OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

spirv/alu: Use vtn_push_ssa_value

Browse Source 
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5278>
This commit is contained in:
Jason Ekstrand
2020-05-29 15:57:42 -05:00
parent dbb4a24413
commit 7560ed279f
1 changed files with 61 additions and 60 deletions
Download Patch File Download Diff File Expand all files Collapse all files

119
src/compiler/spirv/vtn_alu.c
View File

@@ -153,48 +153,46 @@ mat_times_scalar(struct vtn_builder *b,
return dest; return dest;
} }
static void static struct vtn_ssa_value *
vtn_handle_matrix_alu(struct vtn_builder *b, SpvOp opcode, vtn_handle_matrix_alu(struct vtn_builder *b, SpvOp opcode,
struct vtn_value *dest,
struct vtn_ssa_value *src0, struct vtn_ssa_value *src1) struct vtn_ssa_value *src0, struct vtn_ssa_value *src1)
{ {
switch (opcode) { switch (opcode) {
case SpvOpFNegate: { case SpvOpFNegate: {
dest->ssa = vtn_create_ssa_value(b, src0->type); struct vtn_ssa_value *dest = vtn_create_ssa_value(b, src0->type);
unsigned cols = glsl_get_matrix_columns(src0->type); unsigned cols = glsl_get_matrix_columns(src0->type);
for (unsigned i = 0; i < cols; i++) for (unsigned i = 0; i < cols; i++)
dest->ssa->elems[i]->def = nir_fneg(&b->nb, src0->elems[i]->def); dest->elems[i]->def = nir_fneg(&b->nb, src0->elems[i]->def);
break; return dest;
} }
case SpvOpFAdd: { case SpvOpFAdd: {
dest->ssa = vtn_create_ssa_value(b, src0->type); struct vtn_ssa_value *dest = vtn_create_ssa_value(b, src0->type);
unsigned cols = glsl_get_matrix_columns(src0->type); unsigned cols = glsl_get_matrix_columns(src0->type);
for (unsigned i = 0; i < cols; i++) for (unsigned i = 0; i < cols; i++)
dest->ssa->elems[i]->def = dest->elems[i]->def =
nir_fadd(&b->nb, src0->elems[i]->def, src1->elems[i]->def); nir_fadd(&b->nb, src0->elems[i]->def, src1->elems[i]->def);
break; return dest;
} }
case SpvOpFSub: { case SpvOpFSub: {
dest->ssa = vtn_create_ssa_value(b, src0->type); struct vtn_ssa_value *dest = vtn_create_ssa_value(b, src0->type);
unsigned cols = glsl_get_matrix_columns(src0->type); unsigned cols = glsl_get_matrix_columns(src0->type);
for (unsigned i = 0; i < cols; i++) for (unsigned i = 0; i < cols; i++)
dest->ssa->elems[i]->def = dest->elems[i]->def =
nir_fsub(&b->nb, src0->elems[i]->def, src1->elems[i]->def); nir_fsub(&b->nb, src0->elems[i]->def, src1->elems[i]->def);
break; return dest;
} }
case SpvOpTranspose: case SpvOpTranspose:
dest->ssa = vtn_ssa_transpose(b, src0); return vtn_ssa_transpose(b, src0);
break;
case SpvOpMatrixTimesScalar: case SpvOpMatrixTimesScalar:
if (src0->transposed) { if (src0->transposed) {
dest->ssa = vtn_ssa_transpose(b, mat_times_scalar(b, src0->transposed, return vtn_ssa_transpose(b, mat_times_scalar(b, src0->transposed,
src1->def)); src1->def));
} else { } else {
dest->ssa = mat_times_scalar(b, src0, src1->def); return mat_times_scalar(b, src0, src1->def);
} }
break; break;
@@ -202,9 +200,9 @@ vtn_handle_matrix_alu(struct vtn_builder *b, SpvOp opcode,
case SpvOpMatrixTimesVector: case SpvOpMatrixTimesVector:
case SpvOpMatrixTimesMatrix: case SpvOpMatrixTimesMatrix:
if (opcode == SpvOpVectorTimesMatrix) { if (opcode == SpvOpVectorTimesMatrix) {
dest->ssa = matrix_multiply(b, vtn_ssa_transpose(b, src1), src0); return matrix_multiply(b, vtn_ssa_transpose(b, src1), src0);
} else { } else {
dest->ssa = matrix_multiply(b, src0, src1); return matrix_multiply(b, src0, src1);
} }
break; break;
@@ -414,10 +412,10 @@ void
vtn_handle_alu(struct vtn_builder *b, SpvOp opcode, vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count) const uint32_t *w, unsigned count)
{ {
struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa); struct vtn_value *dest_val = vtn_untyped_value(b, w[2]);
const struct glsl_type *type = vtn_get_type(b, w[1])->type; const struct glsl_type *dest_type = vtn_get_type(b, w[1])->type;
vtn_foreach_decoration(b, val, handle_no_contraction, NULL); vtn_foreach_decoration(b, dest_val, handle_no_contraction, NULL);
/* Collect the various SSA sources */ /* Collect the various SSA sources */
const unsigned num_inputs = count - 3; const unsigned num_inputs = count - 3;
@@ -427,12 +425,13 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
if (glsl_type_is_matrix(vtn_src[0]->type) || if (glsl_type_is_matrix(vtn_src[0]->type) ||
(num_inputs >= 2 && glsl_type_is_matrix(vtn_src[1]->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]); vtn_push_ssa_value(b, w[2],
vtn_handle_matrix_alu(b, opcode, vtn_src[0], vtn_src[1]));
b->nb.exact = b->exact; b->nb.exact = b->exact;
return; return;
} }
val->ssa = vtn_create_ssa_value(b, type); struct vtn_ssa_value *dest = vtn_create_ssa_value(b, dest_type);
nir_ssa_def *src[4] = { NULL, }; nir_ssa_def *src[4] = { NULL, };
for (unsigned i = 0; i < num_inputs; i++) { for (unsigned i = 0; i < num_inputs; i++) {
vtn_assert(glsl_type_is_vector_or_scalar(vtn_src[i]->type)); vtn_assert(glsl_type_is_vector_or_scalar(vtn_src[i]->type));
@@ -441,81 +440,81 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
switch (opcode) { switch (opcode) {
case SpvOpAny: case SpvOpAny:
val->ssa->def = nir_bany(&b->nb, src[0]); dest->def = nir_bany(&b->nb, src[0]);
break; break;
case SpvOpAll: case SpvOpAll:
val->ssa->def = nir_ball(&b->nb, src[0]); dest->def = nir_ball(&b->nb, src[0]);
break; break;
case SpvOpOuterProduct: { case SpvOpOuterProduct: {
for (unsigned i = 0; i < src[1]->num_components; i++) { for (unsigned i = 0; i < src[1]->num_components; i++) {
val->ssa->elems[i]->def = dest->elems[i]->def =
nir_fmul(&b->nb, src[0], nir_channel(&b->nb, src[1], i)); nir_fmul(&b->nb, src[0], nir_channel(&b->nb, src[1], i));
} }
break; break;
} }
case SpvOpDot: case SpvOpDot:
val->ssa->def = nir_fdot(&b->nb, src[0], src[1]); dest->def = nir_fdot(&b->nb, src[0], src[1]);
break; break;
case SpvOpIAddCarry: case SpvOpIAddCarry:
vtn_assert(glsl_type_is_struct_or_ifc(val->ssa->type)); vtn_assert(glsl_type_is_struct_or_ifc(dest_type));
val->ssa->elems[0]->def = nir_iadd(&b->nb, src[0], src[1]); dest->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]); dest->elems[1]->def = nir_uadd_carry(&b->nb, src[0], src[1]);
break; break;
case SpvOpISubBorrow: case SpvOpISubBorrow:
vtn_assert(glsl_type_is_struct_or_ifc(val->ssa->type)); vtn_assert(glsl_type_is_struct_or_ifc(dest_type));
val->ssa->elems[0]->def = nir_isub(&b->nb, src[0], src[1]); dest->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]); dest->elems[1]->def = nir_usub_borrow(&b->nb, src[0], src[1]);
break; break;
case SpvOpUMulExtended: { case SpvOpUMulExtended: {
vtn_assert(glsl_type_is_struct_or_ifc(val->ssa->type)); vtn_assert(glsl_type_is_struct_or_ifc(dest_type));
nir_ssa_def *umul = nir_umul_2x32_64(&b->nb, src[0], src[1]); nir_ssa_def *umul = nir_umul_2x32_64(&b->nb, src[0], src[1]);
val->ssa->elems[0]->def = nir_unpack_64_2x32_split_x(&b->nb, umul); dest->elems[0]->def = nir_unpack_64_2x32_split_x(&b->nb, umul);
val->ssa->elems[1]->def = nir_unpack_64_2x32_split_y(&b->nb, umul); dest->elems[1]->def = nir_unpack_64_2x32_split_y(&b->nb, umul);
break; break;
} }
case SpvOpSMulExtended: { case SpvOpSMulExtended: {
vtn_assert(glsl_type_is_struct_or_ifc(val->ssa->type)); vtn_assert(glsl_type_is_struct_or_ifc(dest_type));
nir_ssa_def *smul = nir_imul_2x32_64(&b->nb, src[0], src[1]); nir_ssa_def *smul = nir_imul_2x32_64(&b->nb, src[0], src[1]);
