| ... | ... | @@ -2391,45 +2391,51 @@ const DeclGen = struct { |
| 2391 | 2391 | } |
| 2392 | 2392 | |
| 2393 | 2393 | fn minMax(self: *DeclGen, result_ty: Type, op: std.math.CompareOperator, lhs_id: IdRef, rhs_id: IdRef) !IdRef { |
| 2394 | | const result_ty_ref = try self.resolveType(result_ty, .direct); |
| 2395 | 2394 | const info = self.arithmeticTypeInfo(result_ty); |
| 2396 | 2395 | |
| 2397 | | // TODO: Use fmin for OpenCL |
| 2398 | | const cmp_id = try self.cmp(op, Type.bool, result_ty, lhs_id, rhs_id); |
| 2399 | | const selection_id = switch (info.class) { |
| 2400 | | .float => blk: { |
| 2401 | | // cmp uses OpFOrd. When we have 0 [<>] nan this returns false, |
| 2402 | | // but we want it to pick lhs. Therefore we also have to check if |
| 2403 | | // rhs is nan. We don't need to care about the result when both |
| 2404 | | // are nan. |
| 2405 | | const rhs_is_nan_id = self.spv.allocId(); |
| 2406 | | const bool_ty_ref = try self.resolveType(Type.bool, .direct); |
| 2407 | | try self.func.body.emit(self.spv.gpa, .OpIsNan, .{ |
| 2408 | | .id_result_type = self.typeId(bool_ty_ref), |
| 2409 | | .id_result = rhs_is_nan_id, |
| 2410 | | .x = rhs_id, |
| 2411 | | }); |
| 2412 | | const float_cmp_id = self.spv.allocId(); |
| 2413 | | try self.func.body.emit(self.spv.gpa, .OpLogicalOr, .{ |
| 2414 | | .id_result_type = self.typeId(bool_ty_ref), |
| 2415 | | .id_result = float_cmp_id, |
| 2416 | | .operand_1 = cmp_id, |
| 2417 | | .operand_2 = rhs_is_nan_id, |
| 2418 | | }); |
| 2419 | | break :blk float_cmp_id; |
| 2420 | | }, |
| 2421 | | else => cmp_id, |
| 2422 | | }; |
| 2396 | var wip = try self.elementWise(result_ty); |
| 2397 | defer wip.deinit(); |
| 2398 | for (wip.results, 0..) |*result_id, i| { |
| 2399 | const lhs_elem_id = try wip.elementAt(result_ty, lhs_id, i); |
| 2400 | const rhs_elem_id = try wip.elementAt(result_ty, rhs_id, i); |
| 2401 | |
| 2402 | // TODO: Use fmin for OpenCL |
| 2403 | const cmp_id = try self.cmp(op, Type.bool, wip.scalar_ty, lhs_elem_id, rhs_elem_id); |
| 2404 | const selection_id = switch (info.class) { |
| 2405 | .float => blk: { |
| 2406 | // cmp uses OpFOrd. When we have 0 [<>] nan this returns false, |
| 2407 | // but we want it to pick lhs. Therefore we also have to check if |
| 2408 | // rhs is nan. We don't need to care about the result when both |
| 2409 | // are nan. |
| 2410 | const rhs_is_nan_id = self.spv.allocId(); |
| 2411 | const bool_ty_ref = try self.resolveType(Type.bool, .direct); |
| 2412 | try self.func.body.emit(self.spv.gpa, .OpIsNan, .{ |
| 2413 | .id_result_type = self.typeId(bool_ty_ref), |
| 2414 | .id_result = rhs_is_nan_id, |
| 2415 | .x = rhs_elem_id, |
| 2416 | }); |
| 2417 | const float_cmp_id = self.spv.allocId(); |
| 2418 | try self.func.body.emit(self.spv.gpa, .OpLogicalOr, .{ |
| 2419 | .id_result_type = self.typeId(bool_ty_ref), |
| 2420 | .id_result = float_cmp_id, |
| 2421 | .operand_1 = cmp_id, |
| 2422 | .operand_2 = rhs_is_nan_id, |
| 2423 | }); |
| 2424 | break :blk float_cmp_id; |
| 2425 | }, |
| 2426 | else => cmp_id, |
| 2427 | }; |
| 2423 | 2428 | |
| 2424 | | const result_id = self.spv.allocId(); |
| 2425 | | try self.func.body.emit(self.spv.gpa, .OpSelect, .{ |
| 2426 | | .id_result_type = self.typeId(result_ty_ref), |
| 2427 | | .id_result = result_id, |
| 2428 | | .condition = selection_id, |
| 2429 | | .object_1 = lhs_id, |
| 2430 | | .object_2 = rhs_id, |
| 2431 | | }); |
| 2432 | | return result_id; |
| 2429 | result_id.* = self.spv.allocId(); |
| 2430 | try self.func.body.emit(self.spv.gpa, .OpSelect, .{ |
| 2431 | .id_result_type = wip.scalar_ty_id, |
| 2432 | .id_result = result_id.*, |
| 2433 | .condition = selection_id, |
| 2434 | .object_1 = lhs_elem_id, |
| 2435 | .object_2 = rhs_elem_id, |
| 2436 | }); |
| 2437 | } |
| 2438 | return wip.finalize(); |
| 2433 | 2439 | } |
| 2434 | 2440 | |
| 2435 | 2441 | /// This function normalizes values to a canonical representation |
| ... | ... | @@ -3107,20 +3113,15 @@ const DeclGen = struct { |
| 3107 | 3113 | return result_id; |
| 3108 | 3114 | }, |
| 3109 | 3115 | .Vector => { |
| 3110 | | const child_ty = ty.childType(mod); |
| 3111 | | const vector_len = ty.vectorLen(mod); |
| 3112 | | |
| 3113 | | const constituents = try self.gpa.alloc(IdRef, vector_len); |
| 3114 | | defer self.gpa.free(constituents); |
| 3115 | | |
| 3116 | | for (constituents, 0..) |*constituent, i| { |
| 3117 | | const lhs_index_id = try self.extractField(child_ty, cmp_lhs_id, @intCast(i)); |
| 3118 | | const rhs_index_id = try self.extractField(child_ty, cmp_rhs_id, @intCast(i)); |
| 3119 | | const result_id = try self.cmp(op, Type.bool, child_ty, lhs_index_id, rhs_index_id); |
| 3120 | | constituent.* = try self.convertToIndirect(Type.bool, result_id); |
| 3116 | var wip = try self.elementWise(result_ty); |
| 3117 | defer wip.deinit(); |
| 3118 | const scalar_ty = ty.scalarType(mod); |
| 3119 | for (wip.results, 0..) |*result_id, i| { |
| 3120 | const lhs_elem_id = try wip.elementAt(ty, lhs_id, i); |
| 3121 | const rhs_elem_id = try wip.elementAt(ty, rhs_id, i); |
| 3122 | result_id.* = try self.cmp(op, Type.bool, scalar_ty, lhs_elem_id, rhs_elem_id); |
| 3121 | 3123 | } |
| 3122 | | |
| 3123 | | return try self.constructArray(result_ty, constituents); |
| 3124 | return wip.finalize(); |
| 3124 | 3125 | }, |
| 3125 | 3126 | else => unreachable, |
| 3126 | 3127 | }; |
Robin Voetter <robin@voetter.nl> 2024-01-21 20:12:25+01:00
Signed by SSH key SHA256:ZS52FNyUv2WUXvO4njmVaFVO46RHojFuOrxRc4LuKzg
src