*/
#include "pipe/p_state.h"
-#include "util/u_format.h"
+#include "util/format/u_format.h"
#include "util/u_cpu_detect.h"
#include "gallivm/lp_bld_type.h"
if (format_desc->channel[z_swizzle].type == UTIL_FORMAT_TYPE_FLOAT) {
type.floating = TRUE;
assert(z_swizzle == 0);
- assert(format_desc->channel[z_swizzle].size == format_desc->block.bits);
+ assert(format_desc->channel[z_swizzle].size == 32);
}
else if(format_desc->channel[z_swizzle].type == UTIL_FORMAT_TYPE_UNSIGNED) {
assert(format_desc->block.bits <= 32);
get_z_shift_and_mask(const struct util_format_description *format_desc,
unsigned *shift, unsigned *width, unsigned *mask)
{
- const unsigned total_bits = format_desc->block.bits;
+ unsigned total_bits;
unsigned z_swizzle;
- unsigned chan;
- unsigned padding_left, padding_right;
-
+
assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS);
assert(format_desc->block.width == 1);
assert(format_desc->block.height == 1);
+ /* 64bit d/s format is special already extracted 32 bits */
+ total_bits = format_desc->block.bits > 32 ? 32 : format_desc->block.bits;
+
z_swizzle = format_desc->swizzle[0];
- if (z_swizzle == UTIL_FORMAT_SWIZZLE_NONE)
+ if (z_swizzle == PIPE_SWIZZLE_NONE)
return FALSE;
*width = format_desc->channel[z_swizzle].size;
+ /* & 31 is for the same reason as the 32-bit limit above */
+ *shift = format_desc->channel[z_swizzle].shift & 31;
- padding_right = 0;
- for (chan = 0; chan < z_swizzle; ++chan)
- padding_right += format_desc->channel[chan].size;
-
- padding_left =
- total_bits - (padding_right + *width);
-
- if (padding_left || padding_right) {
- unsigned long long mask_left = (1ULL << (total_bits - padding_left)) - 1;
- unsigned long long mask_right = (1ULL << (padding_right)) - 1;
- *mask = mask_left ^ mask_right;
- }
- else {
+ if (*width == total_bits) {
*mask = 0xffffffff;
+ } else {
+ *mask = ((1 << *width) - 1) << *shift;
}
- *shift = padding_right;
-
return TRUE;
}
unsigned *shift, unsigned *mask)
{
unsigned s_swizzle;
- unsigned chan, sz;
+ unsigned sz;
s_swizzle = format_desc->swizzle[1];
- if (s_swizzle == UTIL_FORMAT_SWIZZLE_NONE)
+ if (s_swizzle == PIPE_SWIZZLE_NONE)
return FALSE;
- *shift = 0;
- for (chan = 0; chan < s_swizzle; chan++)
- *shift += format_desc->channel[chan].size;
+ /* just special case 64bit d/s format */
+ if (format_desc->block.bits > 32) {
+ /* XXX big-endian? */
+ assert(format_desc->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT);
+ *shift = 0;
+ *mask = 0xff;
+ return TRUE;
+ }
+ *shift = format_desc->channel[s_swizzle].shift;
sz = format_desc->channel[s_swizzle].size;
*mask = (1U << sz) - 1U;
* Test the depth mask. Add the number of channel which has none zero mask
* into the occlusion counter. e.g. maskvalue is {-1, -1, -1, -1}.
* The counter will add 4.
+ * TODO: could get that out of the fs loop.
*
* \param type holds element type of the mask vector.
* \param maskvalue is the depth test mask.
