Merge branch 'develop' into driver

test/tf/matmul_test.pb

+
+2
+
0Placeholder*
+dtype0
*
+shape
+:
+2
+
1Placeholder*
+dtype0
*
+shape
+:
+F
+matmul1MatMul
0
1*
+
T0
*
+transpose_a(
*
+transpose_b(
"
\ No newline at end of file

test/tf/mul_test.pb

+
+:
+
0Placeholder*
+shape:


*
+dtype0

+:
+
1Placeholder*
+dtype0
*
+shape:



+
+mul1Mul
0
1*
+
T0
"
\ No newline at end of file

test/tf/pack_test.pb

+
+.
+
0Placeholder*
+shape:*
+dtype0

+.
+
1Placeholder*
+dtype0
*
+shape:
+.
+
2Placeholder*
+dtype0
*
+shape:
+4
+pack1Pack
0
1
2*
+
T0
*
+
+axis
*
+
N"
\ No newline at end of file

test/tf/pack_test_nhwc.pb

+
+:
+
0Placeholder*
+dtype0
*
+shape:



+:
+
1Placeholder*
+dtype0
*
+shape:



+:
+
2Placeholder*
+dtype0
*
+shape:



+4
+pack1Pack
0
1
2*
+
T0
*
+
+axis*
+
N"
\ No newline at end of file

test/tf/pooling_test.pb

+
+:
+
0Placeholder*
+dtype0
*
+shape:

+u
+avg_poolingAvgPool
0*
+ksize
+

*
+paddingVALID*
+
T0
*
+data_formatNHWC*
+strides
+


+u
+max_poolingMaxPool
0*
+data_formatNHWC*
+strides
+

*
+ksize
+

*
+paddingVALID*
+
T0
"
\ No newline at end of file

test/tf/relu6_test.pb

+
+:
+
0Placeholder*
+dtype0
*
+shape:

+
+relu6Relu6
0*
+
T0
"
\ No newline at end of file

test/tf/relu_test.pb

+
+:
+
0Placeholder*
+dtype0
*
+shape:

+
+reluRelu
0*
+
T0
"
\ No newline at end of file

test/tf/softmax_test.pb

+
+2
+
0Placeholder*
+dtype0
*
+shape
+:

