask");
/* Combine with input mask if necessary */
if (mask) {
/* We can be blending floating values but masks are always integer... */
unsigned floating = bld.base.type.floating;
bld.base.type.floating = 0;
mask = lp_build_and(&bld.base, color_mask, mask);
bld.base.type.floating = floating;
} else {
mask = color_mask;
}
}
/* Apply mask, if one exists */
if (mask) {
result = lp_build_select(&bld.base, mask, result, dst);
}
return result;
}
/*
* Do a cached lookup.
*
* Returns (vectors of) 4x8 rgba aos value
*/
LLVMValueRef
lp_build_fetch_cached_texels(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
unsigned n,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j,
LLVMValueRef cache)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned count, low_bit, log2size;
LLVMValueRef color, offset_stored, addr, ptr_addrtrunc, tmp;
LLVMValueRef ij_index, hash_index, hash_mask, block_index;
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef i64t = LLVMInt64TypeInContext(gallivm->context);
struct lp_type type;
struct lp_build_context bld32;
memset(&type, 0, sizeof type);
type.width = 32;
type.length = n;
assert(format_desc->block.width == 4);
assert(format_desc->block.height == 4);
lp_build_context_init(&bld32, gallivm, type);
/*
* compute hash - we use direct mapped cache, the hash function could
* be better but it needs to be simple
* per-element:
* compare offset with offset stored at tag (hash)
* if not equal decode/store block, update tag
* extract color from cache
* assemble result vector
*/
/* TODO: not ideal with 32bit pointers... */
low_bit = util_logbase2(format_desc->block.bits / 8);
log2size = util_logbase2(LP_BUILD_FORMAT_CACHE_SIZE);
addr = LLVMBuildPtrToInt(builder, base_ptr, i64t, "");
ptr_addrtrunc = LLVMBuildPtrToInt(builder, base_ptr, i32t, "");
ptr_addrtrunc = lp_build_broadcast_scalar(&bld32, ptr_addrtrunc);
/* For the hash function, first mask off the unused lowest bits. Then just
do some xor with address bits - only use lower 32bits */
ptr_addrtrunc = LLVMBuildAdd(builder, offset, ptr_addrtrunc, "");
ptr_addrtrunc = LLVMBuildLShr(builder, ptr_addrtrunc,
lp_build_const_int_vec(gallivm, type, low_bit), "");
/* This only really makes sense for size 64,128,256 */
hash_index = ptr_addrtrunc;
ptr_addrtrunc = LLVMBuildLShr(builder, ptr_addrtrunc,
lp_build_const_int_vec(gallivm, type, 2*log2size), "");
hash_index = LLVMBuildXor(builder, ptr_addrtrunc, hash_index, "");
tmp = LLVMBuildLShr(builder, hash_index,
lp_build_const_int_vec(gallivm, type, log2size), "");
hash_index = LLVMBuildXor(builder, hash_index, tmp, "");
hash_mask = lp_build_const_int_vec(gallivm, type, LP_BUILD_FORMAT_CACHE_SIZE - 1);
hash_index = LLVMBuildAnd(builder, hash_index, hash_mask, "");
ij_index = LLVMBuildShl(builder, i, lp_build_const_int_vec(gallivm, type, 2), "");
ij_index = LLVMBuildAdd(builder, ij_index, j, "");
block_index = LLVMBuildShl(builder, hash_index,
lp_build_const_int_vec(gallivm, type, 4), "");
block_index = LLVMBuildAdd(builder, ij_index, block_index, "");
if (n > 1) {
color = LLVMGetUndef(LLVMVectorType(i32t, n));
for (count = 0; count < n; count++) {
LLVMValueRef index, cond, colorx;
LLVMValueRef block_indexx, hash_indexx, addrx, offsetx, ptr_addrx;
struct lp_build_if_state if_ctx;
index = lp_build_const_int32(gallivm, count);
offsetx = LLVMBuildExtractElement(builder, offset, index, "");
addrx = LLVMBuildZExt(builder, offsetx, i64t, "");
addrx = LLVMBuildAdd(builder, addrx, addr, "");
block_indexx = LLVMBuildExtractElement(builder, block_index, index, "");
hash_indexx = LLVMBuildLShr(builder, block_indexx,
lp_build_const_int32(gallivm, 4), "");
offset_stored = lookup_tag_data(gallivm, cache, hash_indexx);
cond = LLVMBuildICmp(builder, LLVMIntNE, offset_stored, addrx, "");
lp_build_if(&if_ctx, gallivm, cond);
{
ptr_addrx = LLVMBuildIntToPtr(builder, addrx,
LLVMPointerType(i8t, 0), "");
update_cached_block(gallivm, format_desc, ptr_addrx, hash_indexx, cache);
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, 1,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
#endif
}
lp_build_endif(&if_ctx);
colorx = lookup_cached_pixel(gallivm, cache, block_indexx);
color = LLVMBuildInsertElement(builder, color, colorx,
lp_build_const_int32(gallivm, count), "");
}
}
else {
LLVMValueRef cond;
struct lp_build_if_state if_ctx;
tmp = LLVMBuildZExt(builder, offset, i64t, "");
addr = LLVMBuildAdd(builder, tmp, addr, "");
offset_stored = lookup_tag_data(gallivm, cache, hash_index);
cond = LLVMBuildICmp(builder, LLVMIntNE, offset_stored, addr, "");
lp_build_if(&if_ctx, gallivm, cond);
{
tmp = LLVMBuildIntToPtr(builder, addr, LLVMPointerType(i8t, 0), "");
update_cached_block(gallivm, format_desc, tmp, hash_index, cache);
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, 1,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
#endif
}
lp_build_endif(&if_ctx);
color = lookup_cached_pixel(gallivm, cache, block_index);
}
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, n,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_TOTAL);
#endif
return LLVMBuildBitCast(builder, color, LLVMVectorType(i8t, n * 4), "");
}
/**
* Sample a single texture image with (bi-)(tri-)linear sampling.
* Return filtered color as two vectors of 16-bit fixed point values.
*/
static void
lp_build_sample_image_linear(struct lp_build_sample_context *bld,
LLVMValueRef int_size,
LLVMValueRef row_stride_vec,
LLVMValueRef img_stride_vec,
LLVMValueRef data_ptr,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
LLVMValueRef *colors_lo,
LLVMValueRef *colors_hi)
{
const unsigned dims = bld->dims;
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_build_context i32, h16, u8n;
LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
LLVMValueRef i32_c8, i32_c128, i32_c255;
LLVMValueRef width_vec, height_vec, depth_vec;
LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
LLVMValueRef t_ipart = NULL, t_fpart = NULL, t_fpart_lo = NULL, t_fpart_hi = NULL;
LLVMValueRef r_ipart = NULL, r_fpart = NULL, r_fpart_lo = NULL, r_fpart_hi = NULL;
LLVMValueRef x_stride, y_stride, z_stride;
LLVMValueRef x_offset0, x_offset1;
LLVMValueRef y_offset0, y_offset1;
LLVMValueRef z_offset0, z_offset1;
LLVMValueRef offset[2][2][2]; /* [z][y][x] */
LLVMValueRef x_subcoord[2], y_subcoord[2], z_subcoord[2];
LLVMValueRef neighbors_lo[2][2][2]; /* [z][y][x] */
LLVMValueRef neighbors_hi[2][2][2]; /* [z][y][x] */
LLVMValueRef packed_lo, packed_hi;
unsigned x, y, z;
unsigned i, j, k;
unsigned numj, numk;
lp_build_context_init(&i32, bld->gallivm, lp_type_int_vec(32));
lp_build_context_init(&h16, bld->gallivm, lp_type_ufixed(16));
lp_build_context_init(&u8n, bld->gallivm, lp_type_unorm(8));
i32_vec_type = lp_build_vec_type(bld->gallivm, i32.type);
h16_vec_type = lp_build_vec_type(bld->gallivm, h16.type);
u8n_vec_type = lp_build_vec_type(bld->gallivm, u8n.type);
lp_build_extract_image_sizes(bld,
bld->int_size_type,
bld->int_coord_type,
int_size,
&width_vec,
&height_vec,
&depth_vec);
if (bld->static_state->normalized_coords) {
LLVMValueRef scaled_size;
LLVMValueRef flt_size;
/* scale size by 256 (8 fractional bits) */
scaled_size = lp_build_shl_imm(&bld->int_size_bld, int_size, 8);
flt_size = lp_build_int_to_float(&bld->float_size_bld, scaled_size);
