static void
calc_offsets(struct lp_build_context *coeff_bld,
unsigned quad_start_index,
LLVMValueRef *pixoffx,
LLVMValueRef *pixoffy)
{
unsigned i;
unsigned num_pix = coeff_bld->type.length;
struct gallivm_state *gallivm = coeff_bld->gallivm;
LLVMBuilderRef builder = coeff_bld->gallivm->builder;
LLVMValueRef nr, pixxf, pixyf;
*pixoffx = coeff_bld->undef;
*pixoffy = coeff_bld->undef;
for (i = 0; i < num_pix; i++) {
nr = lp_build_const_int32(gallivm, i);
pixxf = lp_build_const_float(gallivm, quad_offset_x[i % num_pix] +
(quad_start_index & 1) * 2);
pixyf = lp_build_const_float(gallivm, quad_offset_y[i % num_pix] +
(quad_start_index & 2));
*pixoffx = LLVMBuildInsertElement(builder, *pixoffx, pixxf, nr, "");
*pixoffy = LLVMBuildInsertElement(builder, *pixoffy, pixyf, nr, "");
}
}
static void kill_if_fetch_args(struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
unsigned i;
LLVMValueRef conds[TGSI_NUM_CHANNELS];
for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
LLVMValueRef value = lp_build_emit_fetch(bld_base, inst, 0, i);
conds[i] = LLVMBuildFCmp(builder, LLVMRealOLT, value,
bld_base->base.zero, "");
}
/* Or the conditions together */
for (i = TGSI_NUM_CHANNELS - 1; i > 0; i--) {
conds[i - 1] = LLVMBuildOr(builder, conds[i], conds[i - 1], "");
}
emit_data->dst_type = LLVMVoidTypeInContext(gallivm->context);
emit_data->arg_count = 1;
emit_data->args[0] = LLVMBuildSelect(builder, conds[0],
lp_build_const_float(gallivm, -1.0f),
bld_base->base.zero, "");
}
static LLVMValueRef
llvm_face_select_helper(
struct radeon_llvm_context * ctx,
const char *intrinsic, unsigned face_register,
unsigned frontcolor_register, unsigned backcolor_regiser)
{
LLVMValueRef backcolor = llvm_load_input_helper(
ctx,
intrinsic,
backcolor_regiser);
LLVMValueRef front_color = llvm_load_input_helper(
ctx,
intrinsic,
frontcolor_register);
LLVMValueRef face = llvm_load_input_helper(
ctx,
"llvm.R600.load.input",
face_register);
LLVMValueRef is_face_positive = LLVMBuildFCmp(
ctx->soa.bld_base.base.gallivm->builder,
LLVMRealUGT, face,
lp_build_const_float(ctx->soa.bld_base.base.gallivm, 0.0f),
"");
return LLVMBuildSelect(
ctx->soa.bld_base.base.gallivm->builder,
is_face_positive,
front_color,
backcolor,
"");
}
static LLVMValueRef
llvm_face_select_helper(
struct radeon_llvm_context * ctx,
LLVMValueRef face, LLVMValueRef front_color, LLVMValueRef back_color)
{
const struct lp_build_context * bb = &ctx->soa.bld_base.base;
LLVMValueRef is_front = LLVMBuildFCmp(
bb->gallivm->builder, LLVMRealUGT, face,
lp_build_const_float(bb->gallivm, 0.0f), "");
return LLVMBuildSelect(bb->gallivm->builder, is_front,
front_color, back_color, "");
}
static void emit_cndlt(
const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
LLVMValueRef float_zero = lp_build_const_float(
bld_base->base.gallivm, 0.0f);
LLVMValueRef cmp = LLVMBuildFCmp(
builder, LLVMRealULT, emit_data->args[0], float_zero, "");
emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
cmp, emit_data->args[1], emit_data->args[2], "");
}
LLVMValueRef
lp_build_zero(struct gallivm_state *gallivm, struct lp_type type)
{
if (type.length == 1) {
if (type.floating)
return lp_build_const_float(gallivm, 0.0);
else
return LLVMConstInt(LLVMIntTypeInContext(gallivm->context, type.width), 0, 0);
}
else {
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
return LLVMConstNull(vec_type);
}
}
static void llvm_load_input(
struct radeon_llvm_context * ctx,
unsigned input_index,
const struct tgsi_full_declaration *decl)
{
const struct r600_shader_io * input = &ctx->r600_inputs[input_index];
unsigned chan;
int two_side = (ctx->two_side && input->name == TGSI_SEMANTIC_COLOR);
LLVMValueRef v;
boolean require_interp_intrinsic = ctx->chip_class >= EVERGREEN &&
ctx->type == TGSI_PROCESSOR_FRAGMENT;
if (require_interp_intrinsic && input->spi_sid) {
