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_alpha.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_arit.c73
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_arit.h6
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_blend.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_const.c12
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_const.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_conv.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_debug.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_depth.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_flow.c28
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_flow.h6
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_format.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_interp.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_intr.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_logic.c55
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_logic.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_pack.c2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_pack.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_sample.c24
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_sample.h12
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c853
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_struct.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_swizzle.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_tgsi.h2
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c147
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_type.c13
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_type.h2
27 files changed, 1011 insertions, 250 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_alpha.h b/src/gallium/auxiliary/gallivm/lp_bld_alpha.h
index 634575670d..fe3cedcc48 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_alpha.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_alpha.h
@@ -35,7 +35,7 @@
#define LP_BLD_ALPHA_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
struct pipe_alpha_state;
struct lp_type;
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.c b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
index e2c6788397..aa47338b32 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_arit.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
@@ -232,6 +232,37 @@ lp_build_add(struct lp_build_context *bld,
}
+/** Return the sum of the elements of a */
+LLVMValueRef
+lp_build_sum_vector(struct lp_build_context *bld,
+ LLVMValueRef a)
+{
+ const struct lp_type type = bld->type;
+ LLVMValueRef index, res;
+ int i;
+
+ if (a == bld->zero)
+ return bld->zero;
+ if (a == bld->undef)
+ return bld->undef;
+ assert(type.length > 1);
+
+ assert(!bld->type.norm);
+
+ index = LLVMConstInt(LLVMInt32Type(), 0, 0);
+ res = LLVMBuildExtractElement(bld->builder, a, index, "");
+
+ for (i = 1; i < type.length; i++) {
+ index = LLVMConstInt(LLVMInt32Type(), i, 0);
+ res = LLVMBuildAdd(bld->builder, res,
+ LLVMBuildExtractElement(bld->builder, a, index, ""),
+ "");
+ }
+
+ return res;
+}
+
+
/**
* Generate a - b
*/
@@ -689,12 +720,12 @@ lp_build_negate(struct lp_build_context *bld,
}
+/** Return -1, 0 or +1 depending on the sign of a */
LLVMValueRef
lp_build_sgn(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
- LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMValueRef cond;
LLVMValueRef res;
@@ -704,14 +735,29 @@ lp_build_sgn(struct lp_build_context *bld,
res = bld->one;
}
else if(type.floating) {
- /* Take the sign bit and add it to 1 constant */
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
- LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
+ LLVMTypeRef vec_type;
+ LLVMTypeRef int_type;
+ LLVMValueRef mask;
LLVMValueRef sign;
LLVMValueRef one;
- sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
+ unsigned long long maskBit = (unsigned long long)1 << (type.width - 1);
+
+ if (type.length == 1) {
+ int_type = lp_build_int_elem_type(type);
+ vec_type = lp_build_elem_type(type);
+ mask = LLVMConstInt(int_type, maskBit, 0);
+ }
+ else {
+ /* vector */
+ int_type = lp_build_int_vec_type(type);
+ vec_type = lp_build_vec_type(type);
+ mask = lp_build_int_const_scalar(type, maskBit);
+ }
+
+ /* Take the sign bit and add it to 1 constant */
+ sign = LLVMBuildBitCast(bld->builder, a, int_type, "");
sign = LLVMBuildAnd(bld->builder, sign, mask, "");
- one = LLVMConstBitCast(bld->one, int_vec_type);
+ one = LLVMConstBitCast(bld->one, int_type);
res = LLVMBuildOr(bld->builder, sign, one, "");
res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
}
@@ -883,6 +929,10 @@ lp_build_floor(struct lp_build_context *bld,
assert(type.floating);
+ if (type.length == 1) {
+ return LLVMBuildFPTrunc(bld->builder, a, LLVMFloatType(), "");
+ }
+
if(util_cpu_caps.has_sse4_1)
return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
else {
@@ -953,6 +1003,9 @@ lp_build_itrunc(struct lp_build_context *bld,
}
+/**
+ * Convert float[] to int[] with round().
+ */
LLVMValueRef
lp_build_iround(struct lp_build_context *bld,
LLVMValueRef a)
@@ -1013,6 +1066,14 @@ lp_build_ifloor(struct lp_build_context *bld,
LLVMValueRef res;
assert(type.floating);
+
+ if (type.length == 1) {
+ /* scalar float to int */
+ LLVMTypeRef int_type = LLVMIntType(type.width);
+ res = LLVMBuildFPToSI(bld->builder, a, int_type, "");
+ return res;
+ }
+
assert(lp_check_value(type, a));
if(util_cpu_caps.has_sse4_1) {
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.h b/src/gallium/auxiliary/gallivm/lp_bld_arit.h
index 55385e3a66..7a10fe1220 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_arit.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.h
@@ -37,7 +37,7 @@
#define LP_BLD_ARIT_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
struct lp_type;
@@ -57,6 +57,10 @@ lp_build_add(struct lp_build_context *bld,
LLVMValueRef b);
LLVMValueRef
+lp_build_sum_vector(struct lp_build_context *bld,
+ LLVMValueRef a);
+
+LLVMValueRef
lp_build_sub(struct lp_build_context *bld,
LLVMValueRef a,
LLVMValueRef b);
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_blend.h b/src/gallium/auxiliary/gallivm/lp_bld_blend.h
index da272e549f..5a9e1c1fb2 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_blend.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_blend.h
@@ -40,7 +40,7 @@
* for a standalone example.
*/
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
#include "pipe/p_format.h"
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_const.c b/src/gallium/auxiliary/gallivm/lp_bld_const.c
index 53447757e8..8a275fa72f 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_const.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_const.c
@@ -221,8 +221,16 @@ lp_build_undef(struct lp_type type)
LLVMValueRef
lp_build_zero(struct lp_type type)
{
- LLVMTypeRef vec_type = lp_build_vec_type(type);
- return LLVMConstNull(vec_type);
+ if (type.length == 1) {
+ if (type.floating)
+ return LLVMConstReal(LLVMFloatType(), 0.0);
+ else
+ return LLVMConstInt(LLVMIntType(type.width), 0, 0);
+ }
+ else {
+ LLVMTypeRef vec_type = lp_build_vec_type(type);
+ return LLVMConstNull(vec_type);
+ }
}
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_const.h b/src/gallium/auxiliary/gallivm/lp_bld_const.h
index cb8e1c7b00..4078636103 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_const.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_const.h
@@ -37,7 +37,7 @@
#define LP_BLD_CONST_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
#include <pipe/p_compiler.h>
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_conv.h b/src/gallium/auxiliary/gallivm/lp_bld_conv.h
index 948e68fae4..78e8155ff7 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_conv.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_conv.h
@@ -37,7 +37,7 @@
#define LP_BLD_CONV_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
struct lp_type;
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_debug.h b/src/gallium/auxiliary/gallivm/lp_bld_debug.h
index 583e6132b4..441ad94786 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_debug.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_debug.h
@@ -30,7 +30,7 @@
#define LP_BLD_DEBUG_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
#include "pipe/p_compiler.h"
#include "util/u_string.h"
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_depth.h b/src/gallium/auxiliary/gallivm/lp_bld_depth.h
index 79d6981bb5..8be80024ae 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_depth.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_depth.h
@@ -36,7 +36,7 @@
#define LP_BLD_DEPTH_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
struct pipe_depth_state;
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_flow.c b/src/gallium/auxiliary/gallivm/lp_bld_flow.c
index bc83138908..106fc03e46 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_flow.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_flow.c
@@ -308,7 +308,7 @@ lp_build_flow_scope_end(struct lp_build_flow_context *flow)
* Note: this function has no dependencies on the flow code and could
* be used elsewhere.