LLVMInt32TypeInContext(context), bits);
count = lp_build_intrinsic_unary(builder, popcntintr,
LLVMInt32TypeInContext(context), bits);
+ count = LLVMBuildZExt(builder, count, LLVMIntTypeInContext(context, 64), "");
}
else if(util_cpu_caps.has_avx && type.length == 8) {
const char *movmskintr = "llvm.x86.avx.movmsk.ps.256";
LLVMInt32TypeInContext(context), bits);
count = lp_build_intrinsic_unary(builder, popcntintr,
LLVMInt32TypeInContext(context), bits);
+ count = LLVMBuildZExt(builder, count, LLVMIntTypeInContext(context, 64), "");
}
else {
unsigned i;
}
count = lp_build_intrinsic_unary(builder, popcntintr, counttype, countd);
- if (type.length > 4) {
- count = LLVMBuildTrunc(builder, count, LLVMIntTypeInContext(context, 32), "");
+ if (type.length > 8) {
+ count = LLVMBuildTrunc(builder, count, LLVMIntTypeInContext(context, 64), "");
+ }
+ else if (type.length < 8) {
+ count = LLVMBuildZExt(builder, count, LLVMIntTypeInContext(context, 64), "");
}
}
newcount = LLVMBuildLoad(builder, counter, "origcount");
*
* \param type the data type of the fragment depth/stencil values
* \param format_desc description of the depth/stencil surface
+ * \param is_1d whether this resource has only one dimension
* \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)
*/
-LLVMValueRef
+void
lp_build_depth_stencil_load_swizzled(struct gallivm_state *gallivm,
struct lp_type z_src_type,
const struct util_format_description *format_desc,
+ boolean is_1d,
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_dst, zs_dst1, zs_dst2;
+ LLVMValueRef zs_dst1, zs_dst2;
LLVMValueRef zs_dst_ptr;
LLVMValueRef depth_offset1, depth_offset2;
- unsigned depth_bits = format_desc->block.bits/8;
+ 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 offset2 = LLVMBuildMul(builder, loopmsb,
depth_stride, "");
depth_offset1 = LLVMBuildMul(builder, looplsb,
- lp_build_const_int32(gallivm, depth_bits * 2), "");
+ 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 */
* 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);
}
}
/* 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,
- LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0), "");
+ 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,
- LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0), "");
- zs_dst2 = LLVMBuildLoad(builder, zs_dst_ptr, "");
+ if (is_1d) {
+ zs_dst2 = lp_build_undef(gallivm, zs_load_type);
+ }
+ else {
+ 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, "");
+ }
- zs_dst = LLVMBuildShuffleVector(builder, zs_dst1, zs_dst2,
- LLVMConstVector(shuffles, zs_type.length), "");
+ *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) */
- zs_dst = LLVMBuildZExt(builder, zs_dst, lp_build_int_vec_type(gallivm, z_src_type), "");
+ *z_fb = LLVMBuildZExt(builder, *z_fb,
+ lp_build_int_vec_type(gallivm, z_src_type), "");
}
- lp_build_name(zs_dst, "zs_dst");
+ 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");
+ }
- return zs_dst;
+ lp_build_name(*z_fb, "z_dst");
+ lp_build_name(*s_fb, "s_dst");
+ lp_build_name(*z_fb, "z_dst");
}
/**
* Store depth/stencil values.
* Incoming values are swizzled (typically n 2x2 quads), stored linear.
- * If there's a mask it will do reload/select/store otherwise just store.
+ * If there's a mask it will do select/store otherwise just store.
*
* \param type the data type of the fragment depth/stencil values
* \param format_desc description of the depth/stencil surface
+ * \param is_1d whether this resource has only one dimension
* \param mask the alive/dead pixel mask for the quad (vector)
+ * \param z_fb z values read from fb (with padding)
+ * \param s_fb s values read from fb (with padding)
* \param loop_counter the current loop iteration
* \param depth_ptr pointer to the depth/stencil values of this 4x4 block
* \param depth_stride stride of the depth/stencil buffer
- * \param zs_value the depth/stencil values to store
+ * \param z_value the depth values to store (with padding)
+ * \param s_value the stencil values to store (with padding)
*/
void
lp_build_depth_stencil_write_swizzled(struct gallivm_state *gallivm,
struct lp_type z_src_type,
const struct util_format_description *format_desc,
+ boolean is_1d,
struct lp_build_mask_context *mask,
+ LLVMValueRef z_fb,
+ LLVMValueRef s_fb,
LLVMValueRef loop_counter,
LLVMValueRef depth_ptr,
LLVMValueRef depth_stride,
- LLVMValueRef zs_value)
+ LLVMValueRef z_value,
+ LLVMValueRef s_value)
{
struct lp_build_context z_bld;
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 4];
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef mask_value = NULL;
- LLVMValueRef zs_dst = NULL, zs_dst1, zs_dst2;
+ LLVMValueRef zs_dst1, zs_dst2;
LLVMValueRef zs_dst_ptr1, zs_dst_ptr2;
LLVMValueRef depth_offset1, depth_offset2;
- unsigned depth_bits = format_desc->block.bits/8;
+ LLVMTypeRef load_ptr_type;
+ unsigned depth_bytes = format_desc->block.bits / 8;
struct lp_type zs_type = lp_depth_type(format_desc, z_src_type.length);
+ struct lp_type z_type = zs_type;
struct lp_type zs_load_type = zs_type;
+
zs_load_type.length = zs_load_type.length / 2;
+ load_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0);
- lp_build_context_init(&z_bld, gallivm, zs_type);
+ z_type.width = z_src_type.width;
+
+ lp_build_context_init(&z_bld, gallivm, z_type);
/*
* This is far from ideal, at least for late depth write we should do this
* outside the fs loop to avoid all the swizzle stuff.