+
+softmaxSoftmax
0*
+
T0
"
\ No newline at end of file

test/tf/tf_test.cpp

+#include <iostream>
+#include <vector>
+#include <migraphx/literal.hpp>
+#include <migraphx/operators.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/tf.hpp>
+#include "test.hpp"
+
+TEST_CASE(add_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 2, 2, 3}});
+    auto l1 = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {1, 2, 2, 3}});
+    p.add_instruction(migraphx::op::add{}, l0, l1);
+    auto prog = migraphx::parse_tf("add_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(add_bcast_test)
+{
+
+    migraphx::program p;
+    migraphx::shape s0{migraphx::shape::float_type, {2, 3}};
+    auto l0 = p.add_parameter("0", s0);
+    auto l1 = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {2, 1}});
+    auto l2 = p.add_instruction(migraphx::op::multibroadcast{s0.lens()}, l0);
+    auto l3 = p.add_instruction(migraphx::op::multibroadcast{s0.lens()}, l1);
+    p.add_instruction(migraphx::op::add{}, l2, l3);
+    auto prog = migraphx::parse_tf("add_bcast_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(batchnorm_test)
+{
+    float epsilon  = 1.001e-5f;
+    float momentum = 0.9f;
+
+    migraphx::program p;
+    migraphx::op::batch_norm_inference op{
+        epsilon, momentum, migraphx::op::batch_norm_inference::spatial};
+    migraphx::shape s0{migraphx::shape::float_type, {32}};
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 32, 16, 16}});
+    std::vector<float> const_vals(32);
+    std::fill(const_vals.begin(), const_vals.end(), 1.0f);
+
+    auto l2 = p.add_parameter("2", s0);
+    auto l3 = p.add_parameter("3", s0);
+    auto l4 = p.add_parameter("4", s0);
+    auto l1 = p.add_literal(migraphx::literal{s0, const_vals});
+    p.add_instruction(op, l0, l1, l2, l3, l4);
+    auto prog = migraphx::parse_tf("batchnorm_test.pb", true);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(biasadd_test)
+{
+    migraphx::program p;
+    migraphx::shape s0{migraphx::shape::float_type, {1, 500, 1, 1}};
+    uint64_t axis = 1;
+    auto l0       = p.add_parameter("0", s0);
+    auto l1       = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {500}});
+    auto l2       = p.add_instruction(migraphx::op::broadcast{axis, l0->get_shape().lens()}, l1);
+    p.add_instruction(migraphx::op::add{}, l0, l2);
+    auto prog = migraphx::parse_tf("biasadd_test.pb", true);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(concat_test)
+{
+    migraphx::program p;
+
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {4, 7, 3}});
+    auto l1 = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {4, 2, 3}});
+
+    int axis = 1;
+    // tf uses axis as the third input, and it is in int32 format
+    // add the literal using a vector in order to set stride to 1 (like in tf parser)
+    p.add_literal(migraphx::shape{migraphx::shape::int32_type}, std::vector<int>{axis});
+
+    p.add_instruction(migraphx::op::concat{static_cast<std::size_t>(axis)}, l0, l1);
+    auto prog = migraphx::parse_tf("concat_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(const_test)
+{
+    migraphx::program p;
+    p.add_literal(migraphx::shape{migraphx::shape::float_type}, std::vector<float>{1.0f});
+    auto prog = migraphx::parse_tf("constant_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(conv_test)
+{
+    migraphx::program p;
+
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    std::vector<float> weight_data(3 * 3 * 3 * 32);
+    std::fill(weight_data.begin(), weight_data.end(), 1.0f);
+    auto l1 =
+        p.add_literal(migraphx::shape{migraphx::shape::float_type, {3, 3, 3, 32}}, weight_data);
+
+    migraphx::op::convolution op;
+    op.padding_mode = migraphx::op::padding_mode_t::same;
+    op.stride       = {1, 1};
+    op.dilation     = {1, 1};
+    auto l2         = p.add_instruction(migraphx::op::transpose{{0, 3, 1, 2}}, l1);
+    auto l3         = p.add_instruction(migraphx::op::transpose{{1, 3, 0, 2}}, l2);
+    p.add_instruction(op, l0, l3);