*/
-static LLVMBasicBlockRef
+LLVMBasicBlockRef
lp_build_insert_new_block(LLVMBuilderRef builder, const char *name)
{
LLVMBasicBlockRef current_block;
@@ -648,7 +648,9 @@ lp_build_if(struct lp_build_if_state *ctx,
ifthen->phi[i] = LLVMBuildPhi(builder, LLVMTypeOf(*flow->variables[i]), "");
/* add add the initial value of the var from the entry block */
- LLVMAddIncoming(ifthen->phi[i], flow->variables[i], &ifthen->entry_block, 1);
+ if (!LLVMIsUndef(*flow->variables[i]))
+ LLVMAddIncoming(ifthen->phi[i], flow->variables[i],
+ &ifthen->entry_block, 1);
}
/* create/insert true_block before merge_block */
@@ -695,18 +697,21 @@ lp_build_endif(struct lp_build_if_state *ctx)
{
struct lp_build_flow_context *flow = ctx->flow;
struct lp_build_flow_if *ifthen;
+ LLVMBasicBlockRef curBlock = LLVMGetInsertBlock(ctx->builder);
unsigned i;
ifthen = &lp_build_flow_pop(flow, LP_BUILD_FLOW_IF)->ifthen;
assert(ifthen);
+ /* Insert branch to the merge block from current block */
+ LLVMBuildBr(ctx->builder, ifthen->merge_block);
+
if (ifthen->false_block) {
LLVMPositionBuilderAtEnd(ctx->builder, ifthen->merge_block);
/* for each variable, update the Phi node with a (variable, block) pair */
for (i = 0; i < flow->num_variables; i++) {
assert(*flow->variables[i]);
- LLVMAddIncoming(ifthen->phi[i], flow->variables[i], &ifthen->false_block, 1);
-
+ LLVMAddIncoming(ifthen->phi[i], flow->variables[i], &curBlock, 1);
/* replace the variable ref with the phi function */
*flow->variables[i] = ifthen->phi[i];
}
@@ -742,15 +747,18 @@ lp_build_endif(struct lp_build_if_state *ctx)
ifthen->true_block, ifthen->merge_block);
}
- /* Append an unconditional Br(anch) instruction on the true_block */
- LLVMPositionBuilderAtEnd(ctx->builder, ifthen->true_block);
- LLVMBuildBr(ctx->builder, ifthen->merge_block);
+ /* Insert branch from end of true_block to merge_block */
if (ifthen->false_block) {
- /* Append an unconditional Br(anch) instruction on the false_block */
- LLVMPositionBuilderAtEnd(ctx->builder, ifthen->false_block);
+ /* Append an unconditional Br(anch) instruction on the true_block */
+ LLVMPositionBuilderAtEnd(ctx->builder, ifthen->true_block);
LLVMBuildBr(ctx->builder, ifthen->merge_block);
}
-
+ else {
+ /* No else clause.
+ * Note that we've already inserted the branch at the end of
+ * true_block. See the very first LLVMBuildBr() call in this function.
+ */
+ }
/* Resume building code at end of the ifthen->merge_block */
LLVMPositionBuilderAtEnd(ctx->builder, ifthen->merge_block);
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_flow.h b/src/gallium/auxiliary/gallivm/lp_bld_flow.h
index 4c225a0d4f..e158836549 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_flow.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_flow.h
@@ -35,7 +35,7 @@
#define LP_BLD_FLOW_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
struct lp_type;
@@ -145,7 +145,9 @@ lp_build_else(struct lp_build_if_state *ctx);
void
lp_build_endif(struct lp_build_if_state *ctx);
-
+
+LLVMBasicBlockRef
+lp_build_insert_new_block(LLVMBuilderRef builder, const char *name);
#endif /* !LP_BLD_FLOW_H */
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format.h b/src/gallium/auxiliary/gallivm/lp_bld_format.h
index 970bee379f..8972c0dc17 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_format.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format.h
@@ -34,7 +34,7 @@
* Pixel format helpers.
*/
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
#include "pipe/p_format.h"
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_interp.h b/src/gallium/auxiliary/gallivm/lp_bld_interp.h
index ca958cdf34..177b5e943e 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_interp.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_interp.h
@@ -41,7 +41,7 @@
#define LP_BLD_INTERP_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
#include "tgsi/tgsi_exec.h"
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_intr.h b/src/gallium/auxiliary/gallivm/lp_bld_intr.h
index f813f27074..7d5506c733 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_intr.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_intr.h
@@ -37,7 +37,7 @@
#define LP_BLD_INTR_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
/**
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_logic.c b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
index 7c585fda78..f3df3dd138 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_logic.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
@@ -42,6 +42,26 @@
#include "lp_bld_logic.h"
+/*
+ * XXX
+ *
+ * Selection with vector conditional like
+ *
+ * select <4 x i1> %C, %A, %B
+ *
+ * is valid IR (e.g. llvm/test/Assembler/vector-select.ll), but it is not
+ * supported on any backend.
+ *
+ * Expanding the boolean vector to full SIMD register width, as in
+ *
+ * sext <4 x i1> %C to <4 x i32>
+ *
+ * is valid and supported (e.g., llvm/test/CodeGen/X86/vec_compare.ll), but
+ * it causes assertion failures in LLVM 2.6. It appears to work correctly on
+ * LLVM 2.7.
+ */
+
+
/**
* Build code to compare two values 'a' and 'b' of 'type' using the given func.
* \param func one of PIPE_FUNC_x
@@ -54,13 +74,11 @@ lp_build_compare(LLVMBuilderRef builder,
LLVMValueRef a,
LLVMValueRef b)
{
- LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef zeros = LLVMConstNull(int_vec_type);
LLVMValueRef ones = LLVMConstAllOnes(int_vec_type);
LLVMValueRef cond;
LLVMValueRef res;
- unsigned i;
assert(func >= PIPE_FUNC_NEVER);
assert(func <= PIPE_FUNC_ALWAYS);
@@ -74,10 +92,12 @@ lp_build_compare(LLVMBuilderRef builder,
/* XXX: It is not clear if we should use the ordered or unordered operators */
+#if HAVE_LLVM < 0x0207
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
if(type.width * type.length == 128) {
if(type.floating && util_cpu_caps.has_sse) {
/* float[4] comparison */
+ LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMValueRef args[3];
unsigned cc;
boolean swap;
@@ -147,6 +167,7 @@ lp_build_compare(LLVMBuilderRef builder,
const char *pcmpgt;
LLVMValueRef args[2];
LLVMValueRef res;
+ LLVMTypeRef vec_type = lp_build_vec_type(type);
switch (type.width) {
case 8:
@@ -200,6 +221,7 @@ lp_build_compare(LLVMBuilderRef builder,
}
} /* if (type.width * type.length == 128) */
#endif
+#endif /* HAVE_LLVM < 0x0207 */
if(type.floating) {
LLVMRealPredicate op;
@@ -233,16 +255,19 @@ lp_build_compare(LLVMBuilderRef builder,
return lp_build_undef(type);
}
-#if 0
- /* XXX: Although valid IR, no LLVM target currently support this */
+#if HAVE_LLVM >= 0x0207
cond = LLVMBuildFCmp(builder, op, a, b, "");
- res = LLVMBuildSelect(builder, cond, ones, zeros, "");
+ res = LLVMBuildSExt(builder, cond, int_vec_type, "");
#else
- res = LLVMGetUndef(int_vec_type);
if (type.length == 1) {
- res = LLVMBuildFCmp(builder, op, a, b, "");
+ cond = LLVMBuildFCmp(builder, op, a, b, "");
+ res = LLVMBuildSExt(builder, cond, int_vec_type, "");
}
else {
+ unsigned i;
+
+ res = LLVMGetUndef(int_vec_type);
+
debug_printf("%s: warning: using slow element-wise float"
" vector comparison\n", __FUNCTION__);
for (i = 0; i < type.length; ++i) {
@@ -286,16 +311,19 @@ lp_build_compare(LLVMBuilderRef builder,
return lp_build_undef(type);
}
-#if 0
- /* XXX: Although valid IR, no LLVM target currently support this */
+#if HAVE_LLVM >= 0x0207
cond = LLVMBuildICmp(builder, op, a, b, "");
- res = LLVMBuildSelect(builder, cond, ones, zeros, "");
+ res = LLVMBuildSExt(builder, cond, int_vec_type, "");
#else
- res = LLVMGetUndef(int_vec_type);
if (type.length == 1) {
- res = LLVMBuildICmp(builder, op, a, b, "");
+ cond = LLVMBuildICmp(builder, op, a, b, "");
+ res = LLVMBuildSExt(builder, cond, int_vec_type, "");
}
else {
+ unsigned i;
+
+ res = LLVMGetUndef(int_vec_type);
+
debug_printf("%s: warning: using slow element-wise int"
" vector comparison\n", __FUNCTION__);
@@ -337,6 +365,8 @@ lp_build_cmp(struct lp_build_context *bld,
/**
* Return mask ? a : b;
+ *
+ * mask is a bitwise mask, composed of 0 or ~0 for each element.
