/**
* 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);
}
/**
* For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
* two (adjacent) mipmap level indexes. Later, we'll sample from those
* two mipmap levels and interpolate between them.
*/
void
lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
unsigned unit,
LLVMValueRef lod_ipart,
LLVMValueRef *lod_fpart_inout,
LLVMValueRef *level0_out,
LLVMValueRef *level1_out)
{
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_build_context *int_bld = &bld->int_bld;
struct lp_build_context *float_bld = &bld->float_bld;
LLVMValueRef last_level;
LLVMValueRef clamp_min;
LLVMValueRef clamp_max;
*level0_out = lod_ipart;
*level1_out = lp_build_add(int_bld, lod_ipart, int_bld->one);
last_level = bld->dynamic_state->last_level(bld->dynamic_state,
bld->gallivm, unit);
/*
* Clamp both lod_ipart and lod_ipart + 1 to [0, last_level], with the
* minimum number of comparisons, and zeroing lod_fpart in the extreme
* ends in the process.
*/
/* lod_ipart < 0 */
clamp_min = LLVMBuildICmp(builder, LLVMIntSLT,
lod_ipart, int_bld->zero,
"clamp_lod_to_zero");
*level0_out = LLVMBuildSelect(builder, clamp_min,
int_bld->zero, *level0_out, "");
*level1_out = LLVMBuildSelect(builder, clamp_min,
int_bld->zero, *level1_out, "");
*lod_fpart_inout = LLVMBuildSelect(builder, clamp_min,
float_bld->zero, *lod_fpart_inout, "");
/* lod_ipart >= last_level */
clamp_max = LLVMBuildICmp(builder, LLVMIntSGE,
lod_ipart, last_level,
"clamp_lod_to_last");
*level0_out = LLVMBuildSelect(builder, clamp_max,
last_level, *level0_out, "");
*level1_out = LLVMBuildSelect(builder, clamp_max,
last_level, *level1_out, "");
*lod_fpart_inout = LLVMBuildSelect(builder, clamp_max,
float_bld->zero, *lod_fpart_inout, "");
lp_build_name(*level0_out, "sampler%u_miplevel0", unit);
lp_build_name(*level1_out, "sampler%u_miplevel1", unit);
lp_build_name(*lod_fpart_inout, "sampler%u_mipweight", unit);
}
static void
attrib_name(LLVMValueRef val, unsigned attrib, unsigned chan, const char *suffix)
{
if(attrib == 0)
lp_build_name(val, "pos.%c%s", "xyzw"[chan], suffix);
else
lp_build_name(val, "input%u.%c%s", attrib - 1, "xyzw"[chan], suffix);
}
/**
* 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);
}
static LLVMValueRef
add_blend_test(struct gallivm_state *gallivm,
const struct pipe_blend_state *blend,
struct lp_type type)
{
LLVMModuleRef module = gallivm->module;
LLVMContextRef context = gallivm->context;
LLVMTypeRef vec_type;
LLVMTypeRef args[5];
LLVMValueRef func;
LLVMValueRef src_ptr;
LLVMValueRef src1_ptr;
LLVMValueRef dst_ptr;
LLVMValueRef const_ptr;
LLVMValueRef res_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
const enum pipe_format format = PIPE_FORMAT_R8G8B8A8_UNORM;
const unsigned rt = 0;
const unsigned char swizzle[4] = { 0, 1, 2, 3 };
LLVMValueRef src;
LLVMValueRef src1;
LLVMValueRef dst;
LLVMValueRef con;
LLVMValueRef res;
vec_type = lp_build_vec_type(gallivm, type);
args[4] = args[3] = args[2] = args[1] = args[0] = LLVMPointerType(vec_type, 0);
func = LLVMAddFunction(module, "test", LLVMFunctionType(LLVMVoidTypeInContext(context), args, 5, 0));
LLVMSetFunctionCallConv(func, LLVMCCallConv);
src_ptr = LLVMGetParam(func, 0);
src1_ptr = LLVMGetParam(func, 1);
dst_ptr = LLVMGetParam(func, 2);
const_ptr = LLVMGetParam(func, 3);
res_ptr = LLVMGetParam(func, 4);
block = LLVMAppendBasicBlockInContext(context, func, "entry");
builder = gallivm->builder;
LLVMPositionBuilderAtEnd(builder, block);
src = LLVMBuildLoad(builder, src_ptr, "src");
src1 = LLVMBuildLoad(builder, src1_ptr, "src1");
dst = LLVMBuildLoad(builder, dst_ptr, "dst");
con = LLVMBuildLoad(builder, const_ptr, "const");
res = lp_build_blend_aos(gallivm, blend, format, type, rt, src, NULL,
src1, NULL, dst, NULL, con, NULL, swizzle, 4);
lp_build_name(res, "res");
LLVMBuildStore(builder, res, res_ptr);
LLVMBuildRetVoid(builder);;
gallivm_verify_function(gallivm, func);
return func;
}
/**
* Convert float[] to int[] with floor().
