summaryrefslogtreecommitdiff
path: root/src/gallium/auxiliary/gallivm
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/auxiliary/gallivm')
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_arit.c56
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_arit.h12
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_init.cpp (renamed from src/gallium/auxiliary/gallivm/lp_bld_misc.cpp)46
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_init.h (renamed from src/gallium/auxiliary/gallivm/lp_bld_misc.h)17
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_logic.c5
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_pack.c15
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_sample.c30
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_sample.h14
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c860
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c121
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_type.c20
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_type.h4
12 files changed, 1074 insertions, 126 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.c b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
index bbce31f9eb..32f9e5201c 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_arit.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
@@ -669,6 +669,14 @@ lp_build_abs(struct lp_build_context *bld,
LLVMValueRef
+lp_build_negate(struct lp_build_context *bld,
+ LLVMValueRef a)
+{
+ return LLVMBuildNeg(bld->builder, a, "");
+}
+
+
+LLVMValueRef
lp_build_sgn(struct lp_build_context *bld,
LLVMValueRef a)
{
@@ -710,6 +718,41 @@ lp_build_sgn(struct lp_build_context *bld,
/**
+ * Set the sign of float vector 'a' according to 'sign'.
+ * If sign==0, return abs(a).
+ * If sign==1, return -abs(a);
+ * Other values for sign produce undefined results.
+ */
+LLVMValueRef
+lp_build_set_sign(struct lp_build_context *bld,
+ LLVMValueRef a, LLVMValueRef sign)
+{
+ const struct lp_type type = bld->type;
+ LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
+ LLVMTypeRef vec_type = lp_build_vec_type(type);
+ LLVMValueRef shift = lp_build_int_const_scalar(type, type.width - 1);
+ LLVMValueRef mask = lp_build_int_const_scalar(type,
+ ~((unsigned long long) 1 << (type.width - 1)));
+ LLVMValueRef val, res;
+
+ assert(type.floating);
+
+ /* val = reinterpret_cast<int>(a) */
+ val = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
+ /* val = val & mask */
+ val = LLVMBuildAnd(bld->builder, val, mask, "");
+ /* sign = sign << shift */
+ sign = LLVMBuildShl(bld->builder, sign, shift, "");
+ /* res = val | sign */
+ res = LLVMBuildOr(bld->builder, val, sign, "");
+ /* res = reinterpret_cast<float>(res) */
+ res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
+
+ return res;
+}
+
+
+/**
* Convert vector of int to vector of float.
*/
LLVMValueRef
@@ -857,6 +900,19 @@ lp_build_ceil(struct lp_build_context *bld,
/**
+ * Return fractional part of 'a' computed as a - floor(f)
+ * Typically used in texture coord arithmetic.
+ */
+LLVMValueRef
+lp_build_fract(struct lp_build_context *bld,
+ LLVMValueRef a)
+{
+ assert(bld->type.floating);
+ return lp_build_sub(bld, a, lp_build_floor(bld, a));
+}
+
+
+/**
* Convert to integer, through whichever rounding method that's fastest,
* typically truncating toward zero.
*/
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.h b/src/gallium/auxiliary/gallivm/lp_bld_arit.h
index da84b7ca02..55385e3a66 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_arit.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.h
@@ -117,10 +117,18 @@ lp_build_abs(struct lp_build_context *bld,
LLVMValueRef a);
LLVMValueRef
+lp_build_negate(struct lp_build_context *bld,
+ LLVMValueRef a);
+
+LLVMValueRef
lp_build_sgn(struct lp_build_context *bld,
LLVMValueRef a);
LLVMValueRef
+lp_build_set_sign(struct lp_build_context *bld,
+ LLVMValueRef a, LLVMValueRef sign);
+
+LLVMValueRef
lp_build_int_to_float(struct lp_build_context *bld,
LLVMValueRef a);
@@ -141,6 +149,10 @@ lp_build_trunc(struct lp_build_context *bld,
LLVMValueRef a);
LLVMValueRef
+lp_build_fract(struct lp_build_context *bld,
+ LLVMValueRef a);
+
+LLVMValueRef
lp_build_ifloor(struct lp_build_context *bld,
LLVMValueRef a);
LLVMValueRef
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp b/src/gallium/auxiliary/gallivm/lp_bld_init.cpp
index 6e79438ead..067397a520 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
+++ b/src/gallium/auxiliary/gallivm/lp_bld_init.cpp
@@ -26,39 +26,34 @@
**************************************************************************/
-#include "pipe/p_config.h"
+#include <llvm/Config/config.h>
+#include <llvm/Target/TargetSelect.h>
+#include <llvm/Target/TargetOptions.h>
-#include "lp_bld_misc.h"
+#include "pipe/p_config.h"
+#include "lp_bld_init.h"
-#ifndef LLVM_NATIVE_ARCH
-namespace llvm {
- extern void LinkInJIT();
-}
+extern "C" void LLVMLinkInJIT();
-void
-LLVMLinkInJIT(void)
+extern "C" void
+lp_build_init(void)
{
- llvm::LinkInJIT();
-}
-
-
-extern "C" int X86TargetMachineModule;
-
-
-int
-LLVMInitializeNativeTarget(void)
-{
-#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
- X86TargetMachineModule = 1;
+#if defined(PIPE_OS_WINDOWS) && defined(PIPE_ARCH_X86)
+ /*
+ * This is mis-detected on some hardware / software combinations.
+ */
+ llvm::StackAlignment = 4;
+ llvm::RealignStack = true;
#endif
- return 0;
-}
+ /* Same as LLVMInitializeNativeTarget(); */
+ llvm::InitializeNativeTarget();
-#endif
+ LLVMLinkInJIT();
+}
/*
@@ -69,7 +64,6 @@ LLVMInitializeNativeTarget(void)
*/
#if defined(_MSC_VER) && defined(_DEBUG)
#include <crtdefs.h>
-extern "C" {
- _CRTIMP void __cdecl _invalid_parameter_noinfo(void) {}
-}
+extern "C" _CRTIMP void __cdecl
+_invalid_parameter_noinfo(void) {}
#endif
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_misc.h b/src/gallium/auxiliary/gallivm/lp_bld_init.h
index 0e787e0b9c..07f50d1c43 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_misc.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_init.h
@@ -26,26 +26,17 @@
**************************************************************************/
-#ifndef LP_BLD_MISC_H
-#define LP_BLD_MISC_H
+#ifndef LP_BLD_INIT_H
+#define LP_BLD_INIT_H
-#include "llvm/Config/config.h"
-
#ifdef __cplusplus
extern "C" {
#endif
-#ifndef LLVM_NATIVE_ARCH
-
void
-LLVMLinkInJIT(void);
-
-int
-LLVMInitializeNativeTarget(void);
-
-#endif /* !LLVM_NATIVE_ARCH */
+lp_build_init(void);
#ifdef __cplusplus
@@ -53,4 +44,4 @@ LLVMInitializeNativeTarget(void);
#endif
-#endif /* !LP_BLD_MISC_H */
+#endif /* !LP_BLD_INIT_H */
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_logic.c b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
index 41ac81b744..2726747eae 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_logic.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
@@ -45,6 +45,7 @@
/**
* Build code to compare two values 'a' and 'b' of 'type' using the given func.
* \param func one of PIPE_FUNC_x
+ * The result values will be 0 for false or ~0 for true.