*/
LLVMValueRef
lp_build_select(struct lp_build_context *bld,
@@ -351,6 +381,7 @@ lp_build_select(struct lp_build_context *bld,
return a;
if (type.length == 1) {
+ mask = LLVMBuildTrunc(bld->builder, mask, LLVMInt1Type(), "");
res = LLVMBuildSelect(bld->builder, mask, a, b, "");
}
else {
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_logic.h b/src/gallium/auxiliary/gallivm/lp_bld_logic.h
index a399ebf39e..b54ec13b70 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_logic.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_logic.h
@@ -37,7 +37,7 @@
#define LP_BLD_LOGIC_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
#include "pipe/p_defines.h" /* For PIPE_FUNC_xxx */
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_pack.c b/src/gallium/auxiliary/gallivm/lp_bld_pack.c
index 4c61d10749..23398f41f9 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_pack.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_pack.c
@@ -256,7 +256,7 @@ lp_build_pack2(LLVMBuilderRef builder,
LLVMValueRef lo,
LLVMValueRef hi)
{
-#if !(HAVE_LLVM >= 0x0207)
+#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);
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_pack.h b/src/gallium/auxiliary/gallivm/lp_bld_pack.h
index fb2a34984a..346a17d580 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_pack.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_pack.h
@@ -37,7 +37,7 @@
#define LP_BLD_PACK_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
struct lp_type;
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample.c b/src/gallium/auxiliary/gallivm/lp_bld_sample.c
index 311c9f1b9e..543fd5fea3 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample.c
@@ -155,14 +155,16 @@ lp_build_gather(LLVMBuilderRef builder,
/**
* Compute the offset of a pixel.
*
- * x, y, y_stride are vectors
+ * x, y, z, y_stride, z_stride are vectors
*/
LLVMValueRef
lp_build_sample_offset(struct lp_build_context *bld,
const struct util_format_description *format_desc,
LLVMValueRef x,
LLVMValueRef y,
- LLVMValueRef y_stride)
+ LLVMValueRef z,
+ LLVMValueRef y_stride,
+ LLVMValueRef z_stride)
{
LLVMValueRef x_stride;
LLVMValueRef offset;
@@ -178,6 +180,10 @@ lp_build_sample_offset(struct lp_build_context *bld,
LLVMValueRef y_offset_lo, y_offset_hi;
LLVMValueRef offset_lo, offset_hi;
+ /* XXX 1D & 3D addressing not done yet */
+ assert(!z);
+ assert(!z_stride);
+
x_lo = LLVMBuildAnd(bld->builder, x, bld->one, "");
y_lo = LLVMBuildAnd(bld->builder, y, bld->one, "");
@@ -201,13 +207,17 @@ lp_build_sample_offset(struct lp_build_context *bld,
offset = lp_build_add(bld, offset_hi, offset_lo);
}
else {
- LLVMValueRef x_offset;
- LLVMValueRef y_offset;
+ offset = lp_build_mul(bld, x, x_stride);
- x_offset = lp_build_mul(bld, x, x_stride);
- y_offset = lp_build_mul(bld, y, y_stride);
+ if (y && y_stride) {
+ LLVMValueRef y_offset = lp_build_mul(bld, y, y_stride);
+ offset = lp_build_add(bld, offset, y_offset);
+ }
- offset = lp_build_add(bld, x_offset, y_offset);
+ if (z && z_stride) {
+ LLVMValueRef z_offset = lp_build_mul(bld, z, z_stride);
+ offset = lp_build_add(bld, offset, z_offset);
+ }
}
return offset;
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample.h b/src/gallium/auxiliary/gallivm/lp_bld_sample.h
index 68db91d6fd..7f08bfaac1 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample.h
@@ -36,7 +36,7 @@
#define LP_BLD_SAMPLE_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
struct pipe_texture;
struct pipe_sampler_state;
@@ -113,9 +113,9 @@ struct lp_sampler_dynamic_state
unsigned unit);
LLVMValueRef
- (*stride)( struct lp_sampler_dynamic_state *state,
- LLVMBuilderRef builder,
- unsigned unit);
+ (*row_stride)( struct lp_sampler_dynamic_state *state,
+ LLVMBuilderRef builder,
+ unsigned unit);
LLVMValueRef
(*data_ptr)( struct lp_sampler_dynamic_state *state,
@@ -148,7 +148,9 @@ lp_build_sample_offset(struct lp_build_context *bld,
const struct util_format_description *format_desc,
LLVMValueRef x,
LLVMValueRef y,
- LLVMValueRef y_stride);
+ LLVMValueRef z,
+ LLVMValueRef y_stride,
+ LLVMValueRef z_stride);
void
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
index a965d394f4..72018737a8 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
@@ -48,6 +48,7 @@
#include "lp_bld_logic.h"
#include "lp_bld_swizzle.h"
#include "lp_bld_pack.h"
+#include "lp_bld_flow.h"
#include "lp_bld_format.h"
#include "lp_bld_sample.h"
@@ -139,7 +140,55 @@ lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
/**
- * Gen code to fetch a texel from a texture at int coords (x, y).
+ * Dereference stride_array[mipmap_level] array to get a stride.
+ * Return stride as a vector.
+ */
+static LLVMValueRef
+lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
+ LLVMValueRef stride_array, LLVMValueRef level)
+{
+ LLVMValueRef indexes[2], stride;
+ indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
+ indexes[1] = level;
+ stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, "");
+ stride = LLVMBuildLoad(bld->builder, stride, "");
+ stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
+ return stride;
+}
+
+
+/** Dereference stride_array[0] array to get a stride (as vector). */
+static LLVMValueRef
+lp_build_get_const_level_stride_vec(struct lp_build_sample_context *bld,
+ LLVMValueRef stride_array, int level)
+{
+ LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
+ return lp_build_get_level_stride_vec(bld, stride_array, lvl);
+}
+
+
+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 fetch a texel from a texture at int coords (x, y, z).
+ * The computation depends on whether the texture is 1D, 2D or 3D.
* The result, texel, will be:
* texel[0] = red values
* texel[1] = green values
@@ -150,12 +199,16 @@ static void
lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
LLVMValueRef width,
LLVMValueRef height,
+ LLVMValueRef depth,
LLVMValueRef x,
LLVMValueRef y,
+ LLVMValueRef z,
LLVMValueRef y_stride,
+ LLVMValueRef z_stride,
LLVMValueRef data_ptr,
LLVMValueRef *texel)
{
+ const int dims = texture_dims(bld->static_state->target);
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
LLVMValueRef offset;
LLVMValueRef packed;
@@ -169,7 +222,7 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
}
- if (wrap_mode_uses_border_color(bld->static_state->wrap_t)) {
+ if (dims >= 2 && 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);
@@ -182,6 +235,19 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
}
}
+ if (dims == 3 && wrap_mode_uses_border_color(bld->static_state->wrap_r)) {
+ LLVMValueRef b1, b2;
+ b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
+ b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
+ 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
@@ -197,10 +263,10 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
* the texel color results with the border color.