lp_build_unnormalized_coords(bld, flt_size, &s, &t, &r);
}
else {
/* scale coords by 256 (8 fractional bits) */
s = lp_build_mul_imm(&bld->coord_bld, s, 256);
if (dims >= 2)
t = lp_build_mul_imm(&bld->coord_bld, t, 256);
if (dims >= 3)
r = lp_build_mul_imm(&bld->coord_bld, r, 256);
}
/* convert float to int */
s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
if (dims >= 2)
t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
if (dims >= 3)
r = LLVMBuildFPToSI(builder, r, i32_vec_type, "");
/* subtract 0.5 (add -128) */
i32_c128 = lp_build_const_int_vec(bld->gallivm, i32.type, -128);
s = LLVMBuildAdd(builder, s, i32_c128, "");
if (dims >= 2) {
t = LLVMBuildAdd(builder, t, i32_c128, "");
}
if (dims >= 3) {
r = LLVMBuildAdd(builder, r, i32_c128, "");
}
/* compute floor (shift right 8) */
i32_c8 = lp_build_const_int_vec(bld->gallivm, i32.type, 8);
s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
if (dims >= 2)
t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
if (dims >= 3)
r_ipart = LLVMBuildAShr(builder, r, i32_c8, "");
/* compute fractional part (AND with 0xff) */
i32_c255 = lp_build_const_int_vec(bld->gallivm, i32.type, 255);
s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
if (dims >= 2)
t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
if (dims >= 3)
r_fpart = LLVMBuildAnd(builder, r, i32_c255, "");
/* get pixel, row and image strides */
x_stride = lp_build_const_vec(bld->gallivm, bld->int_coord_bld.type,
bld->format_desc->block.bits/8);
y_stride = row_stride_vec;
z_stride = img_stride_vec;
/* do texcoord wrapping and compute texel offsets */
lp_build_sample_wrap_linear_int(bld,
bld->format_desc->block.width,
s_ipart, width_vec, x_stride,
bld->static_state->pot_width,
bld->static_state->wrap_s,
&x_offset0, &x_offset1,
&x_subcoord[0], &x_subcoord[1]);
for (z = 0; z < 2; z++) {
for (y = 0; y < 2; y++) {
offset[z][y][0] = x_offset0;
offset[z][y][1] = x_offset1;
}
}
if (dims >= 2) {
lp_build_sample_wrap_linear_int(bld,
bld->format_desc->block.height,
t_ipart, height_vec, y_stride,
bld->static_state->pot_height,
bld->static_state->wrap_t,
&y_offset0, &y_offset1,
&y_subcoord[0], &y_subcoord[1]);
for (z = 0; z < 2; z++) {
for (x = 0; x < 2; x++) {
offset[z][0][x] = lp_build_add(&bld->int_coord_bld,
offset[z][0][x], y_offset0);
offset[z][1][x] = lp_build_add(&bld->int_coord_bld,
offset[z][1][x], y_offset1);
}
}
}
if (dims >= 3) {
lp_build_sample_wrap_linear_int(bld,
bld->format_desc->block.height,
r_ipart, depth_vec, z_stride,
bld->static_state->pot_depth,
bld->static_state->wrap_r,
&z_offset0, &z_offset1,
&z_subcoord[0], &z_subcoord[1]);
for (y = 0; y < 2; y++) {
for (x = 0; x < 2; x++) {
offset[0][y][x] = lp_build_add(&bld->int_coord_bld,
offset[0][y][x], z_offset0);
offset[1][y][x] = lp_build_add(&bld->int_coord_bld,
offset[1][y][x], z_offset1);
}
}
}
else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
LLVMValueRef z_offset;
z_offset = lp_build_mul(&bld->int_coord_bld, r, img_stride_vec);
for (y = 0; y < 2; y++) {
for (x = 0; x < 2; x++) {
/* The r coord is the cube face in [0,5] */
offset[0][y][x] = lp_build_add(&bld->int_coord_bld,
offset[0][y][x], z_offset);
}
}
}
/*
* Transform 4 x i32 in
*
* s_fpart = {s0, s1, s2, s3}
*
* into 8 x i16
*
* s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
*
* into two 8 x i16
*
* s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
* s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
*
* and likewise for t_fpart. There is no risk of loosing precision here
* since the fractional parts only use the lower 8bits.
*/
s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
if (dims >= 2)
t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
if (dims >= 3)
r_fpart = LLVMBuildBitCast(builder, r_fpart, h16_vec_type, "");
{
LLVMTypeRef elem_type = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
LLVMValueRef shuffle_lo;
LLVMValueRef shuffle_hi;
for (j = 0; j < h16.type.length; j += 4) {
#ifdef PIPE_ARCH_LITTLE_ENDIAN
unsigned subindex = 0;
#else
unsigned subindex = 1;
#endif
LLVMValueRef index;
index = LLVMConstInt(elem_type, j/2 + subindex, 0);
for (i = 0; i < 4; ++i)
shuffles_lo[j + i] = index;
index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
for (i = 0; i < 4; ++i)
shuffles_hi[j + i] = index;
}
shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef,
shuffle_lo, "");
s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef,
shuffle_hi, "");
if (dims >= 2) {
t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef,
shuffle_lo, "");
t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef,
shuffle_hi, "");
}
if (dims >= 3) {
r_fpart_lo = LLVMBuildShuffleVector(builder, r_fpart, h16.undef,
shuffle_lo, "");
r_fpart_hi = LLVMBuildShuffleVector(builder, r_fpart, h16.undef,
shuffle_hi, "");
}
}
/*
* Fetch the pixels as 4 x 32bit (rgba order might differ):
*
* rgba0 rgba1 rgba2 rgba3
*
* bit cast them into 16 x u8
*
* r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
*
* unpack them into two 8 x i16:
*
* r0 g0 b0 a0 r1 g1 b1 a1
* r2 g2 b2 a2 r3 g3 b3 a3
*
* The higher 8 bits of the resulting elements will be zero.
*/
numj = 1 + (dims >= 2);
numk = 1 + (dims >= 3);
for (k = 0; k < numk; k++) {
for (j = 0; j < numj; j++) {
for (i = 0; i < 2; i++) {
LLVMValueRef rgba8;
if (util_format_is_rgba8_variant(bld->format_desc)) {
/*
* Given the format is a rgba8, just read the pixels as is,
* without any swizzling. Swizzling will be done later.
*/
rgba8 = lp_build_gather(bld->gallivm,
bld->texel_type.length,
bld->format_desc->block.bits,
bld->texel_type.width,
data_ptr, offset[k][j][i]);
rgba8 = LLVMBuildBitCast(builder, rgba8, u8n_vec_type, "");
}
else {
rgba8 = lp_build_fetch_rgba_aos(bld->gallivm,
bld->format_desc,
u8n.type,
data_ptr, offset[k][j][i],
x_subcoord[i],
y_subcoord[j]);
}
/* Expand one 4*rgba8 to two 2*rgba16 */
lp_build_unpack2(bld->gallivm, u8n.type, h16.type,
rgba8,
&neighbors_lo[k][j][i], &neighbors_hi[k][j][i]);
}
}
}
/*
* Linear interpolation with 8.8 fixed point.
*/
if (dims == 1) {
/* 1-D lerp */
packed_lo = lp_build_lerp(&h16,
s_fpart_lo,
neighbors_lo[0][0][0],
neighbors_lo[0][0][1]);
packed_hi = lp_build_lerp(&h16,
s_fpart_hi,
neighbors_hi[0][0][0],
neighbors_hi[0][0][1]);
}
else {
/* 2-D lerp */
packed_lo = lp_build_lerp_2d(&h16,
s_fpart_lo, t_fpart_lo,
neighbors_lo[0][0][0],
neighbors_lo[0][0][1],
neighbors_lo[0][1][0],
neighbors_lo[0][1][1]);
packed_hi = lp_build_lerp_2d(&h16,
s_fpart_hi, t_fpart_hi,
neighbors_hi[0][0][0],
neighbors_hi[0][0][1],
neighbors_hi[0][1][0],
neighbors_hi[0][1][1]);
if (dims >= 3) {
LLVMValueRef packed_lo2, packed_hi2;
/* lerp in the second z slice */
packed_lo2 = lp_build_lerp_2d(&h16,
s_fpart_lo, t_fpart_lo,
neighbors_lo[1][0][0],
neighbors_lo[1][0][1],
neighbors_lo[1][1][0],
neighbors_lo[1][1][1]);
packed_hi2 = lp_build_lerp_2d(&h16,
s_fpart_hi, t_fpart_hi,
neighbors_hi[1][0][0],
neighbors_hi[1][0][1],
neighbors_hi[1][1][0],
neighbors_hi[1][1][1]);
/* interp between two z slices */
packed_lo = lp_build_lerp(&h16, r_fpart_lo,
packed_lo, packed_lo2);
packed_hi = lp_build_lerp(&h16, r_fpart_hi,
packed_hi, packed_hi2);
}
}
*colors_lo = packed_lo;
*colors_hi = packed_hi;
}
/**
* Sample the texture/mipmap using given image filter and mip filter.
