*/
+#include <float.h>
+
#include "util/u_memory.h"
#include "util/u_debug.h"
#include "util/u_math.h"
#include "lp_bld_type.h"
#include "lp_bld_const.h"
+#include "lp_bld_init.h"
#include "lp_bld_intr.h"
#include "lp_bld_logic.h"
#include "lp_bld_pack.h"
#include "lp_bld_debug.h"
+#include "lp_bld_bitarit.h"
#include "lp_bld_arit.h"
+#include "lp_bld_flow.h"
+
+#if defined(PIPE_ARCH_SSE)
+#include <xmmintrin.h>
+#endif
+
+#ifndef _MM_DENORMALS_ZERO_MASK
+#define _MM_DENORMALS_ZERO_MASK 0x0040
+#endif
+#ifndef _MM_FLUSH_ZERO_MASK
+#define _MM_FLUSH_ZERO_MASK 0x8000
+#endif
-#define EXP_POLY_DEGREE 3
+#define EXP_POLY_DEGREE 5
-#define LOG_POLY_DEGREE 5
+#define LOG_POLY_DEGREE 4
/**
* Generate min(a, b)
* No checks for special case values of a or b = 1 or 0 are done.
+ * NaN's are handled according to the behavior specified by the
+ * nan_behavior argument.
*/
static LLVMValueRef
lp_build_min_simple(struct lp_build_context *bld,
LLVMValueRef a,
- LLVMValueRef b)
+ LLVMValueRef b,
+ enum gallivm_nan_behavior nan_behavior)
{
const struct lp_type type = bld->type;
const char *intrinsic = NULL;
+ unsigned intr_size = 0;
LLVMValueRef cond;
assert(lp_check_value(type, a));
/* TODO: optimize the constant case */
- if(type.width * type.length == 128) {
- if(type.floating) {
- if(type.width == 32 && util_cpu_caps.has_sse)
+ if (type.floating && util_cpu_caps.has_sse) {
+ if (type.width == 32) {
+ if (type.length == 1) {
+ intrinsic = "llvm.x86.sse.min.ss";
+ intr_size = 128;
+ }
+ else if (type.length <= 4 || !util_cpu_caps.has_avx) {
intrinsic = "llvm.x86.sse.min.ps";
- if(type.width == 64 && util_cpu_caps.has_sse2)
+ intr_size = 128;
+ }
+ else {
+ intrinsic = "llvm.x86.avx.min.ps.256";
+ intr_size = 256;
+ }
+ }
+ if (type.width == 64 && util_cpu_caps.has_sse2) {
+ if (type.length == 1) {
+ intrinsic = "llvm.x86.sse2.min.sd";
+ intr_size = 128;
+ }
+ else if (type.length == 2 || !util_cpu_caps.has_avx) {
intrinsic = "llvm.x86.sse2.min.pd";
+ intr_size = 128;
+ }
+ else {
+ intrinsic = "llvm.x86.avx.min.pd.256";
+ intr_size = 256;
+ }
}
- else {
- if(type.width == 8 && !type.sign && util_cpu_caps.has_sse2)
- intrinsic = "llvm.x86.sse2.pminu.b";
- if(type.width == 8 && type.sign && util_cpu_caps.has_sse4_1)
+ }
+ else if (type.floating && util_cpu_caps.has_altivec) {
+ if (nan_behavior == GALLIVM_NAN_RETURN_NAN ||
+ nan_behavior == GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN) {
+ debug_printf("%s: altivec doesn't support nan return nan behavior\n",
+ __FUNCTION__);
+ }
+ if (type.width == 32 && type.length == 4) {
+ intrinsic = "llvm.ppc.altivec.vminfp";
+ intr_size = 128;
+ }
+ } else if (util_cpu_caps.has_sse2 && type.length >= 2) {
+ intr_size = 128;
+ if ((type.width == 8 || type.width == 16) &&
+ (type.width * type.length <= 64) &&
+ (gallivm_debug & GALLIVM_DEBUG_PERF)) {
+ debug_printf("%s: inefficient code, bogus shuffle due to packing\n",
+ __FUNCTION__);
+ }
+ if (type.width == 8 && !type.sign) {
+ intrinsic = "llvm.x86.sse2.pminu.b";
+ }
+ else if (type.width == 16 && type.sign) {
+ intrinsic = "llvm.x86.sse2.pmins.w";
+ }
+ if (util_cpu_caps.has_sse4_1) {
+ if (type.width == 8 && type.sign) {
intrinsic = "llvm.x86.sse41.pminsb";
- if(type.width == 16 && !type.sign && util_cpu_caps.has_sse4_1)
+ }
+ if (type.width == 16 && !type.sign) {
intrinsic = "llvm.x86.sse41.pminuw";
- if(type.width == 16 && type.sign && util_cpu_caps.has_sse2)
- intrinsic = "llvm.x86.sse2.pmins.w";
- if(type.width == 32 && !type.sign && util_cpu_caps.has_sse4_1)
+ }
+ if (type.width == 32 && !type.sign) {
intrinsic = "llvm.x86.sse41.pminud";
- if(type.width == 32 && type.sign && util_cpu_caps.has_sse4_1)
+ }
+ if (type.width == 32 && type.sign) {
intrinsic = "llvm.x86.sse41.pminsd";
+ }
+ }
+ } else if (util_cpu_caps.has_altivec) {
+ intr_size = 128;
+ if (type.width == 8) {
+ if (!type.sign) {
+ intrinsic = "llvm.ppc.altivec.vminub";
+ } else {
+ intrinsic = "llvm.ppc.altivec.vminsb";
+ }
+ } else if (type.width == 16) {
+ if (!type.sign) {
+ intrinsic = "llvm.ppc.altivec.vminuh";
+ } else {
+ intrinsic = "llvm.ppc.altivec.vminsh";
+ }
+ } else if (type.width == 32) {
+ if (!type.sign) {
+ intrinsic = "llvm.ppc.altivec.vminuw";
+ } else {
+ intrinsic = "llvm.ppc.altivec.vminsw";
+ }
}
}
- if(intrinsic)
- return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
+ if (intrinsic) {
+ /* We need to handle nan's for floating point numbers. If one of the
+ * inputs is nan the other should be returned (required by both D3D10+
+ * and OpenCL).
+ * The sse intrinsics return the second operator in case of nan by
+ * default so we need to special code to handle those.
+ */
+ if (util_cpu_caps.has_sse && type.floating &&
+ nan_behavior != GALLIVM_NAN_BEHAVIOR_UNDEFINED &&
+ nan_behavior != GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN &&
+ nan_behavior != GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN) {
+ LLVMValueRef isnan, min;
+ min = lp_build_intrinsic_binary_anylength(bld->gallivm, intrinsic,
+ type,
+ intr_size, a, b);
+ if (nan_behavior == GALLIVM_NAN_RETURN_OTHER) {
+ isnan = lp_build_isnan(bld, b);
+ return lp_build_select(bld, isnan, a, min);
+ } else {
+ assert(nan_behavior == GALLIVM_NAN_RETURN_NAN);
+ isnan = lp_build_isnan(bld, a);
+ return lp_build_select(bld, isnan, a, min);
+ }
+ } else {
+ return lp_build_intrinsic_binary_anylength(bld->gallivm, intrinsic,
+ type,
+ intr_size, a, b);
+ }
+ }
- cond = lp_build_cmp(bld, PIPE_FUNC_LESS, a, b);
- return lp_build_select(bld, cond, a, b);
+ if (type.floating) {
+ switch (nan_behavior) {
+ case GALLIVM_NAN_RETURN_NAN: {
+ LLVMValueRef isnan = lp_build_isnan(bld, b);
+ cond = lp_build_cmp(bld, PIPE_FUNC_LESS, a, b);
+ cond = LLVMBuildXor(bld->gallivm->builder, cond, isnan, "");
+ return lp_build_select(bld, cond, a, b);
+ }
+ break;
+ case GALLIVM_NAN_RETURN_OTHER: {
+ LLVMValueRef isnan = lp_build_isnan(bld, a);
+ cond = lp_build_cmp(bld, PIPE_FUNC_LESS, a, b);
+ cond = LLVMBuildXor(bld->gallivm->builder, cond, isnan, "");
+ return lp_build_select(bld, cond, a, b);
+ }
+ break;
+ case GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN:
+ cond = lp_build_cmp_ordered(bld, PIPE_FUNC_LESS, a, b);
+ return lp_build_select(bld, cond, a, b);
+ case GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN:
+ cond = lp_build_cmp(bld, PIPE_FUNC_LESS, b, a);
+ return lp_build_select(bld, cond, b, a);
+ case GALLIVM_NAN_BEHAVIOR_UNDEFINED:
+ cond = lp_build_cmp(bld, PIPE_FUNC_LESS, a, b);
+ return lp_build_select(bld, cond, a, b);
+ break;
+ default:
+ assert(0);
+ cond = lp_build_cmp(bld, PIPE_FUNC_LESS, a, b);
+ return lp_build_select(bld, cond, a, b);
+ }
+ } else {
+ cond = lp_build_cmp(bld, PIPE_FUNC_LESS, a, b);
+ return lp_build_select(bld, cond, a, b);
+ }
}
/**
* Generate max(a, b)
* No checks for special case values of a or b = 1 or 0 are done.
+ * NaN's are handled according to the behavior specified by the
+ * nan_behavior argument.
*/
static LLVMValueRef
lp_build_max_simple(struct lp_build_context *bld,
LLVMValueRef a,
- LLVMValueRef b)
+ LLVMValueRef b,
+ enum gallivm_nan_behavior nan_behavior)
{
const struct lp_type type = bld->type;
const char *intrinsic = NULL;
+ unsigned intr_size = 0;
LLVMValueRef cond;
assert(lp_check_value(type, a));
/* TODO: optimize the constant case */
- if(type.width * type.length == 128) {
- if(type.floating) {
- if(type.width == 32 && util_cpu_caps.has_sse)
+ if (type.floating && util_cpu_caps.has_sse) {
+ if (type.width == 32) {
+ if (type.length == 1) {
+ intrinsic = "llvm.x86.sse.max.ss";
+ intr_size = 128;
+ }
+ else if (type.length <= 4 || !util_cpu_caps.has_avx) {
intrinsic = "llvm.x86.sse.max.ps";
- if(type.width == 64 && util_cpu_caps.has_sse2)
+ intr_size = 128;
+ }
+ else {
+ intrinsic = "llvm.x86.avx.max.ps.256";
+ intr_size = 256;
+ }
+ }
+ if (type.width == 64 && util_cpu_caps.has_sse2) {
+ if (type.length == 1) {
+ intrinsic = "llvm.x86.sse2.max.sd";
+ intr_size = 128;
+ }
+ else if (type.length == 2 || !util_cpu_caps.has_avx) {
intrinsic = "llvm.x86.sse2.max.pd";
+ intr_size = 128;
+ }
+ else {
+ intrinsic = "llvm.x86.avx.max.pd.256";
+ intr_size = 256;
+ }
}
- else {
- if(type.width == 8 && !type.sign && util_cpu_caps.has_sse2)
- intrinsic = "llvm.x86.sse2.pmaxu.b";
- if(type.width == 8 && type.sign && util_cpu_caps.has_sse4_1)
+ }
+ else if (type.floating && util_cpu_caps.has_altivec) {
+ if (nan_behavior == GALLIVM_NAN_RETURN_NAN ||
+ nan_behavior == GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN) {
+ debug_printf("%s: altivec doesn't support nan return nan behavior\n",
+ __FUNCTION__);
+ }
+ if (type.width == 32 || type.length == 4) {
+ intrinsic = "llvm.ppc.altivec.vmaxfp";
+ intr_size = 128;
+ }
+ } else if (util_cpu_caps.has_sse2 && type.length >= 2) {
+ intr_size = 128;
+ if ((type.width == 8 || type.width == 16) &&
+ (type.width * type.length <= 64) &&
+ (gallivm_debug & GALLIVM_DEBUG_PERF)) {
+ debug_printf("%s: inefficient code, bogus shuffle due to packing\n",
+ __FUNCTION__);
+ }
+ if (type.width == 8 && !type.sign) {
+ intrinsic = "llvm.x86.sse2.pmaxu.b";
+ intr_size = 128;
+ }
+ else if (type.width == 16 && type.sign) {
+ intrinsic = "llvm.x86.sse2.pmaxs.w";
+ }
+ if (util_cpu_caps.has_sse4_1) {
+ if (type.width == 8 && type.sign) {
intrinsic = "llvm.x86.sse41.pmaxsb";
- if(type.width == 16 && !type.sign && util_cpu_caps.has_sse4_1)
+ }
+ if (type.width == 16 && !type.sign) {
intrinsic = "llvm.x86.sse41.pmaxuw";
- if(type.width == 16 && type.sign && util_cpu_caps.has_sse2)
- intrinsic = "llvm.x86.sse2.pmaxs.w";
- if(type.width == 32 && !type.sign && util_cpu_caps.has_sse4_1)
+ }
+ if (type.width == 32 && !type.sign) {
intrinsic = "llvm.x86.sse41.pmaxud";
- if(type.width == 32 && type.sign && util_cpu_caps.has_sse4_1)
+ }
+ if (type.width == 32 && type.sign) {
intrinsic = "llvm.x86.sse41.pmaxsd";
+ }
}
+ } else if (util_cpu_caps.has_altivec) {
+ intr_size = 128;
+ if (type.width == 8) {
+ if (!type.sign) {
+ intrinsic = "llvm.ppc.altivec.vmaxub";
+ } else {
+ intrinsic = "llvm.ppc.altivec.vmaxsb";
+ }
+ } else if (type.width == 16) {
+ if (!type.sign) {
+ intrinsic = "llvm.ppc.altivec.vmaxuh";
+ } else {
+ intrinsic = "llvm.ppc.altivec.vmaxsh";
+ }
+ } else if (type.width == 32) {
+ if (!type.sign) {
+ intrinsic = "llvm.ppc.altivec.vmaxuw";
+ } else {
+ intrinsic = "llvm.ppc.altivec.vmaxsw";
+ }
+ }
}
- if(intrinsic)
- return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
+ if (intrinsic) {
+ if (util_cpu_caps.has_sse && type.floating &&
+ nan_behavior != GALLIVM_NAN_BEHAVIOR_UNDEFINED &&
+ nan_behavior != GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN &&
+ nan_behavior != GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN) {
+ LLVMValueRef isnan, max;
+ max = lp_build_intrinsic_binary_anylength(bld->gallivm, intrinsic,
+ type,
+ intr_size, a, b);
+ if (nan_behavior == GALLIVM_NAN_RETURN_OTHER) {
+ isnan = lp_build_isnan(bld, b);
+ return lp_build_select(bld, isnan, a, max);
+ } else {
+ assert(nan_behavior == GALLIVM_NAN_RETURN_NAN);
+ isnan = lp_build_isnan(bld, a);
+ return lp_build_select(bld, isnan, a, max);
+ }
+ } else {
+ return lp_build_intrinsic_binary_anylength(bld->gallivm, intrinsic,
+ type,
+ intr_size, a, b);
+ }
+ }
- cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, b);
- return lp_build_select(bld, cond, a, b);
+ if (type.floating) {
+ switch (nan_behavior) {
+ case GALLIVM_NAN_RETURN_NAN: {
+ LLVMValueRef isnan = lp_build_isnan(bld, b);
+ cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, b);
+ cond = LLVMBuildXor(bld->gallivm->builder, cond, isnan, "");
+ return lp_build_select(bld, cond, a, b);
+ }
+ break;
+ case GALLIVM_NAN_RETURN_OTHER: {
+ LLVMValueRef isnan = lp_build_isnan(bld, a);
+ cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, b);
+ cond = LLVMBuildXor(bld->gallivm->builder, cond, isnan, "");
+ return lp_build_select(bld, cond, a, b);
+ }
+ break;
+ case GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN:
+ cond = lp_build_cmp_ordered(bld, PIPE_FUNC_GREATER, a, b);
+ return lp_build_select(bld, cond, a, b);
+ case GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN:
+ cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, b, a);
+ return lp_build_select(bld, cond, b, a);
+ case GALLIVM_NAN_BEHAVIOR_UNDEFINED:
+ cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, b);
+ return lp_build_select(bld, cond, a, b);
+ break;
+ default:
+ assert(0);
+ cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, b);
+ return lp_build_select(bld, cond, a, b);
+ }
+ } else {
+ cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, b);
+ return lp_build_select(bld, cond, a, b);
+ }
}
lp_build_comp(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
assert(lp_check_value(type, a));
if(LLVMIsConstant(a))
return LLVMConstNot(a);
else
- return LLVMBuildNot(bld->builder, a, "");
+ return LLVMBuildNot(builder, a, "");
}
if(LLVMIsConstant(a))
return LLVMConstSub(bld->one, a);
else
if (type.floating)
- return LLVMBuildFSub(bld->builder, bld->one, a, "");
+ return LLVMBuildFSub(builder, bld->one, a, "");
else
- return LLVMBuildSub(bld->builder, bld->one, a, "");
+ return LLVMBuildSub(builder, bld->one, a, "");
}
LLVMValueRef a,
LLVMValueRef b)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
LLVMValueRef res;
if(a == bld->one || b == bld->one)
return bld->one;
- if(util_cpu_caps.has_sse2 &&
- type.width * type.length == 128 &&
- !type.floating && !type.fixed) {
- if(type.width == 8)
- intrinsic = type.sign ? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b";
- if(type.width == 16)
- intrinsic = type.sign ? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w";
+ if (type.width * type.length == 128 &&
+ !type.floating && !type.fixed) {
+ if(util_cpu_caps.has_sse2) {
+ if(type.width == 8)
+ intrinsic = type.sign ? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b";
+ if(type.width == 16)
+ intrinsic = type.sign ? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w";
+ } else if (util_cpu_caps.has_altivec) {
+ if(type.width == 8)
+ intrinsic = type.sign ? "llvm.ppc.altivec.vaddsbs" : "llvm.ppc.altivec.vaddubs";
+ if(type.width == 16)
+ intrinsic = type.sign ? "llvm.ppc.altivec.vaddshs" : "llvm.ppc.altivec.vadduhs";
+ }
}
- if(intrinsic)
- return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
+ if (intrinsic)
+ return lp_build_intrinsic_binary(builder, intrinsic, lp_build_vec_type(bld->gallivm, bld->type), a, b);
+ }
+
+ if(type.norm && !type.floating && !type.fixed) {
+ if (type.sign) {
+ uint64_t sign = (uint64_t)1 << (type.width - 1);
+ LLVMValueRef max_val = lp_build_const_int_vec(bld->gallivm, type, sign - 1);
+ LLVMValueRef min_val = lp_build_const_int_vec(bld->gallivm, type, sign);
+ /* a_clamp_max is the maximum a for positive b,
+ a_clamp_min is the minimum a for negative b. */
+ LLVMValueRef a_clamp_max = lp_build_min_simple(bld, a, LLVMBuildSub(builder, max_val, b, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED);
+ LLVMValueRef a_clamp_min = lp_build_max_simple(bld, a, LLVMBuildSub(builder, min_val, b, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED);
+ a = lp_build_select(bld, lp_build_cmp(bld, PIPE_FUNC_GREATER, b, bld->zero), a_clamp_max, a_clamp_min);
+ } else {
+ a = lp_build_min_simple(bld, a, lp_build_comp(bld, b), GALLIVM_NAN_BEHAVIOR_UNDEFINED);
+ }
}
if(LLVMIsConstant(a) && LLVMIsConstant(b))
res = LLVMConstAdd(a, b);
else
if (type.floating)
- res = LLVMBuildFAdd(bld->builder, a, b, "");
+ res = LLVMBuildFAdd(builder, a, b, "");
else
- res = LLVMBuildAdd(bld->builder, a, b, "");
+ res = LLVMBuildAdd(builder, a, b, "");
/* clamp to ceiling of 1.0 */
if(bld->type.norm && (bld->type.floating || bld->type.fixed))
- res = lp_build_min_simple(bld, res, bld->one);
+ res = lp_build_min_simple(bld, res, bld->one, GALLIVM_NAN_BEHAVIOR_UNDEFINED);
/* XXX clamp to floor of -1 or 0??? */
}
-/** Return the scalar sum of the elements of a */
+/** Return the scalar sum of the elements of a.