*/
- /* convert x,y coords to linear offset from start of texture, in bytes */
+ /* convert x,y,z 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);
+ x, y, z, y_stride, z_stride);
assert(bld->format_desc->block.width == 1);
assert(bld->format_desc->block.height == 1);
@@ -213,6 +279,8 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
bld->texel_type.width,
data_ptr, offset);
+ texel[0] = texel[1] = texel[2] = texel[3] = NULL;
+
/* convert texels to float rgba */
lp_build_unpack_rgba_soa(bld->builder,
bld->format_desc,
@@ -245,7 +313,7 @@ lp_build_sample_packed(struct lp_build_sample_context *bld,
offset = lp_build_sample_offset(&bld->uint_coord_bld,
bld->format_desc,
- x, y, y_stride);
+ x, y, NULL, y_stride, NULL);
assert(bld->format_desc->block.width == 1);
assert(bld->format_desc->block.height == 1);
@@ -756,24 +824,6 @@ lp_build_minify(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
@@ -794,7 +844,6 @@ lp_build_lod_selector(struct lp_build_sample_context *bld,
{
const int dims = texture_dims(bld->static_state->target);
- struct lp_build_context *coord_bld = &bld->coord_bld;
struct lp_build_context *float_bld = &bld->float_bld;
LLVMValueRef lod_bias = LLVMConstReal(LLVMFloatType(), bld->static_state->lod_bias);
LLVMValueRef min_lod = LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod);
@@ -921,161 +970,635 @@ lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
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;
+ struct lp_build_context *float_bld = &bld->float_bld;
+ struct lp_build_context *int_bld = &bld->int_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);
+ level = lp_build_ifloor(float_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,
+ *level0_out = lp_build_clamp(int_bld, level,
+ int_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);
+ *level1_out = lp_build_add(int_bld, *level0_out, int_bld->one);
+ *level1_out = lp_build_min(int_bld, *level1_out, int_bld->zero);
- *weight_out = lp_build_fract(coord_bld, lod);
+ *weight_out = lp_build_fract(float_bld, lod);
}
-
/**
- * Sample 2D texture with nearest filtering, no mipmapping.
+ * Generate code to sample a mipmap level with nearest filtering.
+ * If sampling a cube texture, r = cube face in [0,5].
*/
static void
-lp_build_sample_2d_nearest_soa(struct lp_build_sample_context *bld,
- LLVMValueRef s,
- LLVMValueRef t,
- LLVMValueRef width,
- LLVMValueRef height,
- LLVMValueRef stride,
- LLVMValueRef data_array,
- LLVMValueRef *texel)
+lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
+ LLVMValueRef width_vec,
+ LLVMValueRef height_vec,
+ LLVMValueRef depth_vec,
+ LLVMValueRef row_stride_vec,
+ LLVMValueRef img_stride_vec,
+ LLVMValueRef data_ptr,
+ LLVMValueRef s,
+ LLVMValueRef t,
+ LLVMValueRef r,
+ LLVMValueRef colors_out[4])
{
- LLVMValueRef x, y;
- LLVMValueRef data_ptr;
+ const int dims = texture_dims(bld->static_state->target);
+ LLVMValueRef x, y, z;
- x = lp_build_sample_wrap_nearest(bld, s, width,
+ /*
+ * Compute integer texcoords.
+ */
+ x = lp_build_sample_wrap_nearest(bld, s, width_vec,
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);
-
lp_build_name(x, "tex.x.wrapped");
- lp_build_name(y, "tex.y.wrapped");
- /* get pointer to mipmap level 0 data */
- data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0);
+ if (dims >= 2) {
+ y = lp_build_sample_wrap_nearest(bld, t, height_vec,
+ bld->static_state->pot_height,
+ bld->static_state->wrap_t);
+ lp_build_name(y, "tex.y.wrapped");
+
+ if (dims == 3) {
+ z = lp_build_sample_wrap_nearest(bld, r, depth_vec,
+ bld->static_state->pot_height,
+ bld->static_state->wrap_r);
+ lp_build_name(z, "tex.z.wrapped");
+ }
+ else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
+ z = r;
+ }
+ else {
+ z = NULL;
+ }
+ }
+ else {
+ y = z = NULL;
+ }
- lp_build_sample_texel_soa(bld, width, height, x, y, stride, data_ptr, texel);
+ /*
+ * Get texture colors.
+ */
+ lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ x, y, z,
+ row_stride_vec, img_stride_vec,
+ data_ptr, colors_out);
}
/**
- * Sample 2D texture with nearest filtering, nearest mipmap.
+ * Generate code to sample a mipmap level with linear filtering.
+ * If sampling a cube texture, r = cube face in [0,5].
*/
static void
-lp_build_sample_2d_nearest_mip_nearest_soa(struct lp_build_sample_context *bld,
- unsigned unit,
- LLVMValueRef s,
- LLVMValueRef t,
- LLVMValueRef width,
- LLVMValueRef height,
- LLVMValueRef width_vec,
- LLVMValueRef height_vec,
- LLVMValueRef stride,
- LLVMValueRef data_array,
- LLVMValueRef *texel)
+lp_build_sample_image_linear(struct lp_build_sample_context *bld,
+ LLVMValueRef width_vec,
+ LLVMValueRef height_vec,
+ LLVMValueRef depth_vec,
+ LLVMValueRef row_stride_vec,
+ LLVMValueRef img_stride_vec,
+ LLVMValueRef data_ptr,
+ LLVMValueRef s,
+ LLVMValueRef t,
+ LLVMValueRef r,
+ LLVMValueRef colors_out[4])
{
- LLVMValueRef x, y;
- LLVMValueRef lod, ilevel, ilevel_vec;
- LLVMValueRef data_ptr;
+ const int dims = texture_dims(bld->static_state->target);
+ LLVMValueRef x0, y0, z0, x1, y1, z1;
+ LLVMValueRef s_fpart, t_fpart, r_fpart;
+ LLVMValueRef neighbors[2][2][4];
+ int chan;
- /* compute float LOD */
- lod = lp_build_lod_selector(bld, s, t, NULL, width, height, NULL);
+ /*
+ * Compute integer texcoords.
+ */
+ lp_build_sample_wrap_linear(bld, s, width_vec,
+ bld->static_state->pot_width,
+ bld->static_state->wrap_s,
+ &x0, &x1, &s_fpart);
+ lp_build_name(x0, "tex.x0.wrapped");
+ lp_build_name(x1, "tex.x1.wrapped");
+
+ if (dims >= 2) {
+ lp_build_sample_wrap_linear(bld, t, height_vec,
+ bld->static_state->pot_height,
+ bld->static_state->wrap_t,
+ &y0, &y1, &t_fpart);
+ lp_build_name(y0, "tex.y0.wrapped");
+ lp_build_name(y1, "tex.y1.wrapped");
+
+ if (dims == 3) {
+ lp_build_sample_wrap_linear(bld, r, depth_vec,
+ bld->static_state->pot_depth,
+ bld->static_state->wrap_r,
+ &z0, &z1, &r_fpart);
+ lp_build_name(z0, "tex.z0.wrapped");
+ lp_build_name(z1, "tex.z1.wrapped");
+ }
+ else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
+ z0 = z1 = r; /* cube face */
+ r_fpart = NULL;
+ }
+ else {
+ z0 = z1 = NULL;
+ r_fpart = NULL;
+ }
+ }
+ else {
+ y0 = y1 = t_fpart = NULL;
+ z0 = z1 = r_fpart = NULL;
+ }
- /* convert LOD to int */
- lp_build_nearest_mip_level(bld, unit, lod, &ilevel);
+ /*
+ * Get texture colors.
