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 struct vtn_ssa_value *
matrix_inverse(struct vtn_builder *b, struct vtn_ssa_value *src)
{
nir_ssa_def *adj_col[4];
unsigned size = glsl_get_vector_elements(src->type);
/* Build up an adjugate matrix */
for (unsigned c = 0; c < size; c++) {
nir_ssa_def *elem[4];
for (unsigned r = 0; r < size; r++) {
elem[r] = build_mat_subdet(&b->nb, src, size, c, r);
if ((r + c) % 2)
elem[r] = nir_fneg(&b->nb, elem[r]);
}
adj_col[c] = nir_vec(&b->nb, elem, size);
}
nir_ssa_def *det_inv = nir_frcp(&b->nb, build_mat_det(b, src));
struct vtn_ssa_value *val = vtn_create_ssa_value(b, src->type);
for (unsigned i = 0; i < size; i++)
val->elems[i]->def = nir_fmul(&b->nb, adj_col[i], det_inv);
return val;
}
static void
handle_glsl450_interpolation(struct vtn_builder *b, enum GLSLstd450 opcode,
const uint32_t *w, unsigned count)
{
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);
nir_intrinsic_op op;
switch (opcode) {
case GLSLstd450InterpolateAtCentroid:
op = nir_intrinsic_interp_var_at_centroid;
break;
case GLSLstd450InterpolateAtSample:
op = nir_intrinsic_interp_var_at_sample;
break;
case GLSLstd450InterpolateAtOffset:
op = nir_intrinsic_interp_var_at_offset;
break;
default:
unreachable("Invalid opcode");
}
nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->nb.shader, op);
nir_deref_var *deref = vtn_nir_deref(b, w[5]);
intrin->variables[0] =
nir_deref_as_var(nir_copy_deref(intrin, &deref->deref));
switch (opcode) {
case GLSLstd450InterpolateAtCentroid:
break;
case GLSLstd450InterpolateAtSample:
case GLSLstd450InterpolateAtOffset:
intrin->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
break;
default:
unreachable("Invalid opcode");
}
intrin->num_components = glsl_get_vector_elements(dest_type);
nir_ssa_dest_init(&intrin->instr, &intrin->dest,
glsl_get_vector_elements(dest_type),
glsl_get_bit_size(dest_type), NULL);
val->ssa->def = &intrin->dest.ssa;
nir_builder_instr_insert(&b->nb, &intrin->instr);
}
static struct vtn_ssa_value *
mat_times_scalar(struct vtn_builder *b,
struct vtn_ssa_value *mat,
nir_ssa_def *scalar)
{
struct vtn_ssa_value *dest = vtn_create_ssa_value(b, mat->type);
for (unsigned i = 0; i < glsl_get_matrix_columns(mat->type); i++) {
if (glsl_get_base_type(mat->type) == GLSL_TYPE_FLOAT)
dest->elems[i]->def = nir_fmul(&b->nb, mat->elems[i]->def, scalar);
else
dest->elems[i]->def = nir_imul(&b->nb, mat->elems[i]->def, scalar);
}
return dest;
}
struct vtn_ssa_value *
vtn_local_load(struct vtn_builder *b, nir_deref_var *src)
{
nir_deref *src_tail = get_deref_tail(src);
struct vtn_ssa_value *val = vtn_create_ssa_value(b, src_tail->type);
_vtn_local_load_store(b, true, src, src_tail, val);
if (src_tail->child) {
nir_deref_array *vec_deref = nir_deref_as_array(src_tail->child);
assert(vec_deref->deref.child == NULL);
val->type = vec_deref->deref.type;
if (vec_deref->deref_array_type == nir_deref_array_type_direct)
val->def = vtn_vector_extract(b, val->def, vec_deref->base_offset);
else
val->def = vtn_vector_extract_dynamic(b, val->def,
vec_deref->indirect.ssa);
}
return val;
}
void
vtn_local_store(struct vtn_builder *b, struct vtn_ssa_value *src,
nir_deref_var *dest)
{