+ * Should avoid this operation whenever possible.
+ */
LLVMValueRef
-lp_build_sum_vector(struct lp_build_context *bld,
- LLVMValueRef a)
+lp_build_horizontal_add(struct lp_build_context *bld,
+ LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
LLVMValueRef index, res;
- unsigned i;
+ unsigned i, length;
+ LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH / 2];
+ LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH / 2];
+ LLVMValueRef vecres, elem2;
assert(lp_check_value(type, a));
assert(!bld->type.norm);
- index = LLVMConstInt(LLVMInt32Type(), 0, 0);
- res = LLVMBuildExtractElement(bld->builder, a, index, "");
+ /*
+ * for byte vectors can do much better with psadbw.
+ * Using repeated shuffle/adds here. Note with multiple vectors
+ * this can be done more efficiently as outlined in the intel
+ * optimization manual.
+ * Note: could cause data rearrangement if used with smaller element
+ * sizes.
+ */
- for (i = 1; i < type.length; i++) {
- index = LLVMConstInt(LLVMInt32Type(), i, 0);
- if (type.floating)
- res = LLVMBuildFAdd(bld->builder, res,
- LLVMBuildExtractElement(bld->builder,
- a, index, ""),
- "");
- else
- res = LLVMBuildAdd(bld->builder, res,
- LLVMBuildExtractElement(bld->builder,
- a, index, ""),
- "");
+ vecres = a;
+ length = type.length / 2;
+ while (length > 1) {
+ LLVMValueRef vec1, vec2;
+ for (i = 0; i < length; i++) {
+ shuffles1[i] = lp_build_const_int32(bld->gallivm, i);
+ shuffles2[i] = lp_build_const_int32(bld->gallivm, i + length);
+ }
+ vec1 = LLVMBuildShuffleVector(builder, vecres, vecres,
+ LLVMConstVector(shuffles1, length), "");
+ vec2 = LLVMBuildShuffleVector(builder, vecres, vecres,
+ LLVMConstVector(shuffles2, length), "");
+ if (type.floating) {
+ vecres = LLVMBuildFAdd(builder, vec1, vec2, "");
+ }
+ else {
+ vecres = LLVMBuildAdd(builder, vec1, vec2, "");
+ }
+ length = length >> 1;
}
+ /* always have vector of size 2 here */
+ assert(length == 1);
+
+ index = lp_build_const_int32(bld->gallivm, 0);
+ res = LLVMBuildExtractElement(builder, vecres, index, "");
+ index = lp_build_const_int32(bld->gallivm, 1);
+ elem2 = LLVMBuildExtractElement(builder, vecres, index, "");
+
+ if (type.floating)
+ res = LLVMBuildFAdd(builder, res, elem2, "");
+ else
+ res = LLVMBuildAdd(builder, res, elem2, "");
+
return res;
}
+/**
+ * Return the horizontal sums of 4 float vectors as a float4 vector.
+ * This uses the technique as outlined in Intel Optimization Manual.
+ */
+static LLVMValueRef
+lp_build_horizontal_add4x4f(struct lp_build_context *bld,
+ LLVMValueRef src[4])
+{
+ struct gallivm_state *gallivm = bld->gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef shuffles[4];
+ LLVMValueRef tmp[4];
+ LLVMValueRef sumtmp[2], shuftmp[2];
+
+ /* lower half of regs */
+ shuffles[0] = lp_build_const_int32(gallivm, 0);
+ shuffles[1] = lp_build_const_int32(gallivm, 1);
+ shuffles[2] = lp_build_const_int32(gallivm, 4);
+ shuffles[3] = lp_build_const_int32(gallivm, 5);
+ tmp[0] = LLVMBuildShuffleVector(builder, src[0], src[1],
+ LLVMConstVector(shuffles, 4), "");
+ tmp[2] = LLVMBuildShuffleVector(builder, src[2], src[3],
+ LLVMConstVector(shuffles, 4), "");
+
+ /* upper half of regs */
+ shuffles[0] = lp_build_const_int32(gallivm, 2);
+ shuffles[1] = lp_build_const_int32(gallivm, 3);
+ shuffles[2] = lp_build_const_int32(gallivm, 6);
+ shuffles[3] = lp_build_const_int32(gallivm, 7);
+ tmp[1] = LLVMBuildShuffleVector(builder, src[0], src[1],
+ LLVMConstVector(shuffles, 4), "");
+ tmp[3] = LLVMBuildShuffleVector(builder, src[2], src[3],
+ LLVMConstVector(shuffles, 4), "");
+
+ sumtmp[0] = LLVMBuildFAdd(builder, tmp[0], tmp[1], "");
+ sumtmp[1] = LLVMBuildFAdd(builder, tmp[2], tmp[3], "");
+
+ shuffles[0] = lp_build_const_int32(gallivm, 0);
+ shuffles[1] = lp_build_const_int32(gallivm, 2);
+ shuffles[2] = lp_build_const_int32(gallivm, 4);
+ shuffles[3] = lp_build_const_int32(gallivm, 6);
+ shuftmp[0] = LLVMBuildShuffleVector(builder, sumtmp[0], sumtmp[1],
+ LLVMConstVector(shuffles, 4), "");
+
+ shuffles[0] = lp_build_const_int32(gallivm, 1);
+ shuffles[1] = lp_build_const_int32(gallivm, 3);
+ shuffles[2] = lp_build_const_int32(gallivm, 5);
+ shuffles[3] = lp_build_const_int32(gallivm, 7);
+ shuftmp[1] = LLVMBuildShuffleVector(builder, sumtmp[0], sumtmp[1],
+ LLVMConstVector(shuffles, 4), "");
+
+ return LLVMBuildFAdd(builder, shuftmp[0], shuftmp[1], "");
+}
+
+
+/*
+ * partially horizontally add 2-4 float vectors with length nx4,
+ * i.e. only four adjacent values in each vector will be added,
+ * assuming values are really grouped in 4 which also determines
+ * output order.
+ *
+ * Return a vector of the same length as the initial vectors,
+ * with the excess elements (if any) being undefined.
+ * The element order is independent of number of input vectors.
+ * For 3 vectors x0x1x2x3x4x5x6x7, y0y1y2y3y4y5y6y7, z0z1z2z3z4z5z6z7
+ * the output order thus will be
+ * sumx0-x3,sumy0-y3,sumz0-z3,undef,sumx4-x7,sumy4-y7,sumz4z7,undef
+ */
+LLVMValueRef
+lp_build_hadd_partial4(struct lp_build_context *bld,
+ LLVMValueRef vectors[],
+ unsigned num_vecs)
+{
+ struct gallivm_state *gallivm = bld->gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef ret_vec;
+ LLVMValueRef tmp[4];
+ const char *intrinsic = NULL;
+
+ assert(num_vecs >= 2 && num_vecs <= 4);
+ assert(bld->type.floating);
+
+ /* only use this with at least 2 vectors, as it is sort of expensive
+ * (depending on cpu) and we always need two horizontal adds anyway,
+ * so a shuffle/add approach might be better.
+ */
+
+ tmp[0] = vectors[0];
+ tmp[1] = vectors[1];
+
+ tmp[2] = num_vecs > 2 ? vectors[2] : vectors[0];
+ tmp[3] = num_vecs > 3 ? vectors[3] : vectors[0];
+
+ if (util_cpu_caps.has_sse3 && bld->type.width == 32 &&
+ bld->type.length == 4) {
+ intrinsic = "llvm.x86.sse3.hadd.ps";
+ }
+ else if (util_cpu_caps.has_avx && bld->type.width == 32 &&
+ bld->type.length == 8) {
+ intrinsic = "llvm.x86.avx.hadd.ps.256";
+ }
+ if (intrinsic) {
+ tmp[0] = lp_build_intrinsic_binary(builder, intrinsic,
+ lp_build_vec_type(gallivm, bld->type),
+ tmp[0], tmp[1]);
+ if (num_vecs > 2) {
+ tmp[1] = lp_build_intrinsic_binary(builder, intrinsic,
+ lp_build_vec_type(gallivm, bld->type),
+ tmp[2], tmp[3]);
+ }
+ else {
+ tmp[1] = tmp[0];
+ }
+ return lp_build_intrinsic_binary(builder, intrinsic,
+ lp_build_vec_type(gallivm, bld->type),
+ tmp[0], tmp[1]);
+ }
+
+ if (bld->type.length == 4) {
+ ret_vec = lp_build_horizontal_add4x4f(bld, tmp);
+ }
+ else {
+ LLVMValueRef partres[LP_MAX_VECTOR_LENGTH/4];
+ unsigned j;
+ unsigned num_iter = bld->type.length / 4;
+ struct lp_type parttype = bld->type;
+ parttype.length = 4;
+ for (j = 0; j < num_iter; j++) {
+ LLVMValueRef partsrc[4];
+ unsigned i;
+ for (i = 0; i < 4; i++) {
+ partsrc[i] = lp_build_extract_range(gallivm, tmp[i], j*4, 4);
+ }
+ partres[j] = lp_build_horizontal_add4x4f(bld, partsrc);
+ }
+ ret_vec = lp_build_concat(gallivm, partres, parttype, num_iter);
+ }
+ return ret_vec;
+}
/**
* Generate a - b
LLVMValueRef a,
LLVMValueRef b)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
LLVMValueRef res;
if(b == bld->one)
return bld->zero;
- if(util_cpu_caps.has_sse2 &&
- type.width * type.length == 128 &&
- !type.floating && !type.fixed) {
- if(type.width == 8)
- intrinsic = type.sign ? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b";
- if(type.width == 16)
- intrinsic = type.sign ? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w";
+ if (type.width * type.length == 128 &&
+ !type.floating && !type.fixed) {
+ if (util_cpu_caps.has_sse2) {
+ if(type.width == 8)
+ intrinsic = type.sign ? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b";
+ if(type.width == 16)
+ intrinsic = type.sign ? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w";
+ } else if (util_cpu_caps.has_altivec) {
+ if(type.width == 8)
+ intrinsic = type.sign ? "llvm.ppc.altivec.vsubsbs" : "llvm.ppc.altivec.vsububs";
+ if(type.width == 16)
+ intrinsic = type.sign ? "llvm.ppc.altivec.vsubshs" : "llvm.ppc.altivec.vsubuhs";
+ }
}
- if(intrinsic)
- return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
+ if (intrinsic)
+ return lp_build_intrinsic_binary(builder, intrinsic, lp_build_vec_type(bld->gallivm, bld->type), a, b);
+ }
+
+ if(type.norm && !type.floating && !type.fixed) {
+ if (type.sign) {
+ uint64_t sign = (uint64_t)1 << (type.width - 1);
+ LLVMValueRef max_val = lp_build_const_int_vec(bld->gallivm, type, sign - 1);
+ LLVMValueRef min_val = lp_build_const_int_vec(bld->gallivm, type, sign);
+ /* a_clamp_max is the maximum a for negative b,
+ a_clamp_min is the minimum a for positive b. */
+ LLVMValueRef a_clamp_max = lp_build_min_simple(bld, a, LLVMBuildAdd(builder, max_val, b, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED);
+ LLVMValueRef a_clamp_min = lp_build_max_simple(bld, a, LLVMBuildAdd(builder, min_val, b, ""), GALLIVM_NAN_BEHAVIOR_UNDEFINED);
+ a = lp_build_select(bld, lp_build_cmp(bld, PIPE_FUNC_GREATER, b, bld->zero), a_clamp_min, a_clamp_max);
+ } else {
+ a = lp_build_max_simple(bld, a, b, GALLIVM_NAN_BEHAVIOR_UNDEFINED);
+ }
}
if(LLVMIsConstant(a) && LLVMIsConstant(b))
res = LLVMConstSub(a, b);
else
if (type.floating)
- res = LLVMBuildFSub(bld->builder, a, b, "");
+ res = LLVMBuildFSub(builder, a, b, "");
else
- res = LLVMBuildSub(bld->builder, a, b, "");
+ res = LLVMBuildSub(builder, a, b, "");
if(bld->type.norm && (bld->type.floating || bld->type.fixed))
- res = lp_build_max_simple(bld, res, bld->zero);
+ res = lp_build_max_simple(bld, res, bld->zero, GALLIVM_NAN_BEHAVIOR_UNDEFINED);
return res;
}
+
/**
- * Normalized 8bit multiplication.
+ * Normalized multiplication.
+ *
+ * There are several approaches for (using 8-bit normalized multiplication as
+ * an example):
*
* - alpha plus one
*
*
* a*b/255 ~= (a*(b + 1)) >> 256
*
- * which is the fastest method that satisfies the following OpenGL criteria
+ * which is the fastest method that satisfies the following OpenGL criteria of
*
* 0*0 = 0 and 255*255 = 255
*
*
* note that just by itself it doesn't satisfies the OpenGL criteria, as
* 255*255 = 254, so the special case b = 255 must be accounted or roundoff
- * must be used
+ * must be used.
*
* - geometric series plus rounding
*
*
* t/255 ~= (t + (t >> 8) + 0x80) >> 8
*
- * achieving the exact results
+ * achieving the exact results.
+ *
+ *
*
* @sa Alvy Ray Smith, Image Compositing Fundamentals, Tech Memo 4, Aug 15, 1995,
* ftp://ftp.alvyray.com/Acrobat/4_Comp.pdf
* http://www.stereopsis.com/doubleblend.html
*/
static LLVMValueRef
-lp_build_mul_u8n(LLVMBuilderRef builder,
- struct lp_type i16_type,
- LLVMValueRef a, LLVMValueRef b)
+lp_build_mul_norm(struct gallivm_state *gallivm,
+ struct lp_type wide_type,
+ LLVMValueRef a, LLVMValueRef b)
{
- LLVMValueRef c8;
+ LLVMBuilderRef builder = gallivm->builder;
+ struct lp_build_context bld;
+ unsigned n;
+ LLVMValueRef half;
LLVMValueRef ab;
- assert(!i16_type.floating);
- assert(lp_check_value(i16_type, a));
- assert(lp_check_value(i16_type, b));
+ assert(!wide_type.floating);
+ assert(lp_check_value(wide_type, a));
+ assert(lp_check_value(wide_type, b));
- c8 = lp_build_const_int_vec(i16_type, 8);
-
-#if 0
-
- /* a*b/255 ~= (a*(b + 1)) >> 256 */
- b = LLVMBuildAdd(builder, b, lp_build_const_int_vec(i16_type, 1), "");
- ab = LLVMBuildMul(builder, a, b, "");
+ lp_build_context_init(&bld, gallivm, wide_type);
+
+ n = wide_type.width / 2;
+ if (wide_type.sign) {
+ --n;
+ }
+
+ /*
+ * TODO: for 16bits normalized SSE2 vectors we could consider using PMULHUW
+ * http://ssp.impulsetrain.com/2011/07/03/multiplying-normalized-16-bit-numbers-with-sse2/
+ */
+
+ /*
+ * a*b / (2**n - 1) ~= (a*b + (a*b >> n) + half) >> n
+ */
-#else
-
- /* ab/255 ~= (ab + (ab >> 8) + 0x80) >> 8 */
ab = LLVMBuildMul(builder, a, b, "");
- ab = LLVMBuildAdd(builder, ab, LLVMBuildLShr(builder, ab, c8, ""), "");
- ab = LLVMBuildAdd(builder, ab, lp_build_const_int_vec(i16_type, 0x80), "");
+ ab = LLVMBuildAdd(builder, ab, lp_build_shr_imm(&bld, ab, n), "");
-#endif
-
- ab = LLVMBuildLShr(builder, ab, c8, "");
+ /*
+ * half = sgn(ab) * 0.5 * (2 ** n) = sgn(ab) * (1 << (n - 1))
+ */
+
+ half = lp_build_const_int_vec(gallivm, wide_type, 1LL << (n - 1));
+ if (wide_type.sign) {
+ LLVMValueRef minus_half = LLVMBuildNeg(builder, half, "");
+ LLVMValueRef sign = lp_build_shr_imm(&bld, ab, wide_type.width - 1);
+ half = lp_build_select(&bld, sign, minus_half, half);
+ }
+ ab = LLVMBuildAdd(builder, ab, half, "");
+
+ /* Final division */
+ ab = lp_build_shr_imm(&bld, ab, n);
return ab;
}
-
/**
* Generate a * b
*/
LLVMValueRef a,
LLVMValueRef b)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
LLVMValueRef shift;
LLVMValueRef res;
if(a == bld->undef || b == bld->undef)
return bld->undef;
- if(!type.floating && !type.fixed && type.norm) {
- if(type.width == 8) {
- struct lp_type i16_type = lp_wider_type(type);
- LLVMValueRef al, ah, bl, bh, abl, abh, ab;
+ if (!type.floating && !type.fixed && type.norm) {
+ struct lp_type wide_type = lp_wider_type(type);
+ LLVMValueRef al, ah, bl, bh, abl, abh, ab;
- lp_build_unpack2(bld->builder, type, i16_type, a, &al, &ah);
- lp_build_unpack2(bld->builder, type, i16_type, b, &bl, &bh);
+ lp_build_unpack2(bld->gallivm, type, wide_type, a, &al, &ah);
+ lp_build_unpack2(bld->gallivm, type, wide_type, b, &bl, &bh);
- /* PMULLW, PSRLW, PADDW */
- abl = lp_build_mul_u8n(bld->builder, i16_type, al, bl);
- abh = lp_build_mul_u8n(bld->builder, i16_type, ah, bh);
+ /* PMULLW, PSRLW, PADDW */
+ abl = lp_build_mul_norm(bld->gallivm, wide_type, al, bl);
+ abh = lp_build_mul_norm(bld->gallivm, wide_type, ah, bh);
- ab = lp_build_pack2(bld->builder, i16_type, type, abl, abh);
-
- return ab;
- }
+ ab = lp_build_pack2(bld->gallivm, wide_type, type, abl, abh);
- /* FIXME */
- assert(0);
+ return ab;
}
if(type.fixed)
- shift = lp_build_const_int_vec(type, type.width/2);
+ shift = lp_build_const_int_vec(bld->gallivm, type, type.width/2);
else
shift = NULL;
}
else {
if (type.floating)
- res = LLVMBuildFMul(bld->builder, a, b, "");
+ res = LLVMBuildFMul(builder, a, b, "");
else
- res = LLVMBuildMul(bld->builder, a, b, "");
+ res = LLVMBuildMul(builder, a, b, "");
if(shift) {
if(type.sign)
- res = LLVMBuildAShr(bld->builder, res, shift, "");
+ res = LLVMBuildAShr(builder, res, shift, "");
else
- res = LLVMBuildLShr(bld->builder, res, shift, "");
+ res = LLVMBuildLShr(builder, res, shift, "");
}
}
LLVMValueRef a,
int b)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef factor;
assert(lp_check_value(bld->type, a));
if(bld->type.floating) {
#if 0
/*
- * Power of two multiplication by directly manipulating the mantissa.