+ */
+ /* get x0/x1 texels */
+ lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ x0, y0, z0,
+ row_stride_vec, img_stride_vec,
+ data_ptr, neighbors[0][0]);
+ lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ x1, y0, z0,
+ row_stride_vec, img_stride_vec,
+ data_ptr, neighbors[0][1]);
+
+ if (dims == 1) {
+ /* Interpolate two samples from 1D image to produce one color */
+ for (chan = 0; chan < 4; chan++) {
+ colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
+ neighbors[0][0][chan],
+ neighbors[0][1][chan]);
+ }
+ }
+ else {
+ /* 2D/3D texture */
+ LLVMValueRef colors0[4];
+
+ /* get x0/x1 texels at y1 */
+ lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ x0, y1, z0,
+ row_stride_vec, img_stride_vec,
+ data_ptr, neighbors[1][0]);
+ lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ x1, y1, z0,
+ row_stride_vec, img_stride_vec,
+ data_ptr, neighbors[1][1]);
+
+ /* Bilinear interpolate the four samples from the 2D image / 3D slice */
+ for (chan = 0; chan < 4; chan++) {
+ colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
+ s_fpart, t_fpart,
+ neighbors[0][0][chan],
+ neighbors[0][1][chan],
+ neighbors[1][0][chan],
+ neighbors[1][1][chan]);
+ }
- ilevel_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel);
+ if (dims == 3) {
+ LLVMValueRef neighbors1[2][2][4];
+ LLVMValueRef colors1[4];
+
+ /* get x0/x1/y0/y1 texels at z1 */
+ lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ x0, y0, z1,
+ row_stride_vec, img_stride_vec,
+ data_ptr, neighbors1[0][0]);
+ lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ x1, y0, z1,
+ row_stride_vec, img_stride_vec,
+ data_ptr, neighbors1[0][1]);
+ lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ x0, y1, z1,
+ row_stride_vec, img_stride_vec,
+ data_ptr, neighbors1[1][0]);
+ lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
+ x1, y1, z1,
+ row_stride_vec, img_stride_vec,
+ data_ptr, neighbors1[1][1]);
+
+ /* Bilinear interpolate the four samples from the second Z slice */
+ for (chan = 0; chan < 4; chan++) {
+ colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
+ s_fpart, t_fpart,
+ neighbors1[0][0][chan],
+ neighbors1[0][1][chan],
+ neighbors1[1][0][chan],
+ neighbors1[1][1][chan]);
+ }
- /* compute width_vec, height at mipmap level 'ilevel' */
- width_vec = lp_build_minify(bld, width_vec, ilevel_vec);
- height_vec = lp_build_minify(bld, height_vec, ilevel_vec);
- stride = lp_build_minify(bld, stride, ilevel_vec);
+ /* Linearly interpolate the two samples from the two 3D slices */
+ for (chan = 0; chan < 4; chan++) {
+ colors_out[chan] = lp_build_lerp(&bld->texel_bld,
+ r_fpart,
+ colors0[chan], colors1[chan]);
+ }
+ }
+ else {
+ /* 2D tex */
+ for (chan = 0; chan < 4; chan++) {
+ colors_out[chan] = colors0[chan];
+ }
+ }
+ }
+}
- x = lp_build_sample_wrap_nearest(bld, s, width_vec,
- bld->static_state->pot_width,
- bld->static_state->wrap_s);
- y = lp_build_sample_wrap_nearest(bld, t, height_vec,
- bld->static_state->pot_height,
- bld->static_state->wrap_t);
- lp_build_name(x, "tex.x.wrapped");
- lp_build_name(y, "tex.y.wrapped");
+/** Helper used by lp_build_cube_lookup() */
+static LLVMValueRef
+lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord)
+{
+ /* ima = -0.5 / abs(coord); */
+ LLVMValueRef negHalf = lp_build_const_scalar(coord_bld->type, -0.5);
+ LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
+ LLVMValueRef ima = lp_build_mul(coord_bld, negHalf,
+ lp_build_rcp(coord_bld, absCoord));
+ return ima;
+}
- /* get pointer to mipmap level [ilevel] data */
- if (0)
- data_ptr = lp_build_get_mipmap_level(bld, data_array, ilevel);
- else
- data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0);
- lp_build_sample_texel_soa(bld, width_vec, height_vec, x, y, stride, data_ptr, texel);
+/**
+ * Helper used by lp_build_cube_lookup()
+ * \param sign scalar +1 or -1
+ * \param coord float vector
+ * \param ima float vector
+ */
+static LLVMValueRef
+lp_build_cube_coord(struct lp_build_context *coord_bld,
+ LLVMValueRef sign, int negate_coord,
+ LLVMValueRef coord, LLVMValueRef ima)
+{
+ /* return negate(coord) * ima * sign + 0.5; */
+ LLVMValueRef half = lp_build_const_scalar(coord_bld->type, 0.5);
+ LLVMValueRef res;
+
+ assert(negate_coord == +1 || negate_coord == -1);
+
+ if (negate_coord == -1) {
+ coord = lp_build_negate(coord_bld, coord);
+ }
+
+ res = lp_build_mul(coord_bld, coord, ima);
+ if (sign) {
+ sign = lp_build_broadcast_scalar(coord_bld, sign);
+ res = lp_build_mul(coord_bld, res, sign);
+ }
+ res = lp_build_add(coord_bld, res, half);
+
+ return res;
}
+/** Helper used by lp_build_cube_lookup()
+ * Return (major_coord >= 0) ? pos_face : neg_face;
+ */
+static LLVMValueRef
+lp_build_cube_face(struct lp_build_sample_context *bld,
+ LLVMValueRef major_coord,
+ unsigned pos_face, unsigned neg_face)
+{
+ LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
+ major_coord,
+ bld->float_bld.zero, "");
+ LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0);
+ LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0);
+ LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, "");
+ return res;
+}
+
+
+
/**
- * Sample 2D texture with bilinear filtering.
+ * Generate code to do cube face selection and per-face texcoords.