*/
LLVMValueRef
lp_build_compare(LLVMBuilderRef builder,
@@ -311,6 +312,7 @@ lp_build_compare(LLVMBuilderRef builder,
/**
* Build code to compare two values 'a' and 'b' using the given func.
* \param func one of PIPE_FUNC_x
+ * The result values will be 0 for false or ~0 for true.
*/
LLVMValueRef
lp_build_cmp(struct lp_build_context *bld,
@@ -322,6 +324,9 @@ lp_build_cmp(struct lp_build_context *bld,
}
+/**
+ * Return mask ? a : b;
+ */
LLVMValueRef
lp_build_select(struct lp_build_context *bld,
LLVMValueRef mask,
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_pack.c b/src/gallium/auxiliary/gallivm/lp_bld_pack.c
index bc360ad77a..4c61d10749 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_pack.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_pack.c
@@ -256,7 +256,9 @@ lp_build_pack2(LLVMBuilderRef builder,
LLVMValueRef lo,
LLVMValueRef hi)
{
+#if !(HAVE_LLVM >= 0x0207)
LLVMTypeRef src_vec_type = lp_build_vec_type(src_type);
+#endif
LLVMTypeRef dst_vec_type = lp_build_vec_type(dst_type);
LLVMValueRef shuffle;
LLVMValueRef res;
@@ -272,11 +274,14 @@ lp_build_pack2(LLVMBuilderRef builder,
switch(src_type.width) {
case 32:
if(dst_type.sign) {
+#if HAVE_LLVM >= 0x0207
+ res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packssdw.128", dst_vec_type, lo, hi);
+#else
res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packssdw.128", src_vec_type, lo, hi);
+#endif
}
else {
if (util_cpu_caps.has_sse4_1) {
- /* PACKUSDW is the only instrinsic with a consistent signature */
return lp_build_intrinsic_binary(builder, "llvm.x86.sse41.packusdw", dst_vec_type, lo, hi);
}
else {
@@ -288,9 +293,17 @@ lp_build_pack2(LLVMBuilderRef builder,
case 16:
if(dst_type.sign)
+#if HAVE_LLVM >= 0x0207
+ res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packsswb.128", dst_vec_type, lo, hi);
+#else
res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packsswb.128", src_vec_type, lo, hi);
+#endif
else
+#if HAVE_LLVM >= 0x0207
+ res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packuswb.128", dst_vec_type, lo, hi);
+#else
res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packuswb.128", src_vec_type, lo, hi);
+#endif
break;
default:
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample.c b/src/gallium/auxiliary/gallivm/lp_bld_sample.c
index a133b56ac5..29cadcc15a 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample.c
@@ -44,6 +44,11 @@
#include "lp_bld_sample.h"
+/**
+ * Initialize lp_sampler_static_state object with the gallium sampler
+ * and texture state.
+ * The former is considered to be static and the later dynamic.
+ */
void
lp_sampler_static_state(struct lp_sampler_static_state *state,
const struct pipe_texture *texture,
@@ -57,6 +62,18 @@ lp_sampler_static_state(struct lp_sampler_static_state *state,
if(!sampler)
return;
+ /*
+ * We don't copy sampler state over unless it is actually enabled, to avoid
+ * spurious recompiles, as the sampler static state is part of the shader
+ * key.
+ *
+ * Ideally the state tracker or cso_cache module would make all state
+ * canonical, but until that happens it's better to be safe than sorry here.
+ *
+ * XXX: Actually there's much more than can be done here, especially
+ * regarding 1D/2D/3D/CUBE textures, wrap modes, etc.
+ */
+
state->format = texture->format;
state->target = texture->target;
state->pot_width = util_is_pot(texture->width0);
@@ -69,11 +86,20 @@ lp_sampler_static_state(struct lp_sampler_static_state *state,
state->min_img_filter = sampler->min_img_filter;
state->min_mip_filter = sampler->min_mip_filter;
state->mag_img_filter = sampler->mag_img_filter;
+
state->compare_mode = sampler->compare_mode;
- if(sampler->compare_mode != PIPE_TEX_COMPARE_NONE) {
- state->compare_func = sampler->compare_func;
+ if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE) {
+ state->compare_func = sampler->compare_func;
}
+
state->normalized_coords = sampler->normalized_coords;
+ state->lod_bias = sampler->lod_bias;
+ state->min_lod = sampler->min_lod;
+ state->max_lod = sampler->max_lod;
+ state->border_color[0] = sampler->border_color[0];
+ state->border_color[1] = sampler->border_color[1];
+ state->border_color[2] = sampler->border_color[2];
+ state->border_color[3] = sampler->border_color[3];
}
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample.h b/src/gallium/auxiliary/gallivm/lp_bld_sample.h
index 39edcf13d1..5ba0925bb6 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample.h
@@ -70,6 +70,8 @@ struct lp_sampler_static_state
unsigned compare_mode:1;
unsigned compare_func:3;
unsigned normalized_coords:1;
+ float lod_bias, min_lod, max_lod;
+ float border_color[4];
};
@@ -98,6 +100,18 @@ struct lp_sampler_dynamic_state
LLVMBuilderRef builder,
unsigned unit);
+ /** Obtain the base texture depth. */
+ LLVMValueRef
+ (*depth)( struct lp_sampler_dynamic_state *state,
+ LLVMBuilderRef builder,
+ unsigned unit);
+
+ /** Obtain the number of mipmap levels (minus one). */
+ LLVMValueRef
+ (*last_level)( struct lp_sampler_dynamic_state *state,
+ LLVMBuilderRef builder,
+ unsigned unit);
+
LLVMValueRef
(*stride)( struct lp_sampler_dynamic_state *state,
LLVMBuilderRef builder,
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
index e268862282..bc2f307c33 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
@@ -69,7 +69,11 @@ struct lp_build_sample_context
struct lp_type coord_type;
struct lp_build_context coord_bld;
- /** Integer coordinates */
+ /** Unsigned integer coordinates */
+ struct lp_type uint_coord_type;
+ struct lp_build_context uint_coord_bld;
+
+ /** Signed integer coordinates */
struct lp_type int_coord_type;
struct lp_build_context int_coord_bld;
@@ -79,18 +83,93 @@ struct lp_build_sample_context
};
+/**
+ * Does the given texture wrap mode allow sampling the texture border color?
+ * XXX maybe move this into gallium util code.
+ */
+static boolean
+wrap_mode_uses_border_color(unsigned mode)
+{
+ switch (mode) {
+ case PIPE_TEX_WRAP_REPEAT:
+ case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
+ case PIPE_TEX_WRAP_MIRROR_REPEAT:
+ case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
+ return FALSE;
+ case PIPE_TEX_WRAP_CLAMP:
+ case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
+ case PIPE_TEX_WRAP_MIRROR_CLAMP:
+ case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
+ return TRUE;
+ default:
+ assert(0 && "unexpected wrap mode");
+ return FALSE;
+ }
+}
+
+
+
+/**
+ * Gen code to fetch a texel from a texture at int coords (x, y).