v = llvm_load_input_vector(ctx, input->lds_pos, input->ij_index,
(input->interpolate > 0));
} else
v = LLVMGetParam(ctx->main_fn, input->gpr);
if (two_side) {
struct r600_shader_io * back_input =
&ctx->r600_inputs[input->back_color_input];
LLVMValueRef v2;
LLVMValueRef face = LLVMGetParam(ctx->main_fn, ctx->face_gpr);
face = LLVMBuildExtractElement(ctx->gallivm.builder, face,
lp_build_const_int32(&(ctx->gallivm), 0), "");
if (require_interp_intrinsic && back_input->spi_sid)
v2 = llvm_load_input_vector(ctx, back_input->lds_pos,
back_input->ij_index, (back_input->interpolate > 0));
else
v2 = LLVMGetParam(ctx->main_fn, back_input->gpr);
v = llvm_face_select_helper(ctx, face, v, v2);
}
for (chan = 0; chan < 4; chan++) {
unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
ctx->inputs[soa_index] = LLVMBuildExtractElement(ctx->gallivm.builder, v,
lp_build_const_int32(&(ctx->gallivm), chan), "");
if (input->name == TGSI_SEMANTIC_POSITION &&
ctx->type == TGSI_PROCESSOR_FRAGMENT && chan == 3) {
/* RCP for fragcoord.w */
ctx->inputs[soa_index] = LLVMBuildFDiv(ctx->gallivm.builder,
lp_build_const_float(&(ctx->gallivm), 1.0f),
ctx->inputs[soa_index], "");
}
}
}
/**
* Pack a single pixel.
*
* @param rgba 4 float vector with the unpacked components.
*
* XXX: This is mostly for reference and testing -- operating a single pixel at
* a time is rarely if ever needed.
*/
LLVMValueRef
lp_build_pack_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *desc,
LLVMValueRef rgba)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef type;
LLVMValueRef packed = NULL;
LLVMValueRef swizzles[4];
LLVMValueRef shifted, casted, scaled, unswizzled;
LLVMValueRef shifts[4];
LLVMValueRef scales[4];
boolean normalized;
unsigned shift;
unsigned i, j;
assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
assert(desc->block.width == 1);
assert(desc->block.height == 1);
type = LLVMIntTypeInContext(gallivm->context, desc->block.bits);
/* Unswizzle the color components into the source vector. */
for (i = 0; i < 4; ++i) {
for (j = 0; j < 4; ++j) {
if (desc->swizzle[j] == i)
break;
}
if (j < 4)
swizzles[i] = lp_build_const_int32(gallivm, j);
else
swizzles[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
}
unswizzled = LLVMBuildShuffleVector(builder, rgba,
LLVMGetUndef(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4)),
LLVMConstVector(swizzles, 4), "");
normalized = FALSE;
shift = 0;
for (i = 0; i < 4; ++i) {
unsigned bits = desc->channel[i].size;
if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
scales[i] = LLVMGetUndef(LLVMFloatTypeInContext(gallivm->context));
}
else {
unsigned mask = (1 << bits) - 1;
assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
assert(bits < 32);
shifts[i] = lp_build_const_int32(gallivm, shift);
if (desc->channel[i].normalized) {
scales[i] = lp_build_const_float(gallivm, mask);
normalized = TRUE;
}
else
scales[i] = lp_build_const_float(gallivm, 1.0);
}
shift += bits;
}
if (normalized)
scaled = LLVMBuildFMul(builder, unswizzled, LLVMConstVector(scales, 4), "");
else
scaled = unswizzled;
casted = LLVMBuildFPToSI(builder, scaled, LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), "");
shifted = LLVMBuildShl(builder, casted, LLVMConstVector(shifts, 4), "");
/* Bitwise or all components */
for (i = 0; i < 4; ++i) {
if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
LLVMValueRef component = LLVMBuildExtractElement(builder, shifted,
lp_build_const_int32(gallivm, i), "");
if (packed)
packed = LLVMBuildOr(builder, packed, component, "");
else
packed = component;
}
}
if (!packed)
packed = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
if (desc->block.bits < 32)
packed = LLVMBuildTrunc(builder, packed, type, "");
return packed;
}
/**
* Unpack a single pixel into its RGBA components.
