#include "lp_bld_intr.h"
#include "lp_bld_logic.h"
#include "lp_bld_pack.h"
-#include "lp_bld_debug.h"
#include "lp_bld_arit.h"
{
const struct lp_type type = bld->type;
LLVMValueRef index, res;
- int i;
+ unsigned i;
if (a == bld->zero)
return bld->zero;
if(type.floating) {
/* Mask out the sign bit */
- if (type.length == 1) {
- LLVMTypeRef int_type = LLVMIntType(type.width);
- LLVMTypeRef float_type = LLVMFloatType();
- unsigned long long absMask = ~(1ULL << (type.width - 1));
- LLVMValueRef mask = LLVMConstInt(int_type, absMask, 0);
- a = LLVMBuildBitCast(bld->builder, a, int_type, "");
- a = LLVMBuildAnd(bld->builder, a, mask, "");
- a = LLVMBuildBitCast(bld->builder, a, float_type, "");
- return a;
- }
- else {
- /* vector of floats */
- 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;
- }
+ 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(type.width*type.length == 128 && util_cpu_caps.has_ssse3) {
LLVMValueRef one;
unsigned long long maskBit = (unsigned long long)1 << (type.width - 1);
- if (type.length == 1) {
- int_type = lp_build_int_elem_type(type);
- vec_type = lp_build_elem_type(type);
- mask = LLVMConstInt(int_type, maskBit, 0);
- }
- else {
- /* vector */
- int_type = lp_build_int_vec_type(type);
- vec_type = lp_build_vec_type(type);
- mask = lp_build_const_int_vec(type, maskBit);
- }
+ int_type = lp_build_int_vec_type(type);
+ vec_type = lp_build_vec_type(type);
+ mask = lp_build_const_int_vec(type, maskBit);
/* Take the sign bit and add it to 1 constant */
sign = LLVMBuildBitCast(bld->builder, a, int_type, "");
LLVMValueRef a)
{
const struct lp_type type = bld->type;
+ LLVMTypeRef vec_type = lp_build_vec_type(type);
assert(type.floating);
- /*assert(lp_check_value(type, a));*/
- if (type.length == 1) {
- LLVMTypeRef float_type = LLVMFloatType();
- return LLVMBuildSIToFP(bld->builder, a, float_type, "");
- }
- else {
- LLVMTypeRef vec_type = lp_build_vec_type(type);
- /*LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);*/
- LLVMValueRef res;
- res = LLVMBuildSIToFP(bld->builder, a, vec_type, "");
- return res;
- }
+ return LLVMBuildSIToFP(bld->builder, a, vec_type, "");
}
}
+/**
+ * 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)
assert(type.floating);
assert(lp_check_value(type, a));
- if(util_cpu_caps.has_sse4_1)
+ if (util_cpu_caps.has_sse4_1 && type.width*type.length == 128)
return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_TRUNCATE);
else {
LLVMTypeRef vec_type = lp_build_vec_type(type);
}
+/**
+ * Return float (vector) rounded to nearest integer (vector). The returned
+ * value is a float (vector).
+ * Ex: round(0.9) = 1.0
+ * Ex: round(-1.5) = -2.0
+ */
LLVMValueRef
lp_build_round(struct lp_build_context *bld,
LLVMValueRef a)
assert(type.floating);
assert(lp_check_value(type, a));
- if(util_cpu_caps.has_sse4_1)
+ if (util_cpu_caps.has_sse4_1 && type.width*type.length == 128)
return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST);
else {
LLVMTypeRef vec_type = lp_build_vec_type(type);
}
+/**
+ * Return floor of float (vector), result is a float (vector)
+ * Ex: floor(1.1) = 1.0
+ * Ex: floor(-1.1) = -2.0
+ */
LLVMValueRef
lp_build_floor(struct lp_build_context *bld,
LLVMValueRef a)
const struct lp_type type = bld->type;
assert(type.floating);
+ assert(lp_check_value(type, a));
- if (type.length == 1) {
- LLVMValueRef res;
- res = lp_build_ifloor(bld, a);
- res = LLVMBuildSIToFP(bld->builder, res, LLVMFloatType(), "");
- return res;
- }
-
- if(util_cpu_caps.has_sse4_1)
+ if (util_cpu_caps.has_sse4_1 && type.width*type.length == 128)
return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
else {
LLVMTypeRef vec_type = lp_build_vec_type(type);
}
+/**
+ * Return ceiling of float (vector), returning float (vector).