+    auto prog = migraphx::parse_tf("conv_test.pb", true);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(depthwiseconv_test)
+{
+    migraphx::program p;
+
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    std::vector<float> weight_data(3 * 3 * 3 * 1);
+    std::fill(weight_data.begin(), weight_data.end(), 1.0f);
+    auto l1 =
+        p.add_literal(migraphx::shape{migraphx::shape::float_type, {3, 3, 3, 1}}, weight_data);
+
+    migraphx::op::convolution op;
+    op.padding_mode = migraphx::op::padding_mode_t::same;
+    op.stride       = {1, 1};
+    op.dilation     = {1, 1};
+    op.group        = 3;
+    auto l2         = p.add_instruction(migraphx::op::transpose{{0, 3, 1, 2}}, l1);
+    auto l3         = p.add_instruction(migraphx::op::transpose{{1, 3, 0, 2}}, l2);
+    auto l4         = p.add_instruction(migraphx::op::contiguous{}, l3);
+    auto l5         = p.add_instruction(migraphx::op::reshape{{3, 1, 3, 3}}, l4);
+    p.add_instruction(op, l0, l5);
+    auto prog = migraphx::parse_tf("depthwise_conv_test.pb", true);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(identity_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    p.add_instruction(migraphx::op::identity{}, l0);
+    auto prog = migraphx::parse_tf("identity_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(matmul_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {8, 4}});
+    auto l1 = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {4, 8}});
+
+    auto trans_l0 = p.add_instruction(migraphx::op::transpose{{1, 0}}, l0);
+    auto trans_l1 = p.add_instruction(migraphx::op::transpose{{1, 0}}, l1);
+
+    p.add_instruction(migraphx::op::dot{}, trans_l0, trans_l1);
+    auto prog = migraphx::parse_tf("matmul_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(mean_test)
+{
+    migraphx::program p;
+    migraphx::literal l{migraphx::shape{migraphx::shape::int32_type, {2}}, {2, 3}};
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    p.add_literal(l);
+    p.add_literal(l);
+    migraphx::op::pooling op;
+    op.lengths = {16, 16};
+    auto l3    = p.add_instruction(op, l0);
+    p.add_instruction(migraphx::op::squeeze{{2, 3}}, l3);
+    p.add_instruction(op, l0);
+    auto prog = migraphx::parse_tf("mean_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(mean_test_nhwc)
+{
+    migraphx::program p;
+    migraphx::literal l{migraphx::shape{migraphx::shape::int32_type, {2}}, {1, 2}};
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    p.add_literal(l);
+    p.add_literal(l);
+    migraphx::op::pooling op;
+    op.lengths = {16, 16};
+    auto l3    = p.add_instruction(op, l0);
+    p.add_instruction(migraphx::op::squeeze{{2, 3}}, l3);
+    p.add_instruction(op, l0);
+    auto prog = migraphx::parse_tf("mean_test_nhwc.pb", true);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(mul_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 1, 1, 16}});
+    auto l1 = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {1, 1, 1, 16}});
+
+    p.add_instruction(migraphx::op::mul{}, l0, l1);
+    auto prog = migraphx::parse_tf("mul_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(pack_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {2}});
+    auto l1 = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {2}});
+    auto l2 = p.add_parameter("2", migraphx::shape{migraphx::shape::float_type, {2}});
+    std::vector<migraphx::instruction_ref> args{l0, l1, l2};
+    std::vector<migraphx::instruction_ref> unsqueezed_args;
+    int64_t axis = 1;
+
+    std::transform(args.begin(),
+                   args.end(),
+                   std::back_inserter(unsqueezed_args),
+                   [&](migraphx::instruction_ref arg) {
+                       return p.add_instruction(migraphx::op::unsqueeze{{axis}}, arg);
+                   });
+    p.add_instruction(migraphx::op::concat{static_cast<size_t>(axis)}, unsqueezed_args);
+    auto prog = migraphx::parse_tf("pack_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(pack_test_nhwc)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 2, 1, 1}});
