build_mat2_det(nir_builder *b, nir_ssa_def *col[2])
{
unsigned swiz[2] = {1, 0 };
- nir_ssa_def *p = nir_fmul(b, col[0], nir_swizzle(b, col[1], swiz, 2, true));
+ nir_ssa_def *p = nir_fmul(b, col[0], nir_swizzle(b, col[1], swiz, 2));
return nir_fsub(b, nir_channel(b, p, 0), nir_channel(b, p, 1));
}
nir_ssa_def *prod0 =
nir_fmul(b, col[0],
- nir_fmul(b, nir_swizzle(b, col[1], yzx, 3, true),
- nir_swizzle(b, col[2], zxy, 3, true)));
+ nir_fmul(b, nir_swizzle(b, col[1], yzx, 3),
+ nir_swizzle(b, col[2], zxy, 3)));
nir_ssa_def *prod1 =
nir_fmul(b, col[0],
- nir_fmul(b, nir_swizzle(b, col[1], zxy, 3, true),
- nir_swizzle(b, col[2], yzx, 3, true)));
+ nir_fmul(b, nir_swizzle(b, col[1], zxy, 3),
+ nir_swizzle(b, col[2], yzx, 3)));
nir_ssa_def *diff = nir_fsub(b, prod0, prod1);
swiz[j] = j + (j >= i);
nir_ssa_def *subcol[3];
- subcol[0] = nir_swizzle(b, col[1], swiz, 3, true);
- subcol[1] = nir_swizzle(b, col[2], swiz, 3, true);
- subcol[2] = nir_swizzle(b, col[3], swiz, 3, true);
+ subcol[0] = nir_swizzle(b, col[1], swiz, 3);
+ subcol[1] = nir_swizzle(b, col[2], swiz, 3);
+ subcol[2] = nir_swizzle(b, col[3], swiz, 3);
subdet[i] = build_mat3_det(b, subcol);
}
return nir_channel(b, src->elems[1 - col]->def, 1 - row);
} else {
/* Swizzle to get all but the specified row */
- unsigned swiz[3];
+ unsigned swiz[NIR_MAX_VEC_COMPONENTS] = {0};
for (unsigned j = 0; j < 3; j++)
swiz[j] = j + (j >= row);
for (unsigned j = 0; j < size; j++) {
if (j != col) {
subcol[j - (j > col)] = nir_swizzle(b, src->elems[j]->def,
- swiz, size - 1, true);
+ swiz, size - 1);
}
}
}
/**
- * Return e^x.
- */
-static nir_ssa_def *
-build_exp(nir_builder *b, nir_ssa_def *x)
-{
- return nir_fexp2(b, nir_fmul_imm(b, x, M_LOG2E));
-}
-
-/**
- * Return ln(x) - the natural logarithm of x.
- */
-static nir_ssa_def *
-build_log(nir_builder *b, nir_ssa_def *x)
-{
- return nir_fmul_imm(b, nir_flog2(b, x), 1.0 / M_LOG2E);
-}
-
-/**
- * Approximate asin(x) by the formula:
- * asin~(x) = sign(x) * (pi/2 - sqrt(1 - |x|) * (pi/2 + |x|(pi/4 - 1 + |x|(p0 + |x|p1))))
+ * Approximate asin(x) by the piecewise formula:
+ * for |x| < 0.5, asin~(x) = x * (1 + x²(pS0 + x²(pS1 + x²*pS2)) / (1 + x²*qS1))
+ * for |x| ≥ 0.5, asin~(x) = sign(x) * (π/2 - sqrt(1 - |x|) * (π/2 + |x|(π/4 - 1 + |x|(p0 + |x|p1))))
*
- * which is correct to first order at x=0 and x=±1 regardless of the p
+ * The latter is correct to first order at x=0 and x=±1 regardless of the p
* coefficients but can be made second-order correct at both ends by selecting
* the fit coefficients appropriately. Different p coefficients can be used
* in the asin and acos implementation to minimize some relative error metric
* in each case.
*/
static nir_ssa_def *
-build_asin(nir_builder *b, nir_ssa_def *x, float p0, float p1)
+build_asin(nir_builder *b, nir_ssa_def *x, float p0, float p1, bool piecewise)
{
if (x->bit_size == 16) {
/* The polynomial approximation isn't precise enough to meet half-float
* approximation in 32-bit math and then we convert the result back to
* 16-bit.