* data0_ptr and data1_ptr point to the two mipmap levels to sample
* from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
* If we're using nearest miplevel sampling the '1' values will be null/unused.
*/
static void
lp_build_sample_mipmap(struct lp_build_sample_context *bld,
unsigned img_filter,
unsigned mip_filter,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
LLVMValueRef ilevel0,
LLVMValueRef ilevel1,
LLVMValueRef lod_fpart,
LLVMValueRef colors_lo_var,
LLVMValueRef colors_hi_var)
{
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef size0;
LLVMValueRef size1;
LLVMValueRef row_stride0_vec;
LLVMValueRef row_stride1_vec;
LLVMValueRef img_stride0_vec;
LLVMValueRef img_stride1_vec;
LLVMValueRef data_ptr0;
LLVMValueRef data_ptr1;
LLVMValueRef colors0_lo, colors0_hi;
LLVMValueRef colors1_lo, colors1_hi;
/* sample the first mipmap level */
lp_build_mipmap_level_sizes(bld, ilevel0,
&size0,
&row_stride0_vec, &img_stride0_vec);
data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
lp_build_sample_image_nearest(bld,
size0,
row_stride0_vec, img_stride0_vec,
data_ptr0, s, t, r,
&colors0_lo, &colors0_hi);
}
else {
assert(img_filter == PIPE_TEX_FILTER_LINEAR);
lp_build_sample_image_linear(bld,
size0,
row_stride0_vec, img_stride0_vec,
data_ptr0, s, t, r,
&colors0_lo, &colors0_hi);
}
/* Store the first level's colors in the output variables */
LLVMBuildStore(builder, colors0_lo, colors_lo_var);
LLVMBuildStore(builder, colors0_hi, colors_hi_var);
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
LLVMValueRef h16_scale = lp_build_const_float(bld->gallivm, 256.0);
LLVMTypeRef i32_type = LLVMIntTypeInContext(bld->gallivm->context, 32);
struct lp_build_if_state if_ctx;
LLVMValueRef need_lerp;
lod_fpart = LLVMBuildFMul(builder, lod_fpart, h16_scale, "");
lod_fpart = LLVMBuildFPToSI(builder, lod_fpart, i32_type, "lod_fpart.fixed16");
/* need_lerp = lod_fpart > 0 */
need_lerp = LLVMBuildICmp(builder, LLVMIntSGT,
lod_fpart, LLVMConstNull(i32_type),
"need_lerp");
lp_build_if(&if_ctx, bld->gallivm, need_lerp);
{
struct lp_build_context h16_bld;
lp_build_context_init(&h16_bld, bld->gallivm, lp_type_ufixed(16));
/* sample the second mipmap level */
lp_build_mipmap_level_sizes(bld, ilevel1,
&size1,
&row_stride1_vec, &img_stride1_vec);
data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
lp_build_sample_image_nearest(bld,
size1,
row_stride1_vec, img_stride1_vec,
data_ptr1, s, t, r,
&colors1_lo, &colors1_hi);
}
else {
lp_build_sample_image_linear(bld,
size1,
row_stride1_vec, img_stride1_vec,
data_ptr1, s, t, r,
&colors1_lo, &colors1_hi);
}
/* interpolate samples from the two mipmap levels */
lod_fpart = LLVMBuildTrunc(builder, lod_fpart, h16_bld.elem_type, "");
lod_fpart = lp_build_broadcast_scalar(&h16_bld, lod_fpart);
#if HAVE_LLVM == 0x208
/* This is a work-around for a bug in LLVM 2.8.
* Evidently, something goes wrong in the construction of the
* lod_fpart short[8] vector. Adding this no-effect shuffle seems
* to force the vector to be properly constructed.
* Tested with mesa-demos/src/tests/mipmap_limits.c (press t, f).
*/
{
LLVMValueRef shuffles[8], shuffle;
int i;
assert(h16_bld.type.length <= Elements(shuffles));
for (i = 0; i < h16_bld.type.length; i++)
shuffles[i] = lp_build_const_int32(bld->gallivm, 2 * (i & 1));
shuffle = LLVMConstVector(shuffles, h16_bld.type.length);
lod_fpart = LLVMBuildShuffleVector(builder,
lod_fpart, lod_fpart,
shuffle, "");
}
#endif
colors0_lo = lp_build_lerp(&h16_bld, lod_fpart,
colors0_lo, colors1_lo);
colors0_hi = lp_build_lerp(&h16_bld, lod_fpart,
colors0_hi, colors1_hi);
LLVMBuildStore(builder, colors0_lo, colors_lo_var);
LLVMBuildStore(builder, colors0_hi, colors_hi_var);
}
lp_build_endif(&if_ctx);
}
}
/**
* Fetch a pixel into a 4 float AoS.
*
* \param format_desc describes format of the image we're fetching from
* \param ptr address of the pixel block (or the texel if uncompressed)
* \param i, j the sub-block pixel coordinates. For non-compressed formats
* these will always be (0, 0).
* \return a 4 element vector with the pixel's RGBA values.
*/
LLVMValueRef
lp_build_fetch_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned num_pixels = type.length / 4;
struct lp_build_context bld;
assert(type.length <= LP_MAX_VECTOR_LENGTH);
assert(type.length % 4 == 0);
lp_build_context_init(&bld, gallivm, type);
/*
* Trivial case
*
* The format matches the type (apart of a swizzle) so no need for
* scaling or converting.
*/
if (format_matches_type(format_desc, type) &&
format_desc->block.bits <= type.width * 4 &&
util_is_power_of_two(format_desc->block.bits)) {
LLVMValueRef packed;
/*
* The format matches the type (apart of a swizzle) so no need for
* scaling or converting.
*/
packed = lp_build_gather(gallivm, type.length/4,
format_desc->block.bits, type.width*4,
base_ptr, offset);
assert(format_desc->block.bits <= type.width * type.length);
packed = LLVMBuildBitCast(gallivm->builder, packed,
lp_build_vec_type(gallivm, type), "");
return lp_build_format_swizzle_aos(format_desc, &bld, packed);
}
/*
* Bit arithmetic
*/
if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
util_is_power_of_two(format_desc->block.bits) &&
format_desc->block.bits <= 32 &&
format_desc->is_bitmask &&
!format_desc->is_mixed &&
(format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED ||
format_desc->channel[1].type == UTIL_FORMAT_TYPE_UNSIGNED)) {
LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
LLVMValueRef res;
unsigned k;
/*
* Unpack a pixel at a time into a <4 x float> RGBA vector
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef packed;
packed = lp_build_gather_elem(gallivm, num_pixels,
format_desc->block.bits, 32,
base_ptr, offset, k);
tmps[k] = lp_build_unpack_arith_rgba_aos(gallivm,
format_desc,
packed);
}
/*
* Type conversion.
*
* TODO: We could avoid floating conversion for integer to
* integer conversions.
*/
if (gallivm_debug & GALLIVM_DEBUG_PERF && !type.floating) {
debug_printf("%s: unpacking %s with floating point\n",
__FUNCTION__, format_desc->short_name);
}
lp_build_conv(gallivm,
lp_float32_vec4_type(),
type,
tmps, num_pixels, &res, 1);
return lp_build_format_swizzle_aos(format_desc, &bld, res);
}
/*
* YUV / subsampled formats
*/
if (format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
struct lp_type tmp_type;
LLVMValueRef tmp;
memset(&tmp_type, 0, sizeof tmp_type);
tmp_type.width = 8;
tmp_type.length = num_pixels * 4;
tmp_type.norm = TRUE;
tmp = lp_build_fetch_subsampled_rgba_aos(gallivm,
format_desc,
num_pixels,
base_ptr,
offset,
i, j);
lp_build_conv(gallivm,
tmp_type, type,
&tmp, 1, &tmp, 1);
return tmp;
}
/*
* Fallback to util_format_description::fetch_rgba_8unorm().
*/
if (format_desc->fetch_rgba_8unorm &&
!type.floating && type.width == 8 && !type.sign && type.norm) {
/*
* Fallback to calling util_format_description::fetch_rgba_8unorm.
*
* This is definitely not the most efficient way of fetching pixels, as
* we miss the opportunity to do vectorization, but this it is a
* convenient for formats or scenarios for which there was no opportunity
* or incentive to optimize.