*/
if (z_src_type.length == 4) {
- unsigned i;
LLVMValueRef looplsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
LLVMValueRef loopmsb = LLVMBuildAnd(builder, loop_counter,
LLVMValueRef offset2 = LLVMBuildMul(builder, loopmsb,
depth_stride, "");
depth_offset1 = LLVMBuildMul(builder, looplsb,
- lp_build_const_int32(gallivm, depth_bits * 2), "");
+ 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;
}
}
-
depth_offset2 = LLVMBuildAdd(builder, depth_offset1, depth_stride, "");
zs_dst_ptr1 = LLVMBuildGEP(builder, depth_ptr, &depth_offset1, 1, "");
- zs_dst_ptr1 = LLVMBuildBitCast(builder,
- zs_dst_ptr1,
- LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0), "");
+ zs_dst_ptr1 = LLVMBuildBitCast(builder, zs_dst_ptr1, load_ptr_type, "");
zs_dst_ptr2 = LLVMBuildGEP(builder, depth_ptr, &depth_offset2, 1, "");
- zs_dst_ptr2 = LLVMBuildBitCast(builder,
- zs_dst_ptr2,
- LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0), "");
+ zs_dst_ptr2 = LLVMBuildBitCast(builder, zs_dst_ptr2, load_ptr_type, "");
- if (mask) {
- zs_dst1 = LLVMBuildLoad(builder, zs_dst_ptr1, "");
- zs_dst2 = LLVMBuildLoad(builder, zs_dst_ptr2, "");
- zs_dst = LLVMBuildShuffleVector(builder, zs_dst1, zs_dst2,
- LLVMConstVector(shuffles, zs_type.length),
- "zsbufval");
+ if (format_desc->block.bits > 32) {
+ s_value = LLVMBuildBitCast(builder, s_value, z_bld.vec_type, "");
+ }
+ if (mask) {
mask_value = lp_build_mask_value(mask);
+ z_value = lp_build_select(&z_bld, mask_value, z_value, z_fb);
+ if (format_desc->block.bits > 32) {
+ s_fb = LLVMBuildBitCast(builder, s_fb, z_bld.vec_type, "");
+ s_value = lp_build_select(&z_bld, mask_value, s_value, s_fb);
+ }
}
if (zs_type.width < z_src_type.width) {
- /* Truncate incoming ZS and mask values (e.g., when writing to Z16_UNORM) */
- zs_value = LLVMBuildTrunc(builder, zs_value, z_bld.vec_type, "");
- if (mask)
- mask_value = LLVMBuildTrunc(builder, mask_value, z_bld.vec_type, "");
+ /* Truncate ZS values (e.g., when writing to Z16_UNORM) */
+ z_value = LLVMBuildTrunc(builder, z_value,
+ lp_build_int_vec_type(gallivm, zs_type), "");
}
- if (mask) {
- zs_value = lp_build_select(&z_bld, mask_value, zs_value, zs_dst);
- }
-
- if (z_src_type.length == 4) {
- zs_dst1 = lp_build_extract_range(gallivm, zs_value, 0, 2);
- zs_dst2 = lp_build_extract_range(gallivm, zs_value, 2, 2);
+ if (format_desc->block.bits <= 32) {
+ if (z_src_type.length == 4) {
+ zs_dst1 = lp_build_extract_range(gallivm, z_value, 0, 2);
+ zs_dst2 = lp_build_extract_range(gallivm, z_value, 2, 2);
+ }
+ else {
+ assert(z_src_type.length == 8);
+ zs_dst1 = LLVMBuildShuffleVector(builder, z_value, z_value,
+ LLVMConstVector(&shuffles[0],
+ zs_load_type.length), "");
+ zs_dst2 = LLVMBuildShuffleVector(builder, z_value, z_value,
+ LLVMConstVector(&shuffles[4],
+ zs_load_type.length), "");
+ }
}
else {
- assert(z_src_type.length == 8);
- zs_dst1 = LLVMBuildShuffleVector(builder, zs_value, zs_value,
- LLVMConstVector(&shuffles[0],
- zs_load_type.length),
- "");
- zs_dst2 = LLVMBuildShuffleVector(builder, zs_value, zs_value,
- LLVMConstVector(&shuffles[4],
- zs_load_type.length),
- "");
-
+ if (z_src_type.length == 4) {
+ zs_dst1 = lp_build_interleave2(gallivm, z_type,
+ z_value, s_value, 0);
+ zs_dst2 = lp_build_interleave2(gallivm, z_type,
+ z_value, s_value, 1);
+ }
+ else {
+ unsigned i;
+ LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 2];
+ assert(z_src_type.length == 8);
+ for (i = 0; i < 8; i++) {
+ shuffles[i*2] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
+ shuffles[i*2+1] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2 +