*/
static void
-lp_build_sample_2d_linear_soa(struct lp_build_sample_context *bld,
- LLVMValueRef s,
- LLVMValueRef t,
- LLVMValueRef width,
- LLVMValueRef height,
- LLVMValueRef stride,
- LLVMValueRef data_array,
- LLVMValueRef *texel)
+lp_build_cube_lookup(struct lp_build_sample_context *bld,
+ LLVMValueRef s,
+ LLVMValueRef t,
+ LLVMValueRef r,
+ LLVMValueRef *face,
+ LLVMValueRef *face_s,
+ LLVMValueRef *face_t)
{
- LLVMValueRef s_fpart;
- LLVMValueRef t_fpart;
- LLVMValueRef x0, x1;
- LLVMValueRef y0, y1;
- LLVMValueRef neighbors[2][2][4];
- LLVMValueRef data_ptr;
- unsigned chan;
+ struct lp_build_context *float_bld = &bld->float_bld;
+ struct lp_build_context *coord_bld = &bld->coord_bld;
+ LLVMValueRef rx, ry, rz;
+ LLVMValueRef arx, ary, arz;
+ LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 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;
+ LLVMValueRef rx_pos, ry_pos, rz_pos;
- 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);
+ assert(bld->coord_bld.type.length == 4);
- /* get pointer to mipmap level 0 data */
- data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0);
+ /*
+ * 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));
- 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]);
+ arx = lp_build_abs(float_bld, rx);
+ ary = lp_build_abs(float_bld, ry);
+ arz = lp_build_abs(float_bld, rz);
- /* TODO: Don't interpolate missing channels */
- for(chan = 0; chan < 4; ++chan) {
- texel[chan] = lp_build_lerp_2d(&bld->texel_bld,
- s_fpart, t_fpart,
- neighbors[0][0][chan],
- neighbors[0][1][chan],
- neighbors[1][0][chan],
- neighbors[1][1][chan]);
+ /*
+ * Compare sign/magnitude of rx,ry,rz to determine face
+ */
+ arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
+ arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
+ ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
+ ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
+
+ arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
+ ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
+
+ rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
+ ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
+ rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
+
+ {
+ struct lp_build_flow_context *flow_ctx;
+ struct lp_build_if_state if_ctx;
+
+ flow_ctx = lp_build_flow_create(bld->builder);
+ lp_build_flow_scope_begin(flow_ctx);
+
+ *face_s = bld->coord_bld.undef;
+ *face_t = bld->coord_bld.undef;
+ *face = bld->int_bld.undef;
+
+ lp_build_name(*face_s, "face_s");
+ lp_build_name(*face_t, "face_t");
+ lp_build_name(*face, "face");
+
+ lp_build_flow_scope_declare(flow_ctx, face_s);
+ lp_build_flow_scope_declare(flow_ctx, face_t);
+ lp_build_flow_scope_declare(flow_ctx, face);
+
+ lp_build_if(&if_ctx, flow_ctx, bld->builder, 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);
+ }
+ lp_build_else(&if_ctx);
+ {
+ struct lp_build_flow_context *flow_ctx2;
+ struct lp_build_if_state if_ctx2;
+
+ LLVMValueRef face_s2 = bld->coord_bld.undef;
+ LLVMValueRef face_t2 = bld->coord_bld.undef;
+ LLVMValueRef face2 = bld->int_bld.undef;
+
+ flow_ctx2 = lp_build_flow_create(bld->builder);
+ lp_build_flow_scope_begin(flow_ctx2);
+ lp_build_flow_scope_declare(flow_ctx2, &face_s2);
+ lp_build_flow_scope_declare(flow_ctx2, &face_t2);
+ lp_build_flow_scope_declare(flow_ctx2, &face2);
+
+ ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
+
+ lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz);
+ {
+ /* +/- Y face */
+ LLVMValueRef sign = lp_build_sgn(float_bld, ry);
+ LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
+ face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
+ face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
+ face2 = lp_build_cube_face(bld, ry,
+ PIPE_TEX_FACE_POS_Y,
+ PIPE_TEX_FACE_NEG_Y);
+ }
+ 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_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
+ face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
+ face2 = lp_build_cube_face(bld, rz,
+ PIPE_TEX_FACE_POS_Z,
+ PIPE_TEX_FACE_NEG_Z);
+ }
+ lp_build_endif(&if_ctx2);
+ lp_build_flow_scope_end(flow_ctx2);
+ lp_build_flow_destroy(flow_ctx2);
+
+ *face_s = face_s2;
+ *face_t = face_t2;
+ *face = face2;
+ }
+
+ lp_build_endif(&if_ctx);
+ lp_build_flow_scope_end(flow_ctx);
+ lp_build_flow_destroy(flow_ctx);
+ }
+}
+
+
+
+/**
+ * General texture sampling codegen.
+ * This function handles texture sampling for all texture targets (1D,
+ * 2D, 3D, cube) and all filtering modes.
+ */
+static void
+lp_build_sample_general(struct lp_build_sample_context *bld,
+ unsigned unit,
+ LLVMValueRef s,
+ LLVMValueRef t,
+ LLVMValueRef r,
+ LLVMValueRef width,
+ LLVMValueRef height,
+ LLVMValueRef depth,
+ LLVMValueRef width_vec,
+ LLVMValueRef height_vec,
+ LLVMValueRef depth_vec,
+ LLVMValueRef row_stride_array,
+ LLVMValueRef img_stride_vec,
+ LLVMValueRef data_array,
+ LLVMValueRef *colors_out)
+{
+ const unsigned mip_filter = bld->static_state->min_mip_filter;
+ const unsigned min_filter = bld->static_state->min_img_filter;
+ const unsigned mag_filter = bld->static_state->mag_img_filter;
+ const int dims = texture_dims(bld->static_state->target);
+ LLVMValueRef lod, lod_fpart;
+ LLVMValueRef ilevel0, ilevel1, ilevel0_vec, ilevel1_vec;
+ LLVMValueRef width0_vec = NULL, height0_vec = NULL, depth0_vec = NULL;
+ LLVMValueRef width1_vec = NULL, height1_vec = NULL, depth1_vec = NULL;
+ LLVMValueRef row_stride0_vec = NULL, row_stride1_vec = NULL;
+ LLVMValueRef img_stride0_vec = NULL, img_stride1_vec = NULL;
+ LLVMValueRef data_ptr0, data_ptr1;
+ int chan;
+
+ /*
+ printf("%s mip %d min %d mag %d\n", __FUNCTION__,
+ mip_filter, min_filter, mag_filter);
+ */
+
+ /*
+ * Compute the level of detail (mipmap level index(es)).
+ */
+ if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
+ /* always use mip level 0 */
+ ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
+ }
+ else {
+ /* compute float LOD */
+ lod = lp_build_lod_selector(bld, s, t, r, width, height, depth);
+
+ if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
+ lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
+ }
+ else {
+ assert(mip_filter == PIPE_TEX_MIPFILTER_LINEAR);
+ lp_build_linear_mip_levels(bld, unit, lod, &ilevel0, &ilevel1,
+ &lod_fpart);
+ lod_fpart = lp_build_broadcast_scalar(&bld->coord_bld, lod_fpart);
+ }
+ }
+
+ /*
+ * Convert scalar integer mipmap levels into vectors.
+ */
+ ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
+ if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
+ ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
+
+ /*
+ * Compute width, height at mipmap level 'ilevel0'
+ */
+ width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
+ if (dims >= 2) {
+ height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
+ row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
+ ilevel0);
+ if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
+ img_stride0_vec = lp_build_mul(&bld->int_coord_bld,
+ row_stride0_vec, height0_vec);
+ if (dims == 3) {
+ depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
+ }
+ }
+ }
+ if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
+ /* compute width, height, depth for second mipmap level at ilevel1 */
+ width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
+ if (dims >= 2) {
+ height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
+ row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
+ ilevel1);
+ if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
+ img_stride1_vec = lp_build_mul(&bld->int_coord_bld,
+ row_stride1_vec, height1_vec);
+ if (dims ==3) {
+ depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
+ }
+ }
+ }
+ }
+
+ /*
+ * Choose cube face, recompute texcoords.
+ */
+ if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
+ LLVMValueRef face, face_s, face_t;
+ lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
+ s = face_s; /* vec */
+ t = face_t; /* vec */
+ /* use 'r' to indicate cube face */
+ r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
+ }
+
+ /*
+ * Get pointer(s) to image data for mipmap level(s).
+ */
+ data_ptr0 = lp_build_get_mipmap_level(bld, data_array, ilevel0);
+ if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
+ data_ptr1 = lp_build_get_mipmap_level(bld, data_array, ilevel1);
+ }
+
+ /*
+ * Get/interpolate texture colors.