*/
LLVMValueRef
lp_build_ifloor(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef res;
assert(type.floating);
assert(lp_check_value(type, a));
if(util_cpu_caps.has_sse4_1) {
res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
}
else {
/* Take the sign bit and add it to 1 constant */
LLVMTypeRef vec_type = lp_build_vec_type(type);
unsigned mantissa = lp_mantissa(type);
LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
LLVMValueRef sign;
LLVMValueRef offset;
/* sign = a < 0 ? ~0 : 0 */
sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
sign = LLVMBuildAnd(bld->builder, sign, mask, "");
sign = LLVMBuildAShr(bld->builder, sign, lp_build_int_const_scalar(type, type.width - 1), "");
lp_build_name(sign, "floor.sign");
/* offset = -0.99999(9)f */
offset = lp_build_const_scalar(type, -(double)(((unsigned long long)1 << mantissa) - 1)/((unsigned long long)1 << mantissa));
offset = LLVMConstBitCast(offset, int_vec_type);
/* offset = a < 0 ? -0.99999(9)f : 0.0f */
offset = LLVMBuildAnd(bld->builder, offset, sign, "");
offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "");
lp_build_name(offset, "floor.offset");
res = LLVMBuildAdd(bld->builder, a, offset, "");
lp_build_name(res, "floor.res");
}
res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
lp_build_name(res, "floor");
return res;
}
LLVMValueRef
lp_build_struct_get(struct gallivm_state *gallivm,
LLVMValueRef ptr,
unsigned member,
const char *name)
{
LLVMValueRef member_ptr;
LLVMValueRef res;
assert(LLVMGetTypeKind(LLVMTypeOf(ptr)) == LLVMPointerTypeKind);
assert(LLVMGetTypeKind(LLVMGetElementType(LLVMTypeOf(ptr))) == LLVMStructTypeKind);
member_ptr = lp_build_struct_get_ptr(gallivm, ptr, member, name);
res = LLVMBuildLoad(gallivm->builder, member_ptr, "");
lp_build_name(res, "%s.%s", LLVMGetValueName(ptr), name);
return res;
}
LLVMValueRef
lp_build_pointer_get(LLVMBuilderRef builder,
LLVMValueRef ptr,
LLVMValueRef index)
{
LLVMValueRef element_ptr;
LLVMValueRef res;
assert(LLVMGetTypeKind(LLVMTypeOf(ptr)) == LLVMPointerTypeKind);
element_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
res = LLVMBuildLoad(builder, element_ptr, "");
#ifdef DEBUG
lp_build_name(res, "%s[%s]", LLVMGetValueName(ptr), LLVMGetValueName(index));
#endif
return res;
}
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);
}
LLVMValueRef
lp_build_struct_get_ptr(struct gallivm_state *gallivm,
LLVMValueRef ptr,
unsigned member,
const char *name)
{
LLVMValueRef indices[2];
LLVMValueRef member_ptr;
assert(LLVMGetTypeKind(LLVMTypeOf(ptr)) == LLVMPointerTypeKind);
assert(LLVMGetTypeKind(LLVMGetElementType(LLVMTypeOf(ptr))) == LLVMStructTypeKind);
indices[0] = lp_build_const_int32(gallivm, 0);
indices[1] = lp_build_const_int32(gallivm, member);
member_ptr = LLVMBuildGEP(gallivm->builder, ptr, indices, Elements(indices), "");
lp_build_name(member_ptr, "%s.%s_ptr", LLVMGetValueName(ptr), name);
return member_ptr;
}
LLVMValueRef
lp_build_array_get(struct gallivm_state *gallivm,
LLVMValueRef ptr,
LLVMValueRef index)
{
LLVMValueRef element_ptr;
LLVMValueRef res;
assert(LLVMGetTypeKind(LLVMTypeOf(ptr)) == LLVMPointerTypeKind);
assert(LLVMGetTypeKind(LLVMGetElementType(LLVMTypeOf(ptr))) == LLVMArrayTypeKind);
element_ptr = lp_build_array_get_ptr(gallivm, ptr, index);
res = LLVMBuildLoad(gallivm->builder, element_ptr, "");
#ifdef DEBUG
lp_build_name(res, "%s[%s]", LLVMGetValueName(ptr), LLVMGetValueName(index));
#endif
return res;
}
LLVMValueRef
lp_build_array_get_ptr(struct gallivm_state *gallivm,
LLVMValueRef ptr,
LLVMValueRef index)
{
LLVMValueRef indices[2];
LLVMValueRef element_ptr;
assert(LLVMGetTypeKind(LLVMTypeOf(ptr)) == LLVMPointerTypeKind);
assert(LLVMGetTypeKind(LLVMGetElementType(LLVMTypeOf(ptr))) == LLVMArrayTypeKind);
indices[0] = lp_build_const_int32(gallivm, 0);
indices[1] = index;
element_ptr = LLVMBuildGEP(gallivm->builder, ptr, indices, Elements(indices), "");
#ifdef DEBUG
lp_build_name(element_ptr, "&%s[%s]",
LLVMGetValueName(ptr), LLVMGetValueName(index));
#endif
return element_ptr;
}
static LLVMValueRef
generate_scissor_test(LLVMBuilderRef builder,
LLVMValueRef context_ptr,
const struct lp_build_interp_soa_context *interp,
struct lp_type type)
{
LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMValueRef xpos = interp->pos[0], ypos = interp->pos[1];
LLVMValueRef xmin, ymin, xmax, ymax;
LLVMValueRef m0, m1, m2, m3, m;
/* xpos, ypos contain the window coords for the four pixels in the quad */
assert(xpos);
assert(ypos);
/* get the current scissor bounds, convert to vectors */
xmin = lp_jit_context_scissor_xmin_value(builder, context_ptr);
xmin = lp_build_broadcast(builder, vec_type, xmin);
ymin = lp_jit_context_scissor_ymin_value(builder, context_ptr);
ymin = lp_build_broadcast(builder, vec_type, ymin);
xmax = lp_jit_context_scissor_xmax_value(builder, context_ptr);
xmax = lp_build_broadcast(builder, vec_type, xmax);
ymax = lp_jit_context_scissor_ymax_value(builder, context_ptr);
ymax = lp_build_broadcast(builder, vec_type, ymax);
/* compare the fragment's position coordinates against the scissor bounds */
m0 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, xpos, xmin);
m1 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, ypos, ymin);
m2 = lp_build_compare(builder, type, PIPE_FUNC_LESS, xpos, xmax);
m3 = lp_build_compare(builder, type, PIPE_FUNC_LESS, ypos, ymax);
/* AND all the masks together */
m = LLVMBuildAnd(builder, m0, m1, "");
m = LLVMBuildAnd(builder, m, m2, "");
m = LLVMBuildAnd(builder, m, m3, "");
lp_build_name(m, "scissormask");
return m;
}
/**
* Fetch the specified member of the lp_jit_texture structure.
* \param emit_load if TRUE, emit the LLVM load instruction to actually
* fetch the field's value. Otherwise, just emit the
* GEP code to address the field.
*
* @sa http://llvm.org/docs/GetElementPtr.html
*/
static LLVMValueRef
lp_llvm_texture_member(const struct lp_sampler_dynamic_state *base,
struct gallivm_state *gallivm,
unsigned unit,
unsigned member_index,
const char *member_name,
boolean emit_load)
{
struct llvmpipe_sampler_dynamic_state *state =
(struct llvmpipe_sampler_dynamic_state *)base;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef indices[4];
LLVMValueRef ptr;
LLVMValueRef res;
assert(unit < PIPE_MAX_SAMPLERS);
/* context[0] */
indices[0] = lp_build_const_int32(gallivm, 0);
/* context[0].textures */
indices[1] = lp_build_const_int32(gallivm, LP_JIT_CTX_TEXTURES);
/* context[0].textures[unit] */
indices[2] = lp_build_const_int32(gallivm, unit);
/* context[0].textures[unit].member */
indices[3] = lp_build_const_int32(gallivm, member_index);
ptr = LLVMBuildGEP(builder, state->context_ptr, indices, Elements(indices), "");
if (emit_load)
res = LLVMBuildLoad(builder, ptr, "");
else
res = ptr;
lp_build_name(res, "context.texture%u.%s", unit, member_name);
return res;
}
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);
}
/**
* Initialize the bld->a, dadq fields. This involves fetching
* those values from the arrays which are passed into the JIT function.