*/
- return nir_f2f16(b, build_asin(b, nir_f2f32(b, x), p0, p1));
+ return nir_f2f16(b, build_asin(b, nir_f2f32(b, x), p0, p1, piecewise));
}
-
nir_ssa_def *one = nir_imm_floatN_t(b, 1.0f, x->bit_size);
+ nir_ssa_def *half = nir_imm_floatN_t(b, 0.5f, x->bit_size);
nir_ssa_def *abs_x = nir_fabs(b, x);
nir_ssa_def *p0_plus_xp1 = nir_fadd_imm(b, nir_fmul_imm(b, abs_x, p1), p0);
M_PI_4f - 1.0f)),
M_PI_2f);
- return nir_fmul(b, nir_fsign(b, x),
+ nir_ssa_def *result0 = nir_fmul(b, nir_fsign(b, x),
nir_fsub(b, nir_imm_floatN_t(b, M_PI_2f, x->bit_size),
nir_fmul(b, nir_fsqrt(b, nir_fsub(b, one, abs_x)),
expr_tail)));
-}
-
-/**
- * Compute xs[0] + xs[1] + xs[2] + ... using fadd.
- */
-static nir_ssa_def *
-build_fsum(nir_builder *b, nir_ssa_def **xs, int terms)
-{
- nir_ssa_def *accum = xs[0];
-
- for (int i = 1; i < terms; i++)
- accum = nir_fadd(b, accum, xs[i]);
-
- return accum;
-}
-
-static nir_ssa_def *
-build_atan(nir_builder *b, nir_ssa_def *y_over_x)
-{
- const uint32_t bit_size = y_over_x->bit_size;
-
- nir_ssa_def *abs_y_over_x = nir_fabs(b, y_over_x);
- nir_ssa_def *one = nir_imm_floatN_t(b, 1.0f, bit_size);
-
- /*
- * range-reduction, first step:
- *
- * / y_over_x if |y_over_x| <= 1.0;
- * x = <
- * \ 1.0 / y_over_x otherwise
- */
- nir_ssa_def *x = nir_fdiv(b, nir_fmin(b, abs_y_over_x, one),
- nir_fmax(b, abs_y_over_x, one));
-
- /*
- * approximate atan by evaluating polynomial:
- *
- * x * 0.9999793128310355 - x^3 * 0.3326756418091246 +
- * x^5 * 0.1938924977115610 - x^7 * 0.1173503194786851 +
- * x^9 * 0.0536813784310406 - x^11 * 0.0121323213173444
- */
- nir_ssa_def *x_2 = nir_fmul(b, x, x);
- nir_ssa_def *x_3 = nir_fmul(b, x_2, x);
- nir_ssa_def *x_5 = nir_fmul(b, x_3, x_2);
- nir_ssa_def *x_7 = nir_fmul(b, x_5, x_2);
- nir_ssa_def *x_9 = nir_fmul(b, x_7, x_2);
- nir_ssa_def *x_11 = nir_fmul(b, x_9, x_2);
-
- nir_ssa_def *polynomial_terms[] = {
- nir_fmul_imm(b, x, 0.9999793128310355f),
- nir_fmul_imm(b, x_3, -0.3326756418091246f),
- nir_fmul_imm(b, x_5, 0.1938924977115610f),
- nir_fmul_imm(b, x_7, -0.1173503194786851f),
- nir_fmul_imm(b, x_9, 0.0536813784310406f),
- nir_fmul_imm(b, x_11, -0.0121323213173444f),
- };
-
- nir_ssa_def *tmp =
- build_fsum(b, polynomial_terms, ARRAY_SIZE(polynomial_terms));
-
- /* range-reduction fixup */
- tmp = nir_fadd(b, tmp,
- nir_fmul(b, nir_b2f(b, nir_flt(b, one, abs_y_over_x), bit_size),
- nir_fadd_imm(b, nir_fmul_imm(b, tmp, -2.0f), M_PI_2f)));
-
- /* sign fixup */
- return nir_fmul(b, tmp, nir_fsign(b, y_over_x));
-}
-
-static nir_ssa_def *
-build_atan2(nir_builder *b, nir_ssa_def *y, nir_ssa_def *x)
-{
- assert(y->bit_size == x->bit_size);
- const uint32_t bit_size = x->bit_size;
-
- nir_ssa_def *zero = nir_imm_floatN_t(b, 0, bit_size);
- nir_ssa_def *one = nir_imm_floatN_t(b, 1, bit_size);
-
- /* If we're on the left half-plane rotate the coordinates π/2 clock-wise
- * for the y=0 discontinuity to end up aligned with the vertical
- * discontinuity of atan(s/t) along t=0. This also makes sure that we
- * don't attempt to divide by zero along the vertical line, which may give
- * unspecified results on non-GLSL 4.1-capable hardware.