+ * Power of two multiplication by directly manipulating the exponent.
*
* XXX: This might not be always faster, it will introduce a small error
* for multiplication by zero, and it will produce wrong results
* for Inf and NaN.
*/
unsigned mantissa = lp_mantissa(bld->type);
- factor = lp_build_const_int_vec(bld->type, (unsigned long long)shift << mantissa);
- a = LLVMBuildBitCast(bld->builder, a, lp_build_int_vec_type(bld->type), "");
- a = LLVMBuildAdd(bld->builder, a, factor, "");
- a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(bld->type), "");
+ factor = lp_build_const_int_vec(bld->gallivm, bld->type, (unsigned long long)shift << mantissa);
+ a = LLVMBuildBitCast(builder, a, lp_build_int_vec_type(bld->type), "");
+ a = LLVMBuildAdd(builder, a, factor, "");
+ a = LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, bld->type), "");
return a;
#endif
}
else {
- factor = lp_build_const_vec(bld->type, shift);
- return LLVMBuildShl(bld->builder, a, factor, "");
+ factor = lp_build_const_vec(bld->gallivm, bld->type, shift);
+ return LLVMBuildShl(builder, a, factor, "");
}
}
- factor = lp_build_const_vec(bld->type, (double)b);
+ factor = lp_build_const_vec(bld->gallivm, bld->type, (double)b);
return lp_build_mul(bld, a, factor);
}
LLVMValueRef a,
LLVMValueRef b)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
assert(lp_check_value(type, a));
if(a == bld->zero)
return bld->zero;
- if(a == bld->one)
+ if(a == bld->one && type.floating)
return lp_build_rcp(bld, b);
if(b == bld->zero)
return bld->undef;
return LLVMConstUDiv(a, b);
}
- if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4)
+ if(((util_cpu_caps.has_sse && type.width == 32 && type.length == 4) ||
+ (util_cpu_caps.has_avx && type.width == 32 && type.length == 8)) &&
+ type.floating)
return lp_build_mul(bld, a, lp_build_rcp(bld, b));
if (type.floating)
- return LLVMBuildFDiv(bld->builder, a, b, "");
+ return LLVMBuildFDiv(builder, a, b, "");
else if (type.sign)
- return LLVMBuildSDiv(bld->builder, a, b, "");
+ return LLVMBuildSDiv(builder, a, b, "");
else
- return LLVMBuildUDiv(bld->builder, a, b, "");
+ return LLVMBuildUDiv(builder, a, b, "");
}
/**
- * Linear interpolation -- without any checks.
+ * Linear interpolation helper.
+ *
+ * @param normalized whether we are interpolating normalized values,
+ * encoded in normalized integers, twice as wide.
*
* @sa http://www.stereopsis.com/doubleblend.html
*/
-static INLINE LLVMValueRef
+static inline LLVMValueRef
lp_build_lerp_simple(struct lp_build_context *bld,
LLVMValueRef x,
LLVMValueRef v0,
- LLVMValueRef v1)
+ LLVMValueRef v1,
+ unsigned flags)
{
+ unsigned half_width = bld->type.width/2;
+ LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef delta;
LLVMValueRef res;
delta = lp_build_sub(bld, v1, v0);
- res = lp_build_mul(bld, x, delta);
+ if (flags & LP_BLD_LERP_WIDE_NORMALIZED) {
+ if (!bld->type.sign) {
+ if (!(flags & LP_BLD_LERP_PRESCALED_WEIGHTS)) {
+ /*
+ * Scale x from [0, 2**n - 1] to [0, 2**n] by adding the
+ * most-significant-bit to the lowest-significant-bit, so that
+ * later we can just divide by 2**n instead of 2**n - 1.
+ */
+
+ x = lp_build_add(bld, x, lp_build_shr_imm(bld, x, half_width - 1));
+ }
+
+ /* (x * delta) >> n */
+ res = lp_build_mul(bld, x, delta);
+ res = lp_build_shr_imm(bld, res, half_width);
+ } else {
+ /*
+ * The rescaling trick above doesn't work for signed numbers, so
+ * use the 2**n - 1 divison approximation in lp_build_mul_norm
+ * instead.
+ */
+ assert(!(flags & LP_BLD_LERP_PRESCALED_WEIGHTS));
+ res = lp_build_mul_norm(bld->gallivm, bld->type, x, delta);
+ }
+ } else {
+ assert(!(flags & LP_BLD_LERP_PRESCALED_WEIGHTS));
+ res = lp_build_mul(bld, x, delta);
+ }
res = lp_build_add(bld, v0, res);
- if (bld->type.fixed) {
+ if (((flags & LP_BLD_LERP_WIDE_NORMALIZED) && !bld->type.sign) ||
+ bld->type.fixed) {
+ /* We need to mask out the high order bits when lerping 8bit normalized colors stored on 16bits */
/* XXX: This step is necessary for lerping 8bit colors stored on 16bits,
- * but it will be wrong for other uses. Basically we need a more
- * powerful lp_type, capable of further distinguishing the values
+ * but it will be wrong for true fixed point use cases. Basically we need
+ * a more powerful lp_type, capable of further distinguishing the values
* interpretation from the value storage. */
- res = LLVMBuildAnd(bld->builder, res, lp_build_const_int_vec(bld->type, (1 << bld->type.width/2) - 1), "");
+ res = LLVMBuildAnd(builder, res, lp_build_const_int_vec(bld->gallivm, bld->type, (1 << half_width) - 1), "");
}
return res;
lp_build_lerp(struct lp_build_context *bld,
LLVMValueRef x,
LLVMValueRef v0,
- LLVMValueRef v1)
+ LLVMValueRef v1,
+ unsigned flags)
{
const struct lp_type type = bld->type;
LLVMValueRef res;
assert(lp_check_value(type, v0));
assert(lp_check_value(type, v1));
+ assert(!(flags & LP_BLD_LERP_WIDE_NORMALIZED));
+
if (type.norm) {
struct lp_type wide_type;
struct lp_build_context wide_bld;
LLVMValueRef xl, xh, v0l, v0h, v1l, v1h, resl, resh;
- LLVMValueRef shift;
assert(type.length >= 2);
- assert(!type.sign);
/*
- * Create a wider type, enough to hold the intermediate result of the
- * multiplication.
+ * Create a wider integer type, enough to hold the
+ * intermediate result of the multiplication.
*/
memset(&wide_type, 0, sizeof wide_type);
- wide_type.fixed = TRUE;
+ wide_type.sign = type.sign;
wide_type.width = type.width*2;
wide_type.length = type.length/2;
- lp_build_context_init(&wide_bld, bld->builder, wide_type);
-
- lp_build_unpack2(bld->builder, type, wide_type, x, &xl, &xh);
- lp_build_unpack2(bld->builder, type, wide_type, v0, &v0l, &v0h);
- lp_build_unpack2(bld->builder, type, wide_type, v1, &v1l, &v1h);
-
- /*
- * Scale x from [0, 255] to [0, 256]
- */
-
- shift = lp_build_const_int_vec(wide_type, type.width - 1);
+ lp_build_context_init(&wide_bld, bld->gallivm, wide_type);
- xl = lp_build_add(&wide_bld, xl,
- LLVMBuildAShr(bld->builder, xl, shift, ""));
- xh = lp_build_add(&wide_bld, xh,
- LLVMBuildAShr(bld->builder, xh, shift, ""));
+ lp_build_unpack2(bld->gallivm, type, wide_type, x, &xl, &xh);
+ lp_build_unpack2(bld->gallivm, type, wide_type, v0, &v0l, &v0h);
+ lp_build_unpack2(bld->gallivm, type, wide_type, v1, &v1l, &v1h);
/*
* Lerp both halves.
*/
- resl = lp_build_lerp_simple(&wide_bld, xl, v0l, v1l);
- resh = lp_build_lerp_simple(&wide_bld, xh, v0h, v1h);
+ flags |= LP_BLD_LERP_WIDE_NORMALIZED;
- res = lp_build_pack2(bld->builder, wide_type, type, resl, resh);
+ resl = lp_build_lerp_simple(&wide_bld, xl, v0l, v1l, flags);
+ resh = lp_build_lerp_simple(&wide_bld, xh, v0h, v1h, flags);
+
+ res = lp_build_pack2(bld->gallivm, wide_type, type, resl, resh);
} else {
- res = lp_build_lerp_simple(bld, x, v0, v1);
+ res = lp_build_lerp_simple(bld, x, v0, v1, flags);
}
return res;
}
+/**
+ * Bilinear interpolation.
+ *
+ * Values indices are in v_{yx}.
+ */
LLVMValueRef
lp_build_lerp_2d(struct lp_build_context *bld,
LLVMValueRef x,
LLVMValueRef v00,
LLVMValueRef v01,
LLVMValueRef v10,
- LLVMValueRef v11)
+ LLVMValueRef v11,
+ unsigned flags)
+{
+ LLVMValueRef v0 = lp_build_lerp(bld, x, v00, v01, flags);
+ LLVMValueRef v1 = lp_build_lerp(bld, x, v10, v11, flags);
+ return lp_build_lerp(bld, y, v0, v1, flags);
+}
+
+
+LLVMValueRef
+lp_build_lerp_3d(struct lp_build_context *bld,
+ LLVMValueRef x,
+ LLVMValueRef y,
+ LLVMValueRef z,
+ LLVMValueRef v000,
+ LLVMValueRef v001,
+ LLVMValueRef v010,
+ LLVMValueRef v011,
+ LLVMValueRef v100,
+ LLVMValueRef v101,
+ LLVMValueRef v110,
+ LLVMValueRef v111,
+ unsigned flags)
{
- LLVMValueRef v0 = lp_build_lerp(bld, x, v00, v01);
- LLVMValueRef v1 = lp_build_lerp(bld, x, v10, v11);
- return lp_build_lerp(bld, y, v0, v1);
+ LLVMValueRef v0 = lp_build_lerp_2d(bld, x, y, v000, v001, v010, v011, flags);
+ LLVMValueRef v1 = lp_build_lerp_2d(bld, x, y, v100, v101, v110, v111, flags);
+ return lp_build_lerp(bld, z, v0, v1, flags);
}
/**
* Generate min(a, b)
- * Do checks for special cases.
+ * Do checks for special cases but not for nans.
*/
LLVMValueRef
lp_build_min(struct lp_build_context *bld,
if(a == b)
return a;
- if(bld->type.norm) {
- if(a == bld->zero || b == bld->zero)
- return bld->zero;
+ if (bld->type.norm) {
+ if (!bld->type.sign) {
+ if (a == bld->zero || b == bld->zero) {
+ return bld->zero;
+ }
+ }
if(a == bld->one)
return b;
if(b == bld->one)
return a;
}
- return lp_build_min_simple(bld, a, b);
+ return lp_build_min_simple(bld, a, b, GALLIVM_NAN_BEHAVIOR_UNDEFINED);
}
+/**
+ * Generate min(a, b)
+ * NaN's are handled according to the behavior specified by the
+ * nan_behavior argument.
+ */
+LLVMValueRef
+lp_build_min_ext(struct lp_build_context *bld,
+ LLVMValueRef a,
+ LLVMValueRef b,
+ enum gallivm_nan_behavior nan_behavior)
+{
+ assert(lp_check_value(bld->type, a));
+ assert(lp_check_value(bld->type, b));
+
+ if(a == bld->undef || b == bld->undef)
+ return bld->undef;
+
+ if(a == b)
+ return a;
+
+ if (bld->type.norm) {
+ if (!bld->type.sign) {
+ if (a == bld->zero || b == bld->zero) {
+ return bld->zero;
+ }
+ }
+ if(a == bld->one)
+ return b;
+ if(b == bld->one)
+ return a;
+ }
+
+ return lp_build_min_simple(bld, a, b, nan_behavior);
+}
+
/**
* Generate max(a, b)
- * Do checks for special cases.
+ * Do checks for special cases, but NaN behavior is undefined.
*/
LLVMValueRef
lp_build_max(struct lp_build_context *bld,
if(bld->type.norm) {
if(a == bld->one || b == bld->one)
return bld->one;
- if(a == bld->zero)
- return b;
- if(b == bld->zero)
- return a;
+ if (!bld->type.sign) {
+ if (a == bld->zero) {
+ return b;
+ }
+ if (b == bld->zero) {
+ return a;
+ }
+ }
}
- return lp_build_max_simple(bld, a, b);
+ return lp_build_max_simple(bld, a, b, GALLIVM_NAN_BEHAVIOR_UNDEFINED);
}
+/**
+ * Generate max(a, b)
+ * Checks for special cases.
+ * NaN's are handled according to the behavior specified by the
+ * nan_behavior argument.
+ */
+LLVMValueRef
+lp_build_max_ext(struct lp_build_context *bld,
+ LLVMValueRef a,
+ LLVMValueRef b,
+ enum gallivm_nan_behavior nan_behavior)
+{
+ assert(lp_check_value(bld->type, a));
+ assert(lp_check_value(bld->type, b));
+
+ if(a == bld->undef || b == bld->undef)
+ return bld->undef;
+
+ if(a == b)
+ return a;
+
+ if(bld->type.norm) {
+ if(a == bld->one || b == bld->one)
+ return bld->one;
+ if (!bld->type.sign) {
+ if (a == bld->zero) {
+ return b;
+ }
+ if (b == bld->zero) {
+ return a;
+ }
+ }
+ }
+
+ return lp_build_max_simple(bld, a, b, nan_behavior);
+}
+
/**
* Generate clamp(a, min, max)
+ * NaN behavior (for any of a, min, max) is undefined.
* Do checks for special cases.
*/
LLVMValueRef
}
+/**
+ * Generate clamp(a, 0, 1)
+ * A NaN will get converted to zero.
