#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 5
}
}
else if (type.floating && util_cpu_caps.has_altivec) {
- debug_printf("%s: altivec doesn't support nan behavior modes\n",
- __FUNCTION__);
+ 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_altivec) {
intr_size = 128;
- debug_printf("%s: altivec doesn't support nan behavior modes\n",
- __FUNCTION__);
if (type.width == 8) {
if (!type.sign) {
intrinsic = "llvm.ppc.altivec.vminub";
*/
if (util_cpu_caps.has_sse && type.floating &&
nan_behavior != GALLIVM_NAN_BEHAVIOR_UNDEFINED &&
- nan_behavior != GALLIVM_NAN_RETURN_SECOND) {
- LLVMValueRef isnan, max;
- max = lp_build_intrinsic_binary_anylength(bld->gallivm, intrinsic,
+ 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, max);
+ 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, max);
+ return lp_build_select(bld, isnan, a, min);
}
} else {
return lp_build_intrinsic_binary_anylength(bld->gallivm, intrinsic,
return lp_build_select(bld, cond, a, b);
}
break;
- case GALLIVM_NAN_RETURN_SECOND:
+ 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);
}
}
else if (type.floating && util_cpu_caps.has_altivec) {
- debug_printf("%s: altivec doesn't support nan behavior modes\n",
- __FUNCTION__);
+ 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_altivec) {
intr_size = 128;
- debug_printf("%s: altivec doesn't support nan behavior modes\n",
- __FUNCTION__);
if (type.width == 8) {
if (!type.sign) {
intrinsic = "llvm.ppc.altivec.vmaxub";
if(intrinsic) {
if (util_cpu_caps.has_sse && type.floating &&
nan_behavior != GALLIVM_NAN_BEHAVIOR_UNDEFINED &&
- nan_behavior != GALLIVM_NAN_RETURN_SECOND) {
- LLVMValueRef isnan, min;
- min = lp_build_intrinsic_binary_anylength(bld->gallivm, intrinsic,
+ 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, min);
+ 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, min);
+ return lp_build_select(bld, isnan, a, max);
}
} else {
return lp_build_intrinsic_binary_anylength(bld->gallivm, intrinsic,
return lp_build_select(bld, cond, a, b);
}
break;
- case GALLIVM_NAN_RETURN_SECOND:
+ 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);
return lp_build_intrinsic_binary(builder, intrinsic, lp_build_vec_type(bld->gallivm, bld->type), a, b);
}
- /* TODO: handle signed case */
- if(type.norm && !type.floating && !type.fixed && !type.sign)
- a = lp_build_min_simple(bld, a, lp_build_comp(bld, b), GALLIVM_NAN_BEHAVIOR_UNDEFINED);
+ 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))
if (type.floating)
return lp_build_intrinsic_binary(builder, intrinsic, lp_build_vec_type(bld->gallivm, bld->type), a, b);
}
- /* TODO: handle signed case */
- if(type.norm && !type.floating && !type.fixed && !type.sign)
- a = lp_build_max_simple(bld, a, b, GALLIVM_NAN_BEHAVIOR_UNDEFINED);
+ 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))
if (type.floating)
* half = sgn(ab) * 0.5 * (2 ** n) = sgn(ab) * (1 << (n - 1))
*/
- half = lp_build_const_int_vec(gallivm, wide_type, 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);
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;
/**
* 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)
*/
assert(lp_check_value(bld->type, a));
-#if HAVE_LLVM >= 0x0207
if (bld->type.floating)
a = LLVMBuildFNeg(builder, a, "");
else
-#endif
a = LLVMBuildNeg(builder, a, "");
return a;
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, 2^24);
+ 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;
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, 2^24);
+ 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;
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, 2^24);
+ 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;
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, 2^24);
+ 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;
/**
- * 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 builder = gallivm->builder;
+ LLVMBuilderRef b = gallivm->builder;
struct lp_type int_type = lp_int_type(bld->type);
- LLVMBuilderRef b = builder;
/*
* take the absolute value,
LLVMValueRef absi = LLVMBuildAnd(b, a_v4si, inv_sig_mask, "absi");
LLVMValueRef x_abs = LLVMBuildBitCast(b, absi, bld->vec_type, "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_int_vec(gallivm, bld->type, 0x80000000);
- LLVMValueRef sign_bit_i = LLVMBuildAnd(b, a_v4si, sig_mask, "sign_bit_i");
-
/*
* scale by 4/Pi
* y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
*/
-
+
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, bld->int_vec_type, "emm2_i");
/*
*/
LLVMValueRef y_2 = LLVMBuildSIToFP(b, emm2_and, bld->vec_type, "y_2");
- /* get the swap sign flag
- * emm0 = _mm_and_si128(emm2, *(v4si*)_pi32_4);
- */
- LLVMValueRef pi32_4 = lp_build_const_int_vec(gallivm, bld->type, 4);
- LLVMValueRef emm0_and = LLVMBuildAnd(b, emm2_add, pi32_4, "emm0_and");
-
- /*
- * emm2 = _mm_slli_epi32(emm0, 29);
- */
+ 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 swap_sign_bit = LLVMBuildShl(b, emm0_and, const_29, "swap_sign_bit");
+ LLVMValueRef sign_mask = lp_build_const_int_vec(gallivm, bld->type, 0x80000000);
+
+ /*
+ * 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_int_vec(gallivm, bld->type, 2);