- */
- nir_ssa_def *flip = nir_fge(b, zero, x);
- nir_ssa_def *s = nir_bcsel(b, flip, nir_fabs(b, x), y);
- nir_ssa_def *t = nir_bcsel(b, flip, y, nir_fabs(b, x));
-
- /* If the magnitude of the denominator exceeds some huge value, scale down
- * the arguments in order to prevent the reciprocal operation from flushing
- * its result to zero, which would cause precision problems, and for s
- * infinite would cause us to return a NaN instead of the correct finite
- * value.
- *
- * If fmin and fmax are respectively the smallest and largest positive
- * normalized floating point values representable by the implementation,
- * the constants below should be in agreement with:
- *
- * huge <= 1 / fmin
- * scale <= 1 / fmin / fmax (for |t| >= huge)
- *
- * In addition scale should be a negative power of two in order to avoid
- * loss of precision. The values chosen below should work for most usual
- * floating point representations with at least the dynamic range of ATI's
- * 24-bit representation.
- */
- const double huge_val = bit_size >= 32 ? 1e18 : 16384;
- nir_ssa_def *huge = nir_imm_floatN_t(b, huge_val, bit_size);
- nir_ssa_def *scale = nir_bcsel(b, nir_fge(b, nir_fabs(b, t), huge),
- nir_imm_floatN_t(b, 0.25, bit_size), one);
- nir_ssa_def *rcp_scaled_t = nir_frcp(b, nir_fmul(b, t, scale));
- nir_ssa_def *s_over_t = nir_fmul(b, nir_fmul(b, s, scale), rcp_scaled_t);
-
- /* For |x| = |y| assume tan = 1 even if infinite (i.e. pretend momentarily
- * that ∞/∞ = 1) in order to comply with the rather artificial rules
- * inherited from IEEE 754-2008, namely:
- *
- * "atan2(±∞, −∞) is ±3π/4
- * atan2(±∞, +∞) is ±π/4"
- *
- * Note that this is inconsistent with the rules for the neighborhood of
- * zero that are based on iterated limits:
- *
- * "atan2(±0, −0) is ±π
- * atan2(±0, +0) is ±0"
- *
- * but GLSL specifically allows implementations to deviate from IEEE rules
- * at (0,0), so we take that license (i.e. pretend that 0/0 = 1 here as
- * well).
- */
- nir_ssa_def *tan = nir_bcsel(b, nir_feq(b, nir_fabs(b, x), nir_fabs(b, y)),
- one, nir_fabs(b, s_over_t));
-
- /* Calculate the arctangent and fix up the result if we had flipped the
- * coordinate system.
- */
- nir_ssa_def *arc =
- nir_fadd(b, nir_fmul_imm(b, nir_b2f(b, flip, bit_size), M_PI_2f),
- build_atan(b, tan));
-
- /* Rather convoluted calculation of the sign of the result. When x < 0 we
- * cannot use fsign because we need to be able to distinguish between
- * negative and positive zero. We don't use bitwise arithmetic tricks for
- * consistency with the GLSL front-end. When x >= 0 rcp_scaled_t will
- * always be non-negative so this won't be able to distinguish between
- * negative and positive zero, but we don't care because atan2 is
- * continuous along the whole positive y = 0 half-line, so it won't affect
- * the result significantly.
- */
- return nir_bcsel(b, nir_flt(b, nir_fmin(b, y, rcp_scaled_t), zero),
- nir_fneg(b, arc), arc);
-}
-
-static nir_ssa_def *
-build_frexp16(nir_builder *b, nir_ssa_def *x, nir_ssa_def **exponent)
-{
- assert(x->bit_size == 16);
-
- nir_ssa_def *abs_x = nir_fabs(b, x);
- nir_ssa_def *zero = nir_imm_floatN_t(b, 0, 16);
-
- /* Half-precision floating-point values are stored as
- * 1 sign bit;
- * 5 exponent bits;
- * 10 mantissa bits.