+ */
+ /* XXX temporarily force this path: */
+ if (1 /*min_filter == mag_filter*/) {
+ /* same filter for minification or magnification */
+ LLVMValueRef colors0[4], colors1[4];
+
+ if (min_filter == PIPE_TEX_FILTER_NEAREST) {
+ lp_build_sample_image_nearest(bld,
+ width0_vec, height0_vec, depth0_vec,
+ row_stride0_vec, img_stride0_vec,
+ data_ptr0, s, t, r, colors0);
+
+ if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
+ /* sample the second mipmap level, and interp */
+ lp_build_sample_image_nearest(bld,
+ width1_vec, height1_vec, depth1_vec,
+ row_stride1_vec, img_stride1_vec,
+ data_ptr1, s, t, r, colors1);
+ }
+ }
+ else {
+ assert(min_filter == PIPE_TEX_FILTER_LINEAR);
+
+ lp_build_sample_image_linear(bld,
+ width0_vec, height0_vec, depth0_vec,
+ row_stride0_vec, img_stride0_vec,
+ data_ptr0, s, t, r, colors0);
+
+
+ if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
+ /* sample the second mipmap level, and interp */
+ lp_build_sample_image_linear(bld,
+ width1_vec, height1_vec, depth1_vec,
+ row_stride1_vec, img_stride1_vec,
+ data_ptr1, s, t, r, colors1);
+ }
+ }
+
+ if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
+ /* interpolate samples from the two mipmap levels */
+ for (chan = 0; chan < 4; chan++) {
+ colors_out[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
+ colors0[chan], colors1[chan]);
+ }
+ }
+ else {
+ /* use first/only level's colors */
+ for (chan = 0; chan < 4; chan++) {
+ colors_out[chan] = colors0[chan];
+ }
+ }
+ }
+ else {
+ /* emit conditional to choose min image filter or mag image filter
+ * depending on the lod being >0 or <= 0, respectively.
+ */
+ abort();
}
}
+
static void
lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder,
struct lp_type dst_type,
@@ -1112,7 +1635,7 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
LLVMValueRef t,
LLVMValueRef width,
LLVMValueRef height,
- LLVMValueRef stride,
+ LLVMValueRef stride_array,
LLVMValueRef data_array,
LLVMValueRef *texel)
{
@@ -1129,6 +1652,7 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
LLVMValueRef neighbors_hi[2][2];
LLVMValueRef packed, packed_lo, packed_hi;
LLVMValueRef unswizzled[4];
+ LLVMValueRef stride;
lp_build_context_init(&i32, builder, lp_type_int_vec(32));
lp_build_context_init(&h16, builder, lp_type_ufixed(16));
@@ -1236,6 +1760,8 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, "");
}
+ stride = lp_build_get_const_level_stride_vec(bld, stride_array, 0);
+
/*
* Fetch the pixels as 4 x 32bit (rgba order might differ):
*
@@ -1359,12 +1885,12 @@ lp_build_sample_soa(LLVMBuilderRef builder,
struct lp_build_sample_context bld;
LLVMValueRef width, width_vec;
LLVMValueRef height, height_vec;
- LLVMValueRef stride, stride_vec;
+ LLVMValueRef depth, depth_vec;
+ LLVMValueRef stride_array;
LLVMValueRef data_array;
LLVMValueRef s;
LLVMValueRef t;
LLVMValueRef r;
- boolean done = FALSE;
(void) lp_build_lod_selector; /* temporary to silence warning */
(void) lp_build_nearest_mip_level;
@@ -1395,7 +1921,8 @@ lp_build_sample_soa(LLVMBuilderRef builder,
/* Get the dynamic state */
width = dynamic_state->width(dynamic_state, builder, unit);
height = dynamic_state->height(dynamic_state, builder, unit);
- stride = dynamic_state->stride(dynamic_state, builder, unit);
+ depth = dynamic_state->depth(dynamic_state, builder, unit);
+ stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
/* Note that data_array is an array[level] of pointers to texture images */
@@ -1405,58 +1932,26 @@ lp_build_sample_soa(LLVMBuilderRef builder,
width_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, width);
height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
- stride_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, stride);
-
- if(static_state->target == PIPE_TEXTURE_1D)
- t = bld.coord_bld.zero;
-
- switch (static_state->min_mip_filter) {
- case PIPE_TEX_MIPFILTER_NONE:
- break;
- case PIPE_TEX_MIPFILTER_NEAREST:
-
- switch (static_state->min_img_filter) {
- case PIPE_TEX_FILTER_NEAREST:
- lp_build_sample_2d_nearest_mip_nearest_soa(&bld, unit,
- s, t,
- width, height,
- width_vec, height_vec,
- stride_vec,
- data_array, texel);
- done = TRUE;
- break;
- }
-
- break;
- case PIPE_TEX_MIPFILTER_LINEAR:
- break;
- default:
- assert(0 && "invalid mip filter");
+ depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth);
+
+ if (lp_format_is_rgba8(bld.format_desc) &&
+ static_state->target == PIPE_TEXTURE_2D &&
+ static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
+ static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
+ static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
+ is_simple_wrap_mode(static_state->wrap_s) &&
+ is_simple_wrap_mode(static_state->wrap_t)) {
+ /* special case */
+ lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec,
+ stride_array, data_array, texel);
}
-
- if (!done) {
- switch (static_state->min_img_filter) {
- case PIPE_TEX_FILTER_NEAREST:
- lp_build_sample_2d_nearest_soa(&bld, s, t, width_vec, height_vec,
- stride_vec, data_array, texel);
- break;
- case PIPE_TEX_FILTER_LINEAR:
- 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_vec, height_vec,
- stride_vec, data_array, texel);
- else
- lp_build_sample_2d_linear_soa(&bld, s, t, width_vec, height_vec,
- stride_vec, data_array, texel);
- break;
- default:
- assert(0);
- }
+ else {
+ lp_build_sample_general(&bld, unit, s, t, r,
+ width, height, depth,
+ width_vec, height_vec, depth_vec,
+ stride_array, NULL, data_array,
+ texel);
}
- /* FIXME: respect static_state->min_mip_filter */;
- /* FIXME: respect static_state->mag_img_filter */;
-
lp_build_sample_compare(&bld, r, texel);
}
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_struct.h b/src/gallium/auxiliary/gallivm/lp_bld_struct.h
index 740392f561..34478c10f5 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_struct.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_struct.h
@@ -37,7 +37,7 @@
#define LP_BLD_STRUCT_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
#include <llvm-c/Target.h>
#include "util/u_debug.h"
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_swizzle.h b/src/gallium/auxiliary/gallivm/lp_bld_swizzle.h
index b9472127a6..57b5cc079f 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_swizzle.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_swizzle.h
@@ -37,7 +37,7 @@
#define LP_BLD_SWIZZLE_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
struct lp_type;
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h b/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h
index eddb7a83fa..0f2f8a65b1 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h
@@ -35,7 +35,7 @@
#ifndef LP_BLD_TGSI_H
#define LP_BLD_TGSI_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
struct tgsi_token;
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
index fbb664d43a..5ec59d636c 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