+ * Ex: ceil( 1.1) = 2.0
+ * Ex: ceil(-1.1) = -1.0
+ */
LLVMValueRef
lp_build_ceil(struct lp_build_context *bld,
LLVMValueRef a)
assert(type.floating);
assert(lp_check_value(type, a));
- if(util_cpu_caps.has_sse4_1)
+ if (util_cpu_caps.has_sse4_1 && type.width*type.length == 128)
return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL);
else {
LLVMTypeRef vec_type = lp_build_vec_type(type);
/**
- * Return fractional part of 'a' computed as a - floor(f)
+ * Return fractional part of 'a' computed as a - floor(a)
* Typically used in texture coord arithmetic.
*/
LLVMValueRef
/**
- * Convert to integer, through whichever rounding method that's fastest,
- * typically truncating toward zero.
+ * Return the integer part of a float (vector) value. The returned value is
+ * an integer (vector).
+ * Ex: itrunc(-1.5) = 1
*/
LLVMValueRef
lp_build_itrunc(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
+ LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
assert(type.floating);
+ assert(lp_check_value(type, a));
- if (type.length == 1) {
- LLVMTypeRef int_type = LLVMIntType(type.width);
- return LLVMBuildFPToSI(bld->builder, a, int_type, "");
- }
- else {
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
- assert(lp_check_value(type, a));
- return LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
- }
+ return LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
}
/**
- * Convert float[] to int[] with round().
+ * Return float (vector) rounded to nearest integer (vector). The returned
+ * value is an integer (vector).
+ * Ex: iround(0.9) = 1
+ * Ex: iround(-1.5) = -2
*/
LLVMValueRef
lp_build_iround(struct lp_build_context *bld,
assert(type.floating);
- if (type.length == 1) {
- /* scalar float to int */
- LLVMTypeRef int_type = LLVMIntType(type.width);
- /* XXX we want rounding here! */
- res = LLVMBuildFPToSI(bld->builder, a, int_type, "");
- return res;
- }
-
assert(lp_check_value(type, a));
- if(util_cpu_caps.has_sse4_1) {
+ if (util_cpu_caps.has_sse4_1 && type.width*type.length == 128) {
res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST);
}
else {
/**
- * Convert float[] to int[] with floor().
+ * Return floor of float (vector), result is an int (vector)
+ * Ex: ifloor(1.1) = 1.0
+ * Ex: ifloor(-1.1) = -2.0
*/
LLVMValueRef
lp_build_ifloor(struct lp_build_context *bld,
LLVMValueRef res;
assert(type.floating);
-
- if (type.length == 1) {
- /* scalar float to int */
- LLVMTypeRef int_type = LLVMIntType(type.width);
- res = LLVMBuildFPToSI(bld->builder, a, int_type, "");
- return res;
- }
-
assert(lp_check_value(type, a));
- if(util_cpu_caps.has_sse4_1) {
+ if (util_cpu_caps.has_sse4_1 && type.width*type.length == 128) {
res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
}
else {
/* 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), "");
- lp_build_name(sign, "floor.sign");
+ sign = LLVMBuildAShr(bld->builder, sign, lp_build_const_int_vec(type, type.width - 1), "ifloor.sign");
/* offset = -0.99999(9)f */
- offset = lp_build_const_vec(type, -(double)(((unsigned long long)1 << mantissa) - 1)/((unsigned long long)1 << mantissa));
+ offset = lp_build_const_vec(type, -(double)(((unsigned long long)1 << mantissa) - 10)/((unsigned long long)1 << mantissa));
offset = LLVMConstBitCast(offset, int_vec_type);
- /* offset = a < 0 ? -0.99999(9)f : 0.0f */
+ /* offset = a < 0 ? offset : 0.0f */
offset = LLVMBuildAnd(bld->builder, offset, sign, "");
- offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "");
- lp_build_name(offset, "floor.offset");
+ offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "ifloor.offset");
- res = LLVMBuildAdd(bld->builder, a, offset, "");
- lp_build_name(res, "floor.res");
+ res = LLVMBuildAdd(bld->builder, a, offset, "ifloor.res");
}
- res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
- lp_build_name(res, "floor");
+ /* round to nearest (toward zero) */
+ res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "ifloor.res");
return res;
}
+/**
+ * Return ceiling of float (vector), returning int (vector).