nir_deref *dest_tail = get_deref_tail(dest);
if (dest_tail->child) {
struct vtn_ssa_value *val = vtn_create_ssa_value(b, dest_tail->type);
_vtn_local_load_store(b, true, dest, dest_tail, val);
nir_deref_array *deref = nir_deref_as_array(dest_tail->child);
assert(deref->deref.child == NULL);
if (deref->deref_array_type == nir_deref_array_type_direct)
val->def = vtn_vector_insert(b, val->def, src->def,
deref->base_offset);
else
val->def = vtn_vector_insert_dynamic(b, val->def, src->def,
deref->indirect.ssa);
_vtn_local_load_store(b, false, dest, dest_tail, val);
} else {
_vtn_local_load_store(b, false, dest, dest_tail, src);
}
}
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;
}
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 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;
}
}
void
vtn_handle_subgroup(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);
val->ssa = vtn_create_ssa_value(b, val->type->type);
switch (opcode) {
case SpvOpGroupNonUniformElect: {
vtn_fail_if(val->type->type != glsl_bool_type(),
"OpGroupNonUniformElect must return a Bool");
nir_intrinsic_instr *elect =
nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_elect);
nir_ssa_dest_init_for_type(&elect->instr, &elect->dest,
val->type->type, NULL);
nir_builder_instr_insert(&b->nb, &elect->instr);
val->ssa->def = &elect->dest.ssa;
break;
}
case SpvOpGroupNonUniformBallot: {
vtn_fail_if(val->type->type != glsl_vector_type(GLSL_TYPE_UINT, 4),
"OpGroupNonUniformBallot must return a uvec4");
nir_intrinsic_instr *ballot =
nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_ballot);
ballot->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
nir_ssa_dest_init(&ballot->instr, &ballot->dest, 4, 32, NULL);
ballot->num_components = 4;
nir_builder_instr_insert(&b->nb, &ballot->instr);
val->ssa->def = &ballot->dest.ssa;
break;
}
case SpvOpGroupNonUniformInverseBallot: {
/* This one is just a BallotBitfieldExtract with subgroup invocation.
* We could add a NIR intrinsic but it's easier to just lower it on the
* spot.
*/
nir_intrinsic_instr *intrin =
nir_intrinsic_instr_create(b->nb.shader,
nir_intrinsic_ballot_bitfield_extract);
intrin->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
intrin->src[1] = nir_src_for_ssa(nir_load_subgroup_invocation(&b->nb));
nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
val->type->type, NULL);
nir_builder_instr_insert(&b->nb, &intrin->instr);
val->ssa->def = &intrin->dest.ssa;
break;
}
case SpvOpGroupNonUniformBallotBitExtract:
case SpvOpGroupNonUniformBallotBitCount:
case SpvOpGroupNonUniformBallotFindLSB:
case SpvOpGroupNonUniformBallotFindMSB: {
nir_ssa_def *src0, *src1 = NULL;
nir_intrinsic_op op;
switch (opcode) {
case SpvOpGroupNonUniformBallotBitExtract:
op = nir_intrinsic_ballot_bitfield_extract;
src0 = vtn_ssa_value(b, w[4])->def;
src1 = vtn_ssa_value(b, w[5])->def;
break;
case SpvOpGroupNonUniformBallotBitCount:
switch ((SpvGroupOperation)w[4]) {
case SpvGroupOperationReduce:
op = nir_intrinsic_ballot_bit_count_reduce;
break;
case SpvGroupOperationInclusiveScan:
op = nir_intrinsic_ballot_bit_count_inclusive;
break;
case SpvGroupOperationExclusiveScan:
op = nir_intrinsic_ballot_bit_count_exclusive;