*/
static void
coeffs_init(struct lp_build_interp_soa_context *bld,
LLVMValueRef a0_ptr,
LLVMValueRef dadx_ptr,
LLVMValueRef dady_ptr)
{
struct lp_build_context *coeff_bld = &bld->coeff_bld;
struct lp_build_context *setup_bld = &bld->setup_bld;
struct gallivm_state *gallivm = coeff_bld->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef pixoffx, pixoffy;
unsigned attrib;
unsigned chan;
unsigned i;
pixoffx = coeff_bld->undef;
pixoffy = coeff_bld->undef;
for (i = 0; i < coeff_bld->type.length; i++) {
LLVMValueRef nr = lp_build_const_int32(gallivm, i);
LLVMValueRef pixxf = lp_build_const_float(gallivm, quad_offset_x[i]);
LLVMValueRef pixyf = lp_build_const_float(gallivm, quad_offset_y[i]);
pixoffx = LLVMBuildInsertElement(builder, pixoffx, pixxf, nr, "");
pixoffy = LLVMBuildInsertElement(builder, pixoffy, pixyf, nr, "");
}
for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
const unsigned mask = bld->mask[attrib];
const unsigned interp = bld->interp[attrib];
LLVMValueRef index = lp_build_const_int32(gallivm,
attrib * TGSI_NUM_CHANNELS);
LLVMValueRef ptr;
LLVMValueRef dadxaos = setup_bld->zero;
LLVMValueRef dadyaos = setup_bld->zero;
LLVMValueRef a0aos = setup_bld->zero;
/* always fetch all 4 values for performance/simplicity */
switch (interp) {
case LP_INTERP_PERSPECTIVE:
/* fall-through */
case LP_INTERP_LINEAR:
ptr = LLVMBuildGEP(builder, dadx_ptr, &index, 1, "");
ptr = LLVMBuildBitCast(builder, ptr,
LLVMPointerType(setup_bld->vec_type, 0), "");
dadxaos = LLVMBuildLoad(builder, ptr, "");
ptr = LLVMBuildGEP(builder, dady_ptr, &index, 1, "");
ptr = LLVMBuildBitCast(builder, ptr,
LLVMPointerType(setup_bld->vec_type, 0), "");
dadyaos = LLVMBuildLoad(builder, ptr, "");
attrib_name(dadxaos, attrib, 0, ".dadxaos");
attrib_name(dadyaos, attrib, 0, ".dadyaos");
/* fall-through */
case LP_INTERP_CONSTANT:
case LP_INTERP_FACING:
ptr = LLVMBuildGEP(builder, a0_ptr, &index, 1, "");
ptr = LLVMBuildBitCast(builder, ptr,
LLVMPointerType(setup_bld->vec_type, 0), "");
a0aos = LLVMBuildLoad(builder, ptr, "");
attrib_name(a0aos, attrib, 0, ".a0aos");
break;
case LP_INTERP_POSITION:
/* Nothing to do as the position coeffs are already setup in slot 0 */
continue;
default:
assert(0);
break;
}
/*
* a = a0 + (x * dadx + y * dady)
* a0aos is the attrib value at top left corner of stamp
*/
if (interp != LP_INTERP_CONSTANT &&
interp != LP_INTERP_FACING) {
LLVMValueRef x = lp_build_broadcast_scalar(setup_bld, bld->x);
LLVMValueRef y = lp_build_broadcast_scalar(setup_bld, bld->y);
a0aos = lp_build_fmuladd(builder, x, dadxaos, a0aos);
a0aos = lp_build_fmuladd(builder, y, dadyaos, a0aos);
}
/*
* dadq = {0, dadx, dady, dadx + dady}
* for two quads (side by side) this is:
* {0, dadx, dady, dadx+dady, 2*dadx, 2*dadx+dady, 3*dadx+dady}
*/
for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
/* this generates a CRAPLOAD of shuffles... */
if (mask & (1 << chan)) {
LLVMValueRef dadx, dady;
LLVMValueRef dadq, dadq2;
LLVMValueRef a;
LLVMValueRef chan_index = lp_build_const_int32(gallivm, chan);
if (attrib == 0 && chan == 0) {
a = bld->x;
if (bld->pos_offset) {
a = LLVMBuildFAdd(builder, a, lp_build_const_float(gallivm, bld->pos_offset), "");
}
a = lp_build_broadcast_scalar(coeff_bld, a);
dadx = coeff_bld->one;
dady = coeff_bld->zero;
}
else if (attrib == 0 && chan == 1) {
a = bld->y;
if (bld->pos_offset) {
a = LLVMBuildFAdd(builder, a, lp_build_const_float(gallivm, bld->pos_offset), "");
}
a = lp_build_broadcast_scalar(coeff_bld, a);
dady = coeff_bld->one;
dadx = coeff_bld->zero;
}
else {
dadx = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, dadxaos, chan_index);
dady = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, dadyaos, chan_index);
/*
* a = {a, a, a, a}
*/
a = lp_build_extract_broadcast(gallivm, setup_bld->type,
coeff_bld->type, a0aos, chan_index);
}
dadx = LLVMBuildFMul(builder, dadx, pixoffx, "");
dady = LLVMBuildFMul(builder, dady, pixoffy, "");
dadq = LLVMBuildFAdd(builder, dadx, dady, "");
/*
* Compute the attrib values on the upper-left corner of each
* group of quads.
* Note that if we process 2 quads at once this doesn't
* really exactly to what we want.
* We need to access elem 0 and 2 respectively later if we process
* 2 quads at once.
*/
if (interp != LP_INTERP_CONSTANT &&
interp != LP_INTERP_FACING) {
dadq2 = LLVMBuildFAdd(builder, dadq, dadq, "");
a = LLVMBuildFAdd(builder, a, dadq2, "");
}
#if PERSPECTIVE_DIVIDE_PER_QUAD
/*
* a *= 1 / w
*/
/*
* XXX since we're only going to access elements 0,2 out of 8
* if we have 8-wide vectors we should do the division only 4-wide.
* a is really a 2-elements in a 4-wide vector disguised as 8-wide
* in this case.