+    auto l1 = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {1, 2, 1, 1}});
+    auto l2 = p.add_parameter("2", migraphx::shape{migraphx::shape::float_type, {1, 2, 1, 1}});
+    std::vector<migraphx::instruction_ref> args{l0, l1, l2};
+    std::vector<migraphx::instruction_ref> unsqueezed_args;
+    int64_t nchw_axis = 1;
+
+    std::transform(args.begin(),
+                   args.end(),
+                   std::back_inserter(unsqueezed_args),
+                   [&](migraphx::instruction_ref arg) {
+                       return p.add_instruction(migraphx::op::unsqueeze{{nchw_axis}}, arg);
+                   });
+    p.add_instruction(migraphx::op::concat{static_cast<size_t>(nchw_axis)}, unsqueezed_args);
+    auto prog = migraphx::parse_tf("pack_test_nhwc.pb", true);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(pooling_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    migraphx::op::pooling avg_pool_op{"average"};
+    migraphx::op::pooling max_pool_op{"max"};
+    avg_pool_op.padding_mode = migraphx::op::padding_mode_t::valid;
+    max_pool_op.padding_mode = migraphx::op::padding_mode_t::valid;
+    avg_pool_op.stride       = {2, 2};
+    max_pool_op.stride       = {2, 2};
+    avg_pool_op.lengths      = {2, 2};
+    max_pool_op.lengths      = {2, 2};
+    p.add_instruction(max_pool_op, l0);
+    p.add_instruction(avg_pool_op, l0);
+    auto prog = migraphx::parse_tf("pooling_test.pb", true);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(relu_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    p.add_instruction(migraphx::op::relu{}, l0);
+    auto prog = migraphx::parse_tf("relu_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(relu6_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    p.add_instruction(migraphx::op::clip{6.0, 0.0}, l0);
+    auto prog = migraphx::parse_tf("relu6_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(reshape_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {16}});
+    migraphx::shape s0{migraphx::shape::int32_type, {4}};
+    // in tf, the second arg is a literal that contains new dimensions
+    p.add_literal(migraphx::literal{s0, {1, 1, 1, 16}});
+    p.add_instruction(migraphx::op::reshape{{1, 1, 1, 16}}, l0);
+    auto prog = migraphx::parse_tf("reshape_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(softmax_test)
+{
+    migraphx::program p;
+    auto l0   = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3}});
+    auto dims = l0->get_shape().lens();
+    auto r    = p.add_instruction(migraphx::op::reshape{{long(dims[0]), long(dims[1]), 1, 1}}, l0);
+    auto s    = p.add_instruction(migraphx::op::softmax{}, r);
+    p.add_instruction(migraphx::op::reshape{{long(dims[0]), long(dims[1])}}, s);
+    auto prog = migraphx::parse_tf("softmax_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(squeeze_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 2, 3, 1}});
+    p.add_instruction(migraphx::op::squeeze{{0, 3}}, l0);
+    auto prog = migraphx::parse_tf("squeeze_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(stridedslice_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 10, 1, 1}});
+    std::size_t num_axes = 4;
+    migraphx::op::slice op;
+    op.starts = {0, 0, 0, 0};
+    op.ends   = {1, 5, 1, 1};
+    op.axes   = std::vector<int64_t>(num_axes);
+    std::iota(op.axes.begin(), op.axes.end(), 0);
+    // add literals for starts, ends, and strides in tf (NHWC format)
+    p.add_literal(migraphx::shape{migraphx::shape::int32_type, {4}}, std::vector<int>{0, 0, 0, 0});
+    p.add_literal(migraphx::shape{migraphx::shape::int32_type, {4}}, std::vector<int>{1, 1, 1, 5});
+    p.add_literal(migraphx::shape{migraphx::shape::int32_type, {4}}, std::vector<int>{1, 1, 1, 1});
+
+    auto l1          = p.add_instruction(op, l0);
+    auto shrink_axis = 2;
+    p.add_instruction(migraphx::op::squeeze{{shrink_axis}}, l1);
+    auto prog = migraphx::parse_tf("stridedslice_test.pb", true);
+
+    EXPECT(p == prog);
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/type_conversion.cpp