+ * The result, texel, will be:
+ * texel[0] = red values
+ * texel[1] = green values
+ * texel[2] = blue values
+ * texel[3] = alpha values
+ */
static void
lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
+ LLVMValueRef width,
+ LLVMValueRef height,
LLVMValueRef x,
LLVMValueRef y,
LLVMValueRef y_stride,
LLVMValueRef data_ptr,
LLVMValueRef *texel)
{
+ struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
LLVMValueRef offset;
LLVMValueRef packed;
+ LLVMValueRef use_border = NULL;
+
+ /* use_border = x < 0 || x >= width || y < 0 || y >= height */
+ if (wrap_mode_uses_border_color(bld->static_state->wrap_s)) {
+ LLVMValueRef b1, b2;
+ b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
+ b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
+ use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
+ }
- offset = lp_build_sample_offset(&bld->int_coord_bld,
+ if (wrap_mode_uses_border_color(bld->static_state->wrap_t)) {
+ LLVMValueRef b1, b2;
+ b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
+ b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
+ if (use_border) {
+ use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
+ use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
+ }
+ else {
+ use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
+ }
+ }
+
+ /*
+ * Note: if we find an app which frequently samples the texture border
+ * we might want to implement a true conditional here to avoid sampling
+ * the texture whenever possible (since that's quite a bit of code).
+ * Ex:
+ * if (use_border) {
+ * texel = border_color;
+ * }
+ * else {
+ * texel = sample_texture(coord);
+ * }
+ * As it is now, we always sample the texture, then selectively replace
+ * the texel color results with the border color.
+ */
+
+ /* convert x,y coords to linear offset from start of texture, in bytes */
+ offset = lp_build_sample_offset(&bld->uint_coord_bld,
bld->format_desc,
x, y, y_stride,
data_ptr);
@@ -99,16 +178,30 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
assert(bld->format_desc->block.height == 1);
assert(bld->format_desc->block.bits <= bld->texel_type.width);
+ /* gather the texels from the texture */
packed = lp_build_gather(bld->builder,
bld->texel_type.length,
bld->format_desc->block.bits,
bld->texel_type.width,
data_ptr, offset);
+ /* convert texels to float rgba */
lp_build_unpack_rgba_soa(bld->builder,
bld->format_desc,
bld->texel_type,
packed, texel);
+
+ if (use_border) {
+ /* select texel color or border color depending on use_border */
+ int chan;
+ for (chan = 0; chan < 4; chan++) {
+ LLVMValueRef border_chan =
+ lp_build_const_scalar(bld->texel_type,
+ bld->static_state->border_color[chan]);
+ texel[chan] = lp_build_select(&bld->texel_bld, use_border,
+ border_chan, texel[chan]);
+ }
+ }
}
@@ -121,7 +214,7 @@ lp_build_sample_packed(struct lp_build_sample_context *bld,
{
LLVMValueRef offset;
- offset = lp_build_sample_offset(&bld->int_coord_bld,
+ offset = lp_build_sample_offset(&bld->uint_coord_bld,
bld->format_desc,
x, y, y_stride,
data_ptr);
@@ -138,17 +231,77 @@ lp_build_sample_packed(struct lp_build_sample_context *bld,
}
+/**
+ * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
+ */
+static LLVMValueRef
+lp_build_coord_mirror(struct lp_build_sample_context *bld,
+ LLVMValueRef coord)
+{
+ struct lp_build_context *coord_bld = &bld->coord_bld;
+ struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
+ LLVMValueRef fract, flr, isOdd;
+
+ /* fract = coord - floor(coord) */
+ fract = lp_build_sub(coord_bld, coord, lp_build_floor(coord_bld, coord));
+
+ /* flr = ifloor(coord); */
+ flr = lp_build_ifloor(coord_bld, coord);
+
+ /* isOdd = flr & 1 */
+ isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
+
+ /* make coord positive or negative depending on isOdd */
+ coord = lp_build_set_sign(coord_bld, fract, isOdd);
+
+ /* convert isOdd to float */
+ isOdd = lp_build_int_to_float(coord_bld, isOdd);
+
+ /* add isOdd to coord */
+ coord = lp_build_add(coord_bld, coord, isOdd);
+
+ return coord;
+}
+
+
+/**
+ * We only support a few wrap modes in lp_build_sample_wrap_int() at this time.
+ * Return whether the given mode is supported by that function.
+ */
+static boolean
+is_simple_wrap_mode(unsigned mode)
+{
+ switch (mode) {
+ case PIPE_TEX_WRAP_REPEAT:
+ case PIPE_TEX_WRAP_CLAMP:
+ case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
+ return TRUE;
+ case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
+ default:
+ return FALSE;
+ }
+}
+
+
+/**
+ * Build LLVM code for texture wrap mode, for scaled integer texcoords.
+ * \param coord the incoming texcoord (s,t,r or q) scaled to the texture size
+ * \param length the texture size along one dimension
+ * \param is_pot if TRUE, length is a power of two
+ * \param wrap_mode one of PIPE_TEX_WRAP_x
+ */
static LLVMValueRef
-lp_build_sample_wrap(struct lp_build_sample_context *bld,
- LLVMValueRef coord,
- LLVMValueRef length,
- boolean is_pot,
- unsigned wrap_mode)
+lp_build_sample_wrap_int(struct lp_build_sample_context *bld,
+ LLVMValueRef coord,
+ LLVMValueRef length,
+ boolean is_pot,
+ unsigned wrap_mode)
{
+ struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
LLVMValueRef length_minus_one;
- length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
+ length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
switch(wrap_mode) {
case PIPE_TEX_WRAP_REPEAT:
@@ -161,12 +314,12 @@ lp_build_sample_wrap(struct lp_build_sample_context *bld,
break;
case PIPE_TEX_WRAP_CLAMP:
+ case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
+ case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
coord = lp_build_min(int_coord_bld, coord, length_minus_one);
break;
- case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
- case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
case PIPE_TEX_WRAP_MIRROR_REPEAT:
case PIPE_TEX_WRAP_MIRROR_CLAMP:
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
@@ -174,8 +327,8 @@ lp_build_sample_wrap(struct lp_build_sample_context *bld,
/* FIXME */
_debug_printf("llvmpipe: failed to translate texture wrap mode %s\n",
util_dump_tex_wrap(wrap_mode, TRUE));
- coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
- coord = lp_build_min(int_coord_bld, coord, length_minus_one);
+ coord = lp_build_max(uint_coord_bld, coord, uint_coord_bld->zero);
+ coord = lp_build_min(uint_coord_bld, coord, length_minus_one);
break;
default:
@@ -186,6 +339,376 @@ lp_build_sample_wrap(struct lp_build_sample_context *bld,
}
+/**
+ * Build LLVM code for texture wrap mode for linear filtering.