- *
- * An exponent shift of 10 will shift the mantissa out, leaving only the
- * exponent and sign bit (which itself may be zero, if the absolute value
- * was taken before the bitcast and shift).
- */
- nir_ssa_def *exponent_shift = nir_imm_int(b, 10);
- nir_ssa_def *exponent_bias = nir_imm_intN_t(b, -14, 16);
-
- nir_ssa_def *sign_mantissa_mask = nir_imm_intN_t(b, 0x83ffu, 16);
-
- /* Exponent of floating-point values in the range [0.5, 1.0). */
- nir_ssa_def *exponent_value = nir_imm_intN_t(b, 0x3800u, 16);
-
- nir_ssa_def *is_not_zero = nir_fne(b, abs_x, zero);
-
- /* Significand return must be of the same type as the input, but the
- * exponent must be a 32-bit integer.
- */
- *exponent =
- nir_i2i32(b,
- nir_iadd(b, nir_ushr(b, abs_x, exponent_shift),
- nir_bcsel(b, is_not_zero, exponent_bias, zero)));
-
- return nir_ior(b, nir_iand(b, x, sign_mantissa_mask),
- nir_bcsel(b, is_not_zero, exponent_value, zero));
-}
-
-static nir_ssa_def *
-build_frexp32(nir_builder *b, nir_ssa_def *x, nir_ssa_def **exponent)
-{
- nir_ssa_def *abs_x = nir_fabs(b, x);
- nir_ssa_def *zero = nir_imm_float(b, 0.0f);
-
- /* Single-precision floating-point values are stored as
- * 1 sign bit;
- * 8 exponent bits;
- * 23 mantissa bits.
- *
- * An exponent shift of 23 will shift the mantissa out, leaving only the
- * exponent and sign bit (which itself may be zero, if the absolute value
- * was taken before the bitcast and shift.
- */
- nir_ssa_def *exponent_shift = nir_imm_int(b, 23);
- nir_ssa_def *exponent_bias = nir_imm_int(b, -126);
-
- nir_ssa_def *sign_mantissa_mask = nir_imm_int(b, 0x807fffffu);
-
- /* Exponent of floating-point values in the range [0.5, 1.0). */
- nir_ssa_def *exponent_value = nir_imm_int(b, 0x3f000000u);
-
- nir_ssa_def *is_not_zero = nir_fne(b, abs_x, zero);
-
- *exponent =
- nir_iadd(b, nir_ushr(b, abs_x, exponent_shift),
- nir_bcsel(b, is_not_zero, exponent_bias, zero));
-
- return nir_ior(b, nir_iand(b, x, sign_mantissa_mask),
- nir_bcsel(b, is_not_zero, exponent_value, zero));
-}
-
-static nir_ssa_def *
-build_frexp64(nir_builder *b, nir_ssa_def *x, nir_ssa_def **exponent)
-{
- nir_ssa_def *abs_x = nir_fabs(b, x);
- nir_ssa_def *zero = nir_imm_double(b, 0.0);
- nir_ssa_def *zero32 = nir_imm_float(b, 0.0f);
-
- /* Double-precision floating-point values are stored as
- * 1 sign bit;
- * 11 exponent bits;
- * 52 mantissa bits.
- *
- * We only need to deal with the exponent so first we extract the upper 32
- * bits using nir_unpack_64_2x32_split_y.
- */
- nir_ssa_def *upper_x = nir_unpack_64_2x32_split_y(b, x);
- nir_ssa_def *abs_upper_x = nir_unpack_64_2x32_split_y(b, abs_x);
-
- /* An exponent shift of 20 will shift the remaining mantissa bits out,
- * leaving only the exponent and sign bit (which itself may be zero, if the
- * absolute value was taken before the bitcast and shift.