@@ -41,6 +41,7 @@
#include "util/u_debug.h"
#include "util/u_math.h"
#include "util/u_memory.h"
+#include "tgsi/tgsi_dump.h"
#include "tgsi/tgsi_info.h"
#include "tgsi/tgsi_parse.h"
#include "tgsi/tgsi_util.h"
@@ -95,6 +96,19 @@ struct lp_exec_mask {
int cond_stack_size;
LLVMValueRef cond_mask;
+ LLVMValueRef break_stack[LP_TGSI_MAX_NESTING];
+ int break_stack_size;
+ LLVMValueRef break_mask;
+
+ LLVMValueRef cont_stack[LP_TGSI_MAX_NESTING];
+ int cont_stack_size;
+ LLVMValueRef cont_mask;
+
+ LLVMBasicBlockRef loop_stack[LP_TGSI_MAX_NESTING];
+ int loop_stack_size;
+ LLVMBasicBlockRef loop_block;
+
+
LLVMValueRef exec_mask;
};
@@ -145,14 +159,33 @@ static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context
mask->bld = bld;
mask->has_mask = FALSE;
mask->cond_stack_size = 0;
+ mask->loop_stack_size = 0;
+ mask->break_stack_size = 0;
+ mask->cont_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;
- mask->has_mask = (mask->cond_stack_size > 0);
+ if (mask->loop_stack_size) {
+ /*for loops we need to update the entire mask at
+ * runtime */
+ LLVMValueRef tmp;
+ tmp = LLVMBuildAnd(mask->bld->builder,
+ mask->cont_mask,
+ mask->break_mask,
+ "maskcb");
+ mask->exec_mask = LLVMBuildAnd(mask->bld->builder,
+ mask->cond_mask,
+ tmp,
+ "maskfull");
+ } else
+ mask->exec_mask = mask->cond_mask;
+
+
+ mask->has_mask = (mask->cond_stack_size > 0 ||
+ mask->loop_stack_size > 0);
}
static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
@@ -189,6 +222,89 @@ static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
lp_exec_mask_update(mask);
}
+static void lp_exec_bgnloop(struct lp_exec_mask *mask)
+{
+
+ if (mask->cont_stack_size == 0)
+ mask->cont_mask = LLVMConstAllOnes(mask->int_vec_type);
+ if (mask->cont_stack_size == 0)
+ mask->break_mask = LLVMConstAllOnes(mask->int_vec_type);
+ if (mask->cond_stack_size == 0)
+ mask->cond_mask = LLVMConstAllOnes(mask->int_vec_type);
+ mask->loop_stack[mask->loop_stack_size++] = mask->loop_block;
+ mask->loop_block = lp_build_insert_new_block(mask->bld->builder, "bgnloop");
+ LLVMBuildBr(mask->bld->builder, mask->loop_block);
+ LLVMPositionBuilderAtEnd(mask->bld->builder, mask->loop_block);
+
+ lp_exec_mask_update(mask);
+}
+
+static void lp_exec_break(struct lp_exec_mask *mask)
+{
+ LLVMValueRef exec_mask = LLVMBuildNot(mask->bld->builder,
+ mask->exec_mask,
+ "break");
+
+ mask->break_stack[mask->break_stack_size++] = mask->break_mask;
+ if (mask->break_stack_size > 1) {
+ mask->break_mask = LLVMBuildAnd(mask->bld->builder,
+ mask->break_mask,
+ exec_mask, "break_full");
+ } else
+ mask->break_mask = exec_mask;
+
+ lp_exec_mask_update(mask);
+}
+
+static void lp_exec_continue(struct lp_exec_mask *mask)
+{
+ LLVMValueRef exec_mask = LLVMBuildNot(mask->bld->builder,
+ mask->exec_mask,
+ "");
+
+ mask->cont_stack[mask->cont_stack_size++] = mask->cont_mask;
+ if (mask->cont_stack_size > 1) {
+ mask->cont_mask = LLVMBuildAnd(mask->bld->builder,
+ mask->cont_mask,
+ exec_mask, "");
+ } else
+ mask->cont_mask = exec_mask;
+
+ lp_exec_mask_update(mask);
+}
+
+
+static void lp_exec_endloop(struct lp_exec_mask *mask)
+{
+ LLVMBasicBlockRef endloop;
+ LLVMTypeRef reg_type = LLVMIntType(mask->bld->type.width*
+ mask->bld->type.length);
+ /* i1cond = (mask == 0) */
+ LLVMValueRef i1cond = LLVMBuildICmp(
+ mask->bld->builder,
+ LLVMIntNE,
+ LLVMBuildBitCast(mask->bld->builder, mask->break_mask, reg_type, ""),
+ LLVMConstNull(reg_type), "");
+
+ endloop = lp_build_insert_new_block(mask->bld->builder, "endloop");
+
+ LLVMBuildCondBr(mask->bld->builder,
+ i1cond, mask->loop_block, endloop);
+
+ LLVMPositionBuilderAtEnd(mask->bld->builder, endloop);
+
+ mask->loop_block = mask->loop_stack[--mask->loop_stack_size];
+ /* pop the break mask */
+ if (mask->cont_stack_size) {
+ mask->cont_mask = mask->cont_stack[--mask->cont_stack_size];
+ }
+ if (mask->break_stack_size) {
+ mask->break_mask = mask->cont_stack[--mask->break_stack_size];
+ }
+
+ lp_exec_mask_update(mask);
+}
+
static void lp_exec_mask_store(struct lp_exec_mask *mask,
LLVMValueRef val,
LLVMValueRef dst)
@@ -1363,14 +1479,15 @@ emit_instruction(
case TGSI_OPCODE_TXP:
emit_tex( bld, inst, FALSE, TRUE, dst0 );
break;
-
+
case TGSI_OPCODE_BRK:
- /* FIXME */
- return 0;
+ lp_exec_break(&bld->exec_mask);
break;
case TGSI_OPCODE_IF:
tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
+ tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL,
+ tmp0, bld->base.zero);
lp_exec_mask_cond_push(&bld->exec_mask, tmp0);
break;
@@ -1380,6 +1497,10 @@ emit_instruction(
return 0;
break;
+ case TGSI_OPCODE_BGNLOOP:
+ lp_exec_bgnloop(&bld->exec_mask);
+ break;
+
case TGSI_OPCODE_REP:
/* deprecated */
assert(0);
@@ -1400,6 +1521,10 @@ emit_instruction(
return 0;
break;
+ case TGSI_OPCODE_ENDLOOP:
+ lp_exec_endloop(&bld->exec_mask);
+ break;
+
case TGSI_OPCODE_ENDREP:
/* deprecated */
assert(0);
@@ -1499,8 +1624,7 @@ emit_instruction(
break;
case TGSI_OPCODE_CONT:
- /* FIXME */
- return 0;
+ lp_exec_continue(&bld->exec_mask);
break;
case TGSI_OPCODE_EMIT:
@@ -1603,7 +1727,14 @@ lp_build_tgsi_soa(LLVMBuilderRef builder,
assert( 0 );
}
}
-
+ if (0) {
+ LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
+ LLVMValueRef function = LLVMGetBasicBlockParent(block);
+ debug_printf("11111111111111111111111111111 \n");
+ tgsi_dump(tokens, 0);
+ LLVMDumpValue(function);
+ debug_printf("2222222222222222222222222222 \n");
+ }
tgsi_parse_free( &parse );
}
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_type.c b/src/gallium/auxiliary/gallivm/lp_bld_type.c
index c327ba045a..796af88caa 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_type.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_type.c
@@ -58,7 +58,10 @@ LLVMTypeRef
lp_build_vec_type(struct lp_type type)
{
LLVMTypeRef elem_type = lp_build_elem_type(type);
- return LLVMVectorType(elem_type, type.length);
+ if (type.length == 1)
+ return elem_type;
+ else
+ return LLVMVectorType(elem_type, type.length);
}
@@ -115,6 +118,9 @@ lp_check_vec_type(struct lp_type type, LLVMTypeRef vec_type)
if(!vec_type)
return FALSE;
+ if (type.length == 1)
+ return lp_check_elem_type(type, vec_type);
+
if(LLVMGetTypeKind(vec_type) != LLVMVectorTypeKind)
return FALSE;
@@ -153,7 +159,10 @@ LLVMTypeRef
lp_build_int_vec_type(struct lp_type type)
{
LLVMTypeRef elem_type = lp_build_int_elem_type(type);
- return LLVMVectorType(elem_type, type.length);
+ if (type.length == 1)
+ return elem_type;
+ else
+ return LLVMVectorType(elem_type, type.length);
}
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_type.h b/src/gallium/auxiliary/gallivm/lp_bld_type.h
index 4daa904e63..5b351476ac 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_type.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_type.h
@@ -37,7 +37,7 @@
#define LP_BLD_TYPE_H
-#include <llvm-c/Core.h>
+#include "os/os_llvm.h"
#include <pipe/p_compiler.h>