val->ssa->elems[0]->def = nir_unpack_64_2x32_split_x(&b->nb, smul); dest->elems[0]->def = nir_unpack_64_2x32_split_x(&b->nb, smul);
val->ssa->elems[1]->def = nir_unpack_64_2x32_split_y(&b->nb, smul); dest->elems[1]->def = nir_unpack_64_2x32_split_y(&b->nb, smul);
break; break;
} }
case SpvOpFwidth: case SpvOpFwidth:
val->ssa->def = nir_fadd(&b->nb, dest->def = nir_fadd(&b->nb,
nir_fabs(&b->nb, nir_fddx(&b->nb, src[0])), nir_fabs(&b->nb, nir_fddx(&b->nb, src[0])),
nir_fabs(&b->nb, nir_fddy(&b->nb, src[0]))); nir_fabs(&b->nb, nir_fddy(&b->nb, src[0])));
break; break;
case SpvOpFwidthFine: case SpvOpFwidthFine:
val->ssa->def = nir_fadd(&b->nb, dest->def = nir_fadd(&b->nb,
nir_fabs(&b->nb, nir_fddx_fine(&b->nb, src[0])), nir_fabs(&b->nb, nir_fddx_fine(&b->nb, src[0])),
nir_fabs(&b->nb, nir_fddy_fine(&b->nb, src[0]))); nir_fabs(&b->nb, nir_fddy_fine(&b->nb, src[0])));
break; break;
case SpvOpFwidthCoarse: case SpvOpFwidthCoarse:
val->ssa->def = nir_fadd(&b->nb, dest->def = nir_fadd(&b->nb,
nir_fabs(&b->nb, nir_fddx_coarse(&b->nb, src[0])), nir_fabs(&b->nb, nir_fddx_coarse(&b->nb, src[0])),
nir_fabs(&b->nb, nir_fddy_coarse(&b->nb, src[0]))); nir_fabs(&b->nb, nir_fddy_coarse(&b->nb, src[0])));
break; break;
case SpvOpVectorTimesScalar: case SpvOpVectorTimesScalar:
/* The builder will take care of splatting for us. */ /* The builder will take care of splatting for us. */
val->ssa->def = nir_fmul(&b->nb, src[0], src[1]); dest->def = nir_fmul(&b->nb, src[0], src[1]);
break; break;
case SpvOpIsNan: case SpvOpIsNan:
val->ssa->def = nir_fne(&b->nb, src[0], src[0]); dest->def = nir_fne(&b->nb, src[0], src[0]);
break; break;
case SpvOpIsInf: { case SpvOpIsInf: {
nir_ssa_def *inf = nir_imm_floatN_t(&b->nb, INFINITY, src[0]->bit_size); nir_ssa_def *inf = nir_imm_floatN_t(&b->nb, INFINITY, src[0]->bit_size);
val->ssa->def = nir_ieq(&b->nb, nir_fabs(&b->nb, src[0]), inf); dest->def = nir_ieq(&b->nb, nir_fabs(&b->nb, src[0]), inf);
break; break;
} }
@@ -527,7 +526,7 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
case SpvOpFUnordGreaterThanEqual: { case SpvOpFUnordGreaterThanEqual: {
bool swap; bool swap;
unsigned src_bit_size = glsl_get_bit_size(vtn_src[0]->type); unsigned src_bit_size = glsl_get_bit_size(vtn_src[0]->type);
unsigned dst_bit_size = glsl_get_bit_size(type); unsigned dst_bit_size = glsl_get_bit_size(dest_type);
nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap, nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
src_bit_size, dst_bit_size); src_bit_size, dst_bit_size);
@@ -537,7 +536,7 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
src[1] = tmp; src[1] = tmp;
} }
val->ssa->def = dest->def =
nir_ior(&b->nb, nir_ior(&b->nb,
nir_build_alu(&b->nb, op, src[0], src[1], NULL, NULL), nir_build_alu(&b->nb, op, src[0], src[1], NULL, NULL),
nir_ior(&b->nb, nir_ior(&b->nb,
@@ -553,13 +552,13 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
*/ */
bool swap; bool swap;
unsigned src_bit_size = glsl_get_bit_size(vtn_src[0]->type); unsigned src_bit_size = glsl_get_bit_size(vtn_src[0]->type);
unsigned dst_bit_size = glsl_get_bit_size(type); unsigned dst_bit_size = glsl_get_bit_size(dest_type);
nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap, nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
src_bit_size, dst_bit_size); src_bit_size, dst_bit_size);
assert(!swap); assert(!swap);
val->ssa->def = dest->def =
nir_iand(&b->nb, nir_iand(&b->nb,
nir_build_alu(&b->nb, op, src[0], src[1], NULL, NULL), nir_build_alu(&b->nb, op, src[0], src[1], NULL, NULL),
nir_iand(&b->nb, nir_iand(&b->nb,
@@ -570,13 +569,13 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