+ * \param x0_out returns first integer texcoord
+ * \param x1_out returns second integer texcoord
+ * \param weight_out returns linear interpolation weight
+ */
+static void
+lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
+ LLVMValueRef coord,
+ LLVMValueRef length,
+ boolean is_pot,
+ unsigned wrap_mode,
+ LLVMValueRef *x0_out,
+ LLVMValueRef *x1_out,
+ LLVMValueRef *weight_out)
+{
+ struct lp_build_context *coord_bld = &bld->coord_bld;
+ struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
+ struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
+ LLVMValueRef two = lp_build_const_scalar(coord_bld->type, 2.0);
+ LLVMValueRef half = lp_build_const_scalar(coord_bld->type, 0.5);
+ LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
+ LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
+ LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
+ LLVMValueRef coord0, coord1, weight;
+
+ switch(wrap_mode) {
+ case PIPE_TEX_WRAP_REPEAT:
+ /* mul by size and subtract 0.5 */
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ coord = lp_build_sub(coord_bld, coord, half);
+ /* convert to int */
+ coord0 = lp_build_ifloor(coord_bld, coord);
+ coord1 = lp_build_add(uint_coord_bld, coord0, uint_coord_bld->one);
+ /* compute lerp weight */
+ weight = lp_build_fract(coord_bld, coord);
+ /* repeat wrap */
+ if (is_pot) {
+ coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
+ coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
+ }
+ else {
+ /* Signed remainder won't give the right results for negative
+ * dividends but unsigned remainder does.*/
+ coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
+ coord1 = LLVMBuildURem(bld->builder, coord1, length, "");
+ }
+ break;
+
+ case PIPE_TEX_WRAP_CLAMP:
+ if (bld->static_state->normalized_coords) {
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ }
+ weight = lp_build_fract(coord_bld, coord);
+ coord0 = lp_build_clamp(coord_bld, coord, coord_bld->zero,
+ length_f_minus_one);
+ coord1 = lp_build_add(coord_bld, coord, coord_bld->one);
+ coord1 = lp_build_clamp(coord_bld, coord1, coord_bld->zero,
+ length_f_minus_one);
+ coord0 = lp_build_ifloor(coord_bld, coord0);
+ coord1 = lp_build_ifloor(coord_bld, coord1);
+ break;
+
+ case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
+ if (bld->static_state->normalized_coords) {
+ /* clamp to [0,1] */
+ coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, coord_bld->one);
+ /* mul by tex size and subtract 0.5 */
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ coord = lp_build_sub(coord_bld, coord, half);
+ }
+ else {
+ LLVMValueRef min, max;
+ /* clamp to [0.5, length - 0.5] */
+ min = lp_build_const_scalar(coord_bld->type, 0.5F);
+ max = lp_build_sub(coord_bld, length_f, min);
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ }
+ /* compute lerp weight */
+ weight = lp_build_fract(coord_bld, coord);
+ /* coord0 = floor(coord); */
+ coord0 = lp_build_ifloor(coord_bld, coord);
+ coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
+ /* coord0 = max(coord0, 0) */
+ coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
+ /* coord1 = min(coord1, length-1) */
+ coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
+ break;
+
+ case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
+ {
+ LLVMValueRef min, max;
+ if (bld->static_state->normalized_coords) {
+ /* min = -1.0 / (2 * length) = -0.5 / length */
+ min = lp_build_mul(coord_bld,
+ lp_build_const_scalar(coord_bld->type, -0.5F),
+ lp_build_rcp(coord_bld, length_f));
+ /* max = 1.0 - min */
+ max = lp_build_sub(coord_bld, coord_bld->one, min);
+ /* coord = clamp(coord, min, max) */
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ /* scale coord to length (and sub 0.5?) */
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ coord = lp_build_sub(coord_bld, coord, half);
+ }
+ else {
+ /* clamp to [-0.5, length + 0.5] */
+ min = lp_build_const_scalar(coord_bld->type, -0.5F);
+ max = lp_build_sub(coord_bld, length_f, min);
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ coord = lp_build_sub(coord_bld, coord, half);
+ }
+ /* compute lerp weight */
+ weight = lp_build_fract(coord_bld, coord);
+ /* convert to int */
+ coord0 = lp_build_ifloor(coord_bld, coord);
+ coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
+ }
+ break;
+
+ case PIPE_TEX_WRAP_MIRROR_REPEAT:
+ /* compute mirror function */
+ coord = lp_build_coord_mirror(bld, coord);
+
+ /* scale coord to length */
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ coord = lp_build_sub(coord_bld, coord, half);
+
+ /* compute lerp weight */
+ weight = lp_build_fract(coord_bld, coord);
+
+ /* convert to int coords */
+ coord0 = lp_build_ifloor(coord_bld, coord);
+ coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
+
+ /* coord0 = max(coord0, 0) */
+ coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
+ /* coord1 = min(coord1, length-1) */
+ coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
+ break;
+
+ case PIPE_TEX_WRAP_MIRROR_CLAMP:
+ {
+ LLVMValueRef min, max;
+ /* min = 1.0 / (2 * length) */
+ min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
+ /* max = 1.0 - min */
+ max = lp_build_sub(coord_bld, coord_bld->one, min);
+
+ coord = lp_build_abs(coord_bld, coord);
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ if(0)coord = lp_build_sub(coord_bld, coord, half);
+ weight = lp_build_fract(coord_bld, coord);
+ coord0 = lp_build_ifloor(coord_bld, coord);
+ coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
+ }
+ break;
+
+ case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
+ {
+ LLVMValueRef min, max;
+ /* min = 1.0 / (2 * length) */
+ min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
+ /* max = 1.0 - min */
+ max = lp_build_sub(coord_bld, coord_bld->one, min);
+
+ coord = lp_build_abs(coord_bld, coord);
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ coord = lp_build_sub(coord_bld, coord, half);
+ weight = lp_build_fract(coord_bld, coord);
+ coord0 = lp_build_ifloor(coord_bld, coord);
+ coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
+ }
+ break;
+
+ case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
+ {
+ LLVMValueRef min, max;
+ /* min = -1.0 / (2 * length) = -0.5 / length */
+ min = lp_build_mul(coord_bld,
+ lp_build_const_scalar(coord_bld->type, -0.5F),
+ lp_build_rcp(coord_bld, length_f));
+ /* max = 1.0 - min */
+ max = lp_build_sub(coord_bld, coord_bld->one, min);
+
+ coord = lp_build_abs(coord_bld, coord);
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ coord = lp_build_sub(coord_bld, coord, half);
+ weight = lp_build_fract(coord_bld, coord);
+ coord0 = lp_build_ifloor(coord_bld, coord);
+ coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
+ }
+ break;
+
+ default:
+ assert(0);
+ }
+
+ *x0_out = coord0;
+ *x1_out = coord1;
+ *weight_out = weight;
+}
+
+
+/**
+ * Build LLVM code for texture wrap mode for nearest filtering.