+ z_src_type.length);
+ }
+ zs_dst1 = LLVMBuildShuffleVector(builder, z_value, s_value,
+ LLVMConstVector(&shuffles[0],
+ z_src_type.length), "");
+ zs_dst2 = LLVMBuildShuffleVector(builder, z_value, s_value,
+ LLVMConstVector(&shuffles[8],
+ z_src_type.length), "");
+ }
+ zs_dst1 = LLVMBuildBitCast(builder, zs_dst1,
+ lp_build_vec_type(gallivm, zs_load_type), "");
+ zs_dst2 = LLVMBuildBitCast(builder, zs_dst2,
+ lp_build_vec_type(gallivm, zs_load_type), "");
}
+
LLVMBuildStore(builder, zs_dst1, zs_dst_ptr1);
- LLVMBuildStore(builder, zs_dst2, zs_dst_ptr2);
+ if (!is_1d) {
+ LLVMBuildStore(builder, zs_dst2, zs_dst_ptr2);
+ }
}
/**
struct lp_build_mask_context *mask,
LLVMValueRef stencil_refs[2],
LLVMValueRef z_src,
- LLVMValueRef zs_dst,
+ LLVMValueRef z_fb,
+ LLVMValueRef s_fb,
LLVMValueRef face,
- LLVMValueRef *zs_value,
+ LLVMValueRef *z_value,
+ LLVMValueRef *s_value,
boolean do_branch)
{
LLVMBuilderRef builder = gallivm->builder;
- struct lp_type zs_type;
struct lp_type z_type;
struct lp_build_context z_bld;
struct lp_build_context s_bld;
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 current_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
}
/* Pick the type matching the depth-stencil format. */
- zs_type = lp_depth_type(format_desc, z_src_type.length);
+ z_type = lp_depth_type(format_desc, z_src_type.length);
/* Pick the intermediate type for depth operations. */
- z_type = zs_type;
- /* FIXME: Cope with a depth test type with higher bit width. */
- assert(zs_type.width <= z_src_type.width);
z_type.width = z_src_type.width;
assert(z_type.length == z_src_type.length);
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(z_swizzle != PIPE_SWIZZLE_NONE ||
+ s_swizzle != PIPE_SWIZZLE_NONE);
assert(depth->enabled || stencil[0].enabled);
if (depth->enabled) {
assert(z_swizzle < 4);
- assert(format_desc->block.bits <= z_type.width);
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 ==
- format_desc->block.bits);
+ assert(format_desc->channel[z_swizzle].size == 32);
}
else {
assert(format_desc->channel[z_swizzle].type ==
{
unsigned s_shift, s_mask;
- if (get_z_shift_and_mask(format_desc, &z_shift, &z_width, &z_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);
}
*/
if (z_shift) {
LLVMValueRef shift = lp_build_const_int_vec(gallivm, z_type, z_shift);
- z_dst = LLVMBuildLShr(builder, zs_dst, shift, "z_dst");
+ z_dst = LLVMBuildLShr(builder, z_dst, shift, "z_dst");
} else if (z_bitmask) {
- /* TODO: Instead of loading a mask from memory and ANDing, it's
- * probably faster to just shake the bits with two shifts. */
- z_dst = LLVMBuildAnd(builder, zs_dst, z_bitmask, "z_dst");
+ z_dst = LLVMBuildAnd(builder, z_dst, z_bitmask, "z_dst");
} else {
- z_dst = zs_dst;
lp_build_name(z_dst, "z_dst");
}
}
- if (get_s_shift_and_mask(format_desc, &s_shift, &s_mask)) {
+ if (have_s) {
if (s_shift) {
LLVMValueRef shift = lp_build_const_int_vec(gallivm, s_type, s_shift);
- stencil_vals = LLVMBuildLShr(builder, zs_dst, shift, "");
+ stencil_vals = LLVMBuildLShr(builder, stencil_vals, shift, "");
stencil_shift = shift; /* used below */
}
- else {
- stencil_vals = zs_dst;
- }
if (s_mask != 0xffffffff) {
LLVMValueRef mask = lp_build_const_int_vec(gallivm, s_type, s_mask);
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, "");
- }
+ if (0) {
+ /*
+ * XXX: the scalar expansion below produces atrocious code
+ * (basically producing a 64bit scalar value, then moving the 2
+ * 32bit pieces separately to simd, plus 4 shuffles, which is
+ * seriously lame). But the scalar-simd transitions are always
+ * tricky, so no big surprise there.
