Mesa (master): gallivm: export unordered/ordered cmp to a common function
Zack Rusin
zack at kemper.freedesktop.org
Fri Jul 19 20:34:23 UTC 2013
Module: Mesa
Branch: master
Commit: 018c69ac56ee0579b5d4f138340444b326c4e6dc
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=018c69ac56ee0579b5d4f138340444b326c4e6dc
Author: Zack Rusin <zackr at vmware.com>
Date: Thu Jul 18 03:05:01 2013 -0400
gallivm: export unordered/ordered cmp to a common function
Only the floating point operarators change everything else
is the same so it makes sense to share the code.
Signed-off-by: Zack Rusin <zackr at vmware.com>
Reviewed-by: Jose Fonseca <jfonseca at vmware.com>
Reviewed-by: Roland Scheidegger <sroland at vmware.com>
---
src/gallium/auxiliary/gallivm/lp_bld_logic.c | 441 +++++++++-----------------
1 files changed, 158 insertions(+), 283 deletions(-)
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_logic.c b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
index 8b800cf..322d385 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_logic.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
@@ -68,14 +68,17 @@
/**
* Build code to compare two values 'a' and 'b' of 'type' using the given func.
* \param func one of PIPE_FUNC_x
+ * If the ordered argument is true the function will use LLVM's ordered
+ * comparisons, otherwise unordered comparisons will be used.
* The result values will be 0 for false or ~0 for true.
*/
-LLVMValueRef
-lp_build_compare(struct gallivm_state *gallivm,
- const struct lp_type type,
- unsigned func,
- LLVMValueRef a,
- LLVMValueRef b)
+static LLVMValueRef
+lp_build_compare_ext(struct gallivm_state *gallivm,
+ const struct lp_type type,
+ unsigned func,
+ LLVMValueRef a,
+ LLVMValueRef b,
+ boolean ordered)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, type);
@@ -94,153 +97,6 @@ lp_build_compare(struct gallivm_state *gallivm,
if(func == PIPE_FUNC_ALWAYS)
return ones;
-#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
- /*
- * There are no unsigned integer comparison instructions in SSE.
- */
-
- if (!type.floating && !type.sign &&
- type.width * type.length == 128 &&
- util_cpu_caps.has_sse2 &&
- (func == PIPE_FUNC_LESS ||
- func == PIPE_FUNC_LEQUAL ||
- func == PIPE_FUNC_GREATER ||
- func == PIPE_FUNC_GEQUAL) &&
- (gallivm_debug & GALLIVM_DEBUG_PERF)) {
- debug_printf("%s: inefficient <%u x i%u> unsigned comparison\n",
- __FUNCTION__, type.length, type.width);
- }
-#endif
-
-#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(gallivm, type);
- LLVMValueRef args[3];
- unsigned cc;
- boolean swap;
-
- swap = FALSE;
- switch(func) {
- case PIPE_FUNC_EQUAL:
- cc = 0;
- break;
- case PIPE_FUNC_NOTEQUAL:
- cc = 4;
- break;
- case PIPE_FUNC_LESS:
- cc = 1;
- break;
- case PIPE_FUNC_LEQUAL:
- cc = 2;
- break;
- case PIPE_FUNC_GREATER:
- cc = 1;
- swap = TRUE;
- break;
- case PIPE_FUNC_GEQUAL:
- cc = 2;
- swap = TRUE;
- break;
- default:
- assert(0);
- return lp_build_undef(gallivm, type);
- }
-
- if(swap) {
- args[0] = b;
- args[1] = a;
- }
- else {
- args[0] = a;
- args[1] = b;
- }
-
- args[2] = LLVMConstInt(LLVMInt8TypeInContext(gallivm->context), cc, 0);
- res = lp_build_intrinsic(builder,
- "llvm.x86.sse.cmp.ps",
- vec_type,
- args, 3);
- res = LLVMBuildBitCast(builder, res, int_vec_type, "");
- return res;
- }
- else if(util_cpu_caps.has_sse2) {
- /* int[4] comparison */
- static const struct {
- unsigned swap:1;
- unsigned eq:1;
- unsigned gt:1;
- unsigned not:1;
- } table[] = {
- {0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
- {1, 0, 1, 0}, /* PIPE_FUNC_LESS */
- {0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
- {0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
- {0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
- {0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
- {1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
- {0, 0, 0, 0} /* PIPE_FUNC_ALWAYS */
- };
- const char *pcmpeq;
- const char *pcmpgt;
- LLVMValueRef args[2];
- LLVMValueRef res;
- LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
-
- switch (type.width) {
- case 8:
- pcmpeq = "llvm.x86.sse2.pcmpeq.b";
- pcmpgt = "llvm.x86.sse2.pcmpgt.b";
- break;
- case 16:
- pcmpeq = "llvm.x86.sse2.pcmpeq.w";
- pcmpgt = "llvm.x86.sse2.pcmpgt.w";
- break;
- case 32:
- pcmpeq = "llvm.x86.sse2.pcmpeq.d";
- pcmpgt = "llvm.x86.sse2.pcmpgt.d";
- break;
- default:
- assert(0);
- return lp_build_undef(gallivm, type);
- }
-
- /* There are no unsigned comparison instructions. So flip the sign bit
- * so that the results match.