+ * Ex: iceil( 1.1) = 2
+ * Ex: iceil(-1.1) = -1
+ */
LLVMValueRef
lp_build_iceil(struct lp_build_context *bld,
LLVMValueRef a)
assert(type.floating);
assert(lp_check_value(type, a));
- if(util_cpu_caps.has_sse4_1) {
+ if (util_cpu_caps.has_sse4_1 && type.width*type.length == 128) {
res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL);
}
else {
- assert(0);
- res = bld->undef;
+ LLVMTypeRef vec_type = lp_build_vec_type(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;
+
+ /* 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 = 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);
+
+ /* offset = a < 0 ? 0.0 : offset */
+ offset = LLVMBuildAnd(bld->builder, offset, sign, "");
+ offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "iceil.offset");
+
+ res = LLVMBuildAdd(bld->builder, a, offset, "iceil.res");
}
- res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
+ /* round to nearest (toward zero) */
+ res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "iceil.res");
return res;
}
}
+static inline LLVMValueRef
+lp_build_const_v4si(unsigned long value)
+{
+ LLVMValueRef element = LLVMConstInt(LLVMInt32Type(), value, 0);
+ LLVMValueRef elements[4] = { element, element, element, element };
+ return LLVMConstVector(elements, 4);
+}
+
+static inline LLVMValueRef
+lp_build_const_v4sf(float value)
+{
+ LLVMValueRef element = LLVMConstReal(LLVMFloatType(), value);
+ LLVMValueRef elements[4] = { element, element, element, element };
+ return LLVMConstVector(elements, 4);
+}
+
+
/**
- * Generate cos(a)
+ * Generate sin(a) using SSE2
*/
LLVMValueRef
-lp_build_cos(struct lp_build_context *bld,
- LLVMValueRef a)
+lp_build_sin(struct lp_build_context *bld,
+ LLVMValueRef a)
{
-#ifdef PIPE_OS_WINDOWS
+ struct lp_type int_type = lp_int_type(bld->type);
+ LLVMBuilderRef b = bld->builder;
+ LLVMTypeRef v4sf = LLVMVectorType(LLVMFloatType(), 4);
+ LLVMTypeRef v4si = LLVMVectorType(LLVMInt32Type(), 4);
+
/*
- * FIXME: X86 backend translates llvm.cos.v4f32 to 4 calls to CRT's cosf()
- * which is neither efficient nor does the CRT linkage work on Windows
- * causing segmentation fault. So simply disable the code for now.
+ * take the absolute value,
+ * x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
*/
- return bld->one;
-#else
- const struct lp_type type = bld->type;
- LLVMTypeRef vec_type = lp_build_vec_type(type);
- char intrinsic[32];
- /* TODO: optimize the constant case */
+ LLVMValueRef inv_sig_mask = lp_build_const_v4si(~0x80000000);
+ LLVMValueRef a_v4si = LLVMBuildBitCast(b, a, v4si, "a_v4si");
- assert(type.floating);
- util_snprintf(intrinsic, sizeof intrinsic, "llvm.cos.v%uf%u", type.length, type.width);
+ LLVMValueRef absi = LLVMBuildAnd(b, a_v4si, inv_sig_mask, "absi");
+ LLVMValueRef x_abs = LLVMBuildBitCast(b, absi, v4sf, "x_abs");
- return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a);
-#endif
+ /*
+ * 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");
+
+ /*
+ * scale by 4/Pi
+ * y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
+ */
+
+ LLVMValueRef FOPi = lp_build_const_v4sf(1.27323954473516);
+ LLVMValueRef scale_y = LLVMBuildMul(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");
+
+ /*
+ * 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 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");
+
+ /* 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");
+
+ /*
+ * 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");