- LLVMValueRef emm2_3 = LLVMBuildAnd(b, emm2_and, pi32_2, "emm2_3");
+ 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));
- /*
- * sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit);
- */
- LLVMValueRef sign_bit_1 = LLVMBuildXor(b, sign_bit_i, swap_sign_bit, "sign_bit");
/*
* _PS_CONST(minus_cephes_DP1, -0.78515625);
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");
* 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_vec(gallivm, bld->type, 0.5);
LLVMValueRef tmp = LLVMBuildFMul(b, z, half, "tmp");
LLVMValueRef y_9 = LLVMBuildFSub(b, y_8, tmp, "y_8");
* update the sign
* y = _mm_xor_ps(y, sign_bit);
*/
- LLVMValueRef y_sign = LLVMBuildXor(b, y_combine, sign_bit_1, "y_sin");
+ 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] */
/**
- * 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 gallivm_state *gallivm = bld->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
- struct lp_type int_type = lp_int_type(bld->type);
- LLVMBuilderRef b = builder;
-
- /*
- * take the absolute value,
- * x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
- */
-
- 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, bld->vec_type, "x_abs");
-
- /*
- * scale by 4/Pi
- * y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
- */
-
- 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, 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_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_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, bld->vec_type, "y_2");
-
-
- /*
- * emm2 = _mm_sub_epi32(emm2, *(v4si*)_pi32_2);
- */
- LLVMValueRef const_2 = lp_build_const_int_vec(gallivm, bld->type, 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_int_vec(gallivm, bld->type, ~0);
- LLVMValueRef emm0_not = LLVMBuildXor(b, emm2_2, inv, "emm0_not");
- LLVMValueRef pi32_4 = lp_build_const_int_vec(gallivm, bld->type, 4);
- LLVMValueRef emm0_and = LLVMBuildAnd(b, emm0_not, pi32_4, "emm0_and");
-
- /*
- * emm2 = _mm_slli_epi32(emm0, 29);
- */
- LLVMValueRef const_29 = lp_build_const_int_vec(gallivm, bld->type, 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_int_vec(gallivm, bld->type, 2);
- LLVMValueRef emm2_3 = LLVMBuildAnd(b, emm2_2, pi32_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_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;
- * 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_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;
- * 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_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_vec(gallivm, bld->type, 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_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)
- *
- * 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_or_ps(y,y2);
- */
- 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 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 = LLVMBuildOr(b, y_and, y2_and, "y_combine");
+ return lp_build_sin_or_cos(bld, a, FALSE);
+}
- /*
- * 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, 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)
+ */
+LLVMValueRef
+lp_build_cos(struct lp_build_context *bld,
+ LLVMValueRef a)
+{
+ return lp_build_sin_or_cos(bld, a, TRUE);
}
};
-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;
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__);
- }
-
- assert(type.floating && type.width == 32);
-
- /* 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_SECOND);
- x = lp_build_max_ext(bld, lp_build_const_vec(bld->gallivm, type, -126.99999), x,
- GALLIVM_NAN_RETURN_SECOND);
-
- /* ipart = floor(x) */
- /* fpart = x - ipart */
- lp_build_ifloor_fract(bld, x, &ipart, &fpart);
- }
-
- if(p_exp2_int_part || p_exp2) {
- /* 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, "");
+ /* TODO: optimize the constant case */
+ if (gallivm_debug & GALLIVM_DEBUG_PERF &&
+ LLVMIsConstant(x)) {
+ debug_printf("%s: inefficient/imprecise constant arithmetic\n",
+ __FUNCTION__);
}
- if(p_exp2) {
- expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial,
- Elements(lp_build_exp2_polynomial));
+ assert(type.floating && type.width == 32);
- res = LLVMBuildFMul(builder, expipart, expfpart, "");
- }
+ /* 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);
- if(p_exp2_int_part)
- *p_exp2_int_part = expipart;
+ /* ipart = floor(x) */
+ /* fpart = x - ipart */
+ lp_build_ifloor_fract(bld, x, &ipart, &fpart);
- if(p_frac_part)
- *p_frac_part = fpart;
+ /* 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)
- *p_exp2 = res;
-}
+ expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial,
+ Elements(lp_build_exp2_polynomial));
+ res = LLVMBuildFMul(builder, expipart, expfpart, "");
-LLVMValueRef
-lp_build_exp2(struct lp_build_context *bld,
- LLVMValueRef x)
-{
- LLVMValueRef res;
- lp_build_exp2_approx(bld, x, NULL, NULL, &res);
return res;
}
+
/**
* Extract the exponent of a IEEE-754 floating point value.
*
return ret;
}
+
+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);
+ 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);
+ }
+}
projects / mesa.git / blobdiff