- */
- if (table[func].gt && !type.sign) {
- LLVMValueRef msb = lp_build_const_int_vec(gallivm, type, (unsigned long long)1 << (type.width - 1));
- a = LLVMBuildXor(builder, a, msb, "");
- b = LLVMBuildXor(builder, b, msb, "");
- }
-
- if(table[func].swap) {
- args[0] = b;
- args[1] = a;
- }
- else {
- args[0] = a;
- args[1] = b;
- }
-
- if(table[func].eq)
- res = lp_build_intrinsic(builder, pcmpeq, vec_type, args, 2);
- else if (table[func].gt)
- res = lp_build_intrinsic(builder, pcmpgt, vec_type, args, 2);
- else
- res = LLVMConstNull(vec_type);
-
- if(table[func].not)
- res = LLVMBuildNot(builder, res, "");
-
- return res;
- }
- } /* if (type.width * type.length == 128) */
-#endif
-#endif /* HAVE_LLVM < 0x0207 */
-
if(type.floating) {
LLVMRealPredicate op;
switch(func) {
@@ -251,22 +107,22 @@ lp_build_compare(struct gallivm_state *gallivm,
op = LLVMRealPredicateTrue;
break;
case PIPE_FUNC_EQUAL:
- op = LLVMRealUEQ;
+ op = ordered ? LLVMRealOEQ : LLVMRealUEQ;
break;
case PIPE_FUNC_NOTEQUAL:
- op = LLVMRealUNE;
+ op = ordered ? LLVMRealONE : LLVMRealUNE;
break;
case PIPE_FUNC_LESS:
- op = LLVMRealULT;
+ op = ordered ? LLVMRealOLT : LLVMRealULT;
break;
case PIPE_FUNC_LEQUAL:
- op = LLVMRealULE;
+ op = ordered ? LLVMRealOLE : LLVMRealULE;
break;
case PIPE_FUNC_GREATER:
- op = LLVMRealUGT;
+ op = ordered ? LLVMRealOGT : LLVMRealUGT;
break;
case PIPE_FUNC_GEQUAL:
- op = LLVMRealUGE;
+ op = ordered ? LLVMRealOGE : LLVMRealUGE;
break;
default:
assert(0);
@@ -367,29 +223,20 @@ lp_build_compare(struct gallivm_state *gallivm,
}
/**
- * Build code to compare two values 'a' and 'b' using the given func.
+ * Build code to compare two values 'a' and 'b' of 'type' using the given func.
* \param func one of PIPE_FUNC_x
- * If the operands are floating point numbers, the function will use
- * ordered comparison which means that it will return true if both
- * operands are not a NaN and the specified condition evaluates to true.
* The result values will be 0 for false or ~0 for true.