*/
if (interp == LP_INTERP_PERSPECTIVE) {
LLVMValueRef w = bld->a[0][3];
assert(attrib != 0);
assert(bld->mask[0] & TGSI_WRITEMASK_W);
if (!bld->oow) {
bld->oow = lp_build_rcp(coeff_bld, w);
lp_build_name(bld->oow, "oow");
}
a = lp_build_mul(coeff_bld, a, bld->oow);
}
#endif
attrib_name(a, attrib, chan, ".a");
attrib_name(dadq, attrib, chan, ".dadq");
bld->a[attrib][chan] = lp_build_alloca(gallivm,
LLVMTypeOf(a), "");
LLVMBuildStore(builder, a, bld->a[attrib][chan]);
bld->dadq[attrib][chan] = dadq;
}
}
}
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
*/
static struct lp_fragment_shader_variant *
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
const struct lp_fragment_shader_variant_key *key)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
struct lp_fragment_shader_variant *variant;
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_vec_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
LLVMTypeRef arg_types[9];
LLVMTypeRef func_type;
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef mask_ptr;
LLVMValueRef color_ptr;
LLVMValueRef depth_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
LLVMValueRef y0;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef blend_in_color[NUM_CHANNELS];
unsigned num_fs;
unsigned i;
unsigned chan;
#ifdef DEBUG
tgsi_dump(shader->base.tokens, 0);
if(key->depth.enabled) {
debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format));
debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
debug_printf("depth.writemask = %u\n", key->depth.writemask);
}
if(key->alpha.enabled) {
debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
}
if(key->blend.logicop_enable) {
debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
}
else if(key->blend.blend_enable) {
debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key->blend.rgb_func, TRUE));
debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
debug_printf("alpha_func = %s\n", debug_dump_blend_func (key->blend.alpha_func, TRUE));
debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
}
debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
if(key->sampler[i].format) {
debug_printf("sampler[%u] = \n", i);
debug_printf(" .format = %s\n",
pf_name(key->sampler[i].format));
debug_printf(" .target = %s\n",
debug_dump_tex_target(key->sampler[i].target, TRUE));
debug_printf(" .pot = %u %u %u\n",
key->sampler[i].pot_width,
key->sampler[i].pot_height,
key->sampler[i].pot_depth);
debug_printf(" .wrap = %s %s %s\n",
debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
debug_printf(" .min_img_filter = %s\n",
debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
debug_printf(" .min_mip_filter = %s\n",
debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
debug_printf(" .mag_img_filter = %s\n",
debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
if(key->sampler[i].compare_mode)
debug_printf(" .compare_mode = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE));
debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
debug_printf(" .prefilter = %u\n", key->sampler[i].prefilter);
}
}
#endif
variant = CALLOC_STRUCT(lp_fragment_shader_variant);
if(!variant)
return NULL;
variant->shader = shader;
memcpy(&variant->key, key, sizeof *key);
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 element per vector */
num_fs = 4;
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_vec_type = lp_build_vec_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
blend_int_vec_type = lp_build_int_vec_type(blend_type);
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
arg_types[7] = LLVMPointerType(blend_vec_type, 0); /* color */
arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
variant->function = LLVMAddFunction(screen->module, "shader", func_type);
LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(variant->function, 0);
x = LLVMGetParam(variant->function, 1);
y = LLVMGetParam(variant->function, 2);
a0_ptr = LLVMGetParam(variant->function, 3);
dadx_ptr = LLVMGetParam(variant->function, 4);
dady_ptr = LLVMGetParam(variant->function, 5);
mask_ptr = LLVMGetParam(variant->function, 6);
color_ptr = LLVMGetParam(variant->function, 7);
depth_ptr = LLVMGetParam(variant->function, 8);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(mask_ptr, "mask");
lp_build_name(color_ptr, "color");
lp_build_name(depth_ptr, "depth");
/*
* Function body
*/
block = LLVMAppendBasicBlock(variant->function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
generate_pos0(builder, x, y, &x0, &y0);
lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x0, y0, 2, 0);
#if 0
/* C texture sampling */
sampler = lp_c_sampler_soa_create(context_ptr);
#else
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
#endif
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
if(i != 0)
lp_build_interp_soa_update(&interp);
fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i],
out_color,
depth_ptr_i);
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[chan][i] = out_color[chan];
}
sampler->destroy(sampler);
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
}
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
/*
* Blending.
*/
generate_blend(&key->blend,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/*
* Translate the LLVM IR into machine code.
*/
if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
LLVMDumpValue(variant->function);
abort();
}
LLVMRunFunctionPassManager(screen->pass, variant->function);
#ifdef DEBUG
LLVMDumpValue(variant->function);
debug_printf("\n");
#endif
variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
#ifdef DEBUG
lp_disassemble(variant->jit_function);
#endif
variant->next = shader->variants;
shader->variants = variant;
return variant;
}
static LLVMValueRef
emit_fetch_constant(
struct lp_build_tgsi_context * bld_base,
const struct tgsi_full_src_register * reg,
enum tgsi_opcode_type stype,
unsigned swizzle)
{
struct lp_build_tgsi_aos_context * bld = lp_aos_context(bld_base);
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
struct lp_type type = bld_base->base.type;
LLVMValueRef res;
unsigned chan;
assert(!reg->Register.Indirect);
/*
* Get the constants components
*/
res = bld->bld_base.base.undef;
for (chan = 0; chan < 4; ++chan) {
LLVMValueRef index;
LLVMValueRef scalar_ptr;
LLVMValueRef scalar;
LLVMValueRef swizzle;
index = lp_build_const_int32(bld->bld_base.base.gallivm,
reg->Register.Index * 4 + chan);
scalar_ptr = LLVMBuildGEP(builder, bld->consts_ptr, &index, 1, "");
scalar = LLVMBuildLoad(builder, scalar_ptr, "");
lp_build_name(scalar, "const[%u].%c", reg->Register.Index, "xyzw"[chan]);
/*
* NOTE: constants array is always assumed to be RGBA
*/
swizzle = lp_build_const_int32(bld->bld_base.base.gallivm,
bld->swizzles[chan]);
res = LLVMBuildInsertElement(builder, res, scalar, swizzle, "");
}
/*
* Broadcast the first quaternion to all others.
*
* XXX: could be factored into a reusable function.