+ * \param coord the incoming texcoord (nominally in [0,1])
+ * \param length the texture size along one dimension, as int
+ * \param is_pot if TRUE, length is a power of two
+ * \param wrap_mode one of PIPE_TEX_WRAP_x
+ */
+static LLVMValueRef
+lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
+ LLVMValueRef coord,
+ LLVMValueRef length,
+ boolean is_pot,
+ unsigned wrap_mode)
+{
+ struct lp_build_context *coord_bld = &bld->coord_bld;
+ struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
+ struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
+ LLVMValueRef two = lp_build_const_scalar(coord_bld->type, 2.0);
+ LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
+ LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
+ LLVMValueRef length_f_minus_one = lp_build_sub(coord_bld, length_f, coord_bld->one);
+ LLVMValueRef icoord;
+
+ switch(wrap_mode) {
+ case PIPE_TEX_WRAP_REPEAT:
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ icoord = lp_build_ifloor(coord_bld, coord);
+ if (is_pot)
+ icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
+ else
+ /* Signed remainder won't give the right results for negative
+ * dividends but unsigned remainder does.*/
+ icoord = LLVMBuildURem(bld->builder, icoord, length, "");
+ break;
+
+ case PIPE_TEX_WRAP_CLAMP:
+ /* mul by size */
+ if (bld->static_state->normalized_coords) {
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ }
+ /* floor */
+ icoord = lp_build_ifloor(coord_bld, coord);
+ /* clamp to [0, size-1]. Note: int coord builder type */
+ icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
+ length_minus_one);
+ break;
+
+ case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
+ {
+ LLVMValueRef min, max;
+ if (bld->static_state->normalized_coords) {
+ /* min = 1.0 / (2 * length) */
+ min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
+ /* max = length - min */
+ max = lp_build_sub(coord_bld, length_f, min);
+ /* scale coord to length */
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ }
+ else {
+ /* clamp to [0.5, length - 0.5] */
+ min = lp_build_const_scalar(coord_bld->type, 0.5F);
+ max = lp_build_sub(coord_bld, length_f, min);
+ }
+ /* coord = clamp(coord, min, max) */
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ icoord = lp_build_ifloor(coord_bld, coord);
+ }
+ break;
+
+ case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
+ /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
+ {
+ LLVMValueRef min, max;
+ if (bld->static_state->normalized_coords) {
+ /* min = -1.0 / (2 * length) = -0.5 / length */
+ min = lp_build_mul(coord_bld,
+ lp_build_const_scalar(coord_bld->type, -0.5F),
+ lp_build_rcp(coord_bld, length_f));
+ /* max = length - min */
+ max = lp_build_sub(coord_bld, length_f, min);
+ /* scale coord to length */
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ }
+ else {
+ /* clamp to [-0.5, length + 0.5] */
+ min = lp_build_const_scalar(coord_bld->type, -0.5F);
+ max = lp_build_sub(coord_bld, length_f, min);
+ }
+ /* coord = clamp(coord, min, max) */
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ icoord = lp_build_ifloor(coord_bld, coord);
+ }
+ break;
+
+ case PIPE_TEX_WRAP_MIRROR_REPEAT:
+ {
+ LLVMValueRef min, max;
+ /* min = 1.0 / (2 * length) */
+ min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
+ /* max = length - min */
+ max = lp_build_sub(coord_bld, length_f, min);
+
+ /* compute mirror function */
+ coord = lp_build_coord_mirror(bld, coord);
+
+ /* scale coord to length */
+ coord = lp_build_mul(coord_bld, coord, length_f);
+
+ /* coord = clamp(coord, min, max) */
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ icoord = lp_build_ifloor(coord_bld, coord);
+ }
+ break;
+
+ case PIPE_TEX_WRAP_MIRROR_CLAMP:
+ coord = lp_build_abs(coord_bld, coord);
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f_minus_one);
+ icoord = lp_build_ifloor(coord_bld, coord);
+ break;
+
+ case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
+ {
+ LLVMValueRef min, max;
+ /* min = 1.0 / (2 * length) */
+ min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
+ /* max = length - min */
+ max = lp_build_sub(coord_bld, length_f, min);
+
+ coord = lp_build_abs(coord_bld, coord);
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ icoord = lp_build_ifloor(coord_bld, coord);
+ }
+ break;
+
+ case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
+ {
+ LLVMValueRef min, max;
+ /* min = 1.0 / (2 * length) */
+ min = lp_build_rcp(coord_bld, lp_build_mul(coord_bld, two, length_f));
+ min = lp_build_negate(coord_bld, min);
+ /* max = length - min */
+ max = lp_build_sub(coord_bld, length_f, min);
+
+ coord = lp_build_abs(coord_bld, coord);
+ coord = lp_build_mul(coord_bld, coord, length_f);
+ coord = lp_build_clamp(coord_bld, coord, min, max);
+ icoord = lp_build_ifloor(coord_bld, coord);
+ }
+ break;
+
+ default:
+ assert(0);
+ }
+
+ return icoord;
+}
+
+
+/**
+ * Sample 2D texture with nearest filtering.
+ */
static void
lp_build_sample_2d_nearest_soa(struct lp_build_sample_context *bld,
LLVMValueRef s,
@@ -196,23 +719,25 @@ lp_build_sample_2d_nearest_soa(struct lp_build_sample_context *bld,
LLVMValueRef data_ptr,
LLVMValueRef *texel)
{
- LLVMValueRef x;
- LLVMValueRef y;
+ LLVMValueRef x, y;
- x = lp_build_ifloor(&bld->coord_bld, s);
- y = lp_build_ifloor(&bld->coord_bld, t);
- lp_build_name(x, "tex.x.floor");
- lp_build_name(y, "tex.y.floor");
+ x = lp_build_sample_wrap_nearest(bld, s, width,
+ bld->static_state->pot_width,
+ bld->static_state->wrap_s);
+ y = lp_build_sample_wrap_nearest(bld, t, height,
+ bld->static_state->pot_height,
+ bld->static_state->wrap_t);
- x = lp_build_sample_wrap(bld, x, width, bld->static_state->pot_width, bld->static_state->wrap_s);
- y = lp_build_sample_wrap(bld, y, height, bld->static_state->pot_height, bld->static_state->wrap_t);
lp_build_name(x, "tex.x.wrapped");
lp_build_name(y, "tex.y.wrapped");
- lp_build_sample_texel_soa(bld, x, y, stride, data_ptr, texel);
+ lp_build_sample_texel_soa(bld, width, height, x, y, stride, data_ptr, texel);
}
+/**
+ * Sample 2D texture with bilinear filtering.
+ */
static void
lp_build_sample_2d_linear_soa(struct lp_build_sample_context *bld,
LLVMValueRef s,
@@ -223,9 +748,6 @@ lp_build_sample_2d_linear_soa(struct lp_build_sample_context *bld,
LLVMValueRef data_ptr,
LLVMValueRef *texel)
{
- LLVMValueRef half;
- LLVMValueRef s_ipart;
- LLVMValueRef t_ipart;
LLVMValueRef s_fpart;
LLVMValueRef t_fpart;
LLVMValueRef x0, x1;
@@ -233,32 +755,15 @@ lp_build_sample_2d_linear_soa(struct lp_build_sample_context *bld,
LLVMValueRef neighbors[2][2][4];
unsigned chan;
- half = lp_build_const_scalar(bld->coord_type, 0.5);
- s = lp_build_sub(&bld->coord_bld, s, half);
- t = lp_build_sub(&bld->coord_bld, t, half);
-
- s_ipart = lp_build_floor(&bld->coord_bld, s);
- t_ipart = lp_build_floor(&bld->coord_bld, t);
-
- s_fpart = lp_build_sub(&bld->coord_bld, s, s_ipart);
- t_fpart = lp_build_sub(&bld->coord_bld, t, t_ipart);
+ lp_build_sample_wrap_linear(bld, s, width, bld->static_state->pot_width,
