void
vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
{
struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
const struct glsl_type *type =
vtn_value(b, w[1], vtn_value_type_type)->type->type;
vtn_foreach_decoration(b, val, handle_no_contraction, NULL);
/* Collect the various SSA sources */
const unsigned num_inputs = count - 3;
struct vtn_ssa_value *vtn_src[4] = { NULL, };
for (unsigned i = 0; i < num_inputs; i++)
vtn_src[i] = vtn_ssa_value(b, w[i + 3]);
if (glsl_type_is_matrix(vtn_src[0]->type) ||
(num_inputs >= 2 && glsl_type_is_matrix(vtn_src[1]->type))) {
vtn_handle_matrix_alu(b, opcode, val, vtn_src[0], vtn_src[1]);
b->nb.exact = false;
return;
}
val->ssa = vtn_create_ssa_value(b, type);
nir_ssa_def *src[4] = { NULL, };
for (unsigned i = 0; i < num_inputs; i++) {
assert(glsl_type_is_vector_or_scalar(vtn_src[i]->type));
src[i] = vtn_src[i]->def;
}
switch (opcode) {
case SpvOpAny:
if (src[0]->num_components == 1) {
val->ssa->def = nir_imov(&b->nb, src[0]);
} else {
nir_op op;
switch (src[0]->num_components) {
case 2: op = nir_op_bany_inequal2; break;
case 3: op = nir_op_bany_inequal3; break;
case 4: op = nir_op_bany_inequal4; break;
default: unreachable("invalid number of components");
}
val->ssa->def = nir_build_alu(&b->nb, op, src[0],
nir_imm_int(&b->nb, NIR_FALSE),
NULL, NULL);
}
break;
case SpvOpAll:
if (src[0]->num_components == 1) {
val->ssa->def = nir_imov(&b->nb, src[0]);
} else {
nir_op op;
switch (src[0]->num_components) {
case 2: op = nir_op_ball_iequal2; break;
case 3: op = nir_op_ball_iequal3; break;
case 4: op = nir_op_ball_iequal4; break;
default: unreachable("invalid number of components");
}
val->ssa->def = nir_build_alu(&b->nb, op, src[0],
nir_imm_int(&b->nb, NIR_TRUE),
NULL, NULL);
}
break;
case SpvOpOuterProduct: {
for (unsigned i = 0; i < src[1]->num_components; i++) {
val->ssa->elems[i]->def =
nir_fmul(&b->nb, src[0], nir_channel(&b->nb, src[1], i));
}
break;
}
case SpvOpDot:
val->ssa->def = nir_fdot(&b->nb, src[0], src[1]);
break;
case SpvOpIAddCarry:
assert(glsl_type_is_struct(val->ssa->type));
val->ssa->elems[0]->def = nir_iadd(&b->nb, src[0], src[1]);
val->ssa->elems[1]->def = nir_uadd_carry(&b->nb, src[0], src[1]);
break;
case SpvOpISubBorrow:
assert(glsl_type_is_struct(val->ssa->type));
val->ssa->elems[0]->def = nir_isub(&b->nb, src[0], src[1]);
val->ssa->elems[1]->def = nir_usub_borrow(&b->nb, src[0], src[1]);
break;
case SpvOpUMulExtended:
assert(glsl_type_is_struct(val->ssa->type));
val->ssa->elems[0]->def = nir_imul(&b->nb, src[0], src[1]);
val->ssa->elems[1]->def = nir_umul_high(&b->nb, src[0], src[1]);
break;
case SpvOpSMulExtended:
assert(glsl_type_is_struct(val->ssa->type));
val->ssa->elems[0]->def = nir_imul(&b->nb, src[0], src[1]);
val->ssa->elems[1]->def = nir_imul_high(&b->nb, src[0], src[1]);
break;
case SpvOpFwidth:
val->ssa->def = nir_fadd(&b->nb,
nir_fabs(&b->nb, nir_fddx(&b->nb, src[0])),
nir_fabs(&b->nb, nir_fddy(&b->nb, src[0])));
break;
case SpvOpFwidthFine:
val->ssa->def = nir_fadd(&b->nb,
nir_fabs(&b->nb, nir_fddx_fine(&b->nb, src[0])),
nir_fabs(&b->nb, nir_fddy_fine(&b->nb, src[0])));
break;
case SpvOpFwidthCoarse:
val->ssa->def = nir_fadd(&b->nb,
nir_fabs(&b->nb, nir_fddx_coarse(&b->nb, src[0])),
nir_fabs(&b->nb, nir_fddy_coarse(&b->nb, src[0])));
break;
case SpvOpVectorTimesScalar:
/* The builder will take care of splatting for us. */
val->ssa->def = nir_fmul(&b->nb, src[0], src[1]);
break;
case SpvOpIsNan:
val->ssa->def = nir_fne(&b->nb, src[0], src[0]);
break;
case SpvOpIsInf:
val->ssa->def = nir_feq(&b->nb, nir_fabs(&b->nb, src[0]),
nir_imm_float(&b->nb, INFINITY));
break;
case SpvOpFUnordEqual:
case SpvOpFUnordNotEqual:
case SpvOpFUnordLessThan:
case SpvOpFUnordGreaterThan:
case SpvOpFUnordLessThanEqual:
case SpvOpFUnordGreaterThanEqual: {
bool swap;
nir_alu_type src_alu_type = nir_get_nir_type_for_glsl_type(vtn_src[0]->type);
nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(type);
nir_op op = vtn_nir_alu_op_for_spirv_opcode(opcode, &swap, src_alu_type, dst_alu_type);
if (swap) {
nir_ssa_def *tmp = src[0];
src[0] = src[1];
src[1] = tmp;
}
val->ssa->def =
nir_ior(&b->nb,
nir_build_alu(&b->nb, op, src[0], src[1], NULL, NULL),
nir_ior(&b->nb,
nir_fne(&b->nb, src[0], src[0]),
nir_fne(&b->nb, src[1], src[1])));
break;
}
case SpvOpFOrdEqual:
case SpvOpFOrdNotEqual:
case SpvOpFOrdLessThan:
case SpvOpFOrdGreaterThan:
case SpvOpFOrdLessThanEqual:
case SpvOpFOrdGreaterThanEqual: {
bool swap;
nir_alu_type src_alu_type = nir_get_nir_type_for_glsl_type(vtn_src[0]->type);
nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(type);
nir_op op = vtn_nir_alu_op_for_spirv_opcode(opcode, &swap, src_alu_type, dst_alu_type);
if (swap) {
nir_ssa_def *tmp = src[0];
src[0] = src[1];
src[1] = tmp;
}
val->ssa->def =
nir_iand(&b->nb,
nir_build_alu(&b->nb, op, src[0], src[1], NULL, NULL),
nir_iand(&b->nb,
nir_feq(&b->nb, src[0], src[0]),
nir_feq(&b->nb, src[1], src[1])));
break;
}
default: {
bool swap;
nir_alu_type src_alu_type = nir_get_nir_type_for_glsl_type(vtn_src[0]->type);
nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(type);
nir_op op = vtn_nir_alu_op_for_spirv_opcode(opcode, &swap, src_alu_type, dst_alu_type);
if (swap) {
nir_ssa_def *tmp = src[0];
src[0] = src[1];
src[1] = tmp;
}
val->ssa->def = nir_build_alu(&b->nb, op, src[0], src[1], src[2], src[3]);
break;
} /* default */
}
b->nb.exact = false;
}
static struct vtn_ssa_value *
matrix_multiply(struct vtn_builder *b,
struct vtn_ssa_value *_src0, struct vtn_ssa_value *_src1)
{
struct vtn_ssa_value *src0 = wrap_matrix(b, _src0);
struct vtn_ssa_value *src1 = wrap_matrix(b, _src1);
struct vtn_ssa_value *src0_transpose = wrap_matrix(b, _src0->transposed);
struct vtn_ssa_value *src1_transpose = wrap_matrix(b, _src1->transposed);
unsigned src0_rows = glsl_get_vector_elements(src0->type);
unsigned src0_columns = glsl_get_matrix_columns(src0->type);
unsigned src1_columns = glsl_get_matrix_columns(src1->type);
const struct glsl_type *dest_type;
if (src1_columns > 1) {
dest_type = glsl_matrix_type(glsl_get_base_type(src0->type),
src0_rows, src1_columns);
} else {
dest_type = glsl_vector_type(glsl_get_base_type(src0->type), src0_rows);
}
struct vtn_ssa_value *dest = vtn_create_ssa_value(b, dest_type);
dest = wrap_matrix(b, dest);
bool transpose_result = false;
if (src0_transpose && src1_transpose) {
/* transpose(A) * transpose(B) = transpose(B * A) */
src1 = src0_transpose;
src0 = src1_transpose;
src0_transpose = NULL;
src1_transpose = NULL;
transpose_result = true;
}
if (src0_transpose && !src1_transpose &&
glsl_get_base_type(src0->type) == GLSL_TYPE_FLOAT) {
/* We already have the rows of src0 and the columns of src1 available,
* so we can just take the dot product of each row with each column to
* get the result.
*/
for (unsigned i = 0; i < src1_columns; i++) {
nir_ssa_def *vec_src[4];
for (unsigned j = 0; j < src0_rows; j++) {
vec_src[j] = nir_fdot(&b->nb, src0_transpose->elems[j]->def,
src1->elems[i]->def);
}
dest->elems[i]->def = nir_vec(&b->nb, vec_src, src0_rows);
}
} else {
/* We don't handle the case where src1 is transposed but not src0, since
* the general case only uses individual components of src1 so the
* optimizer should chew through the transpose we emitted for src1.
*/
for (unsigned i = 0; i < src1_columns; i++) {
/* dest[i] = sum(src0[j] * src1[i][j] for all j) */
dest->elems[i]->def =
nir_fmul(&b->nb, src0->elems[0]->def,
nir_channel(&b->nb, src1->elems[i]->def, 0));
for (unsigned j = 1; j < src0_columns; j++) {
dest->elems[i]->def =
nir_fadd(&b->nb, dest->elems[i]->def,
nir_fmul(&b->nb, src0->elems[j]->def,
nir_channel(&b->nb, src1->elems[i]->def, j)));
}
}
}
dest = unwrap_matrix(dest);
if (transpose_result)
dest = vtn_ssa_transpose(b, dest);
return dest;
}
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;
}
}