+ */
+LLVMValueRef
+lp_build_clamp_zero_one_nanzero(struct lp_build_context *bld,
+ LLVMValueRef a)
+{
+ a = lp_build_max_ext(bld, a, bld->zero, GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
+ a = lp_build_min(bld, a, bld->one);
+ return a;
+}
+
+
/**
* Generate abs(a)
*/
lp_build_abs(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
- LLVMTypeRef vec_type = lp_build_vec_type(type);
+ LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
assert(lp_check_value(type, a));
return a;
if(type.floating) {
- /* Mask out the sign bit */
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
- unsigned long long absMask = ~(1ULL << (type.width - 1));
- LLVMValueRef mask = lp_build_const_int_vec(type, ((unsigned long long) absMask));
- a = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
- a = LLVMBuildAnd(bld->builder, a, mask, "");
- a = LLVMBuildBitCast(bld->builder, a, vec_type, "");
- return a;
+ if (0x0306 <= HAVE_LLVM && HAVE_LLVM < 0x0309) {
+ /* Workaround llvm.org/PR27332 */
+ LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->gallivm, type);
+ unsigned long long absMask = ~(1ULL << (type.width - 1));
+ LLVMValueRef mask = lp_build_const_int_vec(bld->gallivm, type, ((unsigned long long) absMask));
+ a = LLVMBuildBitCast(builder, a, int_vec_type, "");
+ a = LLVMBuildAnd(builder, a, mask, "");
+ a = LLVMBuildBitCast(builder, a, vec_type, "");
+ return a;
+ } else {
+ char intrinsic[32];
+ lp_format_intrinsic(intrinsic, sizeof intrinsic, "llvm.fabs", vec_type);
+ return lp_build_intrinsic_unary(builder, intrinsic, vec_type, a);
+ }
}
if(type.width*type.length == 128 && util_cpu_caps.has_ssse3) {
switch(type.width) {
case 8:
- return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.b.128", vec_type, a);
+ return lp_build_intrinsic_unary(builder, "llvm.x86.ssse3.pabs.b.128", vec_type, a);
case 16:
- return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.w.128", vec_type, a);
+ return lp_build_intrinsic_unary(builder, "llvm.x86.ssse3.pabs.w.128", vec_type, a);
case 32:
- return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.d.128", vec_type, a);
+ return lp_build_intrinsic_unary(builder, "llvm.x86.ssse3.pabs.d.128", vec_type, a);
}
}
+ else if (type.width*type.length == 256 && util_cpu_caps.has_ssse3 &&
+ (gallivm_debug & GALLIVM_DEBUG_PERF) &&
+ (type.width == 8 || type.width == 16 || type.width == 32)) {
+ debug_printf("%s: inefficient code, should split vectors manually\n",
+ __FUNCTION__);
+ }
- return lp_build_max(bld, a, LLVMBuildNeg(bld->builder, a, ""));
+ return lp_build_max(bld, a, LLVMBuildNeg(builder, a, ""));
}
lp_build_negate(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
+
assert(lp_check_value(bld->type, a));
-#if HAVE_LLVM >= 0x0207
if (bld->type.floating)
- a = LLVMBuildFNeg(bld->builder, a, "");
+ a = LLVMBuildFNeg(builder, a, "");
else
-#endif
- a = LLVMBuildNeg(bld->builder, a, "");
+ a = LLVMBuildNeg(builder, a, "");
return a;
}
lp_build_sgn(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
LLVMValueRef cond;
LLVMValueRef res;
LLVMValueRef one;
unsigned long long maskBit = (unsigned long long)1 << (type.width - 1);
- int_type = lp_build_int_vec_type(type);
- vec_type = lp_build_vec_type(type);
- mask = lp_build_const_int_vec(type, maskBit);
+ int_type = lp_build_int_vec_type(bld->gallivm, type);
+ vec_type = lp_build_vec_type(bld->gallivm, type);
+ mask = lp_build_const_int_vec(bld->gallivm, 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, "");
+ sign = LLVMBuildBitCast(builder, a, int_type, "");
+ sign = LLVMBuildAnd(builder, sign, mask, "");
one = LLVMConstBitCast(bld->one, int_type);
- res = LLVMBuildOr(bld->builder, sign, one, "");
- res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
+ res = LLVMBuildOr(builder, sign, one, "");
+ res = LLVMBuildBitCast(builder, res, vec_type, "");
}
else
{
- LLVMValueRef minus_one = lp_build_const_vec(type, -1.0);
+ /* signed int/norm/fixed point */
+ /* could use psign with sse3 and appropriate vectors here */
+ LLVMValueRef minus_one = lp_build_const_vec(bld->gallivm, type, -1.0);
cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, bld->zero);
res = lp_build_select(bld, cond, bld->one, minus_one);
}
lp_build_set_sign(struct lp_build_context *bld,
LLVMValueRef a, LLVMValueRef sign)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
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_const_int_vec(type, type.width - 1);
- LLVMValueRef mask = lp_build_const_int_vec(type,
+ LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->gallivm, type);
+ LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
+ LLVMValueRef shift = lp_build_const_int_vec(bld->gallivm, type, type.width - 1);
+ LLVMValueRef mask = lp_build_const_int_vec(bld->gallivm, type,
~((unsigned long long) 1 << (type.width - 1)));
LLVMValueRef val, res;
assert(lp_check_value(type, a));
/* val = reinterpret_cast<int>(a) */
- val = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
+ val = LLVMBuildBitCast(builder, a, int_vec_type, "");
/* val = val & mask */
- val = LLVMBuildAnd(bld->builder, val, mask, "");
+ val = LLVMBuildAnd(builder, val, mask, "");
/* sign = sign << shift */
- sign = LLVMBuildShl(bld->builder, sign, shift, "");
+ sign = LLVMBuildShl(builder, sign, shift, "");
/* res = val | sign */
- res = LLVMBuildOr(bld->builder, val, sign, "");
+ res = LLVMBuildOr(builder, val, sign, "");
/* res = reinterpret_cast<float>(res) */
- res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
+ res = LLVMBuildBitCast(builder, res, vec_type, "");
return res;
}
lp_build_int_to_float(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
- LLVMTypeRef vec_type = lp_build_vec_type(type);
+ LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
assert(type.floating);
- return LLVMBuildSIToFP(bld->builder, a, vec_type, "");
+ return LLVMBuildSIToFP(builder, a, vec_type, "");
}
+static boolean
+arch_rounding_available(const struct lp_type type)
+{
+ if ((util_cpu_caps.has_sse4_1 &&
+ (type.length == 1 || type.width*type.length == 128)) ||
+ (util_cpu_caps.has_avx && type.width*type.length == 256))
+ return TRUE;
+ else if ((util_cpu_caps.has_altivec &&
+ (type.width == 32 && type.length == 4)))
+ return TRUE;
+
+ return FALSE;
+}
-
-enum lp_build_round_sse41_mode
+enum lp_build_round_mode
{
- LP_BUILD_ROUND_SSE41_NEAREST = 0,
- LP_BUILD_ROUND_SSE41_FLOOR = 1,
- LP_BUILD_ROUND_SSE41_CEIL = 2,
- LP_BUILD_ROUND_SSE41_TRUNCATE = 3
+ LP_BUILD_ROUND_NEAREST = 0,
+ LP_BUILD_ROUND_FLOOR = 1,
+ LP_BUILD_ROUND_CEIL = 2,
+ LP_BUILD_ROUND_TRUNCATE = 3
};
-
-static INLINE LLVMValueRef
-lp_build_round_sse41(struct lp_build_context *bld,
- LLVMValueRef a,
- enum lp_build_round_sse41_mode mode)
+static inline LLVMValueRef
+lp_build_iround_nearest_sse2(struct lp_build_context *bld,
+ LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
- LLVMTypeRef i32t = LLVMInt32Type();
+ LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
+ LLVMTypeRef ret_type = lp_build_int_vec_type(bld->gallivm, type);
const char *intrinsic;
LLVMValueRef res;
assert(type.floating);
+ /* using the double precision conversions is a bit more complicated */
+ assert(type.width == 32);
assert(lp_check_value(type, a));
- assert(util_cpu_caps.has_sse4_1);
+ assert(util_cpu_caps.has_sse2);
+ /* This is relying on MXCSR rounding mode, which should always be nearest. */
if (type.length == 1) {
LLVMTypeRef vec_type;
LLVMValueRef undef;
- LLVMValueRef args[3];
+ LLVMValueRef arg;
LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
- switch(type.width) {
- case 32:
- intrinsic = "llvm.x86.sse41.round.ss";
- break;
- case 64:
- intrinsic = "llvm.x86.sse41.round.sd";
- break;
- default:
- assert(0);
- return bld->undef;
- }
-
vec_type = LLVMVectorType(bld->elem_type, 4);
- undef = LLVMGetUndef(vec_type);
+ intrinsic = "llvm.x86.sse.cvtss2si";
- args[0] = undef;
- args[1] = LLVMBuildInsertElement(bld->builder, undef, a, index0, "");
- args[2] = LLVMConstInt(i32t, mode, 0);
+ undef = LLVMGetUndef(vec_type);
- res = lp_build_intrinsic(bld->builder, intrinsic,
- vec_type, args, Elements(args));
+ arg = LLVMBuildInsertElement(builder, undef, a, index0, "");
- res = LLVMBuildExtractElement(bld->builder, res, index0, "");
+ res = lp_build_intrinsic_unary(builder, intrinsic,
+ ret_type, arg);
}
else {
- assert(type.width*type.length == 128);
-
- switch(type.width) {
- case 32:
- intrinsic = "llvm.x86.sse41.round.ps";
- break;
- case 64:
- intrinsic = "llvm.x86.sse41.round.pd";
- break;
- default:
- assert(0);
- return bld->undef;
+ if (type.width* type.length == 128) {
+ intrinsic = "llvm.x86.sse2.cvtps2dq";
}
+ else {
+ assert(type.width*type.length == 256);
+ assert(util_cpu_caps.has_avx);
- res = lp_build_intrinsic_binary(bld->builder, intrinsic,
- bld->vec_type, a,
- LLVMConstInt(i32t, mode, 0));
+ intrinsic = "llvm.x86.avx.cvt.ps2dq.256";
+ }
+ res = lp_build_intrinsic_unary(builder, intrinsic,
+ ret_type, a);
}
return res;
}
-/**
- * Return the integer part of a float (vector) value. The returned value is
- * a float (vector).
- * Ex: trunc(-1.5) = 1.0
+/*
*/
-LLVMValueRef
-lp_build_trunc(struct lp_build_context *bld,
- LLVMValueRef a)
+static inline LLVMValueRef
+lp_build_round_altivec(struct lp_build_context *bld,
+ LLVMValueRef a,
+ enum lp_build_round_mode mode)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
+ const char *intrinsic = NULL;
assert(type.floating);
- assert(lp_check_value(type, a));
- if (util_cpu_caps.has_sse4_1 &&
- (type.length == 1 || type.width*type.length == 128)) {
- return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_TRUNCATE);
+ assert(lp_check_value(type, a));
+ assert(util_cpu_caps.has_altivec);
+
+ (void)type;
+
+ switch (mode) {
+ case LP_BUILD_ROUND_NEAREST:
+ intrinsic = "llvm.ppc.altivec.vrfin";
+ break;
+ case LP_BUILD_ROUND_FLOOR:
+ intrinsic = "llvm.ppc.altivec.vrfim";
+ break;
+ case LP_BUILD_ROUND_CEIL:
+ intrinsic = "llvm.ppc.altivec.vrfip";
+ break;
+ case LP_BUILD_ROUND_TRUNCATE:
+ intrinsic = "llvm.ppc.altivec.vrfiz";
+ break;
+ }
+
+ return lp_build_intrinsic_unary(builder, intrinsic, bld->vec_type, a);
+}
+
+static inline LLVMValueRef
+lp_build_round_arch(struct lp_build_context *bld,
+ LLVMValueRef a,
+ enum lp_build_round_mode mode)
+{
+ if (util_cpu_caps.has_sse4_1) {
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ const struct lp_type type = bld->type;
+ const char *intrinsic_root;
+ char intrinsic[32];
+
+ assert(type.floating);
+ assert(lp_check_value(type, a));
+ (void)type;
+
+ switch (mode) {
+ case LP_BUILD_ROUND_NEAREST:
+ intrinsic_root = "llvm.nearbyint";
+ break;
+ case LP_BUILD_ROUND_FLOOR:
+ intrinsic_root = "llvm.floor";
+ break;
+ case LP_BUILD_ROUND_CEIL:
+ intrinsic_root = "llvm.ceil";
+ break;
+ case LP_BUILD_ROUND_TRUNCATE:
+ intrinsic_root = "llvm.trunc";
+ break;
+ }
+
+ lp_format_intrinsic(intrinsic, sizeof intrinsic, intrinsic_root, bld->vec_type);
+ return lp_build_intrinsic_unary(builder, intrinsic, bld->vec_type, a);
+ }
+ else /* (util_cpu_caps.has_altivec) */
+ return lp_build_round_altivec(bld, a, mode);
+}
+
+/**
+ * Return the integer part of a float (vector) value (== round toward zero).
+ * The returned value is a float (vector).
+ * Ex: trunc(-1.5) = -1.0
+ */
+LLVMValueRef
+lp_build_trunc(struct lp_build_context *bld,
+ LLVMValueRef a)
+{
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ const struct lp_type type = bld->type;
+
+ assert(type.floating);
+ assert(lp_check_value(type, a));
+
+ if (arch_rounding_available(type)) {
+ return lp_build_round_arch(bld, a, LP_BUILD_ROUND_TRUNCATE);
}
else {
- LLVMTypeRef vec_type = lp_build_vec_type(type);
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
- LLVMValueRef res;
- res = LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
- res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
- return res;
+ const struct lp_type type = bld->type;
+ struct lp_type inttype;
+ struct lp_build_context intbld;
+ LLVMValueRef cmpval = lp_build_const_vec(bld->gallivm, type, 1<<24);
+ LLVMValueRef trunc, res, anosign, mask;
+ LLVMTypeRef int_vec_type = bld->int_vec_type;
+ LLVMTypeRef vec_type = bld->vec_type;
+
+ assert(type.width == 32); /* might want to handle doubles at some point */
+
+ inttype = type;
+ inttype.floating = 0;
+ lp_build_context_init(&intbld, bld->gallivm, inttype);
+
+ /* round by truncation */
+ trunc = LLVMBuildFPToSI(builder, a, int_vec_type, "");
+ res = LLVMBuildSIToFP(builder, trunc, vec_type, "floor.trunc");
+
+ /* mask out sign bit */
+ anosign = lp_build_abs(bld, a);
+ /*
+ * mask out all values if anosign > 2^24
+ * This should work both for large ints (all rounding is no-op for them
+ * because such floats are always exact) as well as special cases like
+ * NaNs, Infs (taking advantage of the fact they use max exponent).
+ * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
+ */
+ anosign = LLVMBuildBitCast(builder, anosign, int_vec_type, "");
+ cmpval = LLVMBuildBitCast(builder, cmpval, int_vec_type, "");
+ mask = lp_build_cmp(&intbld, PIPE_FUNC_GREATER, anosign, cmpval);
+ return lp_build_select(bld, mask, a, res);
}
}
lp_build_round(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
assert(type.floating);
assert(lp_check_value(type, a));
- if (util_cpu_caps.has_sse4_1 &&
- (type.length == 1 || type.width*type.length == 128)) {
- return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST);
+ if (arch_rounding_available(type)) {
+ return lp_build_round_arch(bld, a, LP_BUILD_ROUND_NEAREST);
}
else {
- LLVMTypeRef vec_type = lp_build_vec_type(type);
- LLVMValueRef res;
+ const struct lp_type type = bld->type;
+ struct lp_type inttype;
+ struct lp_build_context intbld;
+ LLVMValueRef cmpval = lp_build_const_vec(bld->gallivm, type, 1<<24);
+ LLVMValueRef res, anosign, mask;
+ LLVMTypeRef int_vec_type = bld->int_vec_type;
+ LLVMTypeRef vec_type = bld->vec_type;
+
+ assert(type.width == 32); /* might want to handle doubles at some point */
+
+ inttype = type;
+ inttype.floating = 0;
+ lp_build_context_init(&intbld, bld->gallivm, inttype);
+
res = lp_build_iround(bld, a);
- res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
- return res;
+ res = LLVMBuildSIToFP(builder, res, vec_type, "");
+
+ /* mask out sign bit */
+ anosign = lp_build_abs(bld, a);
+ /*
+ * mask out all values if anosign > 2^24
+ * This should work both for large ints (all rounding is no-op for them
+ * because such floats are always exact) as well as special cases like
+ * NaNs, Infs (taking advantage of the fact they use max exponent).
+ * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
+ */
+ anosign = LLVMBuildBitCast(builder, anosign, int_vec_type, "");
+ cmpval = LLVMBuildBitCast(builder, cmpval, int_vec_type, "");
+ mask = lp_build_cmp(&intbld, PIPE_FUNC_GREATER, anosign, cmpval);
+ return lp_build_select(bld, mask, a, res);
}
}
lp_build_floor(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
assert(type.floating);
assert(lp_check_value(type, a));
- if (util_cpu_caps.has_sse4_1 &&
- (type.length == 1 || type.width*type.length == 128)) {
- return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
+ if (arch_rounding_available(type)) {
+ return lp_build_round_arch(bld, a, LP_BUILD_ROUND_FLOOR);
}
else {
- LLVMTypeRef vec_type = lp_build_vec_type(type);
- LLVMValueRef res;
- res = lp_build_ifloor(bld, a);
- res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
- return res;
+ const struct lp_type type = bld->type;
+ struct lp_type inttype;
+ struct lp_build_context intbld;
+ LLVMValueRef cmpval = lp_build_const_vec(bld->gallivm, type, 1<<24);
+ LLVMValueRef trunc, res, anosign, mask;
+ LLVMTypeRef int_vec_type = bld->int_vec_type;
+ LLVMTypeRef vec_type = bld->vec_type;
+
+ if (type.width != 32) {
+ char intrinsic[32];
+ lp_format_intrinsic(intrinsic, sizeof intrinsic, "llvm.floor", vec_type);
+ return lp_build_intrinsic_unary(builder, intrinsic, vec_type, a);
+ }
+
+ assert(type.width == 32); /* might want to handle doubles at some point */
+
+ inttype = type;
+ inttype.floating = 0;
+ lp_build_context_init(&intbld, bld->gallivm, inttype);
+
+ /* round by truncation */
+ trunc = LLVMBuildFPToSI(builder, a, int_vec_type, "");
+ res = LLVMBuildSIToFP(builder, trunc, vec_type, "floor.trunc");
+
+ if (type.sign) {
+ LLVMValueRef tmp;
+
+ /*
+ * fix values if rounding is wrong (for non-special cases)
+ * - this is the case if trunc > a
+ */
+ mask = lp_build_cmp(bld, PIPE_FUNC_GREATER, res, a);
+ /* tmp = trunc > a ? 1.0 : 0.0 */
+ tmp = LLVMBuildBitCast(builder, bld->one, int_vec_type, "");
+ tmp = lp_build_and(&intbld, mask, tmp);
+ tmp = LLVMBuildBitCast(builder, tmp, vec_type, "");
+ res = lp_build_sub(bld, res, tmp);
+ }
+
+ /* mask out sign bit */
+ anosign = lp_build_abs(bld, a);
+ /*
+ * mask out all values if anosign > 2^24
+ * This should work both for large ints (all rounding is no-op for them
+ * because such floats are always exact) as well as special cases like
+ * NaNs, Infs (taking advantage of the fact they use max exponent).
+ * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
+ */
+ anosign = LLVMBuildBitCast(builder, anosign, int_vec_type, "");
+ cmpval = LLVMBuildBitCast(builder, cmpval, int_vec_type, "");
+ mask = lp_build_cmp(&intbld, PIPE_FUNC_GREATER, anosign, cmpval);
+ return lp_build_select(bld, mask, a, res);
}
}
lp_build_ceil(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
assert(type.floating);
assert(lp_check_value(type, a));
- if (util_cpu_caps.has_sse4_1 &&
- (type.length == 1 || type.width*type.length == 128)) {
- return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL);
+ if (arch_rounding_available(type)) {
+ return lp_build_round_arch(bld, a, LP_BUILD_ROUND_CEIL);
}
else {
- LLVMTypeRef vec_type = lp_build_vec_type(type);
- LLVMValueRef res;
- res = lp_build_iceil(bld, a);
- res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
- return res;
+ const struct lp_type type = bld->type;
+ struct lp_type inttype;
+ struct lp_build_context intbld;
+ LLVMValueRef cmpval = lp_build_const_vec(bld->gallivm, type, 1<<24);
+ LLVMValueRef trunc, res, anosign, mask, tmp;
+ LLVMTypeRef int_vec_type = bld->int_vec_type;
+ LLVMTypeRef vec_type = bld->vec_type;
+
+ if (type.width != 32) {
+ char intrinsic[32];
+ lp_format_intrinsic(intrinsic, sizeof intrinsic, "llvm.ceil", vec_type);
+ return lp_build_intrinsic_unary(builder, intrinsic, vec_type, a);
+ }
+
+ assert(type.width == 32); /* might want to handle doubles at some point */
+
+ inttype = type;
+ inttype.floating = 0;
+ lp_build_context_init(&intbld, bld->gallivm, inttype);
+
+ /* round by truncation */
+ trunc = LLVMBuildFPToSI(builder, a, int_vec_type, "");
+ trunc = LLVMBuildSIToFP(builder, trunc, vec_type, "ceil.trunc");
+
+ /*
+ * fix values if rounding is wrong (for non-special cases)
+ * - this is the case if trunc < a
+ */
+ mask = lp_build_cmp(bld, PIPE_FUNC_LESS, trunc, a);
+ /* tmp = trunc < a ? 1.0 : 0.0 */
+ tmp = LLVMBuildBitCast(builder, bld->one, int_vec_type, "");
+ tmp = lp_build_and(&intbld, mask, tmp);
+ tmp = LLVMBuildBitCast(builder, tmp, vec_type, "");
+ res = lp_build_add(bld, trunc, tmp);
+
+ /* mask out sign bit */
+ anosign = lp_build_abs(bld, a);
+ /*
+ * mask out all values if anosign > 2^24
+ * This should work both for large ints (all rounding is no-op for them
+ * because such floats are always exact) as well as special cases like
+ * NaNs, Infs (taking advantage of the fact they use max exponent).
+ * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
+ */
+ anosign = LLVMBuildBitCast(builder, anosign, int_vec_type, "");
+ cmpval = LLVMBuildBitCast(builder, cmpval, int_vec_type, "");
+ mask = lp_build_cmp(&intbld, PIPE_FUNC_GREATER, anosign, cmpval);
+ return lp_build_select(bld, mask, a, res);
}
}
/**
- * Return the integer part of a float (vector) value. The returned value is
- * an integer (vector).