+ bld->static_state->wrap_s, &x0, &x1, &s_fpart);
+ lp_build_sample_wrap_linear(bld, t, height, bld->static_state->pot_height,
+ bld->static_state->wrap_t, &y0, &y1, &t_fpart);
- x0 = lp_build_itrunc(&bld->coord_bld, s_ipart);
- y0 = lp_build_itrunc(&bld->coord_bld, t_ipart);
-
- x0 = lp_build_sample_wrap(bld, x0, width, bld->static_state->pot_width, bld->static_state->wrap_s);
- y0 = lp_build_sample_wrap(bld, y0, height, bld->static_state->pot_height, bld->static_state->wrap_t);
-
- x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one);
- y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one);
-
- x1 = lp_build_sample_wrap(bld, x1, width, bld->static_state->pot_width, bld->static_state->wrap_s);
- y1 = lp_build_sample_wrap(bld, y1, height, bld->static_state->pot_height, bld->static_state->wrap_t);
-
- lp_build_sample_texel_soa(bld, x0, y0, stride, data_ptr, neighbors[0][0]);
- lp_build_sample_texel_soa(bld, x1, y0, stride, data_ptr, neighbors[0][1]);
- lp_build_sample_texel_soa(bld, x0, y1, stride, data_ptr, neighbors[1][0]);
- lp_build_sample_texel_soa(bld, x1, y1, stride, data_ptr, neighbors[1][1]);
+ lp_build_sample_texel_soa(bld, width, height, x0, y0, stride, data_ptr, neighbors[0][0]);
+ lp_build_sample_texel_soa(bld, width, height, x1, y0, stride, data_ptr, neighbors[0][1]);
+ lp_build_sample_texel_soa(bld, width, height, x0, y1, stride, data_ptr, neighbors[1][0]);
+ lp_build_sample_texel_soa(bld, width, height, x1, y1, stride, data_ptr, neighbors[1][1]);
/* TODO: Don't interpolate missing channels */
for(chan = 0; chan < 4; ++chan) {
@@ -334,20 +839,33 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
h16_vec_type = lp_build_vec_type(h16.type);
u8n_vec_type = lp_build_vec_type(u8n.type);
+ if (bld->static_state->normalized_coords) {
+ LLVMTypeRef coord_vec_type = lp_build_vec_type(bld->coord_type);
+ LLVMValueRef fp_width = LLVMBuildSIToFP(bld->builder, width, coord_vec_type, "");
+ LLVMValueRef fp_height = LLVMBuildSIToFP(bld->builder, height, coord_vec_type, "");
+ s = lp_build_mul(&bld->coord_bld, s, fp_width);
+ t = lp_build_mul(&bld->coord_bld, t, fp_height);
+ }
+
+ /* scale coords by 256 (8 fractional bits) */
s = lp_build_mul_imm(&bld->coord_bld, s, 256);
t = lp_build_mul_imm(&bld->coord_bld, t, 256);
+ /* convert float to int */
s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
+ /* subtract 0.5 (add -128) */
i32_c128 = lp_build_int_const_scalar(i32.type, -128);
s = LLVMBuildAdd(builder, s, i32_c128, "");
t = LLVMBuildAdd(builder, t, i32_c128, "");
+ /* compute floor (shift right 8) */
i32_c8 = lp_build_int_const_scalar(i32.type, 8);
s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
+ /* compute fractional part (AND with 0xff) */
i32_c255 = lp_build_int_const_scalar(i32.type, 255);
s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
@@ -355,14 +873,18 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
x0 = s_ipart;
y0 = t_ipart;
- x0 = lp_build_sample_wrap(bld, x0, width, bld->static_state->pot_width, bld->static_state->wrap_s);
- y0 = lp_build_sample_wrap(bld, y0, height, bld->static_state->pot_height, bld->static_state->wrap_t);
-
x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one);
y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one);
- x1 = lp_build_sample_wrap(bld, x1, width, bld->static_state->pot_width, bld->static_state->wrap_s);
- y1 = lp_build_sample_wrap(bld, y1, height, bld->static_state->pot_height, bld->static_state->wrap_t);
+ x0 = lp_build_sample_wrap_int(bld, x0, width, bld->static_state->pot_width,
+ bld->static_state->wrap_s);
+ y0 = lp_build_sample_wrap_int(bld, y0, height, bld->static_state->pot_height,
+ bld->static_state->wrap_t);
+
+ x1 = lp_build_sample_wrap_int(bld, x1, width, bld->static_state->pot_width,
+ bld->static_state->wrap_s);
+ y1 = lp_build_sample_wrap_int(bld, y1, height, bld->static_state->pot_height,
+ bld->static_state->wrap_t);
/*
* Transform 4 x i32 in
@@ -518,6 +1040,195 @@ lp_build_sample_compare(struct lp_build_sample_context *bld,
}
+static int
+texture_dims(enum pipe_texture_target tex)
+{
+ switch (tex) {
+ case PIPE_TEXTURE_1D:
+ return 1;
+ case PIPE_TEXTURE_2D:
+ case PIPE_TEXTURE_CUBE:
+ return 2;
+ case PIPE_TEXTURE_3D:
+ return 3;
+ default:
+ assert(0 && "bad texture target in texture_dims()");
+ return 2;
+ }
+}
+
+
+/**
+ * Generate code to compute texture level of detail (lambda).
+ * \param s vector of texcoord s values
+ * \param t vector of texcoord t values
+ * \param r vector of texcoord r values
+ * \param width scalar int texture width
+ * \param height scalar int texture height
+ * \param depth scalar int texture depth
+ */
+static LLVMValueRef
+lp_build_lod_selector(struct lp_build_sample_context *bld,
+ LLVMValueRef s,
+ LLVMValueRef t,
+ LLVMValueRef r,
+ LLVMValueRef width,
+ LLVMValueRef height,
+ LLVMValueRef depth)
+
+{
+ const int dims = texture_dims(bld->static_state->target);
+ struct lp_build_context *coord_bld = &bld->coord_bld;
+
+ LLVMValueRef lod_bias = lp_build_const_scalar(bld->coord_bld.type,
+ bld->static_state->lod_bias);
+ LLVMValueRef min_lod = lp_build_const_scalar(bld->coord_bld.type,
+ bld->static_state->min_lod);
+ LLVMValueRef max_lod = lp_build_const_scalar(bld->coord_bld.type,
+ bld->static_state->max_lod);
+
+ LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
+ LLVMValueRef index1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
+ LLVMValueRef index2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
+
+ LLVMValueRef s0, s1, s2;
+ LLVMValueRef t0, t1, t2;
+ LLVMValueRef r0, r1, r2;
+ LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy;
+ LLVMValueRef rho, lod;
+
+ /*
+ * dsdx = abs(s[1] - s[0]);
+ * dsdy = abs(s[2] - s[0]);
+ * dtdx = abs(t[1] - t[0]);
+ * dtdy = abs(t[2] - t[0]);
+ * drdx = abs(r[1] - r[0]);
+ * drdy = abs(r[2] - r[0]);
+ * XXX we're assuming a four-element quad in 2x2 layout here.
+ */
+ s0 = LLVMBuildExtractElement(bld->builder, s, index0, "s0");
+ s1 = LLVMBuildExtractElement(bld->builder, s, index1, "s1");
+ s2 = LLVMBuildExtractElement(bld->builder, s, index2, "s2");
+ dsdx = lp_build_abs(coord_bld, lp_build_sub(coord_bld, s1, s0));
+ dsdy = lp_build_abs(coord_bld, lp_build_sub(coord_bld, s2, s0));
+ if (dims > 1) {
+ t0 = LLVMBuildExtractElement(bld->builder, t, index0, "t0");
+ t1 = LLVMBuildExtractElement(bld->builder, t, index1, "t1");
+ t2 = LLVMBuildExtractElement(bld->builder, t, index2, "t2");
+ dtdx = lp_build_abs(coord_bld, lp_build_sub(coord_bld, t1, t0));
+ dtdy = lp_build_abs(coord_bld, lp_build_sub(coord_bld, t2, t0));
+ if (dims > 2) {
+ r0 = LLVMBuildExtractElement(bld->builder, r, index0, "r0");
+ r1 = LLVMBuildExtractElement(bld->builder, r, index1, "r1");
+ r2 = LLVMBuildExtractElement(bld->builder, r, index2, "r2");
+ drdx = lp_build_abs(coord_bld, lp_build_sub(coord_bld, r1, r0));
+ drdy = lp_build_abs(coord_bld, lp_build_sub(coord_bld, r2, r0));
+ }
+ }
+
+ /* Compute rho = max of all partial derivatives scaled by texture size.