*/
LLVMValueRef
-lp_build_cmp_ordered(struct lp_build_context *bld,
- unsigned func,
- LLVMValueRef a,
- LLVMValueRef b)
+lp_build_compare(struct gallivm_state *gallivm,
+ const struct lp_type type,
+ unsigned func,
+ LLVMValueRef a,
+ LLVMValueRef b)
{
- struct gallivm_state *gallivm = bld->gallivm;
- const struct lp_type type = bld->type;
-
-
- LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, type);
LLVMValueRef zeros = LLVMConstNull(int_vec_type);
LLVMValueRef ones = LLVMConstAllOnes(int_vec_type);
- LLVMValueRef cond;
- LLVMValueRef res;
assert(func >= PIPE_FUNC_NEVER);
assert(func <= PIPE_FUNC_ALWAYS);
@@ -418,129 +265,157 @@ lp_build_cmp_ordered(struct lp_build_context *bld,
__FUNCTION__, type.length, type.width);
}
#endif
- if(type.floating) {
- LLVMRealPredicate op;
- switch(func) {
- case PIPE_FUNC_NEVER:
- op = LLVMRealPredicateFalse;
- break;
- case PIPE_FUNC_ALWAYS:
- op = LLVMRealPredicateTrue;
- break;
- case PIPE_FUNC_EQUAL:
- op = LLVMRealOEQ;
- break;
- case PIPE_FUNC_NOTEQUAL:
- op = LLVMRealONE;
- break;
- case PIPE_FUNC_LESS:
- op = LLVMRealOLT;
- break;
- case PIPE_FUNC_LEQUAL:
- op = LLVMRealOLE;
- break;
- case PIPE_FUNC_GREATER:
- op = LLVMRealOGT;
- break;
- case PIPE_FUNC_GEQUAL:
- op = LLVMRealOGE;
- break;
- default:
- assert(0);
- return lp_build_undef(gallivm, type);
- }
-#if HAVE_LLVM >= 0x0207
- cond = LLVMBuildFCmp(builder, op, a, b, "");
- res = LLVMBuildSExt(builder, cond, int_vec_type, "");
-#else
- if (type.length == 1) {
- cond = LLVMBuildFCmp(builder, op, a, b, "");
- res = LLVMBuildSExt(builder, cond, int_vec_type, "");
- }
- else {
- unsigned i;
+#if HAVE_LLVM < 0x0207
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+ if(type.width * type.length == 128) {
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef cond;
+ LLVMValueRef res;
+ if(type.floating && util_cpu_caps.has_sse) {
+ /* float[4] comparison */
+ LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
+ LLVMValueRef args[3];
+ unsigned cc;
+ boolean swap;
- res = LLVMGetUndef(int_vec_type);
+ swap = FALSE;
+ switch(func) {
+ case PIPE_FUNC_EQUAL:
+ cc = 0;
+ break;
+ case PIPE_FUNC_NOTEQUAL:
+ cc = 4;
+ break;
+ case PIPE_FUNC_LESS:
+ cc = 1;
+ break;
+ case PIPE_FUNC_LEQUAL:
+ cc = 2;
+ break;
+ case PIPE_FUNC_GREATER:
+ cc = 1;
+ swap = TRUE;
+ break;
+ case PIPE_FUNC_GEQUAL:
+ cc = 2;
+ swap = TRUE;
+ break;
+ default:
+ assert(0);
+ return lp_build_undef(gallivm, type);
+ }
- debug_printf("%s: warning: using slow element-wise float"
- " vector comparison\n", __FUNCTION__);
- for (i = 0; i < type.length; ++i) {
- LLVMValueRef index = lp_build_const_int32(gallivm, i);
- cond = LLVMBuildFCmp(builder, op,
- LLVMBuildExtractElement(builder, a, index, ""),
- LLVMBuildExtractElement(builder, b, index, ""),
- "");
- cond = LLVMBuildSelect(builder, cond,
- LLVMConstExtractElement(ones, index),
- LLVMConstExtractElement(zeros, index),
- "");
- res = LLVMBuildInsertElement(builder, res, cond, index, "");
+ if(swap) {
+ args[0] = b;
+ args[1] = a;
+ }
+ else {
+ args[0] = a;
+ args[1] = b;
}
- }
-#endif
- }
- else {
- LLVMIntPredicate op;
- switch(func) {
- case PIPE_FUNC_EQUAL:
- op = LLVMIntEQ;
- break;
- case PIPE_FUNC_NOTEQUAL:
- op = LLVMIntNE;
- break;
- case PIPE_FUNC_LESS:
- op = type.sign ? LLVMIntSLT : LLVMIntULT;
- break;
- case PIPE_FUNC_LEQUAL:
- op = type.sign ? LLVMIntSLE : LLVMIntULE;
- break;
- case PIPE_FUNC_GREATER:
- op = type.sign ? LLVMIntSGT : LLVMIntUGT;
- break;
- case PIPE_FUNC_GEQUAL:
- op = type.sign ? LLVMIntSGE : LLVMIntUGE;
- break;
- default:
- assert(0);
- return lp_build_undef(gallivm, type);
- }
-#if HAVE_LLVM >= 0x0207
- cond = LLVMBuildICmp(builder, op, a, b, "");