*/
LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
char name[256];
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef function;
LLVMValueRef tmp_ptr;
LLVMValueRef tmp;
LLVMValueRef res;
LLVMValueRef callee;
unsigned k;
util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_8unorm",
format_desc->short_name);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: falling back to %s\n", __FUNCTION__, name);
}
/*
* Declare and bind format_desc->fetch_rgba_8unorm().
*/
function = LLVMGetNamedFunction(module, name);
if (!function) {
/*
* Function to call looks like:
* fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
LLVMTypeRef function_type;
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pi8t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
Elements(arg_types), 0);
function = LLVMAddFunction(module, name, function_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
LLVMSetLinkage(function, LLVMExternalLinkage);
assert(LLVMIsDeclaration(function));
}
/* make const pointer for the C fetch_rgba_float function */
callee = lp_build_const_int_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, callee,
LLVMTypeOf(function), "cast callee");
tmp_ptr = lp_build_alloca(gallivm, i32t, "");
res = LLVMGetUndef(LLVMVectorType(i32t, num_pixels));
/*
* Invoke format_desc->fetch_rgba_8unorm() for each pixel and insert the result
* in the SoA vectors.
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
LLVMValueRef args[4];
args[0] = LLVMBuildBitCast(builder, tmp_ptr, pi8t, "");
args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
base_ptr, offset, k);
if (num_pixels == 1) {
args[2] = i;
args[3] = j;
}
else {
args[2] = LLVMBuildExtractElement(builder, i, index, "");
args[3] = LLVMBuildExtractElement(builder, j, index, "");
}
LLVMBuildCall(builder, function, args, Elements(args), "");
tmp = LLVMBuildLoad(builder, tmp_ptr, "");
if (num_pixels == 1) {
res = tmp;
}
else {
res = LLVMBuildInsertElement(builder, res, tmp, index, "");
}
}
/* Bitcast from <n x i32> to <4n x i8> */
res = LLVMBuildBitCast(builder, res, bld.vec_type, "");
return res;
}
/*
* Fallback to util_format_description::fetch_rgba_float().
*/
if (format_desc->fetch_rgba_float) {
/*
* Fallback to calling util_format_description::fetch_rgba_float.
*
* This is definitely not the most efficient way of fetching pixels, as
* we miss the opportunity to do vectorization, but this it is a
* convenient for formats or scenarios for which there was no opportunity
* or incentive to optimize.
*/
LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)));
char name[256];
LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
LLVMTypeRef f32x4t = LLVMVectorType(f32t, 4);
LLVMTypeRef pf32t = LLVMPointerType(f32t, 0);
LLVMTypeRef pi8t = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef function;
LLVMValueRef tmp_ptr;
LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
LLVMValueRef res;
LLVMValueRef callee;
unsigned k;
util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_float",
format_desc->short_name);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: falling back to %s\n", __FUNCTION__, name);
}
/*
* Declare and bind format_desc->fetch_rgba_float().
*/
function = LLVMGetNamedFunction(module, name);
if (!function) {
/*
* Function to call looks like:
* fetch(float *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
LLVMTypeRef function_type;
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pf32t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
Elements(arg_types), 0);
function = LLVMAddFunction(module, name, function_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
LLVMSetLinkage(function, LLVMExternalLinkage);
assert(LLVMIsDeclaration(function));
}
/* Note: we're using this casting here instead of LLVMAddGlobalMapping()
* to work around a bug in LLVM 2.6.
*/
/* make const pointer for the C fetch_rgba_float function */
callee = lp_build_const_int_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_float));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, callee,
LLVMTypeOf(function), "cast callee");
tmp_ptr = lp_build_alloca(gallivm, f32x4t, "");
/*
* Invoke format_desc->fetch_rgba_float() for each pixel and insert the result
* in the SoA vectors.
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef args[4];
args[0] = LLVMBuildBitCast(builder, tmp_ptr, pf32t, "");
args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
base_ptr, offset, k);
if (num_pixels == 1) {
args[2] = i;
args[3] = j;
}
else {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
args[2] = LLVMBuildExtractElement(builder, i, index, "");
args[3] = LLVMBuildExtractElement(builder, j, index, "");
}
LLVMBuildCall(builder, function, args, Elements(args), "");
tmps[k] = LLVMBuildLoad(builder, tmp_ptr, "");
}
lp_build_conv(gallivm,
lp_float32_vec4_type(),
type,
tmps, num_pixels, &res, 1);
return res;
}
assert(0);
return lp_build_undef(gallivm, type);
}
/**
* Texture sampling in AoS format. Used when sampling common 32-bit/texel
* formats. 1D/2D/3D/cube texture supported. All mipmap sampling modes
* but only limited texture coord wrap modes.
*/
void
lp_build_sample_aos(struct lp_build_sample_context *bld,
unsigned unit,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
const LLVMValueRef *ddx,
const LLVMValueRef *ddy,
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
LLVMValueRef texel_out[4])
{
struct lp_build_context *int_bld = &bld->int_bld;
LLVMBuilderRef builder = bld->gallivm->builder;
const unsigned mip_filter = bld->static_state->min_mip_filter;
const unsigned min_filter = bld->static_state->min_img_filter;
const unsigned mag_filter = bld->static_state->mag_img_filter;
const unsigned dims = bld->dims;
LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
LLVMValueRef ilevel0, ilevel1 = NULL;
LLVMValueRef packed, packed_lo, packed_hi;
LLVMValueRef unswizzled[4];
LLVMValueRef face_ddx[4], face_ddy[4];
struct lp_build_context h16_bld;
LLVMValueRef first_level;
LLVMValueRef i32t_zero = lp_build_const_int32(bld->gallivm, 0);
/* we only support the common/simple wrap modes at this time */
assert(lp_is_simple_wrap_mode(bld->static_state->wrap_s));
if (dims >= 2)
assert(lp_is_simple_wrap_mode(bld->static_state->wrap_t));
if (dims >= 3)
assert(lp_is_simple_wrap_mode(bld->static_state->wrap_r));
/* make 16-bit fixed-pt builder context */
lp_build_context_init(&h16_bld, bld->gallivm, lp_type_ufixed(16));
/* cube face selection, compute pre-face coords, etc. */
if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
LLVMValueRef face, face_s, face_t;
lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
s = face_s; /* vec */
t = face_t; /* vec */
/* use 'r' to indicate cube face */
r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
/* recompute ddx, ddy using the new (s,t) face texcoords */
face_ddx[0] = lp_build_scalar_ddx(&bld->coord_bld, s);
face_ddx[1] = lp_build_scalar_ddx(&bld->coord_bld, t);
face_ddx[2] = NULL;
face_ddx[3] = NULL;
face_ddy[0] = lp_build_scalar_ddy(&bld->coord_bld, s);
face_ddy[1] = lp_build_scalar_ddy(&bld->coord_bld, t);
face_ddy[2] = NULL;
face_ddy[3] = NULL;
ddx = face_ddx;
ddy = face_ddy;
}
/*
* Compute the level of detail (float).
*/
if (min_filter != mag_filter ||
mip_filter != PIPE_TEX_MIPFILTER_NONE) {
/* Need to compute lod either to choose mipmap levels or to
* distinguish between minification/magnification with one mipmap level.
*/
lp_build_lod_selector(bld, unit, ddx, ddy,
lod_bias, explicit_lod,
mip_filter,
&lod_ipart, &lod_fpart);
} else {
lod_ipart = i32t_zero;
}
/*
* Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
*/
switch (mip_filter) {
default:
assert(0 && "bad mip_filter value in lp_build_sample_aos()");
/* fall-through */
case PIPE_TEX_MIPFILTER_NONE:
/* always use mip level 0 */
if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
/* XXX this is a work-around for an apparent bug in LLVM 2.7.
* We should be able to set ilevel0 = const(0) but that causes
* bad x86 code to be emitted.
*/
assert(lod_ipart);
lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
}
else {
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
bld->gallivm, unit);
ilevel0 = first_level;
}
break;
case PIPE_TEX_MIPFILTER_NEAREST:
assert(lod_ipart);
lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0);
break;
case PIPE_TEX_MIPFILTER_LINEAR:
assert(lod_ipart);
assert(lod_fpart);
lp_build_linear_mip_levels(bld, unit,
lod_ipart, &lod_fpart,
&ilevel0, &ilevel1);
break;
}
/*
* Get/interpolate texture colors.
*/
packed_lo = lp_build_alloca(bld->gallivm, h16_bld.vec_type, "packed_lo");
packed_hi = lp_build_alloca(bld->gallivm, h16_bld.vec_type, "packed_hi");
if (min_filter == mag_filter) {
/* no need to distinquish between minification and magnification */
lp_build_sample_mipmap(bld,
min_filter, mip_filter,
s, t, r,
ilevel0, ilevel1, lod_fpart,
packed_lo, packed_hi);
}
else {
/* Emit conditional to choose min image filter or mag image filter
* depending on the lod being > 0 or <= 0, respectively.