+ * XXX this can be vectorized somewhat
+ */
+ rho = lp_build_mul(coord_bld,
+ lp_build_max(coord_bld, dsdx, dsdy),
+ lp_build_int_to_float(coord_bld, width));
+ if (dims > 1) {
+ LLVMValueRef max;
+ max = lp_build_mul(coord_bld,
+ lp_build_max(coord_bld, dtdx, dtdy),
+ lp_build_int_to_float(coord_bld, height));
+ rho = lp_build_max(coord_bld, rho, max);
+ if (dims > 2) {
+ max = lp_build_mul(coord_bld,
+ lp_build_max(coord_bld, drdx, drdy),
+ lp_build_int_to_float(coord_bld, depth));
+ rho = lp_build_max(coord_bld, rho, max);
+ }
+ }
+
+ /* compute lod = log2(rho) */
+ lod = lp_build_log2(coord_bld, rho);
+
+ /* add lod bias */
+ lod = lp_build_add(coord_bld, lod, lod_bias);
+
+ /* clamp lod */
+ lod = lp_build_clamp(coord_bld, lod, min_lod, max_lod);
+
+ return lod;
+}
+
+
+/**
+ * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
+ * mipmap level index.
+ * \param lod scalar float texture level of detail
+ * \param level_out returns integer
+ */
+static void
+lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
+ unsigned unit,
+ LLVMValueRef lod,
+ LLVMValueRef *level_out)
+{
+ struct lp_build_context *coord_bld = &bld->coord_bld;
+ struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
+ LLVMValueRef last_level, level;
+
+ last_level = bld->dynamic_state->last_level(bld->dynamic_state,
+ bld->builder, unit);
+
+ /* convert float lod to integer */
+ level = lp_build_iround(coord_bld, lod);
+
+ /* clamp level to legal range of levels */
+ *level_out = lp_build_clamp(int_coord_bld, level,
+ int_coord_bld->zero,
+ last_level);
+}
+
+
+/**
+ * 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.
+ */
+static void
+lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
+ unsigned unit,
+ LLVMValueRef lod,
+ LLVMValueRef *level0_out,
+ LLVMValueRef *level1_out,
+ LLVMValueRef *weight_out)
+{
+ struct lp_build_context *coord_bld = &bld->coord_bld;
+ struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
+ LLVMValueRef last_level, level;
+
+ last_level = bld->dynamic_state->last_level(bld->dynamic_state,
+ bld->builder, unit);
+
+ /* convert float lod to integer */
+ level = lp_build_ifloor(coord_bld, lod);
+
+ /* compute level 0 and clamp to legal range of levels */
+ *level0_out = lp_build_clamp(int_coord_bld, level,
+ int_coord_bld->zero,
+ last_level);
+ /* compute level 1 and clamp to legal range of levels */
+ *level1_out = lp_build_add(int_coord_bld, *level0_out, int_coord_bld->one);
+ *level1_out = lp_build_min(int_coord_bld, *level1_out, int_coord_bld->zero);
+
+ *weight_out = lp_build_fract(coord_bld, lod);
+}
+
+
+
+/**
+ * Build texture sampling code.
+ * 'texel' will return a vector of four LLVMValueRefs corresponding to
+ * R, G, B, A.
+ */
void
lp_build_sample_soa(LLVMBuilderRef builder,
const struct lp_sampler_static_state *static_state,
@@ -536,7 +1247,11 @@ lp_build_sample_soa(LLVMBuilderRef builder,
LLVMValueRef data_ptr;
LLVMValueRef s;
LLVMValueRef t;
- LLVMValueRef p;
+ LLVMValueRef r;
+
+ (void) lp_build_lod_selector; /* temporary to silence warning */
+ (void) lp_build_nearest_mip_level;
+ (void) lp_build_linear_mip_levels;
/* Setup our build context */
memset(&bld, 0, sizeof bld);
@@ -545,9 +1260,11 @@ lp_build_sample_soa(LLVMBuilderRef builder,
bld.dynamic_state = dynamic_state;
bld.format_desc = util_format_description(static_state->format);
bld.coord_type = type;
+ bld.uint_coord_type = lp_uint_type(type);
bld.int_coord_type = lp_int_type(type);
bld.texel_type = type;
lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
+ lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type);
lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
@@ -559,32 +1276,29 @@ lp_build_sample_soa(LLVMBuilderRef builder,
s = coords[0];
t = coords[1];
- p = coords[2];
+ r = coords[2];
- width = lp_build_broadcast_scalar(&bld.int_coord_bld, width);
- height = lp_build_broadcast_scalar(&bld.int_coord_bld, height);
- stride = lp_build_broadcast_scalar(&bld.int_coord_bld, stride);
+ width = lp_build_broadcast_scalar(&bld.uint_coord_bld, width);
+ height = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
+ stride = lp_build_broadcast_scalar(&bld.uint_coord_bld, stride);
if(static_state->target == PIPE_TEXTURE_1D)
t = bld.coord_bld.zero;
- if(static_state->normalized_coords) {
- LLVMTypeRef coord_vec_type = lp_build_vec_type(bld.coord_type);
- LLVMValueRef fp_width = LLVMBuildSIToFP(builder, width, coord_vec_type, "");
- LLVMValueRef fp_height = LLVMBuildSIToFP(builder, height, coord_vec_type, "");
- s = lp_build_mul(&bld.coord_bld, s, fp_width);
- t = lp_build_mul(&bld.coord_bld, t, fp_height);
- }
-
switch (static_state->min_img_filter) {
case PIPE_TEX_FILTER_NEAREST:
- lp_build_sample_2d_nearest_soa(&bld, s, t, width, height, stride, data_ptr, texel);
+ lp_build_sample_2d_nearest_soa(&bld, s, t, width, height,
+ stride, data_ptr, texel);
break;
case PIPE_TEX_FILTER_LINEAR:
- if(lp_format_is_rgba8(bld.format_desc))
- lp_build_sample_2d_linear_aos(&bld, s, t, width, height, stride, data_ptr, texel);
+ if(lp_format_is_rgba8(bld.format_desc) &&
+ is_simple_wrap_mode(static_state->wrap_s) &&
+ is_simple_wrap_mode(static_state->wrap_t))
+ lp_build_sample_2d_linear_aos(&bld, s, t, width, height,
+ stride, data_ptr, texel);
else
- lp_build_sample_2d_linear_soa(&bld, s, t, width, height, stride, data_ptr, texel);
+ lp_build_sample_2d_linear_soa(&bld, s, t, width, height,
+ stride, data_ptr, texel);
break;
default:
assert(0);
@@ -593,5 +1307,5 @@ lp_build_sample_soa(LLVMBuilderRef builder,
/* FIXME: respect static_state->min_mip_filter */;
/* FIXME: respect static_state->mag_img_filter */;
- lp_build_sample_compare(&bld, p, texel);
+ lp_build_sample_compare(&bld, r, texel);
}
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
index 4cf28a9f93..28ff362cac 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
@@ -52,6 +52,7 @@
#include "lp_bld_swizzle.h"
#include "lp_bld_flow.h"
#include "lp_bld_tgsi.h"
+#include "lp_bld_debug.h"