+#include <iostream>
+#include <vector>
+#include <migraphx/literal.hpp>
+#include <migraphx/operators.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/cpu/target.hpp>
+#include <migraphx/verify.hpp>
+#include <migraphx/quantization.hpp>
+#include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/propagate_constant.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/onnx.hpp>
+#include "test.hpp"
+#include <migraphx/half.hpp>
+
+TEST_CASE(param_add)
+{
+    auto create_program_float = [] {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+        auto p1 = p.add_parameter("x", s);
+        auto p2 = p.add_parameter("y", s);
+        p.add_instruction(migraphx::op::add{}, p1, p2);
+
+        return p;
+    };
+
+    auto create_program_half = [] {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+        auto p1  = p.add_parameter("x", s);
+        auto hp1 = p.insert_instruction(std::next(p1), migraphx::op::convert{}, p1);
+        auto p2  = p.add_parameter("y", s);
+        auto hp2 = p.insert_instruction(std::next(p2), migraphx::op::convert{}, p2);
+        auto hs  = p.add_instruction(migraphx::op::add{}, hp1, hp2);
+        p.add_instruction(migraphx::op::convert{migraphx::shape::float_type}, hs);
+
+        return p;
+    };
+
+    {
+        auto p1 = create_program_float();
+        auto p2 = create_program_half();
+
+        migraphx::quantize(p1);
+        EXPECT(p1 == p2);
+    }
+
+    {
+        auto p1 = create_program_float();
+        auto p2 = create_program_half();
+
+        migraphx::quantize(p1, {"add"});
+        EXPECT(p1 == p2);
+    }
+}
+
+TEST_CASE(param_add_sub)
+{
+    auto create_program_float = [] {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+        auto p1   = p.add_parameter("x", s);
+        auto p2   = p.add_parameter("y", s);
+        auto sum  = p.add_instruction(migraphx::op::add{}, p1, p2);
+        auto diff = p.add_instruction(migraphx::op::sub{}, sum, p2);
+        p.add_instruction(migraphx::op::add{}, diff, p1);
+
+        return p;
+    };
+
+    auto create_program_half_add = [] {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+        auto p1  = p.add_parameter("x", s);
+        auto hp1 = p.insert_instruction(
+            std::next(p1), migraphx::op::convert{migraphx::shape::half_type}, p1);
+        auto p2  = p.add_parameter("y", s);
+        auto hp2 = p.insert_instruction(
+            std::next(p2), migraphx::op::convert{migraphx::shape::half_type}, p2);
+        auto hsum  = p.add_instruction(migraphx::op::add{}, hp1, hp2);
+        auto sum   = p.add_instruction(migraphx::op::convert{migraphx::shape::float_type}, hsum);
+        auto diff  = p.add_instruction(migraphx::op::sub{}, sum, p2);
+        auto hdiff = p.add_instruction(
+            migraphx::op::convert{migraphx::op::convert{migraphx::shape::half_type}}, diff);
+        auto res = p.add_instruction(migraphx::op::add{}, hdiff, hp1);
+        p.add_instruction(migraphx::op::convert{migraphx::shape::float_type}, res);
+
+        return p;
+    };
+
+    auto create_program_half_sub = [] {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+        auto p1  = p.add_parameter("x", s);
+        auto p2  = p.add_parameter("y", s);
+        auto hp2 = p.insert_instruction(
+            std::next(p2), migraphx::op::convert{migraphx::shape::half_type}, p2);
+        auto sum   = p.add_instruction(migraphx::op::add{}, p1, p2);
+        auto hsum  = p.add_instruction(migraphx::op::convert{migraphx::shape::half_type}, sum);
+        auto hdiff = p.add_instruction(migraphx::op::sub{}, hsum, hp2);
+        auto diff  = p.add_instruction(migraphx::op::convert{migraphx::shape::float_type}, hdiff);
+        p.add_instruction(migraphx::op::add{}, diff, p1);
+
+        return p;
+    };
+
+    auto create_program_half_all = [] {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+        auto p1  = p.add_parameter("x", s);
+        auto hp1 = p.insert_instruction(
+            std::next(p1), migraphx::op::convert{migraphx::shape::half_type}, p1);
+        auto p2  = p.add_parameter("y", s);
+        auto hp2 = p.insert_instruction(
+            std::next(p2), migraphx::op::convert{migraphx::shape::half_type}, p2);
+        auto hsum  = p.add_instruction(migraphx::op::add{}, hp1, hp2);
+        auto hdiff = p.add_instruction(migraphx::op::sub{}, hsum, hp2);
+        auto hres  = p.add_instruction(migraphx::op::add{}, hdiff, hp1);
+        p.add_instruction(migraphx::op::convert{migraphx::shape::float_type}, hres);
+
+        return p;
+    };
+
+    {
+        auto p1 = create_program_float();
+        auto p2 = create_program_half_add();
+
+        migraphx::quantize(p1, {"add"});
+        EXPECT(p1 == p2);
+    }
+
+    {
+        auto p1 = create_program_float();
+        auto p2 = create_program_half_sub();
+
+        migraphx::quantize(p1, {"sub"});
+        EXPECT(p1 == p2);
+    }
+
+    {
+        auto p1 = create_program_float();
+        auto p2 = create_program_half_all();
+
+        migraphx::quantize(p1);
+        migraphx::run_passes(p1, {migraphx::dead_code_elimination{}});
+
+        EXPECT(p1 == p2);
+    }
+}
+
+TEST_CASE(literal_add)
+{
+    auto create_program_float = [] {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+        std::vector<float> data(2 * 3);
+        std::iota(data.begin(), data.end(), 1.0f);
+        auto l1 = p.add_literal(migraphx::literal(s, data));
+        auto l2 = p.add_literal(migraphx::literal(s, data));
+        p.add_instruction(migraphx::op::add{}, l1, l2);
+
+        return p;
+    };
+
+    auto create_program_half = [] {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::half_type, {2, 3}};
+        std::vector<migraphx::half> data(2 * 3);
+        std::iota(data.begin(), data.end(), 1.0f);
+        auto l1 = p.add_literal(migraphx::literal(s, data));
+        auto l2 = p.add_literal(migraphx::literal(s, data));
+        auto hs = p.add_instruction(migraphx::op::add{}, l1, l2);
+        p.add_instruction(migraphx::op::convert{migraphx::shape::float_type}, hs);
+
+        return p;
+    };
+
+    {
+        auto p1 = create_program_float();
+        auto p2 = create_program_half();
+
+        migraphx::quantize(p1, {"all"});
+        migraphx::run_passes(p1,
+                             {migraphx::propagate_constant{}, migraphx::dead_code_elimination{}});
+        migraphx::run_passes(p2,
+                             {migraphx::propagate_constant{}, migraphx::dead_code_elimination{}});
+
+        EXPECT(p1 == p2);
+    }
+
+    {
+        auto p1 = create_program_float();
+        auto p2 = create_program_half();
+
+        migraphx::quantize(p1, {"add"});
+        migraphx::run_passes(p1,
+                             {migraphx::propagate_constant{}, migraphx::dead_code_elimination{}});
+        migraphx::run_passes(p2,
+                             {migraphx::propagate_constant{}, migraphx::dead_code_elimination{}});
+        EXPECT(p1 == p2);
+    }
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