- * Ex: itrunc(-1.5) = 1
+ * Prevent returning a fractional part of 1.0 for very small negative values of
+ * 'a' by clamping against 0.99999(9).
+ */
+static inline LLVMValueRef
+clamp_fract(struct lp_build_context *bld, LLVMValueRef fract)
+{
+ LLVMValueRef max;
+
+ /* this is the largest number smaller than 1.0 representable as float */
+ max = lp_build_const_vec(bld->gallivm, bld->type,
+ 1.0 - 1.0/(1LL << (lp_mantissa(bld->type) + 1)));
+ return lp_build_min(bld, fract, max);
+}
+
+
+/**
+ * Same as lp_build_fract, but guarantees that the result is always smaller
+ * than one.
+ */
+LLVMValueRef
+lp_build_fract_safe(struct lp_build_context *bld,
+ LLVMValueRef a)
+{
+ return clamp_fract(bld, lp_build_fract(bld, a));
+}
+
+
+/**
+ * Return the integer part of a float (vector) value (== round toward zero).
+ * The returned value is an integer (vector).
+ * Ex: itrunc(-1.5) = -1
*/
LLVMValueRef
lp_build_itrunc(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
+ LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->gallivm, type);
assert(type.floating);
assert(lp_check_value(type, a));
- return LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
+ return LLVMBuildFPToSI(builder, a, int_vec_type, "");
}
lp_build_iround(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
LLVMTypeRef int_vec_type = bld->int_vec_type;
LLVMValueRef res;
assert(lp_check_value(type, a));
- if (util_cpu_caps.has_sse4_1 &&
- (type.length == 1 || type.width*type.length == 128)) {
- res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST);
+ if ((util_cpu_caps.has_sse2 &&
+ ((type.width == 32) && (type.length == 1 || type.length == 4))) ||
+ (util_cpu_caps.has_avx && type.width == 32 && type.length == 8)) {
+ return lp_build_iround_nearest_sse2(bld, a);
+ }
+ if (arch_rounding_available(type)) {
+ res = lp_build_round_arch(bld, a, LP_BUILD_ROUND_NEAREST);
}
else {
LLVMValueRef half;
- half = lp_build_const_vec(type, 0.5);
+ half = lp_build_const_vec(bld->gallivm, type, 0.5);
if (type.sign) {
LLVMTypeRef vec_type = bld->vec_type;
- LLVMValueRef mask = lp_build_const_int_vec(type, (unsigned long long)1 << (type.width - 1));
+ LLVMValueRef mask = lp_build_const_int_vec(bld->gallivm, type,
+ (unsigned long long)1 << (type.width - 1));
LLVMValueRef sign;
/* get sign bit */
- sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
- sign = LLVMBuildAnd(bld->builder, sign, mask, "");
+ sign = LLVMBuildBitCast(builder, a, int_vec_type, "");
+ sign = LLVMBuildAnd(builder, sign, mask, "");
/* sign * 0.5 */
- half = LLVMBuildBitCast(bld->builder, half, int_vec_type, "");
- half = LLVMBuildOr(bld->builder, sign, half, "");
- half = LLVMBuildBitCast(bld->builder, half, vec_type, "");
+ half = LLVMBuildBitCast(builder, half, int_vec_type, "");
+ half = LLVMBuildOr(builder, sign, half, "");
+ half = LLVMBuildBitCast(builder, half, vec_type, "");
}
- res = LLVMBuildFAdd(bld->builder, a, half, "");
+ res = LLVMBuildFAdd(builder, a, half, "");
}
- res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
+ res = LLVMBuildFPToSI(builder, res, int_vec_type, "");
return res;
}
lp_build_ifloor(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
LLVMTypeRef int_vec_type = bld->int_vec_type;
LLVMValueRef res;
assert(type.floating);
assert(lp_check_value(type, a));
- if (util_cpu_caps.has_sse4_1 &&
- (type.length == 1 || type.width*type.length == 128)) {
- res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
- }
- else {
- res = a;
-
- if (type.sign) {
- /* Take the sign bit and add it to 1 constant */
- LLVMTypeRef vec_type = bld->vec_type;
- unsigned mantissa = lp_mantissa(type);
- LLVMValueRef mask = lp_build_const_int_vec(type, (unsigned long long)1 << (type.width - 1));
- LLVMValueRef sign;
- LLVMValueRef offset;
+ res = a;
+ if (type.sign) {
+ if (arch_rounding_available(type)) {
+ res = lp_build_round_arch(bld, a, LP_BUILD_ROUND_FLOOR);
+ }
+ else {
+ struct lp_type inttype;
+ struct lp_build_context intbld;
+ LLVMValueRef trunc, itrunc, mask;
- /* sign = a < 0 ? ~0 : 0 */
- sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
- sign = LLVMBuildAnd(bld->builder, sign, mask, "");
- sign = LLVMBuildAShr(bld->builder, sign, lp_build_const_int_vec(type, type.width - 1), "ifloor.sign");
+ assert(type.floating);
+ assert(lp_check_value(type, a));
- /* offset = -0.99999(9)f */
- offset = lp_build_const_vec(type, -(double)(((unsigned long long)1 << mantissa) - 10)/((unsigned long long)1 << mantissa));
- offset = LLVMConstBitCast(offset, int_vec_type);
+ inttype = type;
+ inttype.floating = 0;
+ lp_build_context_init(&intbld, bld->gallivm, inttype);
- /* offset = a < 0 ? offset : 0.0f */
- offset = LLVMBuildAnd(bld->builder, offset, sign, "");
- offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "ifloor.offset");
+ /* round by truncation */
+ itrunc = LLVMBuildFPToSI(builder, a, int_vec_type, "");
+ trunc = LLVMBuildSIToFP(builder, itrunc, bld->vec_type, "ifloor.trunc");
- res = LLVMBuildFAdd(bld->builder, res, offset, "ifloor.res");
+ /*
+ * fix values if rounding is wrong (for non-special cases)
+ * - this is the case if trunc > a
+ * The results of doing this with NaNs, very large values etc.
+ * are undefined but this seems to be the case anyway.
+ */
+ mask = lp_build_cmp(bld, PIPE_FUNC_GREATER, trunc, a);
+ /* cheapie minus one with mask since the mask is minus one / zero */
+ return lp_build_add(&intbld, itrunc, mask);
}
}
/* round to nearest (toward zero) */
- res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "ifloor.res");
+ res = LLVMBuildFPToSI(builder, res, int_vec_type, "ifloor.res");
return res;
}
lp_build_iceil(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
LLVMTypeRef int_vec_type = bld->int_vec_type;
LLVMValueRef res;
assert(type.floating);
assert(lp_check_value(type, a));
- if (util_cpu_caps.has_sse4_1 &&
- (type.length == 1 || type.width*type.length == 128)) {
- res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL);
+ if (arch_rounding_available(type)) {
+ res = lp_build_round_arch(bld, a, LP_BUILD_ROUND_CEIL);
}
else {
- LLVMTypeRef vec_type = bld->vec_type;
- unsigned mantissa = lp_mantissa(type);
- LLVMValueRef offset;
+ struct lp_type inttype;
+ struct lp_build_context intbld;
+ LLVMValueRef trunc, itrunc, mask;
- /* offset = 0.99999(9)f */
- offset = lp_build_const_vec(type, (double)(((unsigned long long)1 << mantissa) - 10)/((unsigned long long)1 << mantissa));
+ assert(type.floating);
+ assert(lp_check_value(type, a));
- if (type.sign) {
- LLVMValueRef mask = lp_build_const_int_vec(type, (unsigned long long)1 << (type.width - 1));
- LLVMValueRef sign;
+ inttype = type;
+ inttype.floating = 0;
+ lp_build_context_init(&intbld, bld->gallivm, inttype);
- /* sign = a < 0 ? 0 : ~0 */
- sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
- sign = LLVMBuildAnd(bld->builder, sign, mask, "");
- sign = LLVMBuildAShr(bld->builder, sign, lp_build_const_int_vec(type, type.width - 1), "iceil.sign");
- sign = LLVMBuildNot(bld->builder, sign, "iceil.not");
-
- /* offset = a < 0 ? 0.0 : offset */
- offset = LLVMConstBitCast(offset, int_vec_type);
- offset = LLVMBuildAnd(bld->builder, offset, sign, "");
- offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "iceil.offset");
- }
+ /* round by truncation */
+ itrunc = LLVMBuildFPToSI(builder, a, int_vec_type, "");
+ trunc = LLVMBuildSIToFP(builder, itrunc, bld->vec_type, "iceil.trunc");
- res = LLVMBuildFAdd(bld->builder, a, offset, "iceil.res");
+ /*
+ * fix values if rounding is wrong (for non-special cases)
+ * - this is the case if trunc < a
+ * The results of doing this with NaNs, very large values etc.
+ * are undefined but this seems to be the case anyway.
+ */
+ mask = lp_build_cmp(bld, PIPE_FUNC_LESS, trunc, a);
+ /* cheapie plus one with mask since the mask is minus one / zero */
+ return lp_build_sub(&intbld, itrunc, mask);
}
/* round to nearest (toward zero) */
- res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "iceil.res");
+ res = LLVMBuildFPToSI(builder, res, int_vec_type, "iceil.res");
return res;
}
* Combined ifloor() & fract().
*
* Preferred to calling the functions separately, as it will ensure that the
- * stratergy (floor() vs ifloor()) that results in less redundant work is used.
+ * strategy (floor() vs ifloor()) that results in less redundant work is used.
*/
void
lp_build_ifloor_fract(struct lp_build_context *bld,
LLVMValueRef *out_ipart,
LLVMValueRef *out_fpart)
{
-
-
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
LLVMValueRef ipart;
assert(type.floating);
assert(lp_check_value(type, a));
- if (util_cpu_caps.has_sse4_1 &&
- (type.length == 1 || type.width*type.length == 128)) {
+ if (arch_rounding_available(type)) {
/*
* floor() is easier.
*/
ipart = lp_build_floor(bld, a);
- *out_fpart = LLVMBuildFSub(bld->builder, a, ipart, "fpart");
- *out_ipart = LLVMBuildFPToSI(bld->builder, ipart, bld->int_vec_type, "ipart");
+ *out_fpart = LLVMBuildFSub(builder, a, ipart, "fpart");
+ *out_ipart = LLVMBuildFPToSI(builder, ipart, bld->int_vec_type, "ipart");
}
else {
/*
*/
*out_ipart = lp_build_ifloor(bld, a);
- ipart = LLVMBuildSIToFP(bld->builder, *out_ipart, bld->vec_type, "ipart");
- *out_fpart = LLVMBuildFSub(bld->builder, a, ipart, "fpart");
+ ipart = LLVMBuildSIToFP(builder, *out_ipart, bld->vec_type, "ipart");
+ *out_fpart = LLVMBuildFSub(builder, a, ipart, "fpart");
}
}
+/**
+ * Same as lp_build_ifloor_fract, but guarantees that the fractional part is
+ * always smaller than one.
+ */
+void
+lp_build_ifloor_fract_safe(struct lp_build_context *bld,
+ LLVMValueRef a,
+ LLVMValueRef *out_ipart,
+ LLVMValueRef *out_fpart)
+{
+ lp_build_ifloor_fract(bld, a, out_ipart, out_fpart);
+ *out_fpart = clamp_fract(bld, *out_fpart);
+}
+
+
LLVMValueRef
lp_build_sqrt(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
- LLVMTypeRef vec_type = lp_build_vec_type(type);
+ LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
char intrinsic[32];
assert(lp_check_value(type, a));
- /* TODO: optimize the constant case */
- /* TODO: optimize the constant case */
-
assert(type.floating);
- util_snprintf(intrinsic, sizeof intrinsic, "llvm.sqrt.v%uf%u", type.length, type.width);
+ lp_format_intrinsic(intrinsic, sizeof intrinsic, "llvm.sqrt", vec_type);
- return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a);
+ return lp_build_intrinsic_unary(builder, intrinsic, vec_type, a);
}
* - http://en.wikipedia.org/wiki/Division_(digital)#Newton.E2.80.93Raphson_division
* - http://softwarecommunity.intel.com/articles/eng/1818.htm
*/
-static INLINE LLVMValueRef
+static inline LLVMValueRef
lp_build_rcp_refine(struct lp_build_context *bld,
LLVMValueRef a,
LLVMValueRef rcp_a)
{
- LLVMValueRef two = lp_build_const_vec(bld->type, 2.0);
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ LLVMValueRef two = lp_build_const_vec(bld->gallivm, bld->type, 2.0);
LLVMValueRef res;
- res = LLVMBuildFMul(bld->builder, a, rcp_a, "");
- res = LLVMBuildFSub(bld->builder, two, res, "");
- res = LLVMBuildFMul(bld->builder, rcp_a, res, "");
+ res = LLVMBuildFMul(builder, a, rcp_a, "");
+ res = LLVMBuildFSub(builder, two, res, "");
+ res = LLVMBuildFMul(builder, rcp_a, res, "");
return res;
}
lp_build_rcp(struct lp_build_context *bld,
LLVMValueRef a)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
assert(lp_check_value(type, a));
* - it doesn't even get the reciprocate of 1.0 exactly
* - doing Newton-Rapshon steps yields wrong (NaN) values for 0.0 or Inf
* - for recent processors the benefit over DIVPS is marginal, a case
- * depedent
+ * dependent
*
* We could still use it on certain processors if benchmarks show that the
* RCPPS plus necessary workarounds are still preferrable to DIVPS; or for
* particular uses that require less workarounds.
*/
- if (FALSE && util_cpu_caps.has_sse && type.width == 32 && type.length == 4) {
+ if (FALSE && ((util_cpu_caps.has_sse && type.width == 32 && type.length == 4) ||
+ (util_cpu_caps.has_avx && type.width == 32 && type.length == 8))){
const unsigned num_iterations = 0;
LLVMValueRef res;
unsigned i;
+ const char *intrinsic = NULL;
+
+ if (type.length == 4) {
+ intrinsic = "llvm.x86.sse.rcp.ps";
+ }
+ else {
+ intrinsic = "llvm.x86.avx.rcp.ps.256";
+ }
- res = lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rcp.ps", bld->vec_type, a);
+ res = lp_build_intrinsic_unary(builder, intrinsic, bld->vec_type, a);
for (i = 0; i < num_iterations; ++i) {
res = lp_build_rcp_refine(bld, a, res);
return res;
}
- return LLVMBuildFDiv(bld->builder, bld->one, a, "");
+ return LLVMBuildFDiv(builder, bld->one, a, "");
}
*
* x_{i+1} = 0.5 * x_i * (3.0 - a * x_i * x_i)
*
- * See also:
- * - http://softwarecommunity.intel.com/articles/eng/1818.htm
+ * See also Intel 64 and IA-32 Architectures Optimization Manual.
*/
-static INLINE LLVMValueRef
+static inline LLVMValueRef
lp_build_rsqrt_refine(struct lp_build_context *bld,
LLVMValueRef a,
LLVMValueRef rsqrt_a)
{
- LLVMValueRef half = lp_build_const_vec(bld->type, 0.5);
- LLVMValueRef three = lp_build_const_vec(bld->type, 3.0);
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ LLVMValueRef half = lp_build_const_vec(bld->gallivm, bld->type, 0.5);
+ LLVMValueRef three = lp_build_const_vec(bld->gallivm, bld->type, 3.0);
LLVMValueRef res;
- res = LLVMBuildFMul(bld->builder, rsqrt_a, rsqrt_a, "");
- res = LLVMBuildFMul(bld->builder, a, res, "");
- res = LLVMBuildFSub(bld->builder, three, res, "");
- res = LLVMBuildFMul(bld->builder, rsqrt_a, res, "");
- res = LLVMBuildFMul(bld->builder, half, res, "");
+ res = LLVMBuildFMul(builder, rsqrt_a, rsqrt_a, "");
+ res = LLVMBuildFMul(builder, a, res, "");
+ res = LLVMBuildFSub(builder, three, res, "");
+ res = LLVMBuildFMul(builder, rsqrt_a, res, "");
+ res = LLVMBuildFMul(builder, half, res, "");
return res;
}
/**
- * Generate 1/sqrt(a)
+ * Generate 1/sqrt(a).
+ * Result is undefined for values < 0, infinity for +0.
*/
LLVMValueRef
lp_build_rsqrt(struct lp_build_context *bld,
assert(type.floating);
- if (util_cpu_caps.has_sse && type.width == 32 && type.length == 4) {
- const unsigned num_iterations = 0;
+ /*
+ * This should be faster but all denormals will end up as infinity.
+ */
+ if (0 && lp_build_fast_rsqrt_available(type)) {
+ const unsigned num_iterations = 1;
LLVMValueRef res;
unsigned i;
- res = lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rsqrt.ps", bld->vec_type, a);
+ /* rsqrt(1.0) != 1.0 here */
+ res = lp_build_fast_rsqrt(bld, a);
- for (i = 0; i < num_iterations; ++i) {
- res = lp_build_rsqrt_refine(bld, a, res);
+ if (num_iterations) {
+ /*
+ * Newton-Raphson will result in NaN instead of infinity for zero,
+ * and NaN instead of zero for infinity.
+ * Also, need to ensure rsqrt(1.0) == 1.0.
+ * All numbers smaller than FLT_MIN will result in +infinity
+ * (rsqrtps treats all denormals as zero).
+ */
+ LLVMValueRef cmp;
+ LLVMValueRef flt_min = lp_build_const_vec(bld->gallivm, type, FLT_MIN);
+ LLVMValueRef inf = lp_build_const_vec(bld->gallivm, type, INFINITY);
+
+ for (i = 0; i < num_iterations; ++i) {
+ res = lp_build_rsqrt_refine(bld, a, res);
+ }
+ cmp = lp_build_compare(bld->gallivm, type, PIPE_FUNC_LESS, a, flt_min);
+ res = lp_build_select(bld, cmp, inf, res);
+ cmp = lp_build_compare(bld->gallivm, type, PIPE_FUNC_EQUAL, a, inf);
+ res = lp_build_select(bld, cmp, bld->zero, res);
+ cmp = lp_build_compare(bld->gallivm, type, PIPE_FUNC_EQUAL, a, bld->one);
+ res = lp_build_select(bld, cmp, bld->one, res);
}
return res;
return lp_build_rcp(bld, lp_build_sqrt(bld, a));
}
-
-static inline LLVMValueRef
-lp_build_const_v4si(unsigned long value)
+/**
+ * If there's a fast (inaccurate) rsqrt instruction available
+ * (caller may want to avoid to call rsqrt_fast if it's not available,
+ * i.e. for calculating x^0.5 it may do rsqrt_fast(x) * x but if
+ * unavailable it would result in sqrt/div/mul so obviously
+ * much better to just call sqrt, skipping both div and mul).