*/
if (type.length > 4) {
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
unsigned i;
for (chan = 0; chan < 4; ++chan) {
shuffles[chan] = lp_build_const_int32(bld->bld_base.base.gallivm, chan);
}
for (i = 4; i < type.length; ++i) {
shuffles[i] = shuffles[i % 4];
}
res = LLVMBuildShuffleVector(builder,
res, bld->bld_base.base.undef,
LLVMConstVector(shuffles, type.length),
"");
}
return res;
}
static LLVMValueRef
add_blend_test(struct gallivm_state *gallivm,
const struct pipe_blend_state *blend,
enum vector_mode mode,
struct lp_type type)
{
LLVMModuleRef module = gallivm->module;
LLVMContextRef context = gallivm->context;
LLVMTypeRef vec_type;
LLVMTypeRef args[4];
LLVMValueRef func;
LLVMValueRef src_ptr;
LLVMValueRef dst_ptr;
LLVMValueRef const_ptr;
LLVMValueRef res_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
const enum pipe_format format = PIPE_FORMAT_R8G8B8A8_UNORM;
const unsigned rt = 0;
const unsigned char swizzle[4] = { 0, 1, 2, 3 };
vec_type = lp_build_vec_type(gallivm, type);
args[3] = args[2] = args[1] = args[0] = LLVMPointerType(vec_type, 0);
func = LLVMAddFunction(module, "test", LLVMFunctionType(LLVMVoidTypeInContext(context), args, 4, 0));
LLVMSetFunctionCallConv(func, LLVMCCallConv);
src_ptr = LLVMGetParam(func, 0);
dst_ptr = LLVMGetParam(func, 1);
const_ptr = LLVMGetParam(func, 2);
res_ptr = LLVMGetParam(func, 3);
block = LLVMAppendBasicBlockInContext(context, func, "entry");
builder = gallivm->builder;
LLVMPositionBuilderAtEnd(builder, block);
if (mode == AoS) {
LLVMValueRef src;
LLVMValueRef dst;
LLVMValueRef con;
LLVMValueRef res;
src = LLVMBuildLoad(builder, src_ptr, "src");
dst = LLVMBuildLoad(builder, dst_ptr, "dst");
con = LLVMBuildLoad(builder, const_ptr, "const");
res = lp_build_blend_aos(gallivm, blend, &format, type, rt, src, dst, NULL, con, swizzle);
lp_build_name(res, "res");
LLVMBuildStore(builder, res, res_ptr);
}
if (mode == SoA) {
LLVMValueRef src[4];
LLVMValueRef dst[4];
LLVMValueRef con[4];
LLVMValueRef res[4];
unsigned i;
for(i = 0; i < 4; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32TypeInContext(context), i, 0);
src[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, src_ptr, &index, 1, ""), "");
dst[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
con[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
lp_build_name(src[i], "src.%c", "rgba"[i]);
lp_build_name(con[i], "con.%c", "rgba"[i]);
lp_build_name(dst[i], "dst.%c", "rgba"[i]);
}
lp_build_blend_soa(gallivm, blend, type, rt, src, dst, con, res);
for(i = 0; i < 4; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32TypeInContext(context), i, 0);
lp_build_name(res[i], "res.%c", "rgba"[i]);
LLVMBuildStore(builder, res[i], LLVMBuildGEP(builder, res_ptr, &index, 1, ""));
}
}
LLVMBuildRetVoid(builder);;
return func;
}
/**
* 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;
}
static LLVMValueRef
add_blend_test(LLVMModuleRef module,
const struct pipe_blend_state *blend,
enum vector_mode mode,
struct lp_type type)
{
LLVMTypeRef ret_type;
LLVMTypeRef vec_type;
LLVMTypeRef args[4];
LLVMValueRef func;
LLVMValueRef src_ptr;
LLVMValueRef dst_ptr;
LLVMValueRef const_ptr;
LLVMValueRef res_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
ret_type = LLVMInt64Type();
vec_type = lp_build_vec_type(type);
args[3] = args[2] = args[1] = args[0] = LLVMPointerType(vec_type, 0);
func = LLVMAddFunction(module, "test", LLVMFunctionType(LLVMVoidType(), args, 4, 0));
LLVMSetFunctionCallConv(func, LLVMCCallConv);
src_ptr = LLVMGetParam(func, 0);
dst_ptr = LLVMGetParam(func, 1);
const_ptr = LLVMGetParam(func, 2);
res_ptr = LLVMGetParam(func, 3);
block = LLVMAppendBasicBlock(func, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
if (mode == AoS) {
LLVMValueRef src;
LLVMValueRef dst;
LLVMValueRef con;
LLVMValueRef res;
src = LLVMBuildLoad(builder, src_ptr, "src");
dst = LLVMBuildLoad(builder, dst_ptr, "dst");
con = LLVMBuildLoad(builder, const_ptr, "const");
res = lp_build_blend_aos(builder, blend, type, src, dst, con, 3);
lp_build_name(res, "res");
LLVMBuildStore(builder, res, res_ptr);
}
if (mode == SoA) {
LLVMValueRef src[4];
LLVMValueRef dst[4];
LLVMValueRef con[4];
LLVMValueRef res[4];
unsigned i;
for(i = 0; i < 4; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
src[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, src_ptr, &index, 1, ""), "");
dst[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
con[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
lp_build_name(src[i], "src.%c", "rgba"[i]);
lp_build_name(con[i], "con.%c", "rgba"[i]);
lp_build_name(dst[i], "dst.%c", "rgba"[i]);
}
lp_build_blend_soa(builder, blend, type, src, dst, con, res);
for(i = 0; i < 4; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
lp_build_name(res[i], "res.%c", "rgba"[i]);
LLVMBuildStore(builder, res[i], LLVMBuildGEP(builder, res_ptr, &index, 1, ""));
}
}
LLVMBuildRetVoid(builder);;
LLVMDisposeBuilder(builder);
return func;
}
/**
* Generate the code to do inside/outside triangle testing for the
* four pixels in a 2x2 quad. This will set the four elements of the
* quad mask vector to 0 or ~0.
* \param i which quad of the quad group to test, in [0,3]
*/
static void
generate_tri_edge_mask(LLVMBuilderRef builder,
unsigned i,
LLVMValueRef *mask, /* ivec4, out */
LLVMValueRef c0, /* int32 */
LLVMValueRef c1, /* int32 */
LLVMValueRef c2, /* int32 */
LLVMValueRef step0_ptr, /* ivec4 */
LLVMValueRef step1_ptr, /* ivec4 */
LLVMValueRef step2_ptr) /* ivec4 */
{
#define OPTIMIZE_IN_OUT_TEST 0
#if OPTIMIZE_IN_OUT_TEST
struct lp_build_if_state ifctx;
LLVMValueRef not_draw_all;
#endif
struct lp_build_flow_context *flow;
struct lp_type i32_type;
LLVMTypeRef i32vec4_type, mask_type;
LLVMValueRef c0_vec, c1_vec, c2_vec;
LLVMValueRef in_out_mask;
assert(i < 4);
/* int32 vector type */
memset(&i32_type, 0, sizeof i32_type);
i32_type.floating = FALSE; /* values are integers */
i32_type.sign = TRUE; /* values are signed */
i32_type.norm = FALSE; /* values are not normalized */
i32_type.width = 32; /* 32-bit int values */
i32_type.length = 4; /* 4 elements per vector */
i32vec4_type = lp_build_int32_vec4_type();
mask_type = LLVMIntType(32 * 4);
/*
* Use a conditional here to do detailed pixel in/out testing.
* We only have to do this if c0 != INT_MIN.
*/
flow = lp_build_flow_create(builder);
lp_build_flow_scope_begin(flow);
{
#if OPTIMIZE_IN_OUT_TEST
/* not_draw_all = (c0 != INT_MIN) */
not_draw_all = LLVMBuildICmp(builder,
LLVMIntNE,
c0,
LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
"");
in_out_mask = lp_build_int_const_scalar(i32_type, ~0);
lp_build_flow_scope_declare(flow, &in_out_mask);
/* if (not_draw_all) {... */
lp_build_if(&ifctx, flow, builder, not_draw_all);
#endif
{
LLVMValueRef step0_vec, step1_vec, step2_vec;
LLVMValueRef m0_vec, m1_vec, m2_vec;
LLVMValueRef index, m;
/* c0_vec = {c0, c0, c0, c0}
* Note that we emit this code four times but LLVM optimizes away
* three instances of it.