- res = LLVMBuildSExt(builder, cond, int_vec_type, "");
-#else
- if (type.length == 1) {
- cond = LLVMBuildICmp(builder, op, a, b, "");
- res = LLVMBuildSExt(builder, cond, int_vec_type, "");
+ args[2] = LLVMConstInt(LLVMInt8TypeInContext(gallivm->context), cc, 0);
+ res = lp_build_intrinsic(builder,
+ "llvm.x86.sse.cmp.ps",
+ vec_type,
+ args, 3);
+ res = LLVMBuildBitCast(builder, res, int_vec_type, "");
+ return res;
}
- else {
- unsigned i;
+ else if(util_cpu_caps.has_sse2) {
+ /* int[4] comparison */
+ static const struct {
+ unsigned swap:1;
+ unsigned eq:1;
+ unsigned gt:1;
+ unsigned not:1;
+ } table[] = {
+ {0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
+ {1, 0, 1, 0}, /* PIPE_FUNC_LESS */
+ {0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
+ {0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
+ {0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
+ {0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
+ {1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
+ {0, 0, 0, 0} /* PIPE_FUNC_ALWAYS */
+ };
+ const char *pcmpeq;
+ const char *pcmpgt;
+ LLVMValueRef args[2];
+ LLVMValueRef res;
+ LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
- res = LLVMGetUndef(int_vec_type);
+ switch (type.width) {
+ case 8:
+ pcmpeq = "llvm.x86.sse2.pcmpeq.b";
+ pcmpgt = "llvm.x86.sse2.pcmpgt.b";
+ break;
+ case 16:
+ pcmpeq = "llvm.x86.sse2.pcmpeq.w";
+ pcmpgt = "llvm.x86.sse2.pcmpgt.w";
+ break;
+ case 32:
+ pcmpeq = "llvm.x86.sse2.pcmpeq.d";
+ pcmpgt = "llvm.x86.sse2.pcmpgt.d";
+ break;
+ default:
+ assert(0);
+ return lp_build_undef(gallivm, type);
+ }
- if (gallivm_debug & GALLIVM_DEBUG_PERF) {
- debug_printf("%s: using slow element-wise int"
- " vector comparison\n", __FUNCTION__);
+ /* There are no unsigned comparison instructions. So flip the sign bit
+ * so that the results match.
+ */
+ if (table[func].gt && !type.sign) {
+ LLVMValueRef msb = lp_build_const_int_vec(gallivm, type, (unsigned long long)1 << (type.width - 1));
+ a = LLVMBuildXor(builder, a, msb, "");
+ b = LLVMBuildXor(builder, b, msb, "");
}
- for(i = 0; i < type.length; ++i) {
- LLVMValueRef index = lp_build_const_int32(gallivm, i);
- cond = LLVMBuildICmp(builder, op,
- LLVMBuildExtractElement(builder, a, index, ""),
- LLVMBuildExtractElement(builder, b, index, ""),
- "");
- cond = LLVMBuildSelect(builder, cond,
- LLVMConstExtractElement(ones, index),
- LLVMConstExtractElement(zeros, index),
- "");
- res = LLVMBuildInsertElement(builder, res, cond, index, "");
+ if(table[func].swap) {
+ args[0] = b;
+ args[1] = a;
}
+ else {
+ args[0] = a;
+ args[1] = b;
+ }
+
+ if(table[func].eq)
+ res = lp_build_intrinsic(builder, pcmpeq, vec_type, args, 2);
+ else if (table[func].gt)
+ res = lp_build_intrinsic(builder, pcmpgt, vec_type, args, 2);
+ else
+ res = LLVMConstNull(vec_type);
+
+ if(table[func].not)
+ res = LLVMBuildNot(builder, res, "");
+
+ return res;
}
+ } /* if (type.width * type.length == 128) */
#endif
- }
+#endif /* HAVE_LLVM < 0x0207 */
- return res;
+ return lp_build_compare_ext(gallivm, type, func, a, b, FALSE);
+}
+
+/**
+ * Build code to compare two values 'a' and 'b' using the given func.
+ * \param func one of PIPE_FUNC_x
+ * If the operands are floating point numbers, the function will use
+ * ordered comparison which means that it will return true if both
+ * operands are not a NaN and the specified condition evaluates to true.
+ * The result values will be 0 for false or ~0 for true.
+ */
+LLVMValueRef
+lp_build_cmp_ordered(struct lp_build_context *bld,
+ unsigned func,
+ LLVMValueRef a,
+ LLVMValueRef b)
+{
+ return lp_build_compare_ext(bld->gallivm, bld->type, func, a, b, TRUE);
}
/**
More information about the mesa-commit
mailing list