- */
- nir_ssa_def *exponent_shift = nir_imm_int(b, 20);
- nir_ssa_def *exponent_bias = nir_imm_int(b, -1022);
-
- nir_ssa_def *sign_mantissa_mask = nir_imm_int(b, 0x800fffffu);
-
- /* Exponent of floating-point values in the range [0.5, 1.0). */
- nir_ssa_def *exponent_value = nir_imm_int(b, 0x3fe00000u);
-
- nir_ssa_def *is_not_zero = nir_fne(b, abs_x, zero);
-
- *exponent =
- nir_iadd(b, nir_ushr(b, abs_upper_x, exponent_shift),
- nir_bcsel(b, is_not_zero, exponent_bias, zero32));
-
- nir_ssa_def *new_upper =
- nir_ior(b, nir_iand(b, upper_x, sign_mantissa_mask),
- nir_bcsel(b, is_not_zero, exponent_value, zero32));
-
- nir_ssa_def *lower_x = nir_unpack_64_2x32_split_x(b, x);
-
- return nir_pack_64_2x32_split(b, lower_x, new_upper);
+ if (piecewise) {
+ /* approximation for |x| < 0.5 */
+ const float pS0 = 1.6666586697e-01f;
+ const float pS1 = -4.2743422091e-02f;
+ const float pS2 = -8.6563630030e-03f;
+ const float qS1 = -7.0662963390e-01f;
+
+ nir_ssa_def *x2 = nir_fmul(b, x, x);
+ nir_ssa_def *p = nir_fmul(b,
+ x2,
+ nir_fadd_imm(b,
+ nir_fmul(b,
+ x2,
+ nir_fadd_imm(b, nir_fmul_imm(b, x2, pS2),
+ pS1)),
+ pS0));
+
+ nir_ssa_def *q = nir_fadd(b, one, nir_fmul_imm(b, x2, qS1));
+ nir_ssa_def *result1 = nir_fadd(b, x, nir_fmul(b, x, nir_fdiv(b, p, q)));
+ return nir_bcsel(b, nir_flt(b, abs_x, half), result1, result0);
+ } else {
+ return result0;
+ }
}
static nir_op
vtn_nir_alu_op_for_spirv_glsl_opcode(struct vtn_builder *b,
- enum GLSLstd450 opcode)
+ enum GLSLstd450 opcode,
+ unsigned execution_mode,
+ bool *exact)
{
+ *exact = false;
switch (opcode) {
case GLSLstd450Round: return nir_op_fround_even;
case GLSLstd450RoundEven: return nir_op_fround_even;
case GLSLstd450Log2: return nir_op_flog2;
case GLSLstd450Sqrt: return nir_op_fsqrt;
case GLSLstd450InverseSqrt: return nir_op_frsq;
- case GLSLstd450NMin: return nir_op_fmin;
+ case GLSLstd450NMin: *exact = true; return nir_op_fmin;
case GLSLstd450FMin: return nir_op_fmin;
case GLSLstd450UMin: return nir_op_umin;
case GLSLstd450SMin: return nir_op_imin;
- case GLSLstd450NMax: return nir_op_fmax;
+ case GLSLstd450NMax: *exact = true; return nir_op_fmax;
case GLSLstd450FMax: return nir_op_fmax;
case GLSLstd450UMax: return nir_op_umax;
case GLSLstd450SMax: return nir_op_imax;
case GLSLstd450UnpackUnorm4x8: return nir_op_unpack_unorm_4x8;
case GLSLstd450UnpackSnorm2x16: return nir_op_unpack_snorm_2x16;
case GLSLstd450UnpackUnorm2x16: return nir_op_unpack_unorm_2x16;
- case GLSLstd450UnpackHalf2x16: return nir_op_unpack_half_2x16;
+ case GLSLstd450UnpackHalf2x16:
+ if (execution_mode & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16)
+ return nir_op_unpack_half_2x16_flush_to_zero;
+ else
+ return nir_op_unpack_half_2x16;
case GLSLstd450UnpackDouble2x32: return nir_op_unpack_64_2x32;
default:
val->ssa->def = nir_degrees(nb, src[0]);
return;
case GLSLstd450Tan:
- val->ssa->def = nir_fdiv(nb, nir_fsin(nb, src[0]),
- nir_fcos(nb, src[0]));
+ val->ssa->def = nir_ftan(nb, src[0]);
return;
case GLSLstd450Modf: {
return;
case GLSLstd450Exp:
- val->ssa->def = build_exp(nb, src[0]);