*/
c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
lp_build_name(c0_vec, "edgeconst0vec");
lp_build_name(c1_vec, "edgeconst1vec");
lp_build_name(c2_vec, "edgeconst2vec");
/* load step0vec, step1, step2 vec from memory */
index = LLVMConstInt(LLVMInt32Type(), i, 0);
step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
lp_build_name(step0_vec, "step0vec");
lp_build_name(step1_vec, "step1vec");
lp_build_name(step2_vec, "step2vec");
/* m0_vec = step0_ptr[i] > c0_vec */
m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
/* in_out_mask = m0_vec & m1_vec & m2_vec */
m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
lp_build_name(in_out_mask, "inoutmaskvec");
}
#if OPTIMIZE_IN_OUT_TEST
lp_build_endif(&ifctx);
#endif
}
lp_build_flow_scope_end(flow);
lp_build_flow_destroy(flow);
/* This is the initial alive/dead pixel mask for a quad of four pixels.
* It's an int[4] vector with each word set to 0 or ~0.
* Words will get cleared when pixels faile the Z test, etc.
*/
*mask = in_out_mask;
}
/**
* 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);
}
/**
* Load depth/stencil values.
* The stored values are linear, swizzle them.
*
* \param type the data type of the fragment depth/stencil values
* \param format_desc description of the depth/stencil surface
* \param loop_counter the current loop iteration
* \param depth_ptr pointer to the depth/stencil values of this 4x4 block
* \param depth_stride stride of the depth/stencil buffer
* \param z_fb contains z values loaded from fb (may include padding)
* \param s_fb contains s values loaded from fb (may include padding)
*/
void
lp_build_depth_stencil_load_swizzled(struct gallivm_state *gallivm,
struct lp_type z_src_type,
const struct util_format_description *format_desc,
LLVMValueRef depth_ptr,
LLVMValueRef depth_stride,
LLVMValueRef *z_fb,
LLVMValueRef *s_fb,
LLVMValueRef loop_counter)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 4];
LLVMValueRef zs_dst1, zs_dst2;
LLVMValueRef zs_dst_ptr;
LLVMValueRef depth_offset1, depth_offset2;
LLVMTypeRef load_ptr_type;
unsigned depth_bytes = format_desc->block.bits / 8;
struct lp_type zs_type = lp_depth_type(format_desc, z_src_type.length);
struct lp_type zs_load_type = zs_type;
zs_load_type.length = zs_load_type.length / 2;
load_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0);
if (z_src_type.length == 4) {
unsigned i;
LLVMValueRef looplsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
LLVMValueRef loopmsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 2), "");
LLVMValueRef offset2 = LLVMBuildMul(builder, loopmsb,
depth_stride, "");
depth_offset1 = LLVMBuildMul(builder, looplsb,
lp_build_const_int32(gallivm, depth_bytes * 2), "");
depth_offset1 = LLVMBuildAdd(builder, depth_offset1, offset2, "");
/* just concatenate the loaded 2x2 values into 4-wide vector */
for (i = 0; i < 4; i++) {
shuffles[i] = lp_build_const_int32(gallivm, i);
}
}
else {
unsigned i;
LLVMValueRef loopx2 = LLVMBuildShl(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
assert(z_src_type.length == 8);
depth_offset1 = LLVMBuildMul(builder, loopx2, depth_stride, "");
/*
* We load 2x4 values, and need to swizzle them (order
* 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
*/
for (i = 0; i < 8; i++) {
shuffles[i] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
}
}
depth_offset2 = LLVMBuildAdd(builder, depth_offset1, depth_stride, "");
/* Load current z/stencil values from z/stencil buffer */
zs_dst_ptr = LLVMBuildGEP(builder, depth_ptr, &depth_offset1, 1, "");
zs_dst_ptr = LLVMBuildBitCast(builder, zs_dst_ptr, load_ptr_type, "");
zs_dst1 = LLVMBuildLoad(builder, zs_dst_ptr, "");
zs_dst_ptr = LLVMBuildGEP(builder, depth_ptr, &depth_offset2, 1, "");
zs_dst_ptr = LLVMBuildBitCast(builder, zs_dst_ptr, load_ptr_type, "");
zs_dst2 = LLVMBuildLoad(builder, zs_dst_ptr, "");
*z_fb = LLVMBuildShuffleVector(builder, zs_dst1, zs_dst2,
LLVMConstVector(shuffles, zs_type.length), "");
*s_fb = *z_fb;
if (format_desc->block.bits < z_src_type.width) {
/* Extend destination ZS values (e.g., when reading from Z16_UNORM) */
*z_fb = LLVMBuildZExt(builder, *z_fb,
lp_build_int_vec_type(gallivm, z_src_type), "");
}
else if (format_desc->block.bits > 32) {
/* rely on llvm to handle too wide vector we have here nicely */
unsigned i;
struct lp_type typex2 = zs_type;
struct lp_type s_type = zs_type;
LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH / 4];
LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH / 4];
LLVMValueRef tmp;
typex2.width = typex2.width / 2;
typex2.length = typex2.length * 2;
s_type.width = s_type.width / 2;
s_type.floating = 0;
tmp = LLVMBuildBitCast(builder, *z_fb,
lp_build_vec_type(gallivm, typex2), "");
for (i = 0; i < zs_type.length; i++) {
shuffles1[i] = lp_build_const_int32(gallivm, i * 2);
shuffles2[i] = lp_build_const_int32(gallivm, i * 2 + 1);
}
*z_fb = LLVMBuildShuffleVector(builder, tmp, tmp,
LLVMConstVector(shuffles1, zs_type.length), "");
*s_fb = LLVMBuildShuffleVector(builder, tmp, tmp,
LLVMConstVector(shuffles2, zs_type.length), "");
*s_fb = LLVMBuildBitCast(builder, *s_fb,
lp_build_vec_type(gallivm, s_type), "");
lp_build_name(*s_fb, "s_dst");
}
lp_build_name(*z_fb, "z_dst");
lp_build_name(*s_fb, "s_dst");
lp_build_name(*z_fb, "z_dst");
}
/**
* 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);
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
* Note that the function which we generate operates on a block of 16
* pixels at at time. The block contains 2x2 quads. Each quad contains
* 2x2 pixels.
*/
static void
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
struct lp_fragment_shader_variant *variant,
unsigned do_tri_test)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
const struct lp_fragment_shader_variant_key *key = &variant->key;
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_vec_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
LLVMTypeRef arg_types[14];
LLVMTypeRef func_type;
LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
LLVMValueRef y0;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef blend_in_color[NUM_CHANNELS];
LLVMValueRef function;
unsigned num_fs;
unsigned i;
unsigned chan;
unsigned cbuf;
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 elements per vector */
num_fs = 4; /* number of quads per block */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_vec_type = lp_build_vec_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
blend_int_vec_type = lp_build_int_vec_type(blend_type);
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[6] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
arg_types[8] = LLVMInt32Type(); /* c0 */
arg_types[9] = LLVMInt32Type(); /* c1 */
arg_types[10] = LLVMInt32Type(); /* c2 */
/* Note: the step arrays are built as int32[16] but we interpret
* them here as int32_vec4[4].
*/
arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
function = LLVMAddFunction(screen->module, "shader", func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
variant->function[do_tri_test] = function;
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
*/
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(function, 0);
x = LLVMGetParam(function, 1);
y = LLVMGetParam(function, 2);
a0_ptr = LLVMGetParam(function, 3);
dadx_ptr = LLVMGetParam(function, 4);
dady_ptr = LLVMGetParam(function, 5);
color_ptr_ptr = LLVMGetParam(function, 6);
depth_ptr = LLVMGetParam(function, 7);
c0 = LLVMGetParam(function, 8);
c1 = LLVMGetParam(function, 9);
c2 = LLVMGetParam(function, 10);
step0_ptr = LLVMGetParam(function, 11);
step1_ptr = LLVMGetParam(function, 12);
step2_ptr = LLVMGetParam(function, 13);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr");
lp_build_name(depth_ptr, "depth");
lp_build_name(c0, "c0");
lp_build_name(c1, "c1");
lp_build_name(c2, "c2");
lp_build_name(step0_ptr, "step0");
lp_build_name(step1_ptr, "step1");
lp_build_name(step2_ptr, "step2");
/*
* Function body
*/
block = LLVMAppendBasicBlock(function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
generate_pos0(builder, x, y, &x0, &y0);
lp_build_interp_soa_init(&interp,
shader->base.tokens,
key->flatshade,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x0, y0);
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
/* loop over quads in the block */
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
int cbuf;
if(i != 0)
lp_build_interp_soa_update(&interp, i);
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i], /* output */
out_color,
depth_ptr_i,
do_tri_test,
c0, c1, c2,
step0_ptr, step1_ptr, step2_ptr);
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
}
sampler->destroy(sampler);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
LLVMValueRef color_ptr;
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[cbuf][chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
}
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
color_ptr = LLVMBuildLoad(builder,
LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
"");
lp_build_name(color_ptr, "color_ptr%d", cbuf);
/*
* Blending.
*/
generate_blend(&key->blend,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
}
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/* Verify the LLVM IR. If invalid, dump and abort */
#ifdef DEBUG
if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
if (1)
LLVMDumpValue(function);
abort();
}
#endif
/* Apply optimizations to LLVM IR */
if (1)
LLVMRunFunctionPassManager(screen->pass, function);
if (LP_DEBUG & DEBUG_JIT) {
/* Print the LLVM IR to stderr */
LLVMDumpValue(function);
debug_printf("\n");
}
/*
* Translate the LLVM IR into machine code.
*/
variant->jit_function[do_tri_test] = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, function);
if (LP_DEBUG & DEBUG_ASM)
lp_disassemble(variant->jit_function[do_tri_test]);
}
/**
* Generate the fragment shader, depth/stencil test, and alpha tests.
* \param i which quad in the tile, in range [0,3]
* \param do_tri_test if 1, do triangle edge in/out testing
*/
static void
generate_fs(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
const struct lp_fragment_shader_variant_key *key,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef context_ptr,
unsigned i,
const struct lp_build_interp_soa_context *interp,
struct lp_build_sampler_soa *sampler,
LLVMValueRef *pmask,
LLVMValueRef (*color)[4],
LLVMValueRef depth_ptr,
unsigned do_tri_test,
LLVMValueRef c0,
LLVMValueRef c1,
LLVMValueRef c2,
LLVMValueRef step0_ptr,
LLVMValueRef step1_ptr,
LLVMValueRef step2_ptr)
{
const struct tgsi_token *tokens = shader->base.tokens;
LLVMTypeRef elem_type;
LLVMTypeRef vec_type;
LLVMTypeRef int_vec_type;
LLVMValueRef consts_ptr;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
LLVMValueRef z = interp->pos[2];
struct lp_build_flow_context *flow;
struct lp_build_mask_context mask;
boolean early_depth_test;
unsigned attrib;
unsigned chan;
unsigned cbuf;
assert(i < 4);
elem_type = lp_build_elem_type(type);
vec_type = lp_build_vec_type(type);
int_vec_type = lp_build_int_vec_type(type);
consts_ptr = lp_jit_context_constants(builder, context_ptr);
flow = lp_build_flow_create(builder);
memset(outputs, 0, sizeof outputs);
lp_build_flow_scope_begin(flow);
/* Declare the color and z variables */
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
color[cbuf][chan] = LLVMGetUndef(vec_type);
lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
}
}
lp_build_flow_scope_declare(flow, &z);
/* do triangle edge testing */
if (do_tri_test) {
generate_tri_edge_mask(builder, i, pmask,
c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
}
else {
*pmask = build_int32_vec_const(~0);
}
/* 'mask' will control execution based on quad's pixel alive/killed state */
lp_build_mask_begin(&mask, flow, type, *pmask);
if (key->scissor) {
LLVMValueRef smask =
generate_scissor_test(builder, context_ptr, interp, type);
lp_build_mask_update(&mask, smask);
}
early_depth_test =
key->depth.enabled &&
!key->alpha.enabled &&
!shader->info.uses_kill &&
!shader->info.writes_z;
if(early_depth_test)
generate_depth(builder, key,
type, &mask,
z, depth_ptr);
lp_build_tgsi_soa(builder, tokens, type, &mask,
consts_ptr, interp->pos, interp->inputs,
outputs, sampler);
for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
if(outputs[attrib][chan]) {
LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
switch (shader->info.output_semantic_name[attrib]) {
case TGSI_SEMANTIC_COLOR:
{
unsigned cbuf = shader->info.output_semantic_index[attrib];
lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
/* Alpha test */
/* XXX: should the alpha reference value be passed separately? */
/* XXX: should only test the final assignment to alpha */
if(cbuf == 0 && chan == 3) {
LLVMValueRef alpha = out;
LLVMValueRef alpha_ref_value;
alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
lp_build_alpha_test(builder, &key->alpha, type,
&mask, alpha, alpha_ref_value);
}
color[cbuf][chan] = out;
break;
}
case TGSI_SEMANTIC_POSITION:
if(chan == 2)
z = out;
break;
}
}
}
}
if(!early_depth_test)
generate_depth(builder, key,
type, &mask,
z, depth_ptr);
lp_build_mask_end(&mask);
lp_build_flow_scope_end(flow);
lp_build_flow_destroy(flow);
*pmask = mask.value;
}
/**
* Generate code to compute texture level of detail (lambda).
* \param ddx partial derivatives of (s, t, r, q) with respect to X
* \param ddy partial derivatives of (s, t, r, q) with respect to Y
* \param lod_bias optional float vector with the shader lod bias
* \param explicit_lod optional float vector with the explicit lod
* \param width scalar int texture width
* \param height scalar int texture height
* \param depth scalar int texture depth
*
* XXX: The resulting lod is scalar, so ignore all but the first element of
* derivatives, lod_bias, etc that are passed by the shader.
*/
void
lp_build_lod_selector(struct lp_build_sample_context *bld,
unsigned unit,
const LLVMValueRef ddx[4],
const LLVMValueRef ddy[4],
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
unsigned mip_filter,
LLVMValueRef *out_lod_ipart,
LLVMValueRef *out_lod_fpart)
{
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_build_context *float_bld = &bld->float_bld;
LLVMValueRef lod;
*out_lod_ipart = bld->int_bld.zero;
*out_lod_fpart = bld->float_bld.zero;
if (bld->static_state->min_max_lod_equal) {
/* User is forcing sampling from a particular mipmap level.
* This is hit during mipmap generation.
*/
LLVMValueRef min_lod =
bld->dynamic_state->min_lod(bld->dynamic_state, bld->gallivm, unit);
lod = min_lod;
}
else {
LLVMValueRef sampler_lod_bias =
bld->dynamic_state->lod_bias(bld->dynamic_state, bld->gallivm, unit);
LLVMValueRef index0 = lp_build_const_int32(bld->gallivm, 0);
if (explicit_lod) {
lod = LLVMBuildExtractElement(builder, explicit_lod,
index0, "");
}
else {
LLVMValueRef rho;
rho = lp_build_rho(bld, unit, ddx, ddy);
/*
* Compute lod = log2(rho)
*/
if (!lod_bias &&
!bld->static_state->lod_bias_non_zero &&
!bld->static_state->apply_max_lod &&
!bld->static_state->apply_min_lod) {
/*
* Special case when there are no post-log2 adjustments, which
* saves instructions but keeping the integer and fractional lod
* computations separate from the start.
*/
if (mip_filter == PIPE_TEX_MIPFILTER_NONE ||
mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
*out_lod_ipart = lp_build_ilog2(float_bld, rho);
*out_lod_fpart = bld->float_bld.zero;
return;
}
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR &&
!(gallivm_debug & GALLIVM_DEBUG_NO_BRILINEAR)) {
lp_build_brilinear_rho(float_bld, rho, BRILINEAR_FACTOR,
out_lod_ipart, out_lod_fpart);
return;
}
}
if (0) {
lod = lp_build_log2(float_bld, rho);
}
else {
lod = lp_build_fast_log2(float_bld, rho);
}
/* add shader lod bias */
if (lod_bias) {
lod_bias = LLVMBuildExtractElement(builder, lod_bias,
index0, "");
lod = LLVMBuildFAdd(builder, lod, lod_bias, "shader_lod_bias");
}
}
/* add sampler lod bias */
if (bld->static_state->lod_bias_non_zero)
lod = LLVMBuildFAdd(builder, lod, sampler_lod_bias, "sampler_lod_bias");
/* clamp lod */
if (bld->static_state->apply_max_lod) {
LLVMValueRef max_lod =
bld->dynamic_state->max_lod(bld->dynamic_state, bld->gallivm, unit);
lod = lp_build_min(float_bld, lod, max_lod);
}
if (bld->static_state->apply_min_lod) {
LLVMValueRef min_lod =
bld->dynamic_state->min_lod(bld->dynamic_state, bld->gallivm, unit);
lod = lp_build_max(float_bld, lod, min_lod);
}
}
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
if (!(gallivm_debug & GALLIVM_DEBUG_NO_BRILINEAR)) {
lp_build_brilinear_lod(float_bld, lod, BRILINEAR_FACTOR,
out_lod_ipart, out_lod_fpart);
}
else {
lp_build_ifloor_fract(float_bld, lod, out_lod_ipart, out_lod_fpart);
}
lp_build_name(*out_lod_fpart, "lod_fpart");
}
else {
*out_lod_ipart = lp_build_iround(float_bld, lod);
}
lp_build_name(*out_lod_ipart, "lod_ipart");
return;
}
/**
* 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;
}
/**
* Register fetch.
*/
static LLVMValueRef
emit_fetch(
struct lp_build_tgsi_aos_context *bld,
const struct tgsi_full_instruction *inst,
unsigned src_op)
{
struct lp_type type = bld->base.type;
const struct tgsi_full_src_register *reg = &inst->Src[src_op];
LLVMValueRef res;
unsigned chan;
assert(!reg->Register.Indirect);
/*
* Fetch the from the register file.
*/
switch (reg->Register.File) {
case TGSI_FILE_CONSTANT:
/*
* Get the constants components
*/
res = bld->base.undef;
for (chan = 0; chan < 4; ++chan) {
LLVMValueRef index;
LLVMValueRef scalar_ptr;
LLVMValueRef scalar;
LLVMValueRef swizzle;
index = LLVMConstInt(LLVMInt32Type(),
reg->Register.Index*4 + chan,
0);
scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr,
&index, 1, "");
scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
lp_build_name(scalar, "const[%u].%c", reg->Register.Index, "xyzw"[chan]);
/*
* NOTE: constants array is always assumed to be RGBA
*/
swizzle = LLVMConstInt(LLVMInt32Type(), chan, 0);
res = LLVMBuildInsertElement(bld->base.builder, res, scalar, swizzle, "");
}
/*
* Broadcast the first quaternion to all others.
*
* XXX: could be factored into a reusable function.
*/
if (type.length > 4) {
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
unsigned i;
for (chan = 0; chan < 4; ++chan) {
shuffles[chan] = LLVMConstInt(LLVMInt32Type(), chan, 0);
}
for (i = 4; i < type.length; ++i) {
shuffles[i] = shuffles[i % 4];
}
res = LLVMBuildShuffleVector(bld->base.builder,
res, bld->base.undef,
LLVMConstVector(shuffles, type.length),
"");
}
break;
case TGSI_FILE_IMMEDIATE:
res = bld->immediates[reg->Register.Index];
assert(res);
break;
case TGSI_FILE_INPUT:
res = bld->inputs[reg->Register.Index];
assert(res);
break;
case TGSI_FILE_TEMPORARY:
{
LLVMValueRef temp_ptr;
temp_ptr = bld->temps[reg->Register.Index];
res = LLVMBuildLoad(bld->base.builder, temp_ptr, "");
if (!res)
return bld->base.undef;
}
break;
default:
assert(0 && "invalid src register in emit_fetch()");
return bld->base.undef;
}
/*
* Apply sign modifier.
*/
if (reg->Register.Absolute) {
res = lp_build_abs(&bld->base, res);
}
if(reg->Register.Negate) {
res = lp_build_negate(&bld->base, res);
}
/*
* Swizzle the argument
*/
res = swizzle_aos(bld, res,
reg->Register.SwizzleX,
reg->Register.SwizzleY,
reg->Register.SwizzleZ,
reg->Register.SwizzleW);
return res;
}