#define LP_MAX_TEMPS 256
@@ -81,6 +82,21 @@
#define QUAD_BOTTOM_LEFT 2
#define QUAD_BOTTOM_RIGHT 3
+#define LP_TGSI_MAX_NESTING 16
+
+struct lp_exec_mask {
+ struct lp_build_context *bld;
+
+ boolean has_mask;
+
+ LLVMTypeRef int_vec_type;
+
+ LLVMValueRef cond_stack[LP_TGSI_MAX_NESTING];
+ int cond_stack_size;
+ LLVMValueRef cond_mask;
+
+ LLVMValueRef exec_mask;
+};
struct lp_build_tgsi_soa_context
{
@@ -97,9 +113,9 @@ struct lp_build_tgsi_soa_context
LLVMValueRef temps[LP_MAX_TEMPS][NUM_CHANNELS];
struct lp_build_mask_context *mask;
+ struct lp_exec_mask exec_mask;
};
-
static const unsigned char
swizzle_left[4] = {
QUAD_TOP_LEFT, QUAD_TOP_LEFT,
@@ -124,6 +140,73 @@ swizzle_bottom[4] = {
QUAD_BOTTOM_LEFT, QUAD_BOTTOM_RIGHT
};
+static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
+{
+ mask->bld = bld;
+ mask->has_mask = FALSE;
+ mask->cond_stack_size = 0;
+
+ mask->int_vec_type = lp_build_int_vec_type(mask->bld->type);
+}
+
+static void lp_exec_mask_update(struct lp_exec_mask *mask)
+{
+ mask->exec_mask = mask->cond_mask;
+ if (mask->cond_stack_size > 0)
+ mask->has_mask = TRUE;
+}
+
+static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
+ LLVMValueRef val)
+{
+ mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
+ mask->cond_mask = LLVMBuildBitCast(mask->bld->builder, val,
+ mask->int_vec_type, "");
+
+ lp_exec_mask_update(mask);
+}
+
+static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
+{
+ LLVMValueRef prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
+ LLVMValueRef inv_mask = LLVMBuildNot(mask->bld->builder,
+ mask->cond_mask, "");
+
+ /* means that we didn't have any mask before and that
+ * we were fully enabled */
+ if (mask->cond_stack_size <= 1) {
+ prev_mask = LLVMConstAllOnes(mask->int_vec_type);
+ }
+
+ mask->cond_mask = LLVMBuildAnd(mask->bld->builder,
+ inv_mask,
+ prev_mask, "");
+ lp_exec_mask_update(mask);
+}
+
+static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
+{
+ mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
+ lp_exec_mask_update(mask);
+}
+
+static void lp_exec_mask_store(struct lp_exec_mask *mask,
+ LLVMValueRef val,
+ LLVMValueRef dst)
+{
+ if (mask->has_mask) {
+ LLVMValueRef real_val, dst_val;
+
+ dst_val = LLVMBuildLoad(mask->bld->builder, dst, "");
+ real_val = lp_build_select(mask->bld,
+ mask->exec_mask,
+ val, dst_val);
+
+ LLVMBuildStore(mask->bld->builder, real_val, dst);
+ } else
+ LLVMBuildStore(mask->bld->builder, val, dst);
+}
+
static LLVMValueRef
emit_ddx(struct lp_build_tgsi_soa_context *bld,
@@ -287,13 +370,13 @@ emit_store(
switch( reg->Register.File ) {
case TGSI_FILE_OUTPUT:
- LLVMBuildStore(bld->base.builder, value,
- bld->outputs[reg->Register.Index][chan_index]);
+ lp_exec_mask_store(&bld->exec_mask, value,
+ bld->outputs[reg->Register.Index][chan_index]);
break;
case TGSI_FILE_TEMPORARY:
- LLVMBuildStore(bld->base.builder, value,
- bld->temps[reg->Register.Index][chan_index]);
+ lp_exec_mask_store(&bld->exec_mask, value,
+ bld->temps[reg->Register.Index][chan_index]);
break;
case TGSI_FILE_ADDRESS:
@@ -301,6 +384,11 @@ emit_store(
assert(0);
break;
+ case TGSI_FILE_PREDICATE:
+ /* FIXME */
+ assert(0);
+ break;
+
default:
assert( 0 );
}
@@ -498,6 +586,17 @@ emit_instruction(
if (indirect_temp_reference(inst))
return FALSE;
+ /*
+ * Stores and write masks are handled in a general fashion after the long
+ * instruction opcode switch statement.
+ *
+ * Although not stricitly necessary, we avoid generating instructions for
+ * channels which won't be stored, in cases where's that easy. For some
+ * complex instructions, like texture sampling, it is more convenient to
+ * assume a full writemask and then let LLVM optimization passes eliminate
+ * redundant code.
+ */
+
assert(info->num_dst <= 1);
if(info->num_dst) {
FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
@@ -1272,8 +1371,8 @@ emit_instruction(
break;
case TGSI_OPCODE_IF:
- /* FIXME */
- return 0;
+ tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
+ lp_exec_mask_cond_push(&bld->exec_mask, tmp0);
break;
case TGSI_OPCODE_BGNFOR:
@@ -1289,13 +1388,11 @@ emit_instruction(
break;
case TGSI_OPCODE_ELSE:
- /* FIXME */
- return 0;
+ lp_exec_mask_cond_invert(&bld->exec_mask);
break;
case TGSI_OPCODE_ENDIF:
- /* FIXME */
- return 0;
+ lp_exec_mask_cond_pop(&bld->exec_mask);
break;
case TGSI_OPCODE_ENDFOR:
@@ -1458,6 +1555,8 @@ lp_build_tgsi_soa(LLVMBuilderRef builder,
bld.consts_ptr = consts_ptr;
bld.sampler = sampler;
+ lp_exec_mask_init(&bld.exec_mask, &bld.base);
+
tgsi_parse_init( &parse, tokens );
while( !tgsi_parse_end_of_tokens( &parse ) ) {
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_type.c b/src/gallium/auxiliary/gallivm/lp_bld_type.c
index 8270cd057f..c327ba045a 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_type.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_type.c
@@ -178,6 +178,25 @@ lp_build_int32_vec4_type(void)
}
+/**
+ * Create unsigned integer type variation of given type.
+ */
+struct lp_type
+lp_uint_type(struct lp_type type)
+{
+ struct lp_type res_type;
+
+ memset(&res_type, 0, sizeof res_type);
+ res_type.width = type.width;
+ res_type.length = type.length;
+
+ return res_type;
+}
+
+
+/**
+ * Create signed integer type variation of given type.
+ */
struct lp_type
lp_int_type(struct lp_type type)
{
@@ -186,6 +205,7 @@ lp_int_type(struct lp_type type)
memset(&res_type, 0, sizeof res_type);
res_type.width = type.width;
res_type.length = type.length;
+ res_type.sign = 1;
return res_type;
}
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_type.h b/src/gallium/auxiliary/gallivm/lp_bld_type.h
index 62ee05be4d..16946cc28a 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_type.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_type.h
@@ -257,6 +257,10 @@ lp_build_int32_vec4_type(void);
struct lp_type
+lp_uint_type(struct lp_type type);
+
+
+struct lp_type
lp_int_type(struct lp_type type);
generated by cgit v1.2.3 (git 2.39.1) at 2024-11-07 09:28:48 +0000