+ val->ssa->def = nir_fexp(nb, src[0]);
return;
case GLSLstd450Log:
- val->ssa->def = build_log(nb, src[0]);
+ val->ssa->def = nir_flog(nb, src[0]);
return;
case GLSLstd450FClamp:
+ val->ssa->def = nir_fclamp(nb, src[0], src[1], src[2]);
+ return;
case GLSLstd450NClamp:
+ nb->exact = true;
val->ssa->def = nir_fclamp(nb, src[0], src[1], src[2]);
+ nb->exact = false;
return;
case GLSLstd450UClamp:
val->ssa->def = nir_uclamp(nb, src[0], src[1], src[2]);
case GLSLstd450Sinh:
/* 0.5 * (e^x - e^(-x)) */
val->ssa->def =
- nir_fmul_imm(nb, nir_fsub(nb, build_exp(nb, src[0]),
- build_exp(nb, nir_fneg(nb, src[0]))),
+ nir_fmul_imm(nb, nir_fsub(nb, nir_fexp(nb, src[0]),
+ nir_fexp(nb, nir_fneg(nb, src[0]))),
0.5f);
return;
case GLSLstd450Cosh:
/* 0.5 * (e^x + e^(-x)) */
val->ssa->def =
- nir_fmul_imm(nb, nir_fadd(nb, build_exp(nb, src[0]),
- build_exp(nb, nir_fneg(nb, src[0]))),
+ nir_fmul_imm(nb, nir_fadd(nb, nir_fexp(nb, src[0]),
+ nir_fexp(nb, nir_fneg(nb, src[0]))),
0.5f);
return;
case GLSLstd450Tanh: {
- /* tanh(x) := (0.5 * (e^x - e^(-x))) / (0.5 * (e^x + e^(-x)))
- *
- * With a little algebra this reduces to (e^2x - 1) / (e^2x + 1)
+ /* tanh(x) := (e^x - e^(-x)) / (e^x + e^(-x))
*
- * We clamp x to (-inf, +10] to avoid precision problems. When x > 10,
- * e^2x is so much larger than 1.0 that 1.0 gets flushed to zero in the
- * computation e^2x +/- 1 so it can be ignored.
+ * We clamp x to [-10, +10] to avoid precision problems. When x > 10,
+ * e^x dominates the sum, e^(-x) is lost and tanh(x) is 1.0 for 32 bit
+ * floating point.
*
- * For 16-bit precision we clamp x to (-inf, +4.2] since the maximum
- * representable number is only 65,504 and e^(2*6) exceeds that. Also,
- * if x > 4.2, tanh(x) will return 1.0 in fp16.
+ * For 16-bit precision this we clamp x to [-4.2, +4.2].
*/
const uint32_t bit_size = src[0]->bit_size;
const double clamped_x = bit_size > 16 ? 10.0 : 4.2;
- nir_ssa_def *x = nir_fmin(nb, src[0],
- nir_imm_floatN_t(nb, clamped_x, bit_size));
- nir_ssa_def *exp2x = build_exp(nb, nir_fmul_imm(nb, x, 2.0));
- val->ssa->def = nir_fdiv(nb, nir_fadd_imm(nb, exp2x, -1.0),
- nir_fadd_imm(nb, exp2x, 1.0));
+ nir_ssa_def *x = nir_fclamp(nb, src[0],
+ nir_imm_floatN_t(nb, -clamped_x, bit_size),
+ nir_imm_floatN_t(nb, clamped_x, bit_size));
+ val->ssa->def =
+ nir_fdiv(nb, nir_fsub(nb, nir_fexp(nb, x),
+ nir_fexp(nb, nir_fneg(nb, x))),
+ nir_fadd(nb, nir_fexp(nb, x),
+ nir_fexp(nb, nir_fneg(nb, x))));
return;
}
case GLSLstd450Asinh:
val->ssa->def = nir_fmul(nb, nir_fsign(nb, src[0]),
- build_log(nb, nir_fadd(nb, nir_fabs(nb, src[0]),
- nir_fsqrt(nb, nir_fadd_imm(nb, nir_fmul(nb, src[0], src[0]),
- 1.0f)))));
+ nir_flog(nb, nir_fadd(nb, nir_fabs(nb, src[0]),
+ nir_fsqrt(nb, nir_fadd_imm(nb, nir_fmul(nb, src[0], src[0]),
+ 1.0f)))));
return;
case GLSLstd450Acosh:
- val->ssa->def = build_log(nb, nir_fadd(nb, src[0],
+ val->ssa->def = nir_flog(nb, nir_fadd(nb, src[0],
nir_fsqrt(nb, nir_fadd_imm(nb, nir_fmul(nb, src[0], src[0]),
-1.0f))));
return;
case GLSLstd450Atanh: {