*
* @param desc the pixel format for the packed pixel value
* @param packed integer pixel in a format such as PIPE_FORMAT_B8G8R8A8_UNORM
*
* @return RGBA in a float[4] or ubyte[4] or ushort[4] vector.
*/
static INLINE LLVMValueRef
lp_build_unpack_arith_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *desc,
LLVMValueRef packed)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef shifted, casted, scaled, masked;
LLVMValueRef shifts[4];
LLVMValueRef masks[4];
LLVMValueRef scales[4];
boolean normalized;
boolean needs_uitofp;
unsigned shift;
unsigned i;
/* TODO: Support more formats */
assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
assert(desc->block.width == 1);
assert(desc->block.height == 1);
assert(desc->block.bits <= 32);
/* Do the intermediate integer computations with 32bit integers since it
* matches floating point size */
assert (LLVMTypeOf(packed) == LLVMInt32TypeInContext(gallivm->context));
/* Broadcast the packed value to all four channels
* before: packed = BGRA
* after: packed = {BGRA, BGRA, BGRA, BGRA}
*/
packed = LLVMBuildInsertElement(builder,
LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
packed,
LLVMConstNull(LLVMInt32TypeInContext(gallivm->context)),
"");
packed = LLVMBuildShuffleVector(builder,
packed,
LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
LLVMConstNull(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
"");
/* Initialize vector constants */
normalized = FALSE;
needs_uitofp = FALSE;
shift = 0;
/* Loop over 4 color components */
for (i = 0; i < 4; ++i) {
unsigned bits = desc->channel[i].size;
if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
masks[i] = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
scales[i] = LLVMConstNull(LLVMFloatTypeInContext(gallivm->context));
}
else {
unsigned long long mask = (1ULL << bits) - 1;
assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
if (bits == 32) {
needs_uitofp = TRUE;
}
shifts[i] = lp_build_const_int32(gallivm, shift);
masks[i] = lp_build_const_int32(gallivm, mask);
if (desc->channel[i].normalized) {
scales[i] = lp_build_const_float(gallivm, 1.0 / mask);
normalized = TRUE;
}
else
scales[i] = lp_build_const_float(gallivm, 1.0);
}
shift += bits;
}
/* Ex: convert packed = {BGRA, BGRA, BGRA, BGRA}
* into masked = {B, G, R, A}
*/
shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), "");
masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), "");
if (!needs_uitofp) {
/* UIToFP can't be expressed in SSE2 */
casted = LLVMBuildSIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
} else {
casted = LLVMBuildUIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
}
/* At this point 'casted' may be a vector of floats such as
* {255.0, 255.0, 255.0, 255.0}. Next, if the pixel values are normalized
* we'll scale this to {1.0, 1.0, 1.0, 1.0}.
*/
if (normalized)
scaled = LLVMBuildFMul(builder, casted, LLVMConstVector(scales, 4), "");
else
scaled = casted;
return scaled;
}
static void llvm_emit_epilogue(struct lp_build_tgsi_context * bld_base)
{
struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
struct lp_build_context * base = &bld_base->base;
struct pipe_stream_output_info * so = ctx->stream_outputs;
unsigned i;
unsigned next_pos = 60;
unsigned next_param = 0;
unsigned color_count = 0;
boolean has_color = false;
if (ctx->type == TGSI_PROCESSOR_VERTEX && so->num_outputs) {
for (i = 0; i < so->num_outputs; i++) {
unsigned register_index = so->output[i].register_index;
unsigned start_component = so->output[i].start_component;
unsigned num_components = so->output[i].num_components;
unsigned dst_offset = so->output[i].dst_offset;