+ */
+boolean
+lp_build_fast_rsqrt_available(struct lp_type type)
{
- LLVMValueRef element = LLVMConstInt(LLVMInt32Type(), value, 0);
- LLVMValueRef elements[4] = { element, element, element, element };
- return LLVMConstVector(elements, 4);
+ assert(type.floating);
+
+ if ((util_cpu_caps.has_sse && type.width == 32 && type.length == 4) ||
+ (util_cpu_caps.has_avx && type.width == 32 && type.length == 8)) {
+ return true;
+ }
+ return false;
}
-static inline LLVMValueRef
-lp_build_const_v4sf(float value)
+
+/**
+ * Generate 1/sqrt(a).
+ * Result is undefined for values < 0, infinity for +0.
+ * Precision is limited, only ~10 bits guaranteed
+ * (rsqrt 1.0 may not be 1.0, denorms may be flushed to 0).
+ */
+LLVMValueRef
+lp_build_fast_rsqrt(struct lp_build_context *bld,
+ LLVMValueRef a)
{
- LLVMValueRef element = LLVMConstReal(LLVMFloatType(), value);
- LLVMValueRef elements[4] = { element, element, element, element };
- return LLVMConstVector(elements, 4);
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ const struct lp_type type = bld->type;
+
+ assert(lp_check_value(type, a));
+
+ if (lp_build_fast_rsqrt_available(type)) {
+ const char *intrinsic = NULL;
+
+ if (type.length == 4) {
+ intrinsic = "llvm.x86.sse.rsqrt.ps";
+ }
+ else {
+ intrinsic = "llvm.x86.avx.rsqrt.ps.256";
+ }
+ return lp_build_intrinsic_unary(builder, intrinsic, bld->vec_type, a);
+ }
+ else {
+ debug_printf("%s: emulating fast rsqrt with rcp/sqrt\n", __FUNCTION__);
+ }
+ return lp_build_rcp(bld, lp_build_sqrt(bld, a));
}
/**
- * Generate sin(a) using SSE2
+ * Generate sin(a) or cos(a) using polynomial approximation.
+ * TODO: it might be worth recognizing sin and cos using same source
+ * (i.e. d3d10 sincos opcode). Obviously doing both at the same time
+ * would be way cheaper than calculating (nearly) everything twice...
+ * Not sure it's common enough to be worth bothering however, scs
+ * opcode could also benefit from calculating both though.
*/
-LLVMValueRef
-lp_build_sin(struct lp_build_context *bld,
- LLVMValueRef a)
+static LLVMValueRef
+lp_build_sin_or_cos(struct lp_build_context *bld,
+ LLVMValueRef a,
+ boolean cos)
{
+ struct gallivm_state *gallivm = bld->gallivm;
+ LLVMBuilderRef b = gallivm->builder;
struct lp_type int_type = lp_int_type(bld->type);
- LLVMBuilderRef b = bld->builder;
- LLVMTypeRef v4sf = LLVMVectorType(LLVMFloatType(), 4);
- LLVMTypeRef v4si = LLVMVectorType(LLVMInt32Type(), 4);
/*
* take the absolute value,
* x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
*/
- LLVMValueRef inv_sig_mask = lp_build_const_v4si(~0x80000000);
- LLVMValueRef a_v4si = LLVMBuildBitCast(b, a, v4si, "a_v4si");
+ LLVMValueRef inv_sig_mask = lp_build_const_int_vec(gallivm, bld->type, ~0x80000000);
+ LLVMValueRef a_v4si = LLVMBuildBitCast(b, a, bld->int_vec_type, "a_v4si");
LLVMValueRef absi = LLVMBuildAnd(b, a_v4si, inv_sig_mask, "absi");
- LLVMValueRef x_abs = LLVMBuildBitCast(b, absi, v4sf, "x_abs");
-
- /*
- * extract the sign bit (upper one)
- * sign_bit = _mm_and_ps(sign_bit, *(v4sf*)_ps_sign_mask);
- */
- LLVMValueRef sig_mask = lp_build_const_v4si(0x80000000);
- LLVMValueRef sign_bit_i = LLVMBuildAnd(b, a_v4si, sig_mask, "sign_bit_i");
+ LLVMValueRef x_abs = LLVMBuildBitCast(b, absi, bld->vec_type, "x_abs");
/*
* scale by 4/Pi
* y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
*/
-
- LLVMValueRef FOPi = lp_build_const_v4sf(1.27323954473516);
+
+ LLVMValueRef FOPi = lp_build_const_vec(gallivm, bld->type, 1.27323954473516);
LLVMValueRef scale_y = LLVMBuildFMul(b, x_abs, FOPi, "scale_y");
/*
* store the integer part of y in mm0
* emm2 = _mm_cvttps_epi32(y);
*/
-
- LLVMValueRef emm2_i = LLVMBuildFPToSI(b, scale_y, v4si, "emm2_i");
+
+ LLVMValueRef emm2_i = LLVMBuildFPToSI(b, scale_y, bld->int_vec_type, "emm2_i");
/*
* j=(j+1) & (~1) (see the cephes sources)
* emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
*/
- LLVMValueRef all_one = lp_build_const_v4si(1);
+ LLVMValueRef all_one = lp_build_const_int_vec(gallivm, bld->type, 1);
LLVMValueRef emm2_add = LLVMBuildAdd(b, emm2_i, all_one, "emm2_add");
/*
* emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
*/
- LLVMValueRef inv_one = lp_build_const_v4si(~1);
+ LLVMValueRef inv_one = lp_build_const_int_vec(gallivm, bld->type, ~1);
LLVMValueRef emm2_and = LLVMBuildAnd(b, emm2_add, inv_one, "emm2_and");
/*
* y = _mm_cvtepi32_ps(emm2);
*/
- LLVMValueRef y_2 = LLVMBuildSIToFP(b, emm2_and, v4sf, "y_2");
+ LLVMValueRef y_2 = LLVMBuildSIToFP(b, emm2_and, bld->vec_type, "y_2");
+
+ LLVMValueRef const_2 = lp_build_const_int_vec(gallivm, bld->type, 2);
+ LLVMValueRef const_4 = lp_build_const_int_vec(gallivm, bld->type, 4);
+ LLVMValueRef const_29 = lp_build_const_int_vec(gallivm, bld->type, 29);
+ LLVMValueRef sign_mask = lp_build_const_int_vec(gallivm, bld->type, 0x80000000);
- /* get the swap sign flag
- * emm0 = _mm_and_si128(emm2, *(v4si*)_pi32_4);
- */
- LLVMValueRef pi32_4 = lp_build_const_v4si(4);
- LLVMValueRef emm0_and = LLVMBuildAnd(b, emm2_add, pi32_4, "emm0_and");
-
/*
- * emm2 = _mm_slli_epi32(emm0, 29);
- */
- LLVMValueRef const_29 = lp_build_const_v4si(29);
- LLVMValueRef swap_sign_bit = LLVMBuildShl(b, emm0_and, const_29, "swap_sign_bit");
+ * Argument used for poly selection and sign bit determination
+ * is different for sin vs. cos.
+ */
+ LLVMValueRef emm2_2 = cos ? LLVMBuildSub(b, emm2_and, const_2, "emm2_2") :
+ emm2_and;
+
+ LLVMValueRef sign_bit = cos ? LLVMBuildShl(b, LLVMBuildAnd(b, const_4,
+ LLVMBuildNot(b, emm2_2, ""), ""),
+ const_29, "sign_bit") :
+ LLVMBuildAnd(b, LLVMBuildXor(b, a_v4si,
+ LLVMBuildShl(b, emm2_add,
+ const_29, ""), ""),
+ sign_mask, "sign_bit");
/*
- * get the polynom selection mask
+ * get the polynom selection mask
* there is one polynom for 0 <= x <= Pi/4
* and another one for Pi/4<x<=Pi/2
* Both branches will be computed.
- *
+ *
* emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
* emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
*/
- LLVMValueRef pi32_2 = lp_build_const_v4si(2);
- LLVMValueRef emm2_3 = LLVMBuildAnd(b, emm2_and, pi32_2, "emm2_3");
- LLVMValueRef poly_mask = lp_build_compare(b, int_type, PIPE_FUNC_EQUAL,
- emm2_3, lp_build_const_v4si(0));
- /*
- * sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit);
- */
- LLVMValueRef sign_bit_1 = LLVMBuildXor(b, sign_bit_i, swap_sign_bit, "sign_bit");
+ LLVMValueRef emm2_3 = LLVMBuildAnd(b, emm2_2, const_2, "emm2_3");
+ LLVMValueRef poly_mask = lp_build_compare(gallivm,
+ int_type, PIPE_FUNC_EQUAL,
+ emm2_3, lp_build_const_int_vec(gallivm, bld->type, 0));
/*
* _PS_CONST(minus_cephes_DP1, -0.78515625);
* _PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4);
* _PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8);
*/
- LLVMValueRef DP1 = lp_build_const_v4sf(-0.78515625);
- LLVMValueRef DP2 = lp_build_const_v4sf(-2.4187564849853515625e-4);
- LLVMValueRef DP3 = lp_build_const_v4sf(-3.77489497744594108e-8);
+ LLVMValueRef DP1 = lp_build_const_vec(gallivm, bld->type, -0.78515625);
+ LLVMValueRef DP2 = lp_build_const_vec(gallivm, bld->type, -2.4187564849853515625e-4);
+ LLVMValueRef DP3 = lp_build_const_vec(gallivm, bld->type, -3.77489497744594108e-8);
/*
- * The magic pass: "Extended precision modular arithmetic"
- * x = ((x - y * DP1) - y * DP2) - y * DP3;
+ * The magic pass: "Extended precision modular arithmetic"
+ * x = ((x - y * DP1) - y * DP2) - y * DP3;
* xmm1 = _mm_mul_ps(y, xmm1);
* xmm2 = _mm_mul_ps(y, xmm2);
* xmm3 = _mm_mul_ps(y, xmm3);
* x = _mm_add_ps(x, xmm1);
* x = _mm_add_ps(x, xmm2);
* x = _mm_add_ps(x, xmm3);
- */
+ */
LLVMValueRef x_1 = LLVMBuildFAdd(b, x_abs, xmm1, "x_1");
LLVMValueRef x_2 = LLVMBuildFAdd(b, x_1, xmm2, "x_2");
* _PS_CONST(coscof_p1, -1.388731625493765E-003);
* _PS_CONST(coscof_p2, 4.166664568298827E-002);
*/
- LLVMValueRef coscof_p0 = lp_build_const_v4sf(2.443315711809948E-005);
- LLVMValueRef coscof_p1 = lp_build_const_v4sf(-1.388731625493765E-003);
- LLVMValueRef coscof_p2 = lp_build_const_v4sf(4.166664568298827E-002);
+ LLVMValueRef coscof_p0 = lp_build_const_vec(gallivm, bld->type, 2.443315711809948E-005);
+ LLVMValueRef coscof_p1 = lp_build_const_vec(gallivm, bld->type, -1.388731625493765E-003);
+ LLVMValueRef coscof_p2 = lp_build_const_vec(gallivm, bld->type, 4.166664568298827E-002);
/*
* y = *(v4sf*)_ps_coscof_p0;
* tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
* y = _mm_sub_ps(y, tmp);
* y = _mm_add_ps(y, *(v4sf*)_ps_1);
- */
- LLVMValueRef half = lp_build_const_v4sf(0.5);
+ */
+ LLVMValueRef half = lp_build_const_vec(gallivm, bld->type, 0.5);
LLVMValueRef tmp = LLVMBuildFMul(b, z, half, "tmp");
LLVMValueRef y_9 = LLVMBuildFSub(b, y_8, tmp, "y_8");
- LLVMValueRef one = lp_build_const_v4sf(1.0);
+ LLVMValueRef one = lp_build_const_vec(gallivm, bld->type, 1.0);
LLVMValueRef y_10 = LLVMBuildFAdd(b, y_9, one, "y_9");
/*
* _PS_CONST(sincof_p1, 8.3321608736E-3);
* _PS_CONST(sincof_p2, -1.6666654611E-1);
*/
- LLVMValueRef sincof_p0 = lp_build_const_v4sf(-1.9515295891E-4);
- LLVMValueRef sincof_p1 = lp_build_const_v4sf(8.3321608736E-3);
- LLVMValueRef sincof_p2 = lp_build_const_v4sf(-1.6666654611E-1);
+ LLVMValueRef sincof_p0 = lp_build_const_vec(gallivm, bld->type, -1.9515295891E-4);
+ LLVMValueRef sincof_p1 = lp_build_const_vec(gallivm, bld->type, 8.3321608736E-3);
+ LLVMValueRef sincof_p2 = lp_build_const_vec(gallivm, bld->type, -1.6666654611E-1);
/*
* Evaluate the second polynom (Pi/4 <= x <= 0)
* xmm3 = poly_mask;
* y2 = _mm_and_ps(xmm3, y2); //, xmm3);
* y = _mm_andnot_ps(xmm3, y);
- * y = _mm_add_ps(y,y2);
+ * y = _mm_or_ps(y,y2);
*/
- LLVMValueRef y2_i = LLVMBuildBitCast(b, y2_9, v4si, "y2_i");
- LLVMValueRef y_i = LLVMBuildBitCast(b, y_10, v4si, "y_i");
+ LLVMValueRef y2_i = LLVMBuildBitCast(b, y2_9, bld->int_vec_type, "y2_i");
+ LLVMValueRef y_i = LLVMBuildBitCast(b, y_10, bld->int_vec_type, "y_i");
LLVMValueRef y2_and = LLVMBuildAnd(b, y2_i, poly_mask, "y2_and");
- LLVMValueRef inv = lp_build_const_v4si(~0);
- LLVMValueRef poly_mask_inv = LLVMBuildXor(b, poly_mask, inv, "poly_mask_inv");
+ LLVMValueRef poly_mask_inv = LLVMBuildNot(b, poly_mask, "poly_mask_inv");
LLVMValueRef y_and = LLVMBuildAnd(b, y_i, poly_mask_inv, "y_and");
- LLVMValueRef y_combine = LLVMBuildAdd(b, y_and, y2_and, "y_combine");
+ LLVMValueRef y_combine = LLVMBuildOr(b, y_and, y2_and, "y_combine");
/*
* update the sign
* y = _mm_xor_ps(y, sign_bit);
*/
- LLVMValueRef y_sign = LLVMBuildXor(b, y_combine, sign_bit_1, "y_sin");
- LLVMValueRef y_result = LLVMBuildBitCast(b, y_sign, v4sf, "y_result");
+ LLVMValueRef y_sign = LLVMBuildXor(b, y_combine, sign_bit, "y_sign");
+ LLVMValueRef y_result = LLVMBuildBitCast(b, y_sign, bld->vec_type, "y_result");
+
+ LLVMValueRef isfinite = lp_build_isfinite(bld, a);
+
+ /* clamp output to be within [-1, 1] */
+ y_result = lp_build_clamp(bld, y_result,
+ lp_build_const_vec(bld->gallivm, bld->type, -1.f),
+ lp_build_const_vec(bld->gallivm, bld->type, 1.f));
+ /* If a is -inf, inf or NaN then return NaN */
+ y_result = lp_build_select(bld, isfinite, y_result,
+ lp_build_const_vec(bld->gallivm, bld->type, NAN));
return y_result;
}
/**
- * Generate cos(a) using SSE2
+ * Generate sin(a)
*/
LLVMValueRef
-lp_build_cos(struct lp_build_context *bld,
+lp_build_sin(struct lp_build_context *bld,
LLVMValueRef a)
{
- struct lp_type int_type = lp_int_type(bld->type);
- LLVMBuilderRef b = bld->builder;
- LLVMTypeRef v4sf = LLVMVectorType(LLVMFloatType(), 4);
- LLVMTypeRef v4si = LLVMVectorType(LLVMInt32Type(), 4);
-
- /*
- * take the absolute value,
- * x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
- */
+ return lp_build_sin_or_cos(bld, a, FALSE);
+}
- LLVMValueRef inv_sig_mask = lp_build_const_v4si(~0x80000000);
- LLVMValueRef a_v4si = LLVMBuildBitCast(b, a, v4si, "a_v4si");
- LLVMValueRef absi = LLVMBuildAnd(b, a_v4si, inv_sig_mask, "absi");
- LLVMValueRef x_abs = LLVMBuildBitCast(b, absi, v4sf, "x_abs");
+/**
+ * Generate cos(a)
+ */
+LLVMValueRef
+lp_build_cos(struct lp_build_context *bld,
+ LLVMValueRef a)
+{
+ return lp_build_sin_or_cos(bld, a, TRUE);
+}
- /*
- * scale by 4/Pi
- * y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
- */
-
- LLVMValueRef FOPi = lp_build_const_v4sf(1.27323954473516);
- LLVMValueRef scale_y = LLVMBuildFMul(b, x_abs, FOPi, "scale_y");
- /*
- * store the integer part of y in mm0
- * emm2 = _mm_cvttps_epi32(y);
- */
-
- LLVMValueRef emm2_i = LLVMBuildFPToSI(b, scale_y, v4si, "emm2_i");
+/**
+ * Generate pow(x, y)
+ */
+LLVMValueRef
+lp_build_pow(struct lp_build_context *bld,
+ LLVMValueRef x,
+ LLVMValueRef y)
+{
+ /* TODO: optimize the constant case */
+ if (gallivm_debug & GALLIVM_DEBUG_PERF &&
+ LLVMIsConstant(x) && LLVMIsConstant(y)) {
+ debug_printf("%s: inefficient/imprecise constant arithmetic\n",
+ __FUNCTION__);
+ }
- /*
- * j=(j+1) & (~1) (see the cephes sources)
- * emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
- */
+ return lp_build_exp2(bld, lp_build_mul(bld, lp_build_log2(bld, x), y));
+}
- LLVMValueRef all_one = lp_build_const_v4si(1);
- LLVMValueRef emm2_add = LLVMBuildAdd(b, emm2_i, all_one, "emm2_add");
- /*
- * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
- */
- LLVMValueRef inv_one = lp_build_const_v4si(~1);
- LLVMValueRef emm2_and = LLVMBuildAnd(b, emm2_add, inv_one, "emm2_and");
- /*
- * y = _mm_cvtepi32_ps(emm2);
- */
- LLVMValueRef y_2 = LLVMBuildSIToFP(b, emm2_and, v4sf, "y_2");
+/**
+ * Generate exp(x)
+ */
+LLVMValueRef
+lp_build_exp(struct lp_build_context *bld,
+ LLVMValueRef x)
+{
+ /* log2(e) = 1/log(2) */
+ LLVMValueRef log2e = lp_build_const_vec(bld->gallivm, bld->type,
+ 1.4426950408889634);
+ assert(lp_check_value(bld->type, x));
- /*
- * emm2 = _mm_sub_epi32(emm2, *(v4si*)_pi32_2);
- */
- LLVMValueRef const_2 = lp_build_const_v4si(2);
- LLVMValueRef emm2_2 = LLVMBuildSub(b, emm2_and, const_2, "emm2_2");
-
-
- /* get the swap sign flag
- * emm0 = _mm_andnot_si128(emm2, *(v4si*)_pi32_4);
- */
- LLVMValueRef inv = lp_build_const_v4si(~0);
- LLVMValueRef emm0_not = LLVMBuildXor(b, emm2_2, inv, "emm0_not");
- LLVMValueRef pi32_4 = lp_build_const_v4si(4);
- LLVMValueRef emm0_and = LLVMBuildAnd(b, emm0_not, pi32_4, "emm0_and");
-
- /*
- * emm2 = _mm_slli_epi32(emm0, 29);
- */
- LLVMValueRef const_29 = lp_build_const_v4si(29);
- LLVMValueRef sign_bit = LLVMBuildShl(b, emm0_and, const_29, "sign_bit");
-
- /*
- * get the polynom selection mask
- * there is one polynom for 0 <= x <= Pi/4
- * and another one for Pi/4<x<=Pi/2
- * Both branches will be computed.