+
+ /*
+ * _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 = LLVMBuildMul(b, y_2, DP1, "xmm1");
+ LLVMValueRef xmm2 = LLVMBuildMul(b, y_2, DP2, "xmm2");
+ LLVMValueRef xmm3 = LLVMBuildMul(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 = LLVMBuildAdd(b, x_abs, xmm1, "x_1");
+ LLVMValueRef x_2 = LLVMBuildAdd(b, x_1, xmm2, "x_2");
+ LLVMValueRef x_3 = LLVMBuildAdd(b, x_2, xmm3, "x_3");
+
+ /*
+ * Evaluate the first polynom (0 <= x <= Pi/4)
+ *
+ * z = _mm_mul_ps(x,x);
+ */
+ LLVMValueRef z = LLVMBuildMul(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 = LLVMBuildMul(b, z, coscof_p0, "y_3");
+ LLVMValueRef y_4 = LLVMBuildAdd(b, y_3, coscof_p1, "y_4");
+ LLVMValueRef y_5 = LLVMBuildMul(b, y_4, z, "y_5");
+ LLVMValueRef y_6 = LLVMBuildAdd(b, y_5, coscof_p2, "y_6");
+ LLVMValueRef y_7 = LLVMBuildMul(b, y_6, z, "y_7");
+ LLVMValueRef y_8 = LLVMBuildMul(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 = LLVMBuildMul(b, z, half, "tmp");
+ LLVMValueRef y_9 = LLVMBuildSub(b, y_8, tmp, "y_8");
+ LLVMValueRef one = lp_build_const_v4sf(1.0);
+ LLVMValueRef y_10 = LLVMBuildAdd(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 = LLVMBuildMul(b, z, sincof_p0, "y2_3");
+ LLVMValueRef y2_4 = LLVMBuildAdd(b, y2_3, sincof_p1, "y2_4");
+ LLVMValueRef y2_5 = LLVMBuildMul(b, y2_4, z, "y2_5");
+ LLVMValueRef y2_6 = LLVMBuildAdd(b, y2_5, sincof_p2, "y2_6");
+ LLVMValueRef y2_7 = LLVMBuildMul(b, y2_6, z, "y2_7");
+ LLVMValueRef y2_8 = LLVMBuildMul(b, y2_7, x_3, "y2_8");
+ LLVMValueRef y2_9 = LLVMBuildAdd(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 inv = lp_build_const_v4si(~0);
+ 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_1, "y_sin");
+ LLVMValueRef y_result = LLVMBuildBitCast(b, y_sign, v4sf, "y_result");
+ return y_result;
}
/**
- * Generate sin(a)
+ * Generate cos(a) using SSE2
*/
LLVMValueRef
-lp_build_sin(struct lp_build_context *bld,
- LLVMValueRef a)
+lp_build_cos(struct lp_build_context *bld,
+ LLVMValueRef a)
{
-#ifdef PIPE_OS_WINDOWS
+ struct lp_type int_type = lp_int_type(bld->type);
+ LLVMBuilderRef b = bld->builder;
+ LLVMTypeRef v4sf = LLVMVectorType(LLVMFloatType(), 4);
+ LLVMTypeRef v4si = LLVMVectorType(LLVMInt32Type(), 4);
+
/*
- * FIXME: X86 backend translates llvm.sin.v4f32 to 4 calls to CRT's sinf()
- * which is neither efficient nor does the CRT linkage work on Windows
- * causing segmentation fault. So simply disable the code for now.
+ * take the absolute value,
+ * x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
*/
- return bld->zero;
-#else
- const struct lp_type type = bld->type;
- LLVMTypeRef vec_type = lp_build_vec_type(type);
- char intrinsic[32];
- /* TODO: optimize the constant case */
+ LLVMValueRef inv_sig_mask = lp_build_const_v4si(~0x80000000);
+ LLVMValueRef a_v4si = LLVMBuildBitCast(b, a, v4si, "a_v4si");
- assert(type.floating);
- util_snprintf(intrinsic, sizeof intrinsic, "llvm.sin.v%uf%u", type.length, type.width);
+ LLVMValueRef absi = LLVMBuildAnd(b, a_v4si, inv_sig_mask, "absi");
+ LLVMValueRef x_abs = LLVMBuildBitCast(b, absi, v4sf, "x_abs");
- return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a);
-#endif
+ /*
+ * scale by 4/Pi
+ * y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
+ */
+
+ LLVMValueRef FOPi = lp_build_const_v4sf(1.27323954473516);
+ LLVMValueRef scale_y = LLVMBuildMul(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");
+
+ /*