break;
default:
unreachable("Invalid group operation");
}
src0 = vtn_ssa_value(b, w[5])->def;
break;
case SpvOpGroupNonUniformBallotFindLSB:
op = nir_intrinsic_ballot_find_lsb;
src0 = vtn_ssa_value(b, w[4])->def;
break;
case SpvOpGroupNonUniformBallotFindMSB:
op = nir_intrinsic_ballot_find_msb;
src0 = vtn_ssa_value(b, w[4])->def;
break;
default:
unreachable("Unhandled opcode");
}
nir_intrinsic_instr *intrin =
nir_intrinsic_instr_create(b->nb.shader, op);
intrin->src[0] = nir_src_for_ssa(src0);
if (src1)
intrin->src[1] = nir_src_for_ssa(src1);
nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
val->type->type, NULL);
nir_builder_instr_insert(&b->nb, &intrin->instr);
val->ssa->def = &intrin->dest.ssa;
break;
}
case SpvOpGroupNonUniformBroadcastFirst:
vtn_build_subgroup_instr(b, nir_intrinsic_read_first_invocation,
val->ssa, vtn_ssa_value(b, w[4]), NULL, 0, 0);
break;
case SpvOpGroupNonUniformBroadcast:
vtn_build_subgroup_instr(b, nir_intrinsic_read_invocation,
val->ssa, vtn_ssa_value(b, w[4]),
vtn_ssa_value(b, w[5])->def, 0, 0);
break;
case SpvOpGroupNonUniformAll:
case SpvOpGroupNonUniformAny:
case SpvOpGroupNonUniformAllEqual: {
vtn_fail_if(val->type->type != glsl_bool_type(),
"OpGroupNonUniform(All|Any|AllEqual) must return a bool");
nir_intrinsic_op op;
switch (opcode) {
case SpvOpGroupNonUniformAll:
op = nir_intrinsic_vote_all;
break;
case SpvOpGroupNonUniformAny:
op = nir_intrinsic_vote_any;
break;
case SpvOpGroupNonUniformAllEqual: {
switch (glsl_get_base_type(val->type->type)) {
case GLSL_TYPE_FLOAT:
case GLSL_TYPE_DOUBLE:
op = nir_intrinsic_vote_feq;
break;
case GLSL_TYPE_UINT:
case GLSL_TYPE_INT:
case GLSL_TYPE_UINT64:
case GLSL_TYPE_INT64:
case GLSL_TYPE_BOOL:
op = nir_intrinsic_vote_ieq;
break;
default:
unreachable("Unhandled type");
}
break;
}
default:
unreachable("Unhandled opcode");
}
nir_ssa_def *src0 = vtn_ssa_value(b, w[4])->def;
nir_intrinsic_instr *intrin =
nir_intrinsic_instr_create(b->nb.shader, op);
intrin->num_components = src0->num_components;
intrin->src[0] = nir_src_for_ssa(src0);
nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
val->type->type, NULL);
nir_builder_instr_insert(&b->nb, &intrin->instr);
val->ssa->def = &intrin->dest.ssa;
break;
}
case SpvOpGroupNonUniformShuffle:
case SpvOpGroupNonUniformShuffleXor:
case SpvOpGroupNonUniformShuffleUp:
case SpvOpGroupNonUniformShuffleDown: {
nir_intrinsic_op op;
switch (opcode) {
case SpvOpGroupNonUniformShuffle:
op = nir_intrinsic_shuffle;
break;
case SpvOpGroupNonUniformShuffleXor:
op = nir_intrinsic_shuffle_xor;
break;
case SpvOpGroupNonUniformShuffleUp:
op = nir_intrinsic_shuffle_up;
break;
case SpvOpGroupNonUniformShuffleDown:
op = nir_intrinsic_shuffle_down;
break;
default:
unreachable("Invalid opcode");
}
vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[4]),
vtn_ssa_value(b, w[5])->def, 0, 0);
break;
}
case SpvOpGroupNonUniformQuadBroadcast:
vtn_build_subgroup_instr(b, nir_intrinsic_quad_broadcast,
val->ssa, vtn_ssa_value(b, w[4]),
vtn_ssa_value(b, w[5])->def, 0, 0);
break;
case SpvOpGroupNonUniformQuadSwap: {
unsigned direction = vtn_constant_uint(b, w[5]);
nir_intrinsic_op op;
switch (direction) {
case 0:
op = nir_intrinsic_quad_swap_horizontal;
break;
case 1:
op = nir_intrinsic_quad_swap_vertical;
break;
case 2:
op = nir_intrinsic_quad_swap_diagonal;
break;
default:
vtn_fail("Invalid constant value in OpGroupNonUniformQuadSwap");
}
vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[4]),
NULL, 0, 0);
break;
}
case SpvOpGroupNonUniformIAdd:
case SpvOpGroupNonUniformFAdd:
case SpvOpGroupNonUniformIMul:
case SpvOpGroupNonUniformFMul:
case SpvOpGroupNonUniformSMin:
case SpvOpGroupNonUniformUMin:
case SpvOpGroupNonUniformFMin:
case SpvOpGroupNonUniformSMax:
case SpvOpGroupNonUniformUMax:
case SpvOpGroupNonUniformFMax:
case SpvOpGroupNonUniformBitwiseAnd:
case SpvOpGroupNonUniformBitwiseOr:
case SpvOpGroupNonUniformBitwiseXor:
case SpvOpGroupNonUniformLogicalAnd:
case SpvOpGroupNonUniformLogicalOr:
case SpvOpGroupNonUniformLogicalXor: {
nir_op reduction_op;
switch (opcode) {
case SpvOpGroupNonUniformIAdd:
reduction_op = nir_op_iadd;
break;
case SpvOpGroupNonUniformFAdd:
reduction_op = nir_op_fadd;
break;
case SpvOpGroupNonUniformIMul:
reduction_op = nir_op_imul;
break;
case SpvOpGroupNonUniformFMul:
reduction_op = nir_op_fmul;
break;
case SpvOpGroupNonUniformSMin:
reduction_op = nir_op_imin;
break;
case SpvOpGroupNonUniformUMin:
reduction_op = nir_op_umin;
break;
case SpvOpGroupNonUniformFMin:
reduction_op = nir_op_fmin;
break;
case SpvOpGroupNonUniformSMax:
reduction_op = nir_op_imax;
break;
case SpvOpGroupNonUniformUMax:
reduction_op = nir_op_umax;
break;
case SpvOpGroupNonUniformFMax:
reduction_op = nir_op_fmax;
break;
case SpvOpGroupNonUniformBitwiseAnd:
case SpvOpGroupNonUniformLogicalAnd:
reduction_op = nir_op_iand;
break;
case SpvOpGroupNonUniformBitwiseOr:
case SpvOpGroupNonUniformLogicalOr:
reduction_op = nir_op_ior;
break;
case SpvOpGroupNonUniformBitwiseXor:
case SpvOpGroupNonUniformLogicalXor:
reduction_op = nir_op_ixor;
break;
default:
unreachable("Invalid reduction operation");
}
nir_intrinsic_op op;
unsigned cluster_size = 0;
switch ((SpvGroupOperation)w[4]) {
case SpvGroupOperationReduce:
op = nir_intrinsic_reduce;
break;
case SpvGroupOperationInclusiveScan:
op = nir_intrinsic_inclusive_scan;
break;
case SpvGroupOperationExclusiveScan:
op = nir_intrinsic_exclusive_scan;
break;
case SpvGroupOperationClusteredReduce:
op = nir_intrinsic_reduce;
assert(count == 7);
cluster_size = vtn_constant_uint(b, w[6]);
break;
default:
unreachable("Invalid group operation");
}
vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[5]),
NULL, reduction_op, cluster_size);
break;
}
default:
unreachable("Invalid SPIR-V opcode");
}
}
static void
_vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,
nir_ssa_def *index, nir_ssa_def *offset,
struct vtn_access_chain *chain, unsigned chain_idx,
struct vtn_type *type, struct vtn_ssa_value **inout)
{
if (chain && chain_idx >= chain->length)
chain = NULL;
if (load && chain == NULL && *inout == NULL)
*inout = vtn_create_ssa_value(b, type->type);
enum glsl_base_type base_type = glsl_get_base_type(type->type);
switch (base_type) {
case GLSL_TYPE_UINT:
case GLSL_TYPE_INT:
case GLSL_TYPE_FLOAT:
case GLSL_TYPE_BOOL:
/* This is where things get interesting. At this point, we've hit
* a vector, a scalar, or a matrix.
*/
if (glsl_type_is_matrix(type->type)) {
if (chain == NULL) {
/* Loading the whole matrix */
struct vtn_ssa_value *transpose;
unsigned num_ops, vec_width;
if (type->row_major) {
num_ops = glsl_get_vector_elements(type->type);
vec_width = glsl_get_matrix_columns(type->type);
if (load) {
const struct glsl_type *transpose_type =
glsl_matrix_type(base_type, vec_width, num_ops);
*inout = vtn_create_ssa_value(b, transpose_type);
} else {
transpose = vtn_ssa_transpose(b, *inout);
inout = &transpose;
}
} else {
num_ops = glsl_get_matrix_columns(type->type);
vec_width = glsl_get_vector_elements(type->type);
}
for (unsigned i = 0; i < num_ops; i++) {
nir_ssa_def *elem_offset =
nir_iadd(&b->nb, offset,
nir_imm_int(&b->nb, i * type->stride));
_vtn_load_store_tail(b, op, load, index, elem_offset,
&(*inout)->elems[i],
glsl_vector_type(base_type, vec_width));
}
if (load && type->row_major)
*inout = vtn_ssa_transpose(b, *inout);
} else if (type->row_major) {
/* Row-major but with an access chiain. */
nir_ssa_def *col_offset =
vtn_access_link_as_ssa(b, chain->link[chain_idx],
type->array_element->stride);
offset = nir_iadd(&b->nb, offset, col_offset);
if (chain_idx + 1 < chain->length) {
/* Picking off a single element */
nir_ssa_def *row_offset =
vtn_access_link_as_ssa(b, chain->link[chain_idx + 1],
type->stride);
offset = nir_iadd(&b->nb, offset, row_offset);
if (load)
*inout = vtn_create_ssa_value(b, glsl_scalar_type(base_type));
_vtn_load_store_tail(b, op, load, index, offset, inout,
glsl_scalar_type(base_type));
} else {
/* Grabbing a column; picking one element off each row */
unsigned num_comps = glsl_get_vector_elements(type->type);
const struct glsl_type *column_type =
glsl_get_column_type(type->type);
nir_ssa_def *comps[4];
for (unsigned i = 0; i < num_comps; i++) {
nir_ssa_def *elem_offset =
nir_iadd(&b->nb, offset,
nir_imm_int(&b->nb, i * type->stride));
struct vtn_ssa_value *comp, temp_val;
if (!load) {
temp_val.def = nir_channel(&b->nb, (*inout)->def, i);
temp_val.type = glsl_scalar_type(base_type);
}
comp = &temp_val;
_vtn_load_store_tail(b, op, load, index, elem_offset,
&comp, glsl_scalar_type(base_type));
comps[i] = comp->def;
}
if (load) {
if (*inout == NULL)
*inout = vtn_create_ssa_value(b, column_type);
(*inout)->def = nir_vec(&b->nb, comps, num_comps);
}
}
} else {
/* Column-major with a deref. Fall through to array case. */
nir_ssa_def *col_offset =
vtn_access_link_as_ssa(b, chain->link[chain_idx], type->stride);
offset = nir_iadd(&b->nb, offset, col_offset);
_vtn_block_load_store(b, op, load, index, offset,
chain, chain_idx + 1,
type->array_element, inout);
}
} else if (chain == NULL) {
/* Single whole vector */
assert(glsl_type_is_vector_or_scalar(type->type));
_vtn_load_store_tail(b, op, load, index, offset, inout, type->type);
} else {
/* Single component of a vector. Fall through to array case. */
nir_ssa_def *elem_offset =
vtn_access_link_as_ssa(b, chain->link[chain_idx], type->stride);
offset = nir_iadd(&b->nb, offset, elem_offset);
_vtn_block_load_store(b, op, load, index, offset, NULL, 0,
type->array_element, inout);
}
return;
case GLSL_TYPE_ARRAY: {
unsigned elems = glsl_get_length(type->type);
for (unsigned i = 0; i < elems; i++) {
nir_ssa_def *elem_off =
nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * type->stride));
_vtn_block_load_store(b, op, load, index, elem_off, NULL, 0,
type->array_element, &(*inout)->elems[i]);
}
return;
}
case GLSL_TYPE_STRUCT: {
unsigned elems = glsl_get_length(type->type);
for (unsigned i = 0; i < elems; i++) {
nir_ssa_def *elem_off =
nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, type->offsets[i]));
_vtn_block_load_store(b, op, load, index, elem_off, NULL, 0,
type->members[i], &(*inout)->elems[i]);
}
return;
}
default:
unreachable("Invalid block member type");
}
}