- *
- * emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
- * emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
- */
-
- LLVMValueRef pi32_2 = lp_build_const_v4si(2);
- LLVMValueRef emm2_3 = LLVMBuildAnd(b, emm2_2, pi32_2, "emm2_3");
- LLVMValueRef poly_mask = lp_build_compare(b, int_type, PIPE_FUNC_EQUAL,
- emm2_3, lp_build_const_v4si(0));
-
- /*
- * _PS_CONST(minus_cephes_DP1, -0.78515625);
- * _PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4);
- * _PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8);
- */
- LLVMValueRef DP1 = lp_build_const_v4sf(-0.78515625);
- LLVMValueRef DP2 = lp_build_const_v4sf(-2.4187564849853515625e-4);
- LLVMValueRef DP3 = lp_build_const_v4sf(-3.77489497744594108e-8);
-
- /*
- * The magic pass: "Extended precision modular arithmetic"
- * x = ((x - y * DP1) - y * DP2) - y * DP3;
- * xmm1 = _mm_mul_ps(y, xmm1);
- * xmm2 = _mm_mul_ps(y, xmm2);
- * xmm3 = _mm_mul_ps(y, xmm3);
- */
- LLVMValueRef xmm1 = LLVMBuildFMul(b, y_2, DP1, "xmm1");
- LLVMValueRef xmm2 = LLVMBuildFMul(b, y_2, DP2, "xmm2");
- LLVMValueRef xmm3 = LLVMBuildFMul(b, y_2, DP3, "xmm3");
-
- /*
- * x = _mm_add_ps(x, xmm1);
- * x = _mm_add_ps(x, xmm2);
- * x = _mm_add_ps(x, xmm3);
- */
-
- LLVMValueRef x_1 = LLVMBuildFAdd(b, x_abs, xmm1, "x_1");
- LLVMValueRef x_2 = LLVMBuildFAdd(b, x_1, xmm2, "x_2");
- LLVMValueRef x_3 = LLVMBuildFAdd(b, x_2, xmm3, "x_3");
-
- /*
- * Evaluate the first polynom (0 <= x <= Pi/4)
- *
- * z = _mm_mul_ps(x,x);
- */
- LLVMValueRef z = LLVMBuildFMul(b, x_3, x_3, "z");
-
- /*
- * _PS_CONST(coscof_p0, 2.443315711809948E-005);
- * _PS_CONST(coscof_p1, -1.388731625493765E-003);
- * _PS_CONST(coscof_p2, 4.166664568298827E-002);
- */
- LLVMValueRef coscof_p0 = lp_build_const_v4sf(2.443315711809948E-005);
- LLVMValueRef coscof_p1 = lp_build_const_v4sf(-1.388731625493765E-003);
- LLVMValueRef coscof_p2 = lp_build_const_v4sf(4.166664568298827E-002);
-
- /*
- * y = *(v4sf*)_ps_coscof_p0;
- * y = _mm_mul_ps(y, z);
- */
- LLVMValueRef y_3 = LLVMBuildFMul(b, z, coscof_p0, "y_3");
- LLVMValueRef y_4 = LLVMBuildFAdd(b, y_3, coscof_p1, "y_4");
- LLVMValueRef y_5 = LLVMBuildFMul(b, y_4, z, "y_5");
- LLVMValueRef y_6 = LLVMBuildFAdd(b, y_5, coscof_p2, "y_6");
- LLVMValueRef y_7 = LLVMBuildFMul(b, y_6, z, "y_7");
- LLVMValueRef y_8 = LLVMBuildFMul(b, y_7, z, "y_8");
-
-
- /*
- * tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
- * y = _mm_sub_ps(y, tmp);
- * y = _mm_add_ps(y, *(v4sf*)_ps_1);
- */
- LLVMValueRef half = lp_build_const_v4sf(0.5);
- LLVMValueRef tmp = LLVMBuildFMul(b, z, half, "tmp");
- LLVMValueRef y_9 = LLVMBuildFSub(b, y_8, tmp, "y_8");
- LLVMValueRef one = lp_build_const_v4sf(1.0);
- LLVMValueRef y_10 = LLVMBuildFAdd(b, y_9, one, "y_9");
-
- /*
- * _PS_CONST(sincof_p0, -1.9515295891E-4);
- * _PS_CONST(sincof_p1, 8.3321608736E-3);
- * _PS_CONST(sincof_p2, -1.6666654611E-1);
- */
- LLVMValueRef sincof_p0 = lp_build_const_v4sf(-1.9515295891E-4);
- LLVMValueRef sincof_p1 = lp_build_const_v4sf(8.3321608736E-3);
- LLVMValueRef sincof_p2 = lp_build_const_v4sf(-1.6666654611E-1);
-
- /*
- * Evaluate the second polynom (Pi/4 <= x <= 0)
- *
- * y2 = *(v4sf*)_ps_sincof_p0;
- * y2 = _mm_mul_ps(y2, z);
- * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p1);
- * y2 = _mm_mul_ps(y2, z);
- * y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p2);
- * y2 = _mm_mul_ps(y2, z);
- * y2 = _mm_mul_ps(y2, x);
- * y2 = _mm_add_ps(y2, x);
- */
-
- LLVMValueRef y2_3 = LLVMBuildFMul(b, z, sincof_p0, "y2_3");
- LLVMValueRef y2_4 = LLVMBuildFAdd(b, y2_3, sincof_p1, "y2_4");
- LLVMValueRef y2_5 = LLVMBuildFMul(b, y2_4, z, "y2_5");
- LLVMValueRef y2_6 = LLVMBuildFAdd(b, y2_5, sincof_p2, "y2_6");
- LLVMValueRef y2_7 = LLVMBuildFMul(b, y2_6, z, "y2_7");
- LLVMValueRef y2_8 = LLVMBuildFMul(b, y2_7, x_3, "y2_8");
- LLVMValueRef y2_9 = LLVMBuildFAdd(b, y2_8, x_3, "y2_9");
-
- /*
- * select the correct result from the two polynoms
- * xmm3 = poly_mask;
- * y2 = _mm_and_ps(xmm3, y2); //, xmm3);
- * y = _mm_andnot_ps(xmm3, y);
- * y = _mm_add_ps(y,y2);
- */
- LLVMValueRef y2_i = LLVMBuildBitCast(b, y2_9, v4si, "y2_i");
- LLVMValueRef y_i = LLVMBuildBitCast(b, y_10, v4si, "y_i");
- LLVMValueRef y2_and = LLVMBuildAnd(b, y2_i, poly_mask, "y2_and");
- LLVMValueRef poly_mask_inv = LLVMBuildXor(b, poly_mask, inv, "poly_mask_inv");
- LLVMValueRef y_and = LLVMBuildAnd(b, y_i, poly_mask_inv, "y_and");
- LLVMValueRef y_combine = LLVMBuildAdd(b, y_and, y2_and, "y_combine");
-
- /*
- * update the sign
- * y = _mm_xor_ps(y, sign_bit);
- */
- LLVMValueRef y_sign = LLVMBuildXor(b, y_combine, sign_bit, "y_sin");
- LLVMValueRef y_result = LLVMBuildBitCast(b, y_sign, v4sf, "y_result");
- return y_result;
-}
-
-
-/**
- * Generate pow(x, y)
- */
-LLVMValueRef
-lp_build_pow(struct lp_build_context *bld,
- LLVMValueRef x,
- LLVMValueRef y)
-{
- /* TODO: optimize the constant case */
- if (gallivm_debug & GALLIVM_DEBUG_PERF &&
- LLVMIsConstant(x) && LLVMIsConstant(y)) {
- debug_printf("%s: inefficient/imprecise constant arithmetic\n",
- __FUNCTION__);
- }
-
- return lp_build_exp2(bld, lp_build_mul(bld, lp_build_log2(bld, x), y));
-}
+ return lp_build_exp2(bld, lp_build_mul(bld, log2e, x));
+}
/**
- * Generate exp(x)
+ * Generate log(x)
+ * Behavior is undefined with infs, 0s and nans
*/
LLVMValueRef
-lp_build_exp(struct lp_build_context *bld,
+lp_build_log(struct lp_build_context *bld,
LLVMValueRef x)
{
- /* log2(e) = 1/log(2) */
- LLVMValueRef log2e = lp_build_const_vec(bld->type, 1.4426950408889634);
+ /* log(2) */
+ LLVMValueRef log2 = lp_build_const_vec(bld->gallivm, bld->type,
+ 0.69314718055994529);
assert(lp_check_value(bld->type, x));
- return lp_build_mul(bld, log2e, lp_build_exp2(bld, x));
+ return lp_build_mul(bld, log2, lp_build_log2(bld, x));
}
-
/**
- * Generate log(x)
+ * Generate log(x) that handles edge cases (infs, 0s and nans)
*/
LLVMValueRef
-lp_build_log(struct lp_build_context *bld,
- LLVMValueRef x)
+lp_build_log_safe(struct lp_build_context *bld,
+ LLVMValueRef x)
{
/* log(2) */
- LLVMValueRef log2 = lp_build_const_vec(bld->type, 0.69314718055994529);
+ LLVMValueRef log2 = lp_build_const_vec(bld->gallivm, bld->type,
+ 0.69314718055994529);
assert(lp_check_value(bld->type, x));
- return lp_build_mul(bld, log2, lp_build_exp2(bld, x));
+ return lp_build_mul(bld, log2, lp_build_log2_safe(bld, x));
}
* Generate polynomial.
* Ex: coeffs[0] + x * coeffs[1] + x^2 * coeffs[2].
*/
-static LLVMValueRef
+LLVMValueRef
lp_build_polynomial(struct lp_build_context *bld,
LLVMValueRef x,
const double *coeffs,
unsigned num_coeffs)
{
const struct lp_type type = bld->type;
- LLVMValueRef res = NULL;
+ LLVMValueRef even = NULL, odd = NULL;
+ LLVMValueRef x2;
unsigned i;
assert(lp_check_value(bld->type, x));
__FUNCTION__);
}
+ /*
+ * Calculate odd and even terms seperately to decrease data dependency
+ * Ex:
+ * c[0] + x^2 * c[2] + x^4 * c[4] ...
+ * + x * (c[1] + x^2 * c[3] + x^4 * c[5]) ...
+ */
+ x2 = lp_build_mul(bld, x, x);
+
for (i = num_coeffs; i--; ) {
LLVMValueRef coeff;
- coeff = lp_build_const_vec(type, coeffs[i]);
+ coeff = lp_build_const_vec(bld->gallivm, type, coeffs[i]);
- if(res)
- res = lp_build_add(bld, coeff, lp_build_mul(bld, x, res));
- else
- res = coeff;
+ if (i % 2 == 0) {
+ if (even)
+ even = lp_build_add(bld, coeff, lp_build_mul(bld, x2, even));
+ else
+ even = coeff;
+ } else {
+ if (odd)
+ odd = lp_build_add(bld, coeff, lp_build_mul(bld, x2, odd));
+ else
+ odd = coeff;
+ }
}
- if(res)
- return res;
+ if (odd)
+ return lp_build_add(bld, lp_build_mul(bld, odd, x), even);
+ else if (even)
+ return even;
else
return bld->undef;
}
*/
const double lp_build_exp2_polynomial[] = {
#if EXP_POLY_DEGREE == 5
- 0.999999999690134838155,
- 0.583974334321735217258,
- 0.164553105719676828492,
- 0.0292811063701710962255,
- 0.00354944426657875141846,
- 0.000296253726543423377365
+ 1.000000000000000000000, /*XXX: was 0.999999925063526176901, recompute others */
+ 0.693153073200168932794,
+ 0.240153617044375388211,
+ 0.0558263180532956664775,
+ 0.00898934009049466391101,
+ 0.00187757667519147912699
#elif EXP_POLY_DEGREE == 4
- 1.00000001502262084505,
- 0.563586057338685991394,
- 0.150436017652442413623,
- 0.0243220604213317927308,
- 0.0025359088446580436489
+ 1.00000259337069434683,
+ 0.693003834469974940458,
+ 0.24144275689150793076,
+ 0.0520114606103070150235,
+ 0.0135341679161270268764
#elif EXP_POLY_DEGREE == 3
0.999925218562710312959,
0.695833540494823811697,
};
-void
-lp_build_exp2_approx(struct lp_build_context *bld,
- LLVMValueRef x,
- LLVMValueRef *p_exp2_int_part,
- LLVMValueRef *p_frac_part,
- LLVMValueRef *p_exp2)
+LLVMValueRef
+lp_build_exp2(struct lp_build_context *bld,
+ LLVMValueRef x)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
- LLVMTypeRef vec_type = lp_build_vec_type(type);
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
+ LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
LLVMValueRef ipart = NULL;
LLVMValueRef fpart = NULL;
LLVMValueRef expipart = NULL;
assert(lp_check_value(bld->type, x));
- if(p_exp2_int_part || p_frac_part || p_exp2) {
- /* TODO: optimize the constant case */
- if (gallivm_debug & GALLIVM_DEBUG_PERF &&
- LLVMIsConstant(x)) {
- debug_printf("%s: inefficient/imprecise constant arithmetic\n",
- __FUNCTION__);
- }
+ /* TODO: optimize the constant case */
+ if (gallivm_debug & GALLIVM_DEBUG_PERF &&
+ LLVMIsConstant(x)) {
+ debug_printf("%s: inefficient/imprecise constant arithmetic\n",
+ __FUNCTION__);
+ }
- assert(type.floating && type.width == 32);
+ assert(type.floating && type.width == 32);
- x = lp_build_min(bld, x, lp_build_const_vec(type, 129.0));
- x = lp_build_max(bld, x, lp_build_const_vec(type, -126.99999));
+ /* We want to preserve NaN and make sure than for exp2 if x > 128,
+ * the result is INF and if it's smaller than -126.9 the result is 0 */
+ x = lp_build_min_ext(bld, lp_build_const_vec(bld->gallivm, type, 128.0), x,
+ GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN);
+ x = lp_build_max_ext(bld, lp_build_const_vec(bld->gallivm, type, -126.99999),
+ x, GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN);
- /* ipart = floor(x) */
- ipart = lp_build_floor(bld, x);
+ /* ipart = floor(x) */
+ /* fpart = x - ipart */
+ lp_build_ifloor_fract(bld, x, &ipart, &fpart);
- /* fpart = x - ipart */
- fpart = LLVMBuildFSub(bld->builder, x, ipart, "");
- }
+ /* expipart = (float) (1 << ipart) */
+ expipart = LLVMBuildAdd(builder, ipart,
+ lp_build_const_int_vec(bld->gallivm, type, 127), "");
+ expipart = LLVMBuildShl(builder, expipart,
+ lp_build_const_int_vec(bld->gallivm, type, 23), "");
+ expipart = LLVMBuildBitCast(builder, expipart, vec_type, "");
- if(p_exp2_int_part || p_exp2) {
- /* expipart = (float) (1 << ipart) */
- ipart = LLVMBuildFPToSI(bld->builder, ipart, int_vec_type, "");
- expipart = LLVMBuildAdd(bld->builder, ipart, lp_build_const_int_vec(type, 127), "");
- expipart = LLVMBuildShl(bld->builder, expipart, lp_build_const_int_vec(type, 23), "");
- expipart = LLVMBuildBitCast(bld->builder, expipart, vec_type, "");
- }
+ expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial,
+ Elements(lp_build_exp2_polynomial));
- if(p_exp2) {
- expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial,
- Elements(lp_build_exp2_polynomial));
+ res = LLVMBuildFMul(builder, expipart, expfpart, "");
- res = LLVMBuildFMul(bld->builder, expipart, expfpart, "");
- }
+ return res;
+}
- if(p_exp2_int_part)
- *p_exp2_int_part = expipart;
- if(p_frac_part)
- *p_frac_part = fpart;
- if(p_exp2)
- *p_exp2 = res;
+/**
+ * Extract the exponent of a IEEE-754 floating point value.
+ *
+ * Optionally apply an integer bias.
+ *
+ * Result is an integer value with
+ *
+ * ifloor(log2(x)) + bias
+ */
+LLVMValueRef
+lp_build_extract_exponent(struct lp_build_context *bld,
+ LLVMValueRef x,
+ int bias)
+{
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ const struct lp_type type = bld->type;
+ unsigned mantissa = lp_mantissa(type);
+ LLVMValueRef res;
+
+ assert(type.floating);
+
+ assert(lp_check_value(bld->type, x));
+
+ x = LLVMBuildBitCast(builder, x, bld->int_vec_type, "");
+
+ res = LLVMBuildLShr(builder, x,
+ lp_build_const_int_vec(bld->gallivm, type, mantissa), "");
+ res = LLVMBuildAnd(builder, res,
+ lp_build_const_int_vec(bld->gallivm, type, 255), "");
+ res = LLVMBuildSub(builder, res,
+ lp_build_const_int_vec(bld->gallivm, type, 127 - bias), "");
+
+ return res;
}
+/**
+ * Extract the mantissa of the a floating.