case SpvOpFConvert: { case SpvOpFConvert: {
nir_alu_type src_alu_type = nir_get_nir_type_for_glsl_type(vtn_src[0]->type); 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_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(dest_type);
nir_rounding_mode rounding_mode = nir_rounding_mode_undef; nir_rounding_mode rounding_mode = nir_rounding_mode_undef;
vtn_foreach_decoration(b, val, handle_rounding_mode, &rounding_mode); vtn_foreach_decoration(b, dest_val, handle_rounding_mode, &rounding_mode);
nir_op op = nir_type_conversion_op(src_alu_type, dst_alu_type, rounding_mode); nir_op op = nir_type_conversion_op(src_alu_type, dst_alu_type, rounding_mode);
val->ssa->def = nir_build_alu(&b->nb, op, src[0], src[1], NULL, NULL); dest->def = nir_build_alu(&b->nb, op, src[0], src[1], NULL, NULL);
break; break;
} }
@@ -588,7 +587,7 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
case SpvOpShiftRightLogical: { case SpvOpShiftRightLogical: {
bool swap; bool swap;
unsigned src0_bit_size = glsl_get_bit_size(vtn_src[0]->type); unsigned src0_bit_size = glsl_get_bit_size(vtn_src[0]->type);
unsigned dst_bit_size = glsl_get_bit_size(type); unsigned dst_bit_size = glsl_get_bit_size(dest_type);
nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap, nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
src0_bit_size, dst_bit_size); src0_bit_size, dst_bit_size);
@@ -611,17 +610,17 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
src[i] = nir_u2u32(&b->nb, src[i]); src[i] = nir_u2u32(&b->nb, src[i]);
} }
} }
val->ssa->def = nir_build_alu(&b->nb, op, src[0], src[1], src[2], src[3]); dest->def = nir_build_alu(&b->nb, op, src[0], src[1], src[2], src[3]);
break; break;
} }
case SpvOpSignBitSet: case SpvOpSignBitSet:
val->ssa->def = nir_i2b(&b->nb, dest->def = nir_i2b(&b->nb,
nir_ushr(&b->nb, src[0], nir_imm_int(&b->nb, src[0]->bit_size - 1))); nir_ushr(&b->nb, src[0], nir_imm_int(&b->nb, src[0]->bit_size - 1)));
break; break;
case SpvOpUCountTrailingZerosINTEL: case SpvOpUCountTrailingZerosINTEL:
val->ssa->def = nir_umin(&b->nb, dest->def = nir_umin(&b->nb,
nir_find_lsb(&b->nb, src[0]), nir_find_lsb(&b->nb, src[0]),
nir_imm_int(&b->nb, 32u)); nir_imm_int(&b->nb, 32u));
break; break;
@@ -629,7 +628,7 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
default: { default: {
bool swap; bool swap;
unsigned src_bit_size = glsl_get_bit_size(vtn_src[0]->type); unsigned src_bit_size = glsl_get_bit_size(vtn_src[0]->type);
unsigned dst_bit_size = glsl_get_bit_size(type); unsigned dst_bit_size = glsl_get_bit_size(dest_type);
nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap, nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
src_bit_size, dst_bit_size); src_bit_size, dst_bit_size);
@@ -650,7 +649,7 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
break; break;
} }
val->ssa->def = nir_build_alu(&b->nb, op, src[0], src[1], src[2], src[3]); dest->def = nir_build_alu(&b->nb, op, src[0], src[1], src[2], src[3]);
break; break;
} /* default */ } /* default */
} }
@@ -661,8 +660,8 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
case SpvOpISub: case SpvOpISub:
case SpvOpShiftLeftLogical: case SpvOpShiftLeftLogical:
case SpvOpSNegate: { case SpvOpSNegate: {
nir_alu_instr *alu = nir_instr_as_alu(val->ssa->def->parent_instr); nir_alu_instr *alu = nir_instr_as_alu(dest->def->parent_instr);
vtn_foreach_decoration(b, val, handle_no_wrap, alu); vtn_foreach_decoration(b, dest_val, handle_no_wrap, alu);
break; break;
} }
default: default:
@@ -670,6 +669,8 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
break; break;
} }
vtn_push_ssa_value(b, w[2], dest);
b->nb.exact = b->exact; b->nb.exact = b->exact;
} }