+ * This here would be way better, however llvm has some serious
+ * trouble later using it in the select, probably because it will
+ * recognize the expression as constant and move the simd value
+ * away (out of the loop) - and then it will suddenly try
+ * constructing i1 high-bit masks out of it later...
+ * (Try piglit stencil-twoside.)
+ * Note this is NOT due to using SExt/Trunc, it fails exactly the
+ * same even when using native compare/select.
+ * I cannot reproduce this problem when using stand-alone compiler
+ * though, suggesting some problem with optimization passes...
+ * (With stand-alone compilation, the construction of this mask
+ * value, no matter if the easy 3 instruction here or the complex
+ * 16+ one below, never gets separated from where it's used.)
+ * The scalar code still has the same problem, but the generated
+ * code looks a bit better at least for some reason, even if
+ * mostly by luck (the fundamental issue clearly is the same).
+ */
+ front_facing = lp_build_broadcast(gallivm, s_bld.vec_type, face);
+ /* front_facing = face != 0 ? ~0 : 0 */
+ front_facing = lp_build_compare(gallivm, s_bld.type,
+ PIPE_FUNC_NOTEQUAL,
+ front_facing, s_bld.zero);
+ } else {
+ LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
- /* 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]);
+ /* 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, "");
+
+ }
+ }
s_pass_mask = lp_build_stencil_test(&s_bld, stencil,
stencil_refs, stencil_vals,
/* apply stencil-fail operator */
{
- LLVMValueRef s_fail_mask = lp_build_andnot(&s_bld, orig_mask, s_pass_mask);
+ LLVMValueRef s_fail_mask = lp_build_andnot(&s_bld, current_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);
/* compare src Z to dst Z, returning 'pass' mask */
z_pass = lp_build_cmp(&z_bld, depth->func, z_src, z_dst);
+ /* mask off bits that failed stencil test */
+ if (s_pass_mask) {
+ current_mask = LLVMBuildAnd(builder, current_mask, s_pass_mask, "");
+ }
+
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
if (do_branch) {
lp_build_mask_check(mask);
- do_branch = FALSE;
}
}
if (depth->writemask) {
- LLVMValueRef zselectmask;
+ LLVMValueRef z_pass_mask;
/* 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, "");
- }
+ z_pass_mask = LLVMBuildAnd(builder, current_mask, z_pass, "");
/* 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);
+ z_dst = lp_build_select(&z_bld, z_pass_mask, z_src, z_dst);
}
if (stencil[0].enabled) {
LLVMValueRef z_fail_mask, z_pass_mask;
/* apply Z-fail operator */
- z_fail_mask = lp_build_andnot(&z_bld, orig_mask, z_pass);
+ z_fail_mask = lp_build_andnot(&s_bld, current_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, "");
+ z_pass_mask = LLVMBuildAnd(builder, current_mask, z_pass, "");
stencil_vals = lp_build_stencil_op(&s_bld, stencil, Z_PASS_OP,
stencil_refs, stencil_vals,
z_pass_mask, front_facing);
/* 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, "");
+ s_pass_mask = LLVMBuildAnd(builder, current_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 ztencil bits in the right place */
- if (z_dst && z_shift) {
+ /* 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, "");
}
stencil_vals = LLVMBuildShl(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(builder, z_dst, stencil_vals, "");
- else if (z_dst)
- zs_dst = z_dst;
+ /* 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
- zs_dst = stencil_vals;
-
- *zs_value = zs_dst;
+ *z_value = stencil_vals;
+ *s_value = *z_value;
+ }
+ else {
+ *z_value = z_dst;
+ *s_value = stencil_vals;
}
if (s_pass_mask)
if (depth->enabled && stencil[0].enabled)
lp_build_mask_update(mask, z_pass);
-
- if (do_branch)
- lp_build_mask_check(mask);
-
}
projects / mesa.git / blobdiff