+ *
+ * Result is a floating point value with
+ *
+ * x / floor(log2(x))
+ */
LLVMValueRef
-lp_build_exp2(struct lp_build_context *bld,
- LLVMValueRef x)
+lp_build_extract_mantissa(struct lp_build_context *bld,
+ LLVMValueRef x)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ const struct lp_type type = bld->type;
+ unsigned mantissa = lp_mantissa(type);
+ LLVMValueRef mantmask = lp_build_const_int_vec(bld->gallivm, type,
+ (1ULL << mantissa) - 1);
+ LLVMValueRef one = LLVMConstBitCast(bld->one, bld->int_vec_type);
LLVMValueRef res;
- lp_build_exp2_approx(bld, x, NULL, NULL, &res);
+
+ assert(lp_check_value(bld->type, x));
+
+ assert(type.floating);
+
+ x = LLVMBuildBitCast(builder, x, bld->int_vec_type, "");
+
+ /* res = x / 2**ipart */
+ res = LLVMBuildAnd(builder, x, mantmask, "");
+ res = LLVMBuildOr(builder, res, one, "");
+ res = LLVMBuildBitCast(builder, res, bld->vec_type, "");
+
return res;
}
+
/**
- * Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[
+ * Minimax polynomial fit of log2((1.0 + sqrt(x))/(1.0 - sqrt(x)))/sqrt(x) ,for x in range of [0, 1/9[
* These coefficients can be generate with
* http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
*/
const double lp_build_log2_polynomial[] = {
-#if LOG_POLY_DEGREE == 6
- 3.11578814719469302614,
- -3.32419399085241980044,
- 2.59883907202499966007,
- -1.23152682416275988241,
- 0.318212422185251071475,
- -0.0344359067839062357313
-#elif LOG_POLY_DEGREE == 5
- 2.8882704548164776201,
- -2.52074962577807006663,
- 1.48116647521213171641,
- -0.465725644288844778798,
- 0.0596515482674574969533
+#if LOG_POLY_DEGREE == 5
+ 2.88539008148777786488L,
+ 0.961796878841293367824L,
+ 0.577058946784739859012L,
+ 0.412914355135828735411L,
+ 0.308591899232910175289L,
+ 0.352376952300281371868L,
#elif LOG_POLY_DEGREE == 4
- 2.61761038894603480148,
- -1.75647175389045657003,
- 0.688243882994381274313,
- -0.107254423828329604454
+ 2.88539009343309178325L,
+ 0.961791550404184197881L,
+ 0.577440339438736392009L,
+ 0.403343858251329912514L,
+ 0.406718052498846252698L,
#elif LOG_POLY_DEGREE == 3
- 2.28330284476918490682,
- -1.04913055217340124191,
- 0.204446009836232697516
+ 2.88538959748872753838L,
+ 0.961932915889597772928L,
+ 0.571118517972136195241L,
+ 0.493997535084709500285L,
#else
#error
#endif
};
-
/**
* See http://www.devmaster.net/forums/showthread.php?p=43580
+ * http://en.wikipedia.org/wiki/Logarithm#Calculation
+ * http://www.nezumi.demon.co.uk/consult/logx.htm
+ *
+ * If handle_edge_cases is true the function will perform computations
+ * to match the required D3D10+ behavior for each of the edge cases.
+ * That means that if input is:
+ * - less than zero (to and including -inf) then NaN will be returned
+ * - equal to zero (-denorm, -0, +0 or +denorm), then -inf will be returned
+ * - +infinity, then +infinity will be returned
+ * - NaN, then NaN will be returned
+ *
+ * Those checks are fairly expensive so if you don't need them make sure
+ * handle_edge_cases is false.
*/
void
lp_build_log2_approx(struct lp_build_context *bld,
LLVMValueRef x,
LLVMValueRef *p_exp,
LLVMValueRef *p_floor_log2,
- LLVMValueRef *p_log2)
+ LLVMValueRef *p_log2,
+ boolean handle_edge_cases)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
- LLVMTypeRef vec_type = lp_build_vec_type(type);
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
+ LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
+ LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->gallivm, type);
- LLVMValueRef expmask = lp_build_const_int_vec(type, 0x7f800000);
- LLVMValueRef mantmask = lp_build_const_int_vec(type, 0x007fffff);
+ LLVMValueRef expmask = lp_build_const_int_vec(bld->gallivm, type, 0x7f800000);
+ LLVMValueRef mantmask = lp_build_const_int_vec(bld->gallivm, type, 0x007fffff);
LLVMValueRef one = LLVMConstBitCast(bld->one, int_vec_type);
LLVMValueRef i = NULL;
+ LLVMValueRef y = NULL;
+ LLVMValueRef z = NULL;
LLVMValueRef exp = NULL;
LLVMValueRef mant = NULL;
LLVMValueRef logexp = NULL;
assert(type.floating && type.width == 32);
- i = LLVMBuildBitCast(bld->builder, x, int_vec_type, "");
+ /*
+ * We don't explicitly handle denormalized numbers. They will yield a
+ * result in the neighbourhood of -127, which appears to be adequate
+ * enough.
+ */
+
+ i = LLVMBuildBitCast(builder, x, int_vec_type, "");
/* exp = (float) exponent(x) */
- exp = LLVMBuildAnd(bld->builder, i, expmask, "");
+ exp = LLVMBuildAnd(builder, i, expmask, "");
}
if(p_floor_log2 || p_log2) {
- logexp = LLVMBuildLShr(bld->builder, exp, lp_build_const_int_vec(type, 23), "");
- logexp = LLVMBuildSub(bld->builder, logexp, lp_build_const_int_vec(type, 127), "");
- logexp = LLVMBuildSIToFP(bld->builder, logexp, vec_type, "");
+ logexp = LLVMBuildLShr(builder, exp, lp_build_const_int_vec(bld->gallivm, type, 23), "");
+ logexp = LLVMBuildSub(builder, logexp, lp_build_const_int_vec(bld->gallivm, type, 127), "");
+ logexp = LLVMBuildSIToFP(builder, logexp, vec_type, "");
}
- if(p_log2) {
- /* mant = (float) mantissa(x) */
- mant = LLVMBuildAnd(bld->builder, i, mantmask, "");
- mant = LLVMBuildOr(bld->builder, mant, one, "");
- mant = LLVMBuildBitCast(bld->builder, mant, vec_type, "");
+ if (p_log2) {
+ /* mant = 1 + (float) mantissa(x) */
+ mant = LLVMBuildAnd(builder, i, mantmask, "");
+ mant = LLVMBuildOr(builder, mant, one, "");
+ mant = LLVMBuildBitCast(builder, mant, vec_type, "");
- logmant = lp_build_polynomial(bld, mant, lp_build_log2_polynomial,
- Elements(lp_build_log2_polynomial));
+ /* y = (mant - 1) / (mant + 1) */
+ y = lp_build_div(bld,
+ lp_build_sub(bld, mant, bld->one),
+ lp_build_add(bld, mant, bld->one)
+ );
- /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
- logmant = LLVMBuildFMul(bld->builder, logmant, LLVMBuildFSub(bld->builder, mant, bld->one, ""), "");
+ /* z = y^2 */
+ z = lp_build_mul(bld, y, y);
+
+ /* compute P(z) */
+ logmant = lp_build_polynomial(bld, z, lp_build_log2_polynomial,
+ Elements(lp_build_log2_polynomial));
- res = LLVMBuildFAdd(bld->builder, logmant, logexp, "");
+ /* logmant = y * P(z) */
+ logmant = lp_build_mul(bld, y, logmant);
+
+ res = lp_build_add(bld, logmant, logexp);
+
+ if (type.floating && handle_edge_cases) {
+ LLVMValueRef negmask, infmask, zmask;
+ negmask = lp_build_cmp(bld, PIPE_FUNC_LESS, x,
+ lp_build_const_vec(bld->gallivm, type, 0.0f));
+ zmask = lp_build_cmp(bld, PIPE_FUNC_EQUAL, x,
+ lp_build_const_vec(bld->gallivm, type, 0.0f));
+ infmask = lp_build_cmp(bld, PIPE_FUNC_GEQUAL, x,
+ lp_build_const_vec(bld->gallivm, type, INFINITY));
+
+ /* If x is qual to inf make sure we return inf */
+ res = lp_build_select(bld, infmask,
+ lp_build_const_vec(bld->gallivm, type, INFINITY),
+ res);
+ /* If x is qual to 0, return -inf */
+ res = lp_build_select(bld, zmask,
+ lp_build_const_vec(bld->gallivm, type, -INFINITY),
+ res);
+ /* If x is nan or less than 0, return nan */
+ res = lp_build_select(bld, negmask,
+ lp_build_const_vec(bld->gallivm, type, NAN),
+ res);
+ }
}
- if(p_exp) {
- exp = LLVMBuildBitCast(bld->builder, exp, vec_type, "");
+ if (p_exp) {
+ exp = LLVMBuildBitCast(builder, exp, vec_type, "");
*p_exp = exp;
}
- if(p_floor_log2)
+ if (p_floor_log2)
*p_floor_log2 = logexp;
- if(p_log2)
+ if (p_log2)
*p_log2 = res;
}
+/*
+ * log2 implementation which doesn't have special code to
+ * handle edge cases (-inf, 0, inf, NaN). It's faster but
+ * the results for those cases are undefined.
+ */
LLVMValueRef
lp_build_log2(struct lp_build_context *bld,
LLVMValueRef x)
{
LLVMValueRef res;
- lp_build_log2_approx(bld, x, NULL, NULL, &res);
+ lp_build_log2_approx(bld, x, NULL, NULL, &res, FALSE);
+ return res;
+}
+
+/*
+ * Version of log2 which handles all edge cases.
+ * Look at documentation of lp_build_log2_approx for
+ * description of the behavior for each of the edge cases.
+ */
+LLVMValueRef
+lp_build_log2_safe(struct lp_build_context *bld,
+ LLVMValueRef x)
+{
+ LLVMValueRef res;
+ lp_build_log2_approx(bld, x, NULL, NULL, &res, TRUE);
return res;
}
/**
* Faster (and less accurate) log2.
*
- * log2(x) = floor(log2(x)) + frac(x)
+ * log2(x) = floor(log2(x)) - 1 + x / 2**floor(log2(x))
+ *
+ * Piece-wise linear approximation, with exact results when x is a
+ * power of two.
*
* See http://www.flipcode.com/archives/Fast_log_Function.shtml
*/
lp_build_fast_log2(struct lp_build_context *bld,
LLVMValueRef x)
{
- const struct lp_type type = bld->type;
- LLVMTypeRef vec_type = bld->vec_type;
- LLVMTypeRef int_vec_type = bld->int_vec_type;
-
- unsigned mantissa = lp_mantissa(type);
- LLVMValueRef mantmask = lp_build_const_int_vec(type, (1ULL << mantissa) - 1);
- LLVMValueRef one = LLVMConstBitCast(bld->one, int_vec_type);
-
+ LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef ipart;
LLVMValueRef fpart;
assert(lp_check_value(bld->type, x));
- assert(type.floating);
-
- x = LLVMBuildBitCast(bld->builder, x, int_vec_type, "");
+ assert(bld->type.floating);
/* ipart = floor(log2(x)) - 1 */
- ipart = LLVMBuildLShr(bld->builder, x, lp_build_const_int_vec(type, mantissa), "");
- ipart = LLVMBuildAnd(bld->builder, ipart, lp_build_const_int_vec(type, 255), "");
- ipart = LLVMBuildSub(bld->builder, ipart, lp_build_const_int_vec(type, 128), "");
- ipart = LLVMBuildSIToFP(bld->builder, ipart, vec_type, "");
+ ipart = lp_build_extract_exponent(bld, x, -1);
+ ipart = LLVMBuildSIToFP(builder, ipart, bld->vec_type, "");
- /* fpart = 1.0 + frac(x) */
- fpart = LLVMBuildAnd(bld->builder, x, mantmask, "");
- fpart = LLVMBuildOr(bld->builder, fpart, one, "");
- fpart = LLVMBuildBitCast(bld->builder, fpart, vec_type, "");
+ /* fpart = x / 2**ipart */
+ fpart = lp_build_extract_mantissa(bld, x);
- /* floor(log2(x)) + frac(x) */
- return LLVMBuildFAdd(bld->builder, ipart, fpart, "");
+ /* ipart + fpart */
+ return LLVMBuildFAdd(builder, ipart, fpart, "");
}
lp_build_ilog2(struct lp_build_context *bld,
LLVMValueRef x)
{
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ LLVMValueRef sqrt2 = lp_build_const_vec(bld->gallivm, bld->type, M_SQRT2);
+ LLVMValueRef ipart;
+
+ assert(bld->type.floating);
+
+ assert(lp_check_value(bld->type, x));
+
+ /* x * 2^(0.5) i.e., add 0.5 to the log2(x) */
+ x = LLVMBuildFMul(builder, x, sqrt2, "");
+
+ /* ipart = floor(log2(x) + 0.5) */
+ ipart = lp_build_extract_exponent(bld, x, 0);
+
+ return ipart;
+}
+
+LLVMValueRef
+lp_build_mod(struct lp_build_context *bld,
+ LLVMValueRef x,
+ LLVMValueRef y)
+{
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ LLVMValueRef res;
const struct lp_type type = bld->type;
- LLVMTypeRef int_vec_type = bld->int_vec_type;
- unsigned mantissa = lp_mantissa(type);
- LLVMValueRef sqrt2 = lp_build_const_vec(type, 1.4142135623730951);
+ assert(lp_check_value(type, x));
+ assert(lp_check_value(type, y));
- LLVMValueRef ipart;
+ if (type.floating)
+ res = LLVMBuildFRem(builder, x, y, "");
+ else if (type.sign)
+ res = LLVMBuildSRem(builder, x, y, "");
+ else
+ res = LLVMBuildURem(builder, x, y, "");
+ return res;
+}
+
+
+/*
+ * For floating inputs it creates and returns a mask
+ * which is all 1's for channels which are NaN.
+ * Channels inside x which are not NaN will be 0.
+ */
+LLVMValueRef
+lp_build_isnan(struct lp_build_context *bld,
+ LLVMValueRef x)
+{
+ LLVMValueRef mask;
+ LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->gallivm, bld->type);
+ assert(bld->type.floating);
assert(lp_check_value(bld->type, x));
+ mask = LLVMBuildFCmp(bld->gallivm->builder, LLVMRealOEQ, x, x,
+ "isnotnan");
+ mask = LLVMBuildNot(bld->gallivm->builder, mask, "");
+ mask = LLVMBuildSExt(bld->gallivm->builder, mask, int_vec_type, "isnan");
+ return mask;
+}
+
+/* Returns all 1's for floating point numbers that are
+ * finite numbers and returns all zeros for -inf,
+ * inf and nan's */
+LLVMValueRef
+lp_build_isfinite(struct lp_build_context *bld,
+ LLVMValueRef x)
+{
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->gallivm, bld->type);
+ struct lp_type int_type = lp_int_type(bld->type);
+ LLVMValueRef intx = LLVMBuildBitCast(builder, x, int_vec_type, "");
+ LLVMValueRef infornan32 = lp_build_const_int_vec(bld->gallivm, bld->type,
+ 0x7f800000);
+
+ if (!bld->type.floating) {
+ return lp_build_const_int_vec(bld->gallivm, bld->type, 0);
+ }
+ assert(bld->type.floating);
+ assert(lp_check_value(bld->type, x));
+ assert(bld->type.width == 32);
+
+ intx = LLVMBuildAnd(builder, intx, infornan32, "");
+ return lp_build_compare(bld->gallivm, int_type, PIPE_FUNC_NOTEQUAL,
+ intx, infornan32);
+}
+
+/*
+ * Returns true if the number is nan or inf and false otherwise.
+ * The input has to be a floating point vector.
+ */
+LLVMValueRef
+lp_build_is_inf_or_nan(struct gallivm_state *gallivm,
+ const struct lp_type type,
+ LLVMValueRef x)
+{
+ LLVMBuilderRef builder = gallivm->builder;
+ struct lp_type int_type = lp_int_type(type);
+ LLVMValueRef const0 = lp_build_const_int_vec(gallivm, int_type,
+ 0x7f800000);
+ LLVMValueRef ret;
+
assert(type.floating);
- /* x * 2^(0.5) i.e., add 0.5 to the log2(x) */
- x = LLVMBuildFMul(bld->builder, x, sqrt2, "");
+ ret = LLVMBuildBitCast(builder, x, lp_build_vec_type(gallivm, int_type), "");
+ ret = LLVMBuildAnd(builder, ret, const0, "");
+ ret = lp_build_compare(gallivm, int_type, PIPE_FUNC_EQUAL,
+ ret, const0);
- x = LLVMBuildBitCast(bld->builder, x, int_vec_type, "");
+ return ret;
+}
- /* ipart = floor(log2(x) + 0.5) */
- ipart = LLVMBuildLShr(bld->builder, x, lp_build_const_int_vec(type, mantissa), "");
- ipart = LLVMBuildAnd(bld->builder, ipart, lp_build_const_int_vec(type, 255), "");
- ipart = LLVMBuildSub(bld->builder, ipart, lp_build_const_int_vec(type, 127), "");
- return ipart;
+LLVMValueRef
+lp_build_fpstate_get(struct gallivm_state *gallivm)
+{
+ if (util_cpu_caps.has_sse) {
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef mxcsr_ptr = lp_build_alloca(
+ gallivm,
+ LLVMInt32TypeInContext(gallivm->context),
+ "mxcsr_ptr");
+ LLVMValueRef mxcsr_ptr8 = LLVMBuildPointerCast(builder, mxcsr_ptr,
+ LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0), "");
+ lp_build_intrinsic(builder,
+ "llvm.x86.sse.stmxcsr",
+ LLVMVoidTypeInContext(gallivm->context),
+ &mxcsr_ptr8, 1, 0);
+ return mxcsr_ptr;
+ }
+ return 0;
+}
+
+void
+lp_build_fpstate_set_denorms_zero(struct gallivm_state *gallivm,
+ boolean zero)
+{
+ if (util_cpu_caps.has_sse) {
+ /* turn on DAZ (64) | FTZ (32768) = 32832 if available */
+ int daz_ftz = _MM_FLUSH_ZERO_MASK;
+
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef mxcsr_ptr = lp_build_fpstate_get(gallivm);
+ LLVMValueRef mxcsr =
+ LLVMBuildLoad(builder, mxcsr_ptr, "mxcsr");
+
+ if (util_cpu_caps.has_daz) {
+ /* Enable denormals are zero mode */
+ daz_ftz |= _MM_DENORMALS_ZERO_MASK;
+ }
+ if (zero) {
+ mxcsr = LLVMBuildOr(builder, mxcsr,
+ LLVMConstInt(LLVMTypeOf(mxcsr), daz_ftz, 0), "");
+ } else {
+ mxcsr = LLVMBuildAnd(builder, mxcsr,
+ LLVMConstInt(LLVMTypeOf(mxcsr), ~daz_ftz, 0), "");
+ }
+
+ LLVMBuildStore(builder, mxcsr, mxcsr_ptr);
+ lp_build_fpstate_set(gallivm, mxcsr_ptr);
+ }
+}
+
+void
+lp_build_fpstate_set(struct gallivm_state *gallivm,
+ LLVMValueRef mxcsr_ptr)
+{
+ if (util_cpu_caps.has_sse) {
+ LLVMBuilderRef builder = gallivm->builder;
+ mxcsr_ptr = LLVMBuildPointerCast(builder, mxcsr_ptr,
+ LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0), "");
+ lp_build_intrinsic(builder,
+ "llvm.x86.sse.ldmxcsr",
+ LLVMVoidTypeInContext(gallivm->context),
+ &mxcsr_ptr, 1, 0);
+ }
}