nir_ssa_def *one = nir_imm_floatN_t(nb, 1.0, src[0]->bit_size);
val->ssa->def =
- nir_fmul_imm(nb, build_log(nb, nir_fdiv(nb, nir_fadd(nb, src[0], one),
- nir_fsub(nb, one, src[0]))),
+ nir_fmul_imm(nb, nir_flog(nb, nir_fdiv(nb, nir_fadd(nb, src[0], one),
+ nir_fsub(nb, one, src[0]))),
0.5f);
return;
}
case GLSLstd450Asin:
- val->ssa->def = build_asin(nb, src[0], 0.086566724, -0.03102955);
+ val->ssa->def = build_asin(nb, src[0], 0.086566724, -0.03102955, true);
return;
case GLSLstd450Acos:
val->ssa->def =
nir_fsub(nb, nir_imm_floatN_t(nb, M_PI_2f, src[0]->bit_size),
- build_asin(nb, src[0], 0.08132463, -0.02363318));
+ build_asin(nb, src[0], 0.08132463, -0.02363318, false));
return;
case GLSLstd450Atan:
- val->ssa->def = build_atan(nb, src[0]);
+ val->ssa->def = nir_atan(nb, src[0]);
return;
case GLSLstd450Atan2:
- val->ssa->def = build_atan2(nb, src[0], src[1]);
+ val->ssa->def = nir_atan2(nb, src[0], src[1]);
return;
case GLSLstd450Frexp: {
- nir_ssa_def *exponent;
- if (src[0]->bit_size == 64)
- val->ssa->def = build_frexp64(nb, src[0], &exponent);
- else if (src[0]->bit_size == 32)
- val->ssa->def = build_frexp32(nb, src[0], &exponent);
- else
- val->ssa->def = build_frexp16(nb, src[0], &exponent);
+ nir_ssa_def *exponent = nir_frexp_exp(nb, src[0]);
+ val->ssa->def = nir_frexp_sig(nb, src[0]);
nir_store_deref(nb, vtn_nir_deref(b, w[6]), exponent, 0xf);
return;
}
case GLSLstd450FrexpStruct: {
vtn_assert(glsl_type_is_struct_or_ifc(val->ssa->type));
- if (src[0]->bit_size == 64)
- val->ssa->elems[0]->def = build_frexp64(nb, src[0],
- &val->ssa->elems[1]->def);
- else if (src[0]->bit_size == 32)
- val->ssa->elems[0]->def = build_frexp32(nb, src[0],
- &val->ssa->elems[1]->def);
- else
- val->ssa->elems[0]->def = build_frexp16(nb, src[0],
- &val->ssa->elems[1]->def);
+ val->ssa->elems[0]->def = nir_frexp_sig(nb, src[0]);
+ val->ssa->elems[1]->def = nir_frexp_exp(nb, src[0]);
return;
}
- default:
- val->ssa->def =
- nir_build_alu(&b->nb,
- vtn_nir_alu_op_for_spirv_glsl_opcode(b, entrypoint),
- src[0], src[1], src[2], NULL);
+ default: {
+ unsigned execution_mode =
+ b->shader->info.float_controls_execution_mode;
+ bool exact;
+ nir_op op = vtn_nir_alu_op_for_spirv_glsl_opcode(b, entrypoint, execution_mode, &exact);
+ b->nb.exact = exact;
+ val->ssa->def = nir_build_alu(&b->nb, op, src[0], src[1], src[2], NULL);
+ b->nb.exact = false;
return;
}
+ }
}
static void
if (vec_array_deref) {
assert(vec_deref);
- if (nir_src_is_const(vec_deref->arr.index)) {
- val->ssa->def = vtn_vector_extract(b, &intrin->dest.ssa,
- nir_src_as_uint(vec_deref->arr.index));
- } else {
- val->ssa->def = vtn_vector_extract_dynamic(b, &intrin->dest.ssa,
- vec_deref->arr.index.ssa);
- }
+ val->ssa->def = nir_vector_extract(&b->nb, &intrin->dest.ssa,
+ vec_deref->arr.index.ssa);
} else {
val->ssa->def = &intrin->dest.ssa;
}
case GLSLstd450InterpolateAtCentroid:
case GLSLstd450InterpolateAtSample:
case GLSLstd450InterpolateAtOffset:
- handle_glsl450_interpolation(b, ext_opcode, w, count);
+ handle_glsl450_interpolation(b, (enum GLSLstd450)ext_opcode, w, count);
break;
default: