/**
* 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))))
*
* which 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)
{
nir_ssa_def *abs_x = nir_fabs(b, x);
return nir_fmul(b, nir_fsign(b, x),
nir_fsub(b, nir_imm_float(b, M_PI_2f),
nir_fmul(b, nir_fsqrt(b, nir_fsub(b, nir_imm_float(b, 1.0f), abs_x)),
nir_fadd(b, nir_imm_float(b, M_PI_2f),
nir_fmul(b, abs_x,
nir_fadd(b, nir_imm_float(b, M_PI_4f - 1.0f),
nir_fmul(b, abs_x,
nir_fadd(b, nir_imm_float(b, p0),
nir_fmul(b, abs_x,
nir_imm_float(b, p1))))))))));
}
static nir_ssa_def *
build_mat2_det(nir_builder *b, nir_ssa_def *col[2])
{
unsigned swiz[4] = {1, 0, 0, 0};
nir_ssa_def *p = nir_fmul(b, col[0], nir_swizzle(b, col[1], swiz, 2, true));
return nir_fsub(b, nir_channel(b, p, 0), nir_channel(b, p, 1));
}
static void
vtn_handle_matrix_alu(struct vtn_builder *b, SpvOp opcode,
struct vtn_value *dest,
struct vtn_ssa_value *src0, struct vtn_ssa_value *src1)
{
switch (opcode) {
case SpvOpFNegate: {
dest->ssa = vtn_create_ssa_value(b, src0->type);
unsigned cols = glsl_get_matrix_columns(src0->type);
for (unsigned i = 0; i < cols; i++)
dest->ssa->elems[i]->def = nir_fneg(&b->nb, src0->elems[i]->def);
break;
}
case SpvOpFAdd: {
dest->ssa = vtn_create_ssa_value(b, src0->type);
unsigned cols = glsl_get_matrix_columns(src0->type);
for (unsigned i = 0; i < cols; i++)
dest->ssa->elems[i]->def =
nir_fadd(&b->nb, src0->elems[i]->def, src1->elems[i]->def);
break;
}
case SpvOpFSub: {
dest->ssa = vtn_create_ssa_value(b, src0->type);
unsigned cols = glsl_get_matrix_columns(src0->type);
for (unsigned i = 0; i < cols; i++)
dest->ssa->elems[i]->def =
nir_fsub(&b->nb, src0->elems[i]->def, src1->elems[i]->def);
break;
}
case SpvOpTranspose:
dest->ssa = vtn_ssa_transpose(b, src0);
break;
case SpvOpMatrixTimesScalar:
if (src0->transposed) {
dest->ssa = vtn_ssa_transpose(b, mat_times_scalar(b, src0->transposed,
src1->def));
} else {
dest->ssa = mat_times_scalar(b, src0, src1->def);
}
break;
case SpvOpVectorTimesMatrix:
case SpvOpMatrixTimesVector:
case SpvOpMatrixTimesMatrix:
if (opcode == SpvOpVectorTimesMatrix) {
dest->ssa = matrix_multiply(b, vtn_ssa_transpose(b, src1), src0);
} else {
dest->ssa = matrix_multiply(b, src0, src1);
}
break;
default: unreachable("unknown matrix opcode");
}
}
static nir_ssa_def *
build_mat4_det(nir_builder *b, nir_ssa_def **col)
{
nir_ssa_def *subdet[4];
for (unsigned i = 0; i < 4; i++) {
unsigned swiz[3];
for (unsigned j = 0; j < 3; j++)
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);
subdet[i] = build_mat3_det(b, subcol);
}
nir_ssa_def *prod = nir_fmul(b, col[0], nir_vec(b, subdet, 4));
return nir_fadd(b, nir_fsub(b, nir_channel(b, prod, 0),
nir_channel(b, prod, 1)),
nir_fsub(b, nir_channel(b, prod, 2),
nir_channel(b, prod, 3)));
}
static nir_ssa_def *
build_mat3_det(nir_builder *b, nir_ssa_def *col[3])
{
unsigned yzx[4] = {1, 2, 0, 0};
unsigned zxy[4] = {2, 0, 1, 0};
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_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_ssa_def *diff = nir_fsub(b, prod0, prod1);
return nir_fadd(b, nir_channel(b, diff, 0),
nir_fadd(b, nir_channel(b, diff, 1),
nir_channel(b, diff, 2)));
}
static void
handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint,
const uint32_t *w, unsigned count)
{
struct nir_builder *nb = &b->nb;
const struct glsl_type *dest_type =
vtn_value(b, w[1], vtn_value_type_type)->type->type;
struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
val->ssa = vtn_create_ssa_value(b, dest_type);
/* Collect the various SSA sources */
unsigned num_inputs = count - 5;
nir_ssa_def *src[3] = { NULL, };
for (unsigned i = 0; i < num_inputs; i++)
src[i] = vtn_ssa_value(b, w[i + 5])->def;
switch (entrypoint) {
case GLSLstd450Radians:
val->ssa->def = nir_fmul(nb, src[0], nir_imm_float(nb, 0.01745329251));
return;
case GLSLstd450Degrees:
val->ssa->def = nir_fmul(nb, src[0], nir_imm_float(nb, 57.2957795131));
return;
case GLSLstd450Tan:
val->ssa->def = nir_fdiv(nb, nir_fsin(nb, src[0]),
nir_fcos(nb, src[0]));
return;
case GLSLstd450Modf: {
nir_ssa_def *sign = nir_fsign(nb, src[0]);
nir_ssa_def *abs = nir_fabs(nb, src[0]);
val->ssa->def = nir_fmul(nb, sign, nir_ffract(nb, abs));
nir_store_deref_var(nb, vtn_nir_deref(b, w[6]),
nir_fmul(nb, sign, nir_ffloor(nb, abs)), 0xf);
return;
}
case GLSLstd450ModfStruct: {
nir_ssa_def *sign = nir_fsign(nb, src[0]);
nir_ssa_def *abs = nir_fabs(nb, src[0]);
assert(glsl_type_is_struct(val->ssa->type));
val->ssa->elems[0]->def = nir_fmul(nb, sign, nir_ffract(nb, abs));
val->ssa->elems[1]->def = nir_fmul(nb, sign, nir_ffloor(nb, abs));
return;
}
case GLSLstd450Step:
val->ssa->def = nir_sge(nb, src[1], src[0]);
return;
case GLSLstd450Length:
val->ssa->def = build_length(nb, src[0]);
return;
case GLSLstd450Distance:
val->ssa->def = build_length(nb, nir_fsub(nb, src[0], src[1]));
return;
case GLSLstd450Normalize:
val->ssa->def = nir_fdiv(nb, src[0], build_length(nb, src[0]));
return;
case GLSLstd450Exp:
val->ssa->def = build_exp(nb, src[0]);
return;
case GLSLstd450Log:
val->ssa->def = build_log(nb, src[0]);
return;
case GLSLstd450FClamp:
val->ssa->def = build_fclamp(nb, src[0], src[1], src[2]);
return;
case GLSLstd450UClamp:
val->ssa->def = nir_umin(nb, nir_umax(nb, src[0], src[1]), src[2]);
return;
case GLSLstd450SClamp:
val->ssa->def = nir_imin(nb, nir_imax(nb, src[0], src[1]), src[2]);
return;
case GLSLstd450Cross: {
unsigned yzx[4] = { 1, 2, 0, 0 };
unsigned zxy[4] = { 2, 0, 1, 0 };
val->ssa->def =
nir_fsub(nb, nir_fmul(nb, nir_swizzle(nb, src[0], yzx, 3, true),
nir_swizzle(nb, src[1], zxy, 3, true)),
nir_fmul(nb, nir_swizzle(nb, src[0], zxy, 3, true),
nir_swizzle(nb, src[1], yzx, 3, true)));
return;
}
case GLSLstd450SmoothStep: {
/* t = clamp((x - edge0) / (edge1 - edge0), 0, 1) */
nir_ssa_def *t =
build_fclamp(nb, nir_fdiv(nb, nir_fsub(nb, src[2], src[0]),
nir_fsub(nb, src[1], src[0])),
nir_imm_float(nb, 0.0), nir_imm_float(nb, 1.0));
/* result = t * t * (3 - 2 * t) */
val->ssa->def =
nir_fmul(nb, t, nir_fmul(nb, t,
nir_fsub(nb, nir_imm_float(nb, 3.0),
nir_fmul(nb, nir_imm_float(nb, 2.0), t))));
return;
}
case GLSLstd450FaceForward:
val->ssa->def =
nir_bcsel(nb, nir_flt(nb, nir_fdot(nb, src[2], src[1]),
nir_imm_float(nb, 0.0)),
src[0], nir_fneg(nb, src[0]));
return;
case GLSLstd450Reflect:
/* I - 2 * dot(N, I) * N */
val->ssa->def =
nir_fsub(nb, src[0], nir_fmul(nb, nir_imm_float(nb, 2.0),
nir_fmul(nb, nir_fdot(nb, src[0], src[1]),
src[1])));
return;
case GLSLstd450Refract: {
nir_ssa_def *I = src[0];
nir_ssa_def *N = src[1];
nir_ssa_def *eta = src[2];
nir_ssa_def *n_dot_i = nir_fdot(nb, N, I);
nir_ssa_def *one = nir_imm_float(nb, 1.0);
nir_ssa_def *zero = nir_imm_float(nb, 0.0);
/* k = 1.0 - eta * eta * (1.0 - dot(N, I) * dot(N, I)) */
nir_ssa_def *k =
nir_fsub(nb, one, nir_fmul(nb, eta, nir_fmul(nb, eta,
nir_fsub(nb, one, nir_fmul(nb, n_dot_i, n_dot_i)))));
nir_ssa_def *result =
nir_fsub(nb, nir_fmul(nb, eta, I),
nir_fmul(nb, nir_fadd(nb, nir_fmul(nb, eta, n_dot_i),
nir_fsqrt(nb, k)), N));
/* XXX: bcsel, or if statement? */
val->ssa->def = nir_bcsel(nb, nir_flt(nb, k, zero), zero, result);
return;
}
case GLSLstd450Sinh:
/* 0.5 * (e^x - e^(-x)) */
val->ssa->def =
nir_fmul(nb, nir_imm_float(nb, 0.5f),
nir_fsub(nb, build_exp(nb, src[0]),
build_exp(nb, nir_fneg(nb, src[0]))));
return;
case GLSLstd450Cosh:
/* 0.5 * (e^x + e^(-x)) */
val->ssa->def =
nir_fmul(nb, nir_imm_float(nb, 0.5f),
nir_fadd(nb, build_exp(nb, src[0]),
build_exp(nb, nir_fneg(nb, src[0]))));
return;
case GLSLstd450Tanh:
/* (0.5 * (e^x - e^(-x))) / (0.5 * (e^x + e^(-x))) */
val->ssa->def =
nir_fdiv(nb, nir_fmul(nb, nir_imm_float(nb, 0.5f),
nir_fsub(nb, build_exp(nb, src[0]),
build_exp(nb, nir_fneg(nb, src[0])))),
nir_fmul(nb, nir_imm_float(nb, 0.5f),
nir_fadd(nb, build_exp(nb, src[0]),
build_exp(nb, nir_fneg(nb, src[0])))));
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(nb, nir_fmul(nb, src[0], src[0]),
nir_imm_float(nb, 1.0f))))));
return;
case GLSLstd450Acosh:
val->ssa->def = build_log(nb, nir_fadd(nb, src[0],
nir_fsqrt(nb, nir_fsub(nb, nir_fmul(nb, src[0], src[0]),
nir_imm_float(nb, 1.0f)))));
return;
case GLSLstd450Atanh: {
nir_ssa_def *one = nir_imm_float(nb, 1.0);
val->ssa->def = nir_fmul(nb, nir_imm_float(nb, 0.5f),
build_log(nb, nir_fdiv(nb, nir_fadd(nb, one, src[0]),
nir_fsub(nb, one, src[0]))));
return;
}
case GLSLstd450Asin:
val->ssa->def = build_asin(nb, src[0], 0.086566724, -0.03102955);
return;
case GLSLstd450Acos:
val->ssa->def = nir_fsub(nb, nir_imm_float(nb, M_PI_2f),
build_asin(nb, src[0], 0.08132463, -0.02363318));
return;
case GLSLstd450Atan:
val->ssa->def = build_atan(nb, src[0]);
return;
case GLSLstd450Atan2:
val->ssa->def = build_atan2(nb, src[0], src[1]);
return;
case GLSLstd450Frexp: {
nir_ssa_def *exponent;
val->ssa->def = build_frexp(nb, src[0], &exponent);
nir_store_deref_var(nb, vtn_nir_deref(b, w[6]), exponent, 0xf);
return;
}
case GLSLstd450FrexpStruct: {
assert(glsl_type_is_struct(val->ssa->type));
val->ssa->elems[0]->def = build_frexp(nb, src[0],
&val->ssa->elems[1]->def);
return;
}
default:
val->ssa->def =
nir_build_alu(&b->nb, vtn_nir_alu_op_for_spirv_glsl_opcode(entrypoint),
src[0], src[1], src[2], NULL);
return;
}
}