*/
struct lp_build_if_state if_ctx;
LLVMValueRef minify;
/* minify = lod >= 0.0 */
minify = LLVMBuildICmp(builder, LLVMIntSGE,
lod_ipart, int_bld->zero, "");
lp_build_if(&if_ctx, bld->gallivm, minify);
{
/* Use the minification filter */
lp_build_sample_mipmap(bld,
min_filter, mip_filter,
s, t, r,
ilevel0, ilevel1, lod_fpart,
packed_lo, packed_hi);
}
lp_build_else(&if_ctx);
{
/* Use the magnification filter */
lp_build_sample_mipmap(bld,
mag_filter, PIPE_TEX_MIPFILTER_NONE,
s, t, r,
ilevel0, NULL, NULL,
packed_lo, packed_hi);
}
lp_build_endif(&if_ctx);
}
/*
* combine the values stored in 'packed_lo' and 'packed_hi' variables
* into 'packed'
*/
packed = lp_build_pack2(bld->gallivm,
h16_bld.type, lp_type_unorm(8),
LLVMBuildLoad(builder, packed_lo, ""),
LLVMBuildLoad(builder, packed_hi, ""));
/*
* Convert to SoA and swizzle.
*/
lp_build_rgba8_to_f32_soa(bld->gallivm,
bld->texel_type,
packed, unswizzled);
if (util_format_is_rgba8_variant(bld->format_desc)) {
lp_build_format_swizzle_soa(bld->format_desc,
&bld->texel_bld,
unswizzled, texel_out);
}
else {
texel_out[0] = unswizzled[0];
texel_out[1] = unswizzled[1];
texel_out[2] = unswizzled[2];
texel_out[3] = unswizzled[3];
}
}
/**
* Performs blending of src and dst pixels
*
* @param blend the blend state of the shader variant
* @param cbuf_format format of the colour buffer
* @param type data type of the pixel vector
* @param rt rt number
* @param src blend src
* @param dst blend dst
* @param mask optional mask to apply to the blending result
* @param const_ const blend color
* @param swizzle swizzle values for RGBA
*
* @return the result of blending src and dst
*/
LLVMValueRef
lp_build_blend_aos(struct gallivm_state *gallivm,
const struct pipe_blend_state *blend,
const enum pipe_format *cbuf_format,
struct lp_type type,
unsigned rt,
LLVMValueRef src,
LLVMValueRef dst,
LLVMValueRef mask,
LLVMValueRef const_,
const unsigned char swizzle[4])
{
const struct pipe_rt_blend_state * state = &blend->rt[rt];
struct lp_build_blend_aos_context bld;
LLVMValueRef src_factor, dst_factor;
LLVMValueRef result;
unsigned alpha_swizzle = swizzle[3];
boolean fullcolormask;
/* Setup build context */
memset(&bld, 0, sizeof bld);
lp_build_context_init(&bld.base, gallivm, type);
bld.src = src;
bld.dst = dst;
bld.const_ = const_;
if (swizzle[3] > UTIL_FORMAT_SWIZZLE_W || swizzle[3] == swizzle[0])
alpha_swizzle = UTIL_FORMAT_SWIZZLE_NONE;
if (!state->blend_enable) {
result = src;
} else {
boolean rgb_alpha_same = state->rgb_src_factor == state->rgb_dst_factor && state->alpha_src_factor == state->alpha_dst_factor;
assert(rgb_alpha_same || alpha_swizzle != UTIL_FORMAT_SWIZZLE_NONE);
src_factor = lp_build_blend_factor(&bld, state->rgb_src_factor,
state->alpha_src_factor, alpha_swizzle);
dst_factor = lp_build_blend_factor(&bld, state->rgb_dst_factor,
state->alpha_dst_factor, alpha_swizzle);
result = lp_build_blend(&bld.base,
state->rgb_func,
state->rgb_src_factor,
state->rgb_dst_factor,
src,
dst,
src_factor,
dst_factor,
rgb_alpha_same,
false);
if(state->rgb_func != state->alpha_func && alpha_swizzle != UTIL_FORMAT_SWIZZLE_NONE) {
LLVMValueRef alpha;
alpha = lp_build_blend(&bld.base,
state->alpha_func,
state->alpha_src_factor,
state->alpha_dst_factor,
src,
dst,
src_factor,
dst_factor,
rgb_alpha_same,
false);
result = lp_build_blend_swizzle(&bld,
result,
alpha,
LP_BUILD_BLEND_SWIZZLE_RGBA,
alpha_swizzle);
}
}
/* Check if color mask is necessary */
fullcolormask = util_format_colormask_full(util_format_description(cbuf_format[rt]), state->colormask);
if (!fullcolormask) {
LLVMValueRef color_mask;
color_mask = lp_build_const_mask_aos_swizzled(gallivm, bld.base.type, state->colormask, swizzle);
lp_build_name(color_mask, "color_mask");
/* Combine with input mask if necessary */
if (mask) {
mask = lp_build_and(&bld.base, color_mask, mask);
} else {
mask = color_mask;
}
}
/* Apply mask, if one exists */
if (mask) {
result = lp_build_select(&bld.base, mask, result, dst);
}
return result;
}
/**
* Generic type conversion.
*
* TODO: Take a precision argument, or even better, add a new precision member
* to the lp_type union.
*/
void
lp_build_conv(LLVMBuilderRef builder,
struct lp_type src_type,
struct lp_type dst_type,
const LLVMValueRef *src, unsigned num_srcs,
LLVMValueRef *dst, unsigned num_dsts)
{
struct lp_type tmp_type;
LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
unsigned num_tmps;
unsigned i;
/* We must not loose or gain channels. Only precision */
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
assert(num_srcs <= LP_MAX_VECTOR_LENGTH);
assert(num_dsts <= LP_MAX_VECTOR_LENGTH);
tmp_type = src_type;
for(i = 0; i < num_srcs; ++i) {
assert(lp_check_value(src_type, src[i]));
tmp[i] = src[i];
}
num_tmps = num_srcs;
/*
* Clamp if necessary
*/
if(memcmp(&src_type, &dst_type, sizeof src_type) != 0) {
struct lp_build_context bld;
double src_min = lp_const_min(src_type);
double dst_min = lp_const_min(dst_type);
double src_max = lp_const_max(src_type);
double dst_max = lp_const_max(dst_type);
LLVMValueRef thres;
lp_build_context_init(&bld, builder, tmp_type);
if(src_min < dst_min) {
if(dst_min == 0.0)
thres = bld.zero;
else
thres = lp_build_const_vec(src_type, dst_min);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_max(&bld, tmp[i], thres);
}
if(src_max > dst_max) {
if(dst_max == 1.0)
thres = bld.one;
else
thres = lp_build_const_vec(src_type, dst_max);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_min(&bld, tmp[i], thres);
}
}
/*
* Scale to the narrowest range
*/
if(dst_type.floating) {
/* Nothing to do */
}
else if(tmp_type.floating) {
if(!dst_type.fixed && !dst_type.sign && dst_type.norm) {
for(i = 0; i < num_tmps; ++i) {
tmp[i] = lp_build_clamped_float_to_unsigned_norm(builder,
tmp_type,
dst_type.width,
tmp[i]);
}
tmp_type.floating = FALSE;
}
else {
double dst_scale = lp_const_scale(dst_type);
LLVMTypeRef tmp_vec_type;
if (dst_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(tmp_type, dst_scale);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = FALSE;
tmp_vec_type = lp_build_vec_type(tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
else
tmp[i] = LLVMBuildFPToUI(builder, tmp[i], tmp_vec_type, "");
#else
/* FIXME: there is no SSE counterpart for LLVMBuildFPToUI */
tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
#endif
}
}
}
else {
unsigned src_shift = lp_const_shift(src_type);
unsigned dst_shift = lp_const_shift(dst_type);
/* FIXME: compensate different offsets too */
if(src_shift > dst_shift) {
LLVMValueRef shift = lp_build_const_int_vec(tmp_type, src_shift - dst_shift);
for(i = 0; i < num_tmps; ++i)
if(src_type.sign)
tmp[i] = LLVMBuildAShr(builder, tmp[i], shift, "");
else
tmp[i] = LLVMBuildLShr(builder, tmp[i], shift, "");
}
}
/*
* Truncate or expand bit width
*
* No data conversion should happen here, although the sign bits are
* crucial to avoid bad clamping.
*/
{
struct lp_type new_type;
new_type = tmp_type;
new_type.sign = dst_type.sign;
new_type.width = dst_type.width;
new_type.length = dst_type.length;
lp_build_resize(builder, tmp_type, new_type, tmp, num_srcs, tmp, num_dsts);
tmp_type = new_type;
num_tmps = num_dsts;
}
/*
* Scale to the widest range
*/
if(src_type.floating) {
/* Nothing to do */
}
else if(!src_type.floating && dst_type.floating) {
if(!src_type.fixed && !src_type.sign && src_type.norm) {
for(i = 0; i < num_tmps; ++i) {
tmp[i] = lp_build_unsigned_norm_to_float(builder,
src_type.width,
dst_type,
tmp[i]);
}
tmp_type.floating = TRUE;
}
else {
double src_scale = lp_const_scale(src_type);
LLVMTypeRef tmp_vec_type;
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = TRUE;
tmp_type.sign = TRUE;
tmp_vec_type = lp_build_vec_type(tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
else
tmp[i] = LLVMBuildUIToFP(builder, tmp[i], tmp_vec_type, "");
#else
/* FIXME: there is no SSE counterpart for LLVMBuildUIToFP */
tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
#endif
}
if (src_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(tmp_type, 1.0/src_scale);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
}
}
else {
unsigned src_shift = lp_const_shift(src_type);
unsigned dst_shift = lp_const_shift(dst_type);
/* FIXME: compensate different offsets too */
if(src_shift < dst_shift) {
LLVMValueRef shift = lp_build_const_int_vec(tmp_type, dst_shift - src_shift);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildShl(builder, tmp[i], shift, "");
}
}
for(i = 0; i < num_dsts; ++i) {
dst[i] = tmp[i];
assert(lp_check_value(dst_type, dst[i]));
}
}
/**
* Generic type conversion.
*
* TODO: Take a precision argument, or even better, add a new precision member
* to the lp_type union.
*/
void
lp_build_conv(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
const LLVMValueRef *src, unsigned num_srcs,
LLVMValueRef *dst, unsigned num_dsts)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_type tmp_type;
LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
unsigned num_tmps;
unsigned i;
/* We must not loose or gain channels. Only precision */
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
assert(num_srcs <= LP_MAX_VECTOR_LENGTH);
assert(num_dsts <= LP_MAX_VECTOR_LENGTH);
tmp_type = src_type;
for(i = 0; i < num_srcs; ++i) {
assert(lp_check_value(src_type, src[i]));
tmp[i] = src[i];
}
num_tmps = num_srcs;
/* Special case 4x4f --> 1x16ub
*/
if (src_type.floating == 1 &&
src_type.fixed == 0 &&
src_type.sign == 1 &&
src_type.norm == 0 &&
src_type.width == 32 &&
src_type.length == 4 &&
dst_type.floating == 0 &&
dst_type.fixed == 0 &&
dst_type.sign == 0 &&
dst_type.norm == 1 &&
dst_type.width == 8 &&
dst_type.length == 16 &&
4 * num_dsts == num_srcs &&
util_cpu_caps.has_sse2)
{
struct lp_build_context bld;
struct lp_type int16_type = dst_type;
struct lp_type int32_type = dst_type;
LLVMValueRef const_255f;
unsigned i, j;
lp_build_context_init(&bld, gallivm, src_type);
int16_type.width *= 2;
int16_type.length /= 2;
int16_type.sign = 1;
int32_type.width *= 4;
int32_type.length /= 4;
int32_type.sign = 1;
const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);
for (i = 0; i < num_dsts; ++i, src += 4) {
LLVMValueRef lo, hi;
for (j = 0; j < 4; ++j) {
tmp[j] = LLVMBuildFMul(builder, src[j], const_255f, "");
tmp[j] = lp_build_iround(&bld, tmp[j]);
}
/* relying on clamping behavior of sse2 intrinsics here */
lo = lp_build_pack2(gallivm, int32_type, int16_type, tmp[0], tmp[1]);
hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
}
return;
}
/* Special case 2x8f --> 1x16ub
*/
else if (src_type.floating == 1 &&
src_type.fixed == 0 &&
src_type.sign == 1 &&
src_type.norm == 0 &&
src_type.width == 32 &&
src_type.length == 8 &&
dst_type.floating == 0 &&
dst_type.fixed == 0 &&
dst_type.sign == 0 &&
dst_type.norm == 1 &&
dst_type.width == 8 &&
dst_type.length == 16 &&
2 * num_dsts == num_srcs &&
util_cpu_caps.has_avx) {
struct lp_build_context bld;
struct lp_type int16_type = dst_type;
struct lp_type int32_type = dst_type;
LLVMValueRef const_255f;
unsigned i;
lp_build_context_init(&bld, gallivm, src_type);
int16_type.width *= 2;
int16_type.length /= 2;
int16_type.sign = 1;
int32_type.width *= 4;
int32_type.length /= 4;
int32_type.sign = 1;
const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);
for (i = 0; i < num_dsts; ++i, src += 2) {
LLVMValueRef lo, hi, a, b;
a = LLVMBuildFMul(builder, src[0], const_255f, "");
b = LLVMBuildFMul(builder, src[1], const_255f, "");
a = lp_build_iround(&bld, a);
b = lp_build_iround(&bld, b);
tmp[0] = lp_build_extract_range(gallivm, a, 0, 4);
tmp[1] = lp_build_extract_range(gallivm, a, 4, 4);
tmp[2] = lp_build_extract_range(gallivm, b, 0, 4);
tmp[3] = lp_build_extract_range(gallivm, b, 4, 4);
/* relying on clamping behavior of sse2 intrinsics here */
lo = lp_build_pack2(gallivm, int32_type, int16_type, tmp[0], tmp[1]);
hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
}
return;
}
/* Pre convert half-floats to floats
*/
else if (src_type.floating && src_type.width == 16)
{
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_half_to_float(gallivm, src_type, tmp[i]);
tmp_type.width = 32;
}
/*
* Clamp if necessary
*/
if(memcmp(&src_type, &dst_type, sizeof src_type) != 0) {
struct lp_build_context bld;
double src_min = lp_const_min(src_type);
double dst_min = lp_const_min(dst_type);
double src_max = lp_const_max(src_type);
double dst_max = lp_const_max(dst_type);
LLVMValueRef thres;
lp_build_context_init(&bld, gallivm, tmp_type);
if(src_min < dst_min) {
if(dst_min == 0.0)
thres = bld.zero;
else
thres = lp_build_const_vec(gallivm, src_type, dst_min);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_max(&bld, tmp[i], thres);
}
if(src_max > dst_max) {
if(dst_max == 1.0)
thres = bld.one;
else
thres = lp_build_const_vec(gallivm, src_type, dst_max);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_min(&bld, tmp[i], thres);
}
}
/*
* Scale to the narrowest range
*/
if(dst_type.floating) {
/* Nothing to do */
}
else if(tmp_type.floating) {
if(!dst_type.fixed && !dst_type.sign && dst_type.norm) {
for(i = 0; i < num_tmps; ++i) {
tmp[i] = lp_build_clamped_float_to_unsigned_norm(gallivm,
tmp_type,
dst_type.width,
tmp[i]);
}
tmp_type.floating = FALSE;
}
else {
double dst_scale = lp_const_scale(dst_type);
LLVMTypeRef tmp_vec_type;
if (dst_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, dst_scale);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = FALSE;
tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
else
tmp[i] = LLVMBuildFPToUI(builder, tmp[i], tmp_vec_type, "");
#else
/* FIXME: there is no SSE counterpart for LLVMBuildFPToUI */
tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
#endif
}
}
}
else {
unsigned src_shift = lp_const_shift(src_type);
unsigned dst_shift = lp_const_shift(dst_type);
unsigned src_offset = lp_const_offset(src_type);
unsigned dst_offset = lp_const_offset(dst_type);
/* Compensate for different offsets */
if (dst_offset > src_offset && src_type.width > dst_type.width) {
for (i = 0; i < num_tmps; ++i) {
LLVMValueRef shifted;
LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, src_shift - 1);
if(src_type.sign)
shifted = LLVMBuildAShr(builder, tmp[i], shift, "");
else
shifted = LLVMBuildLShr(builder, tmp[i], shift, "");
tmp[i] = LLVMBuildSub(builder, tmp[i], shifted, "");
}
}
if(src_shift > dst_shift) {
LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type,
src_shift - dst_shift);
for(i = 0; i < num_tmps; ++i)
if(src_type.sign)
tmp[i] = LLVMBuildAShr(builder, tmp[i], shift, "");
else
tmp[i] = LLVMBuildLShr(builder, tmp[i], shift, "");
}
}
/*
* Truncate or expand bit width
*
* No data conversion should happen here, although the sign bits are
* crucial to avoid bad clamping.
*/
{
struct lp_type new_type;
new_type = tmp_type;
new_type.sign = dst_type.sign;
new_type.width = dst_type.width;
new_type.length = dst_type.length;
lp_build_resize(gallivm, tmp_type, new_type, tmp, num_srcs, tmp, num_dsts);
tmp_type = new_type;
num_tmps = num_dsts;
}
/*
* Scale to the widest range
*/
if(src_type.floating) {
/* Nothing to do */
}
else if(!src_type.floating && dst_type.floating) {
if(!src_type.fixed && !src_type.sign && src_type.norm) {
for(i = 0; i < num_tmps; ++i) {
tmp[i] = lp_build_unsigned_norm_to_float(gallivm,
src_type.width,
dst_type,
tmp[i]);
}
tmp_type.floating = TRUE;
}
else {
double src_scale = lp_const_scale(src_type);
LLVMTypeRef tmp_vec_type;
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = TRUE;
tmp_type.sign = TRUE;
tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
else
tmp[i] = LLVMBuildUIToFP(builder, tmp[i], tmp_vec_type, "");
#else
/* FIXME: there is no SSE counterpart for LLVMBuildUIToFP */
tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
#endif
}
if (src_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, 1.0/src_scale);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
}
}
else {
unsigned src_shift = lp_const_shift(src_type);
unsigned dst_shift = lp_const_shift(dst_type);
unsigned src_offset = lp_const_offset(src_type);
unsigned dst_offset = lp_const_offset(dst_type);
if (src_shift < dst_shift) {
LLVMValueRef pre_shift[LP_MAX_VECTOR_LENGTH];
LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, dst_shift - src_shift);
for (i = 0; i < num_tmps; ++i) {
pre_shift[i] = tmp[i];
tmp[i] = LLVMBuildShl(builder, tmp[i], shift, "");
}
/* Compensate for different offsets */
if (dst_offset > src_offset) {
for (i = 0; i < num_tmps; ++i) {
tmp[i] = LLVMBuildSub(builder, tmp[i], pre_shift[i], "");
}
}
}
}
for(i = 0; i < num_dsts; ++i) {
dst[i] = tmp[i];
assert(lp_check_value(dst_type, dst[i]));
}
}
/**
* Generate code for performing depth and/or stencil tests.
* We operate on a vector of values (typically a 2x2 quad).
*
* \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 (a 2x2 quad)
* \param zs_dst_ptr pointer to depth/stencil values in framebuffer
* \param facing contains float value indicating front/back facing polygon
*/
void
lp_build_depth_stencil_test(LLVMBuilderRef builder,
const struct pipe_depth_state *depth,
const struct pipe_stencil_state stencil[2],
struct lp_type type,
const struct util_format_description *format_desc,
struct lp_build_mask_context *mask,
LLVMValueRef stencil_refs[2],
LLVMValueRef z_src,
LLVMValueRef zs_dst_ptr,
LLVMValueRef face,
LLVMValueRef counter)
{
struct lp_build_context bld;
struct lp_build_context sbld;
struct lp_type s_type;
LLVMValueRef zs_dst, 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 = mask->value;
/* 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(format_desc->format == PIPE_FORMAT_Z24_UNORM_S8_USCALED ||
format_desc->format == PIPE_FORMAT_S8_USCALED_Z24_UNORM);
}
assert(z_swizzle < 4);
assert(format_desc->block.bits == type.width);
if (type.floating) {
assert(z_swizzle == 0);
assert(format_desc->channel[z_swizzle].type ==
UTIL_FORMAT_TYPE_FLOAT);
assert(format_desc->channel[z_swizzle].size ==
format_desc->block.bits);
}
else {
assert(format_desc->channel[z_swizzle].type ==
UTIL_FORMAT_TYPE_UNSIGNED);
assert(format_desc->channel[z_swizzle].normalized);
assert(!type.fixed);
assert(!type.sign);
assert(type.norm);
}
}
/* Setup build context for Z vals */
lp_build_context_init(&bld, builder, type);
/* Setup build context for stencil vals */
s_type = lp_type_int_vec(type.width);
lp_build_context_init(&sbld, builder, s_type);
/* Load current z/stencil value from z/stencil buffer */
zs_dst = LLVMBuildLoad(builder, zs_dst_ptr, "");
lp_build_name(zs_dst, "zsbufval");
/* Compute and apply the Z/stencil bitmasks and shifts.
*/
{
unsigned z_shift, z_mask;
unsigned s_shift, s_mask;
if (get_z_shift_and_mask(format_desc, &z_shift, &z_mask)) {
if (z_shift) {
LLVMValueRef shift = lp_build_const_int_vec(type, z_shift);
z_src = LLVMBuildLShr(builder, z_src, shift, "");
}
if (z_mask != 0xffffffff) {
LLVMValueRef mask = lp_build_const_int_vec(type, z_mask);
z_src = LLVMBuildAnd(builder, z_src, mask, "");
z_dst = LLVMBuildAnd(builder, zs_dst, mask, "");
z_bitmask = mask; /* used below */
}
else {
z_dst = zs_dst;
}
lp_build_name(z_dst, "zsbuf.z");
}
if (get_s_shift_and_mask(format_desc, &s_shift, &s_mask)) {
if (s_shift) {
LLVMValueRef shift = lp_build_const_int_vec(type, s_shift);
stencil_vals = LLVMBuildLShr(builder, zs_dst, shift, "");
stencil_shift = shift; /* used below */
}
else {
stencil_vals = zs_dst;
}
if (s_mask != 0xffffffff) {
LLVMValueRef mask = lp_build_const_int_vec(type, s_mask);
stencil_vals = LLVMBuildAnd(builder, stencil_vals, mask, "");
}
lp_build_name(stencil_vals, "stencil");
}
}
if (stencil[0].enabled) {
/* convert scalar stencil refs into vectors */
stencil_refs[0] = lp_build_broadcast_scalar(&bld, stencil_refs[0]);
stencil_refs[1] = lp_build_broadcast_scalar(&bld, stencil_refs[1]);
s_pass_mask = lp_build_stencil_test(&sbld, stencil,
stencil_refs, stencil_vals, face);
/* apply stencil-fail operator */
{
LLVMValueRef s_fail_mask = lp_build_andc(&bld, orig_mask, s_pass_mask);
stencil_vals = lp_build_stencil_op(&sbld, stencil, S_FAIL_OP,
stencil_refs, stencil_vals,
s_fail_mask, face);
}
}
if (depth->enabled) {
/* compare src Z to dst Z, returning 'pass' mask */
z_pass = lp_build_cmp(&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 (depth->writemask) {
LLVMValueRef zselectmask = mask->value;
/* mask off bits that failed Z test */
zselectmask = LLVMBuildAnd(builder, zselectmask, z_pass, "");
/* mask off bits that failed stencil test */
if (s_pass_mask) {
zselectmask = LLVMBuildAnd(builder, zselectmask, s_pass_mask, "");
}
/* if combined Z/stencil format, mask off the stencil bits */
if (z_bitmask) {
zselectmask = LLVMBuildAnd(builder, zselectmask, z_bitmask, "");
}
/* Mix the old and new Z buffer values.
* z_dst[i] = (zselectmask[i] & z_src[i]) | (~zselectmask[i] & z_dst[i])
*/
z_dst = lp_build_select_bitwise(&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_andc(&bld, orig_mask, z_pass);
stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_FAIL_OP,
stencil_refs, stencil_vals,
z_fail_mask, face);
/* apply Z-pass operator */
z_pass_mask = LLVMBuildAnd(bld.builder, orig_mask, z_pass, "");
stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_PASS_OP,
stencil_refs, stencil_vals,
z_pass_mask, face);
}
}
else {
/* No depth test: apply Z-pass operator to stencil buffer values which
* passed the stencil test.
*/
s_pass_mask = LLVMBuildAnd(bld.builder, orig_mask, s_pass_mask, "");
stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_PASS_OP,
stencil_refs, stencil_vals,
s_pass_mask, face);
}
/* The Z bits are already in the right place but we may need to shift the
* stencil bits before ORing Z with Stencil to make the final pixel value.
*/
if (stencil_vals && stencil_shift)
stencil_vals = LLVMBuildShl(bld.builder, stencil_vals,
stencil_shift, "");
/* Finally, merge/store the z/stencil values */
if ((depth->enabled && depth->writemask) ||
(stencil[0].enabled && stencil[0].writemask)) {
if (z_dst && stencil_vals)
zs_dst = LLVMBuildOr(bld.builder, z_dst, stencil_vals, "");
else if (z_dst)
zs_dst = z_dst;
else
zs_dst = stencil_vals;
LLVMBuildStore(builder, zs_dst, zs_dst_ptr);
}
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);
if (counter)
lp_build_occlusion_count(builder, type, mask->value, counter);
}
/**
* Unpack several pixels in SoA.
*
* It takes a vector of packed pixels:
*
* packed = {P0, P1, P2, P3, ..., Pn}
*
* And will produce four vectors:
*
* red = {R0, R1, R2, R3, ..., Rn}
* green = {G0, G1, G2, G3, ..., Gn}
* blue = {B0, B1, B2, B3, ..., Bn}
* alpha = {A0, A1, A2, A3, ..., An}
*
* It requires that a packed pixel fits into an element of the output
* channels. The common case is when converting pixel with a depth of 32 bit or
* less into floats.
*
* \param format_desc the format of the 'packed' incoming pixel vector
* \param type the desired type for rgba_out (type.length = n, above)
* \param packed the incoming vector of packed pixels
* \param rgba_out returns the SoA R,G,B,A vectors
*/
void
lp_build_unpack_rgba_soa(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
LLVMValueRef packed,
LLVMValueRef rgba_out[4])
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_build_context bld;
LLVMValueRef inputs[4];
unsigned chan;
assert(format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
assert(format_desc->block.width == 1);
assert(format_desc->block.height == 1);
assert(format_desc->block.bits <= type.width);
/* FIXME: Support more output types */
assert(type.width == 32);
lp_build_context_init(&bld, gallivm, type);
/* Decode the input vector components */
for (chan = 0; chan < format_desc->nr_channels; ++chan) {
const unsigned width = format_desc->channel[chan].size;
const unsigned start = format_desc->channel[chan].shift;
const unsigned stop = start + width;
LLVMValueRef input;
input = packed;
switch(format_desc->channel[chan].type) {
case UTIL_FORMAT_TYPE_VOID:
input = lp_build_undef(gallivm, type);
break;
case UTIL_FORMAT_TYPE_UNSIGNED:
/*
* Align the LSB
*/
if (start) {
input = LLVMBuildLShr(builder, input, lp_build_const_int_vec(gallivm, type, start), "");
}
/*
* Zero the MSBs
*/
if (stop < format_desc->block.bits) {
unsigned mask = ((unsigned long long)1 << width) - 1;
input = LLVMBuildAnd(builder, input, lp_build_const_int_vec(gallivm, type, mask), "");
}
/*
* Type conversion
*/
if (type.floating) {
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
assert(width == 8);
if (format_desc->swizzle[3] == chan) {
input = lp_build_unsigned_norm_to_float(gallivm, width, type, input);
}
else {
struct lp_type conv_type = lp_uint_type(type);
input = lp_build_srgb_to_linear(gallivm, conv_type, input);
}
}
else {
if(format_desc->channel[chan].normalized)
input = lp_build_unsigned_norm_to_float(gallivm, width, type, input);
else
input = LLVMBuildSIToFP(builder, input,
lp_build_vec_type(gallivm, type), "");
}
}
else if (format_desc->channel[chan].pure_integer) {
/* Nothing to do */
} else {
/* FIXME */
assert(0);
}
break;
case UTIL_FORMAT_TYPE_SIGNED:
/*
* Align the sign bit first.
*/
if (stop < type.width) {
unsigned bits = type.width - stop;
LLVMValueRef bits_val = lp_build_const_int_vec(gallivm, type, bits);
input = LLVMBuildShl(builder, input, bits_val, "");
}
/*
* Align the LSB (with an arithmetic shift to preserve the sign)
*/
if (format_desc->channel[chan].size < type.width) {
unsigned bits = type.width - format_desc->channel[chan].size;
LLVMValueRef bits_val = lp_build_const_int_vec(gallivm, type, bits);
input = LLVMBuildAShr(builder, input, bits_val, "");
}
/*
* Type conversion
*/
if (type.floating) {
input = LLVMBuildSIToFP(builder, input, lp_build_vec_type(gallivm, type), "");
if (format_desc->channel[chan].normalized) {
double scale = 1.0 / ((1 << (format_desc->channel[chan].size - 1)) - 1);
LLVMValueRef scale_val = lp_build_const_vec(gallivm, type, scale);
input = LLVMBuildFMul(builder, input, scale_val, "");
/* the formula above will produce value below -1.0 for most negative
* value but everything seems happy with that hence disable for now */
if (0)
input = lp_build_max(&bld, input,
lp_build_const_vec(gallivm, type, -1.0f));
}
}
else if (format_desc->channel[chan].pure_integer) {
/* Nothing to do */
} else {
/* FIXME */
assert(0);
}
break;
case UTIL_FORMAT_TYPE_FLOAT:
if (type.floating) {
assert(start == 0);
assert(stop == 32);
assert(type.width == 32);
input = LLVMBuildBitCast(builder, input, lp_build_vec_type(gallivm, type), "");
}
else {
/* FIXME */
assert(0);
input = lp_build_undef(gallivm, type);
}
break;
case UTIL_FORMAT_TYPE_FIXED:
if (type.floating) {
double scale = 1.0 / ((1 << (format_desc->channel[chan].size/2)) - 1);
LLVMValueRef scale_val = lp_build_const_vec(gallivm, type, scale);
input = LLVMBuildSIToFP(builder, input, lp_build_vec_type(gallivm, type), "");
input = LLVMBuildFMul(builder, input, scale_val, "");
}
else {
/* FIXME */
assert(0);
input = lp_build_undef(gallivm, type);
}
break;
default:
assert(0);
input = lp_build_undef(gallivm, type);
break;
}
inputs[chan] = input;
}
lp_build_format_swizzle_soa(format_desc, &bld, inputs, rgba_out);
}
/**
* 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);
}
static INLINE void
yuv_to_rgb_soa(struct gallivm_state *gallivm,
unsigned n,
LLVMValueRef y, LLVMValueRef u, LLVMValueRef v,
LLVMValueRef *r, LLVMValueRef *g, LLVMValueRef *b)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_type type;
struct lp_build_context bld;
LLVMValueRef c0;
LLVMValueRef c8;
LLVMValueRef c16;
LLVMValueRef c128;
LLVMValueRef c255;
LLVMValueRef cy;
LLVMValueRef cug;
LLVMValueRef cub;
LLVMValueRef cvr;
LLVMValueRef cvg;
memset(&type, 0, sizeof type);
type.sign = TRUE;
type.width = 32;
type.length = n;
lp_build_context_init(&bld, gallivm, type);
assert(lp_check_value(type, y));
assert(lp_check_value(type, u));
assert(lp_check_value(type, v));
/*
* Constants
*/
c0 = lp_build_const_int_vec(gallivm, type, 0);
c8 = lp_build_const_int_vec(gallivm, type, 8);
c16 = lp_build_const_int_vec(gallivm, type, 16);
c128 = lp_build_const_int_vec(gallivm, type, 128);
c255 = lp_build_const_int_vec(gallivm, type, 255);
cy = lp_build_const_int_vec(gallivm, type, 298);
cug = lp_build_const_int_vec(gallivm, type, -100);
cub = lp_build_const_int_vec(gallivm, type, 516);
cvr = lp_build_const_int_vec(gallivm, type, 409);
cvg = lp_build_const_int_vec(gallivm, type, -208);
/*
* y -= 16;
* u -= 128;
* v -= 128;
*/
y = LLVMBuildSub(builder, y, c16, "");
u = LLVMBuildSub(builder, u, c128, "");
v = LLVMBuildSub(builder, v, c128, "");
/*
* r = 298 * _y + 409 * _v + 128;
* g = 298 * _y - 100 * _u - 208 * _v + 128;
* b = 298 * _y + 516 * _u + 128;
*/
y = LLVMBuildMul(builder, y, cy, "");
y = LLVMBuildAdd(builder, y, c128, "");
*r = LLVMBuildMul(builder, v, cvr, "");
*g = LLVMBuildAdd(builder,
LLVMBuildMul(builder, u, cug, ""),
LLVMBuildMul(builder, v, cvg, ""),
"");
*b = LLVMBuildMul(builder, u, cub, "");
*r = LLVMBuildAdd(builder, *r, y, "");
*g = LLVMBuildAdd(builder, *g, y, "");
*b = LLVMBuildAdd(builder, *b, y, "");
/*
* r >>= 8;
* g >>= 8;
* b >>= 8;
*/
*r = LLVMBuildAShr(builder, *r, c8, "r");
*g = LLVMBuildAShr(builder, *g, c8, "g");
*b = LLVMBuildAShr(builder, *b, c8, "b");
/*
* Clamp
*/
*r = lp_build_clamp(&bld, *r, c0, c255);
*g = lp_build_clamp(&bld, *g, c0, c255);
*b = lp_build_clamp(&bld, *b, c0, c255);
}
/**
* 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);
}