+ * 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 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");
+
+
+ /*
+ * 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 = LLVMBuildMul(b, y_2, DP1, "xmm1");
+ LLVMValueRef xmm2 = LLVMBuildMul(b, y_2, DP2, "xmm2");
+ LLVMValueRef xmm3 = LLVMBuildMul(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 = LLVMBuildAdd(b, x_abs, xmm1, "x_1");
+ LLVMValueRef x_2 = LLVMBuildAdd(b, x_1, xmm2, "x_2");
+ LLVMValueRef x_3 = LLVMBuildAdd(b, x_2, xmm3, "x_3");
+
+ /*
+ * Evaluate the first polynom (0 <= x <= Pi/4)
+ *
+ * z = _mm_mul_ps(x,x);
+ */
+ LLVMValueRef z = LLVMBuildMul(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 = LLVMBuildMul(b, z, coscof_p0, "y_3");
+ LLVMValueRef y_4 = LLVMBuildAdd(b, y_3, coscof_p1, "y_4");
+ LLVMValueRef y_5 = LLVMBuildMul(b, y_4, z, "y_5");
+ LLVMValueRef y_6 = LLVMBuildAdd(b, y_5, coscof_p2, "y_6");
+ LLVMValueRef y_7 = LLVMBuildMul(b, y_6, z, "y_7");
+ LLVMValueRef y_8 = LLVMBuildMul(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 = LLVMBuildMul(b, z, half, "tmp");
+ LLVMValueRef y_9 = LLVMBuildSub(b, y_8, tmp, "y_8");
+ LLVMValueRef one = lp_build_const_v4sf(1.0);
+ LLVMValueRef y_10 = LLVMBuildAdd(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 = LLVMBuildMul(b, z, sincof_p0, "y2_3");
+ LLVMValueRef y2_4 = LLVMBuildAdd(b, y2_3, sincof_p1, "y2_4");
+ LLVMValueRef y2_5 = LLVMBuildMul(b, y2_4, z, "y2_5");
+ LLVMValueRef y2_6 = LLVMBuildAdd(b, y2_5, sincof_p2, "y2_6");
+ LLVMValueRef y2_7 = LLVMBuildMul(b, y2_6, z, "y2_7");
+ LLVMValueRef y2_8 = LLVMBuildMul(b, y2_7, x_3, "y2_8");
+ LLVMValueRef y2_9 = LLVMBuildAdd(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;
}
unsigned num_coeffs)
{
const struct lp_type type = bld->type;
- LLVMTypeRef float_type = LLVMFloatType();
LLVMValueRef res = NULL;
unsigned i;
for (i = num_coeffs; i--; ) {
LLVMValueRef coeff;
- if (type.length == 1)
- coeff = LLVMConstReal(float_type, coeffs[i]);
- else
- coeff = lp_build_const_vec(type, coeffs[i]);
+ coeff = lp_build_const_vec(type, coeffs[i]);
if(res)
res = lp_build_add(bld, coeff, lp_build_mul(bld, x, res));
/**
- * Minimax polynomial fit of 2**x, in range [-0.5, 0.5[
+ * Minimax polynomial fit of 2**x, in range [0, 1[
*/
const double lp_build_exp2_polynomial[] = {
#if EXP_POLY_DEGREE == 5
- 9.9999994e-1, 6.9315308e-1, 2.4015361e-1, 5.5826318e-2, 8.9893397e-3, 1.8775767e-3
+ 0.999999999690134838155,
+ 0.583974334321735217258,
+ 0.164553105719676828492,
+ 0.0292811063701710962255,
+ 0.00354944426657875141846,
+ 0.000296253726543423377365
#elif EXP_POLY_DEGREE == 4
- 1.0000026, 6.9300383e-1, 2.4144275e-1, 5.2011464e-2, 1.3534167e-2
+ 1.00000001502262084505,
+ 0.563586057338685991394,
+ 0.150436017652442413623,
+ 0.0243220604213317927308,
+ 0.0025359088446580436489
#elif EXP_POLY_DEGREE == 3
- 9.9992520e-1, 6.9583356e-1, 2.2606716e-1, 7.8024521e-2
+ 0.999925218562710312959,
+ 0.695833540494823811697,
+ 0.226067155427249155588,
+ 0.0780245226406372992967
#elif EXP_POLY_DEGREE == 2
- 1.0017247, 6.5763628e-1, 3.3718944e-1
+ 1.00172476321474503578,
+ 0.657636275736077639316,
+ 0.33718943461968720704
#else
#error
#endif
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));
- /* ipart = int(x - 0.5) */
- ipart = LLVMBuildSub(bld->builder, x, lp_build_const_vec(type, 0.5f), "");
- ipart = LLVMBuildFPToSI(bld->builder, ipart, int_vec_type, "");
+ /* ipart = floor(x) */
+ ipart = lp_build_floor(bld, x);