unsigned chan;
LLVMValueRef elements[4];
if (dst_offset < start_component) {
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
elements[chan] = LLVMBuildLoad(base->gallivm->builder,
ctx->soa.outputs[register_index][(chan + start_component) % TGSI_NUM_CHANNELS], "");
}
start_component = 0;
} else {
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
elements[chan] = LLVMBuildLoad(base->gallivm->builder,
ctx->soa.outputs[register_index][chan], "");
}
}
LLVMValueRef output = lp_build_gather_values(base->gallivm, elements, 4);
LLVMValueRef args[4];
args[0] = output;
args[1] = lp_build_const_int32(base->gallivm, dst_offset - start_component);
args[2] = lp_build_const_int32(base->gallivm, so->output[i].output_buffer);
args[3] = lp_build_const_int32(base->gallivm, ((1 << num_components) - 1) << start_component);
lp_build_intrinsic(base->gallivm->builder, "llvm.R600.store.stream.output",
LLVMVoidTypeInContext(base->gallivm->context), args, 4);
}
}
/* Add the necessary export instructions */
for (i = 0; i < ctx->output_reg_count; i++) {
unsigned chan;
LLVMValueRef elements[4];
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
elements[chan] = LLVMBuildLoad(base->gallivm->builder,
ctx->soa.outputs[i][chan], "");
}
if (ctx->alpha_to_one && ctx->type == TGSI_PROCESSOR_FRAGMENT && ctx->r600_outputs[i].name == TGSI_SEMANTIC_COLOR)
elements[3] = lp_build_const_float(base->gallivm, 1.0f);
LLVMValueRef output = lp_build_gather_values(base->gallivm, elements, 4);
if (ctx->type == TGSI_PROCESSOR_VERTEX) {
switch (ctx->r600_outputs[i].name) {
case TGSI_SEMANTIC_POSITION:
case TGSI_SEMANTIC_PSIZE: {
LLVMValueRef args[3];
args[0] = output;
args[1] = lp_build_const_int32(base->gallivm, next_pos++);
args[2] = lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_POS);
build_intrinsic(
base->gallivm->builder,
"llvm.R600.store.swizzle",
LLVMVoidTypeInContext(base->gallivm->context),
args, 3, 0);
break;
}
case TGSI_SEMANTIC_CLIPVERTEX: {
LLVMValueRef args[3];
unsigned reg_index;
unsigned base_vector_chan;
LLVMValueRef adjusted_elements[4];
for (reg_index = 0; reg_index < 2; reg_index ++) {
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
LLVMValueRef offset = lp_build_const_int32(bld_base->base.gallivm, reg_index * 4 + chan);
LLVMValueRef base_vector = llvm_load_const_buffer(bld_base, offset, CONSTANT_BUFFER_1_ADDR_SPACE);
args[0] = output;
args[1] = base_vector;
adjusted_elements[chan] = build_intrinsic(base->gallivm->builder,
"llvm.AMDGPU.dp4", bld_base->base.elem_type,
args, 2, LLVMReadNoneAttribute);
}
args[0] = lp_build_gather_values(base->gallivm,
adjusted_elements, 4);
args[1] = lp_build_const_int32(base->gallivm, next_pos++);
args[2] = lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_POS);
build_intrinsic(
base->gallivm->builder,
"llvm.R600.store.swizzle",
LLVMVoidTypeInContext(base->gallivm->context),
args, 3, 0);
}
break;
}
case TGSI_SEMANTIC_CLIPDIST : {
LLVMValueRef args[3];
args[0] = output;
args[1] = lp_build_const_int32(base->gallivm, next_pos++);
args[2] = lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_POS);
build_intrinsic(
base->gallivm->builder,
"llvm.R600.store.swizzle",
LLVMVoidTypeInContext(base->gallivm->context),
args, 3, 0);
args[1] = lp_build_const_int32(base->gallivm, next_param++);
args[2] = lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PARAM);
build_intrinsic(
base->gallivm->builder,
"llvm.R600.store.swizzle",
LLVMVoidTypeInContext(base->gallivm->context),
args, 3, 0);
break;
}
case TGSI_SEMANTIC_FOG: {
elements[0] = LLVMBuildLoad(base->gallivm->builder,
ctx->soa.outputs[i][0], "");
elements[1] = elements[2] = lp_build_const_float(base->gallivm, 0.0f);
elements[3] = lp_build_const_float(base->gallivm, 1.0f);
LLVMValueRef args[3];
args[0] = lp_build_gather_values(base->gallivm, elements, 4);
args[1] = lp_build_const_int32(base->gallivm, next_param++);
args[2] = lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PARAM);
build_intrinsic(
base->gallivm->builder,
"llvm.R600.store.swizzle",
LLVMVoidTypeInContext(base->gallivm->context),
args, 3, 0);
break;
}
default: {
LLVMValueRef args[3];
args[0] = output;
args[1] = lp_build_const_int32(base->gallivm, next_param++);
args[2] = lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PARAM);
build_intrinsic(
base->gallivm->builder,
"llvm.R600.store.swizzle",
LLVMVoidTypeInContext(base->gallivm->context),
args, 3, 0);
break;
}
}
} else if (ctx->type == TGSI_PROCESSOR_FRAGMENT) {
switch (ctx->r600_outputs[i].name) {
case TGSI_SEMANTIC_COLOR:
has_color = true;
if ( color_count < ctx->color_buffer_count) {
LLVMValueRef args[3];
args[0] = output;
if (ctx->fs_color_all) {
for (unsigned j = 0; j < ctx->color_buffer_count; j++) {
args[1] = lp_build_const_int32(base->gallivm, j);
args[2] = lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL);
build_intrinsic(
base->gallivm->builder,
"llvm.R600.store.swizzle",
LLVMVoidTypeInContext(base->gallivm->context),
args, 3, 0);
}
} else {
args[1] = lp_build_const_int32(base->gallivm, color_count++);
args[2] = lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL);
build_intrinsic(
base->gallivm->builder,
"llvm.R600.store.swizzle",
LLVMVoidTypeInContext(base->gallivm->context),
args, 3, 0);
}
}
break;
case TGSI_SEMANTIC_POSITION:
lp_build_intrinsic_unary(
base->gallivm->builder,
"llvm.R600.store.pixel.depth",
LLVMVoidTypeInContext(base->gallivm->context),
LLVMBuildLoad(base->gallivm->builder, ctx->soa.outputs[i][2], ""));
break;
case TGSI_SEMANTIC_STENCIL:
lp_build_intrinsic_unary(
base->gallivm->builder,
"llvm.R600.store.pixel.stencil",
LLVMVoidTypeInContext(base->gallivm->context),
LLVMBuildLoad(base->gallivm->builder, ctx->soa.outputs[i][1], ""));
break;
}
}
}
// Add dummy exports
if (ctx->type == TGSI_PROCESSOR_VERTEX) {
if (!next_param) {
lp_build_intrinsic_unary(base->gallivm->builder, "llvm.R600.store.dummy",
LLVMVoidTypeInContext(base->gallivm->context),
lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PARAM));
}
if (!(next_pos-60)) {
lp_build_intrinsic_unary(base->gallivm->builder, "llvm.R600.store.dummy",
LLVMVoidTypeInContext(base->gallivm->context),
lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_POS));
}
}
if (ctx->type == TGSI_PROCESSOR_FRAGMENT) {
if (!has_color) {
lp_build_intrinsic_unary(base->gallivm->builder, "llvm.R600.store.dummy",
LLVMVoidTypeInContext(base->gallivm->context),
lp_build_const_int32(base->gallivm, V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL));
}
}
}
/**
* Generate code to do cube face selection and compute per-face texcoords.
*/
void
lp_build_cube_lookup(struct lp_build_sample_context *bld,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
LLVMValueRef *face,
LLVMValueRef *face_s,
LLVMValueRef *face_t)
{
struct lp_build_context *float_bld = &bld->float_bld;
struct lp_build_context *coord_bld = &bld->coord_bld;
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef rx, ry, rz;
LLVMValueRef arx, ary, arz;
LLVMValueRef c25 = lp_build_const_float(bld->gallivm, 0.25);
LLVMValueRef arx_ge_ary, arx_ge_arz;
LLVMValueRef ary_ge_arx, ary_ge_arz;
LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
assert(bld->coord_bld.type.length == 4);
/*
* Use the average of the four pixel's texcoords to choose the face.
*/
rx = lp_build_mul(float_bld, c25,
lp_build_sum_vector(&bld->coord_bld, s));
ry = lp_build_mul(float_bld, c25,
lp_build_sum_vector(&bld->coord_bld, t));
rz = lp_build_mul(float_bld, c25,
lp_build_sum_vector(&bld->coord_bld, r));
arx = lp_build_abs(float_bld, rx);
ary = lp_build_abs(float_bld, ry);
arz = lp_build_abs(float_bld, rz);
/*
* Compare sign/magnitude of rx,ry,rz to determine face
*/
arx_ge_ary = LLVMBuildFCmp(builder, LLVMRealUGE, arx, ary, "");
arx_ge_arz = LLVMBuildFCmp(builder, LLVMRealUGE, arx, arz, "");
ary_ge_arx = LLVMBuildFCmp(builder, LLVMRealUGE, ary, arx, "");
ary_ge_arz = LLVMBuildFCmp(builder, LLVMRealUGE, ary, arz, "");
arx_ge_ary_arz = LLVMBuildAnd(builder, arx_ge_ary, arx_ge_arz, "");
ary_ge_arx_arz = LLVMBuildAnd(builder, ary_ge_arx, ary_ge_arz, "");
{
struct lp_build_if_state if_ctx;
LLVMValueRef face_s_var;
LLVMValueRef face_t_var;
LLVMValueRef face_var;
face_s_var = lp_build_alloca(bld->gallivm, bld->coord_bld.vec_type, "face_s_var");
face_t_var = lp_build_alloca(bld->gallivm, bld->coord_bld.vec_type, "face_t_var");
face_var = lp_build_alloca(bld->gallivm, bld->int_bld.vec_type, "face_var");
lp_build_if(&if_ctx, bld->gallivm, arx_ge_ary_arz);
{
/* +/- X face */
LLVMValueRef sign = lp_build_sgn(float_bld, rx);
LLVMValueRef ima = lp_build_cube_ima(coord_bld, s);
*face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
*face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
*face = lp_build_cube_face(bld, rx,
PIPE_TEX_FACE_POS_X,
PIPE_TEX_FACE_NEG_X);
LLVMBuildStore(builder, *face_s, face_s_var);
LLVMBuildStore(builder, *face_t, face_t_var);
LLVMBuildStore(builder, *face, face_var);
}
lp_build_else(&if_ctx);
{
struct lp_build_if_state if_ctx2;
lp_build_if(&if_ctx2, bld->gallivm, ary_ge_arx_arz);
{
/* +/- Y face */
LLVMValueRef sign = lp_build_sgn(float_bld, ry);
LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
*face_s = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
*face_t = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
*face = lp_build_cube_face(bld, ry,
PIPE_TEX_FACE_POS_Y,
PIPE_TEX_FACE_NEG_Y);
LLVMBuildStore(builder, *face_s, face_s_var);
LLVMBuildStore(builder, *face_t, face_t_var);
LLVMBuildStore(builder, *face, face_var);
}
lp_build_else(&if_ctx2);
{
/* +/- Z face */
LLVMValueRef sign = lp_build_sgn(float_bld, rz);
LLVMValueRef ima = lp_build_cube_ima(coord_bld, r);
*face_s = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
*face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
*face = lp_build_cube_face(bld, rz,
PIPE_TEX_FACE_POS_Z,
PIPE_TEX_FACE_NEG_Z);
LLVMBuildStore(builder, *face_s, face_s_var);
LLVMBuildStore(builder, *face_t, face_t_var);
LLVMBuildStore(builder, *face, face_var);
}
lp_build_endif(&if_ctx2);
}
lp_build_endif(&if_ctx);
*face_s = LLVMBuildLoad(builder, face_s_var, "face_s");
*face_t = LLVMBuildLoad(builder, face_t_var, "face_t");
*face = LLVMBuildLoad(builder, face_var, "face");
}
}
/**
* 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;
}
}
}
}
void si_prepare_cube_coords(struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data,
LLVMValueRef *coords_arg,
LLVMValueRef *derivs_arg)
{
unsigned target = emit_data->inst->Texture.Texture;
unsigned opcode = emit_data->inst->Instruction.Opcode;
struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef coords[4];
unsigned i;
si_llvm_cube_to_2d_coords(bld_base, coords_arg, coords);
if (opcode == TGSI_OPCODE_TXD && derivs_arg) {
LLVMValueRef derivs[4];
int axis;
/* Convert cube derivatives to 2D derivatives. */
for (axis = 0; axis < 2; axis++) {
LLVMValueRef shifted_cube_coords[4], shifted_coords[4];
/* Shift the cube coordinates by the derivatives to get
* the cube coordinates of the "neighboring pixel".
*/
for (i = 0; i < 3; i++)
shifted_cube_coords[i] =
LLVMBuildFAdd(builder, coords_arg[i],
derivs_arg[axis*3+i], "");
shifted_cube_coords[3] = LLVMGetUndef(bld_base->base.elem_type);
/* Project the shifted cube coordinates onto the face. */
si_llvm_cube_to_2d_coords(bld_base, shifted_cube_coords,
shifted_coords);
/* Subtract both sets of 2D coordinates to get 2D derivatives.
* This won't work if the shifted coordinates ended up
* in a different face.
*/
for (i = 0; i < 2; i++)
derivs[axis * 2 + i] =
LLVMBuildFSub(builder, shifted_coords[i],
coords[i], "");
}
memcpy(derivs_arg, derivs, sizeof(derivs));
}
if (target == TGSI_TEXTURE_CUBE_ARRAY ||
target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
/* for cube arrays coord.z = coord.w(array_index) * 8 + face */
/* coords_arg.w component - array_index for cube arrays */
coords[2] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
coords_arg[3], lp_build_const_float(gallivm, 8.0), coords[2]);
}
/* Preserve compare/lod/bias. Put it in coords.w. */
if (opcode == TGSI_OPCODE_TEX2 ||
opcode == TGSI_OPCODE_TXB2 ||
opcode == TGSI_OPCODE_TXL2) {
coords[3] = coords_arg[4];
} else if (opcode == TGSI_OPCODE_TXB ||
opcode == TGSI_OPCODE_TXL ||
target == TGSI_TEXTURE_SHADOWCUBE) {
coords[3] = coords_arg[3];
}
memcpy(coords_arg, coords, sizeof(coords));
}
/**
* 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);
}
}
static void declare_input_fs(
struct si_shader_context * si_shader_ctx,
unsigned input_index,
const struct tgsi_full_declaration *decl)
{
struct si_shader *shader = &si_shader_ctx->shader->shader;
struct lp_build_context * base =
&si_shader_ctx->radeon_bld.soa.bld_base.base;
struct gallivm_state * gallivm = base->gallivm;
LLVMTypeRef input_type = LLVMFloatTypeInContext(gallivm->context);
LLVMValueRef main_fn = si_shader_ctx->radeon_bld.main_fn;
LLVMValueRef interp_param;
const char * intr_name;
/* This value is:
* [15:0] NewPrimMask (Bit mask for each quad. It is set it the
* quad begins a new primitive. Bit 0 always needs
* to be unset)
* [32:16] ParamOffset
*
*/
LLVMValueRef params = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_PRIM_MASK);
LLVMValueRef attr_number;
unsigned chan;
if (decl->Semantic.Name == TGSI_SEMANTIC_POSITION) {
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
unsigned soa_index =
radeon_llvm_reg_index_soa(input_index, chan);
si_shader_ctx->radeon_bld.inputs[soa_index] =
LLVMGetParam(main_fn, SI_PARAM_POS_X_FLOAT + chan);
if (chan == 3)
/* RCP for fragcoord.w */
si_shader_ctx->radeon_bld.inputs[soa_index] =
LLVMBuildFDiv(gallivm->builder,
lp_build_const_float(gallivm, 1.0f),
si_shader_ctx->radeon_bld.inputs[soa_index],
"");
}
return;
}
if (decl->Semantic.Name == TGSI_SEMANTIC_FACE) {
LLVMValueRef face, is_face_positive;
face = LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE);
is_face_positive = LLVMBuildFCmp(gallivm->builder,
LLVMRealUGT, face,
lp_build_const_float(gallivm, 0.0f),
"");
si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 0)] =
LLVMBuildSelect(gallivm->builder,
is_face_positive,
lp_build_const_float(gallivm, 1.0f),
lp_build_const_float(gallivm, 0.0f),
"");
si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 1)] =
si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 2)] =
lp_build_const_float(gallivm, 0.0f);
si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 3)] =
lp_build_const_float(gallivm, 1.0f);
return;
}
shader->input[input_index].param_offset = shader->ninterp++;
attr_number = lp_build_const_int32(gallivm,
shader->input[input_index].param_offset);
/* XXX: Handle all possible interpolation modes */
switch (decl->Interp.Interpolate) {
case TGSI_INTERPOLATE_COLOR:
if (si_shader_ctx->shader->key.ps.flatshade) {
interp_param = 0;
} else {
if (decl->Interp.Centroid)
interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTROID);
else
interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTER);
}
break;
case TGSI_INTERPOLATE_CONSTANT:
interp_param = 0;
break;
case TGSI_INTERPOLATE_LINEAR:
if (decl->Interp.Centroid)
interp_param = LLVMGetParam(main_fn, SI_PARAM_LINEAR_CENTROID);
else
interp_param = LLVMGetParam(main_fn, SI_PARAM_LINEAR_CENTER);
break;
case TGSI_INTERPOLATE_PERSPECTIVE:
if (decl->Interp.Centroid)
interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTROID);
else
interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTER);
break;
default:
fprintf(stderr, "Warning: Unhandled interpolation mode.\n");
return;
}
intr_name = interp_param ? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
/* XXX: Could there be more than TGSI_NUM_CHANNELS (4) ? */
if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR &&
si_shader_ctx->shader->key.ps.color_two_side) {
LLVMValueRef args[4];
LLVMValueRef face, is_face_positive;
LLVMValueRef back_attr_number =
lp_build_const_int32(gallivm,
shader->input[input_index].param_offset + 1);
face = LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE);
is_face_positive = LLVMBuildFCmp(gallivm->builder,
LLVMRealUGT, face,
lp_build_const_float(gallivm, 0.0f),
"");
args[2] = params;
args[3] = interp_param;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
LLVMValueRef front, back;
args[0] = llvm_chan;
args[1] = attr_number;
front = build_intrinsic(base->gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
args[1] = back_attr_number;
back = build_intrinsic(base->gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
si_shader_ctx->radeon_bld.inputs[soa_index] =
LLVMBuildSelect(gallivm->builder,
is_face_positive,
front,
back,
"");
}
shader->ninterp++;
} else {
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
LLVMValueRef args[4];
LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
args[0] = llvm_chan;
args[1] = attr_number;
args[2] = params;
args[3] = interp_param;
si_shader_ctx->radeon_bld.inputs[soa_index] =
build_intrinsic(base->gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
}
}
}
static void llvm_load_input(
struct radeon_llvm_context * ctx,
unsigned input_index,
const struct tgsi_full_declaration *decl)
{
unsigned chan;
const char *intrinsics = "llvm.R600.load.input";
unsigned offset = 4 * ctx->reserved_reg_count;
if (ctx->type == TGSI_PROCESSOR_FRAGMENT && ctx->chip_class >= EVERGREEN) {
switch (decl->Interp.Interpolate) {
case TGSI_INTERPOLATE_COLOR:
case TGSI_INTERPOLATE_PERSPECTIVE:
offset = 0;
intrinsics = "llvm.R600.load.input.perspective";
break;
case TGSI_INTERPOLATE_LINEAR:
offset = 0;
intrinsics = "llvm.R600.load.input.linear";
break;
case TGSI_INTERPOLATE_CONSTANT:
offset = 0;
intrinsics = "llvm.R600.load.input.constant";
break;
default:
assert(0 && "Unknow Interpolate mode");
}
}
for (chan = 0; chan < 4; chan++) {
unsigned soa_index = radeon_llvm_reg_index_soa(input_index,
chan);
switch (decl->Semantic.Name) {
case TGSI_SEMANTIC_FACE:
ctx->inputs[soa_index] = llvm_load_input_helper(ctx,
"llvm.R600.load.input",
4 * ctx->face_input);
break;
case TGSI_SEMANTIC_POSITION:
if (ctx->type != TGSI_PROCESSOR_FRAGMENT || chan != 3) {
ctx->inputs[soa_index] = llvm_load_input_helper(ctx,
"llvm.R600.load.input",
soa_index + (ctx->reserved_reg_count * 4));
} else {
LLVMValueRef w_coord = llvm_load_input_helper(ctx,
"llvm.R600.load.input",
soa_index + (ctx->reserved_reg_count * 4));
ctx->inputs[soa_index] = LLVMBuildFDiv(ctx->gallivm.builder,
lp_build_const_float(&(ctx->gallivm), 1.0f), w_coord, "");
}
break;
case TGSI_SEMANTIC_COLOR:
if (ctx->two_side) {
unsigned front_location, back_location;
unsigned back_reg = ctx->r600_inputs[input_index]
.potential_back_facing_reg;
if (ctx->chip_class >= EVERGREEN) {
front_location = 4 * ctx->r600_inputs[input_index].lds_pos + chan;
back_location = 4 * ctx->r600_inputs[back_reg].lds_pos + chan;
} else {
front_location = soa_index + 4 * ctx->reserved_reg_count;
back_location = radeon_llvm_reg_index_soa(
ctx->r600_inputs[back_reg].gpr,
chan);
}
ctx->inputs[soa_index] = llvm_face_select_helper(ctx,
intrinsics,
4 * ctx->face_input, front_location, back_location);
break;
}
default:
{
unsigned location;
if (ctx->chip_class >= EVERGREEN) {
location = 4 * ctx->r600_inputs[input_index].lds_pos + chan;
} else {
location = soa_index + 4 * ctx->reserved_reg_count;
}
/* The * 4 is assuming that we are in soa mode. */
ctx->inputs[soa_index] = llvm_load_input_helper(ctx,
intrinsics, location);
break;
}
}
}
}