def blockmatrix_irs(self):
scalar_ir = ir.F64(2)
vector_ir = ir.MakeArray([ir.F64(3), ir.F64(2)],
hl.tarray(hl.tfloat64))
read = ir.BlockMatrixRead(
ir.BlockMatrixNativeReader(resource('blockmatrix_example/0')))
add_two_bms = ir.BlockMatrixMap2(
read, read, 'l', 'r',
ir.ApplyBinaryPrimOp('+', ir.Ref('l'), ir.Ref('r')), "Union")
negate_bm = ir.BlockMatrixMap(
read, 'element', ir.ApplyUnaryPrimOp('-', ir.Ref('element')),
False)
sqrt_bm = ir.BlockMatrixMap(
read, 'element',
hl.sqrt(construct_expr(ir.Ref('element'), hl.tfloat64))._ir, False)
persisted = ir.BlockMatrixRead(ir.BlockMatrixPersistReader('x', read))
scalar_to_bm = ir.ValueToBlockMatrix(scalar_ir, [1, 1], 1)
col_vector_to_bm = ir.ValueToBlockMatrix(vector_ir, [2, 1], 1)
row_vector_to_bm = ir.ValueToBlockMatrix(vector_ir, [1, 2], 1)
broadcast_scalar = ir.BlockMatrixBroadcast(scalar_to_bm, [], [2, 2],
256)
broadcast_col = ir.BlockMatrixBroadcast(col_vector_to_bm, [0], [2, 2],
256)
broadcast_row = ir.BlockMatrixBroadcast(row_vector_to_bm, [1], [2, 2],
256)
transpose = ir.BlockMatrixBroadcast(broadcast_scalar, [1, 0], [2, 2],
256)
matmul = ir.BlockMatrixDot(broadcast_scalar, transpose)
rectangle = ir.Literal(hl.tarray(hl.tint64), [0, 1, 5, 6])
band = ir.Literal(hl.ttuple(hl.tint64, hl.tint64), (-1, 1))
intervals = ir.Literal(
hl.ttuple(hl.tarray(hl.tint64), hl.tarray(hl.tint64)),
([0, 1, 5, 6], [5, 6, 8, 9]))
sparsify1 = ir.BlockMatrixSparsify(read, rectangle,
ir.RectangleSparsifier)
sparsify2 = ir.BlockMatrixSparsify(read, band, ir.BandSparsifier(True))
sparsify3 = ir.BlockMatrixSparsify(read, intervals,
ir.RowIntervalSparsifier(True))
densify = ir.BlockMatrixDensify(read)
pow_ir = (construct_expr(ir.Ref('l'), hl.tfloat64)**construct_expr(
ir.Ref('r'), hl.tfloat64))._ir
squared_bm = ir.BlockMatrixMap2(scalar_to_bm, scalar_to_bm, 'l', 'r',
pow_ir, "NeedsDense")
slice_bm = ir.BlockMatrixSlice(
matmul, [slice(0, 2, 1), slice(0, 1, 1)])
return [
read, persisted, add_two_bms, negate_bm, sqrt_bm, scalar_to_bm,
col_vector_to_bm, row_vector_to_bm, broadcast_scalar,
broadcast_col, broadcast_row, squared_bm, transpose, sparsify1,
sparsify2, sparsify3, densify, matmul, slice_bm
]
def blockmatrix_irs(self):
scalar_ir = ir.F64(2)
vector_ir = ir.MakeArray([ir.F64(3), ir.F64(2)],
hl.tarray(hl.tfloat64))
read = ir.BlockMatrixRead(
ir.BlockMatrixNativeReader(resource('blockmatrix_example/0')))
add_two_bms = ir.BlockMatrixMap2(
read, read, 'l', 'r',
ir.ApplyBinaryPrimOp('+', ir.Ref('l'), ir.Ref('r')))
negate_bm = ir.BlockMatrixMap(
read, 'element', ir.ApplyUnaryPrimOp('-', ir.Ref('element')))
sqrt_bm = ir.BlockMatrixMap(
read, 'element',
hl.sqrt(construct_expr(ir.Ref('element'), hl.tfloat64))._ir)
scalar_to_bm = ir.ValueToBlockMatrix(scalar_ir, [1, 1], 1)
col_vector_to_bm = ir.ValueToBlockMatrix(vector_ir, [2, 1], 1)
row_vector_to_bm = ir.ValueToBlockMatrix(vector_ir, [1, 2], 1)
broadcast_scalar = ir.BlockMatrixBroadcast(scalar_to_bm, [], [2, 2],
256)
broadcast_col = ir.BlockMatrixBroadcast(col_vector_to_bm, [0], [2, 2],
256)
broadcast_row = ir.BlockMatrixBroadcast(row_vector_to_bm, [1], [2, 2],
256)
transpose = ir.BlockMatrixBroadcast(broadcast_scalar, [1, 0], [2, 2],
256)
matmul = ir.BlockMatrixDot(broadcast_scalar, transpose)
pow_ir = (construct_expr(ir.Ref('l'), hl.tfloat64)**construct_expr(
ir.Ref('r'), hl.tfloat64))._ir
squared_bm = ir.BlockMatrixMap2(scalar_to_bm, scalar_to_bm, 'l', 'r',
pow_ir)
slice_bm = ir.BlockMatrixSlice(
matmul, [slice(0, 2, 1), slice(0, 1, 1)])
return [
read, add_two_bms, negate_bm, sqrt_bm, scalar_to_bm,
col_vector_to_bm, row_vector_to_bm, broadcast_scalar,
broadcast_col, broadcast_row, squared_bm, transpose, matmul,
slice_bm
]