/* fpart = x - ipart */
- fpart = LLVMBuildSIToFP(bld->builder, ipart, vec_type, "");
- fpart = LLVMBuildSub(bld->builder, x, fpart, "");
+ fpart = LLVMBuildSub(bld->builder, x, ipart, "");
}
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, "");
*/
const double lp_build_log2_polynomial[] = {
#if LOG_POLY_DEGREE == 6
- 3.11578814719469302614, -3.32419399085241980044, 2.59883907202499966007, -1.23152682416275988241, 0.318212422185251071475, -0.0344359067839062357313
+ 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
+ 2.8882704548164776201,
+ -2.52074962577807006663,
+ 1.48116647521213171641,
+ -0.465725644288844778798,
+ 0.0596515482674574969533
#elif LOG_POLY_DEGREE == 4
- 2.61761038894603480148, -1.75647175389045657003, 0.688243882994381274313, -0.107254423828329604454
+ 2.61761038894603480148,
+ -1.75647175389045657003,
+ 0.688243882994381274313,
+ -0.107254423828329604454
#elif LOG_POLY_DEGREE == 3
- 2.28330284476918490682, -1.04913055217340124191, 0.204446009836232697516
+ 2.28330284476918490682,
+ -1.04913055217340124191,
+ 0.204446009836232697516
#else
#error
#endif
}
-/** scalar version of above function */
-static void
-lp_build_float_log2_approx(struct lp_build_context *bld,
- LLVMValueRef x,
- LLVMValueRef *p_exp,
- LLVMValueRef *p_floor_log2,
- LLVMValueRef *p_log2)
-{
- const struct lp_type type = bld->type;
- LLVMTypeRef float_type = LLVMFloatType();
- LLVMTypeRef int_type = LLVMIntType(type.width);
-
- LLVMValueRef expmask = LLVMConstInt(int_type, 0x7f800000, 0);
- LLVMValueRef mantmask = LLVMConstInt(int_type, 0x007fffff, 0);
- LLVMValueRef one = LLVMConstBitCast(bld->one, int_type);
-
- LLVMValueRef i = NULL;
- LLVMValueRef exp = NULL;
- LLVMValueRef mant = NULL;
- LLVMValueRef logexp = NULL;
- LLVMValueRef logmant = NULL;
- LLVMValueRef res = NULL;
-
- if(p_exp || p_floor_log2 || p_log2) {
- /* TODO: optimize the constant case */
- if(LLVMIsConstant(x))
- debug_printf("%s: inefficient/imprecise constant arithmetic\n",
- __FUNCTION__);
-
- assert(type.floating && type.width == 32);
-
- i = LLVMBuildBitCast(bld->builder, x, int_type, "");
-
- /* exp = (float) exponent(x) */
- exp = LLVMBuildAnd(bld->builder, i, expmask, "");
- }
-
- if(p_floor_log2 || p_log2) {
- LLVMValueRef c23 = LLVMConstInt(int_type, 23, 0);
- LLVMValueRef c127 = LLVMConstInt(int_type, 127, 0);
- logexp = LLVMBuildLShr(bld->builder, exp, c23, "");
- logexp = LLVMBuildSub(bld->builder, logexp, c127, "");
- logexp = LLVMBuildSIToFP(bld->builder, logexp, float_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, float_type, "");
-
- logmant = lp_build_polynomial(bld, mant, lp_build_log2_polynomial,
- Elements(lp_build_log2_polynomial));
-
- /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
- logmant = LLVMBuildMul(bld->builder, logmant, LLVMBuildSub(bld->builder, mant, bld->one, ""), "");
-
- res = LLVMBuildAdd(bld->builder, logmant, logexp, "");
- }
-
- if(p_exp) {
- exp = LLVMBuildBitCast(bld->builder, exp, float_type, "");
- *p_exp = exp;
- }
-
- if(p_floor_log2)
- *p_floor_log2 = logexp;
-
- if(p_log2)
- *p_log2 = res;
-}
-
-
LLVMValueRef
lp_build_log2(struct lp_build_context *bld,
LLVMValueRef x)
{
LLVMValueRef res;
- if (bld->type.length == 1) {
- lp_build_float_log2_approx(bld, x, NULL, NULL, &res);
- }
- else {
- lp_build_log2_approx(bld, x, NULL, NULL, &res);
- }
+ lp_build_log2_approx(bld, x, NULL, NULL, &res);
return res;
}
projects / mesa.git / blobdiff