tools/include/concat_opt.hpp

@@ -9,7 +9,7 @@
 #include <utility>
 
 #include <migraphx/operation.hpp>
-#include <migraphx/operators.hpp>
+#include <migraphx/op/concat.hpp>
 #include <migraphx/config.hpp>
 
 namespace migraphx {

tools/include/operation.hpp

@@ -49,7 +49,7 @@ struct operation
     argument compute(context& ctx, const shape& output, const std::vector<argument>& input) const;
     /// An optional method to return which argument the output will alias. If
     /// there is no aliased output then -1 can be returned.
-    int output_alias(const std::vector<shape>& input) const;
+    std::ptrdiff_t output_alias(const std::vector<shape>& input) const;
     /// An optional stream operator to print the operation. When this is not
     /// implemented, it will just print the operation's name.
     friend std::ostream& operator<<(std::ostream& os, const operation& op);
@@ -69,7 +69,7 @@ auto operator<<(std::ostream& os, const T& x) -> decltype(os << x.name())
 {
     os << x.name();
     char delim = '[';
-    reflect_each(x, [&](auto& y, auto name) {
+    reflect_each(x, [&](auto&& y, auto name) {
         os << delim;
         os << name << "=";
         stream_write_value(os, y);
@@ -87,6 +87,8 @@ namespace operation_equal {
 template <class T, class U>
 auto operator==(const T& x, const U& y) -> decltype(x.name() == y.name())
 {
+    static_assert(is_reflectable<T>{} or sizeof(T) <= 1,
+                  "Missing equality operator or reflect method.");
     if(x.name() != y.name())
         return false;
     const auto& yy = any_cast<T>(y);
@@ -175,7 +177,7 @@ auto is_context_free_op(const T& x) -> decltype(is_context_free_op(
 }
 
 template <class T>
-int output_alias_op(rank<0>, const T&, const std::vector<shape>&)
+std::ptrdiff_t output_alias_op(rank<0>, const T&, const std::vector<shape>&)
 {
     return -1;
 }
@@ -188,7 +190,7 @@ auto output_alias_op(rank<1>, const T& x, const std::vector<shape>& shapes)
 }
 
 template <class T>
-int output_alias_op(const T& x, const std::vector<shape>& shapes)
+std::ptrdiff_t output_alias_op(const T& x, const std::vector<shape>& shapes)
 {
     return output_alias_op(rank<1>{}, x, shapes);
 }
@@ -238,7 +240,7 @@ auto has_finalize_op(const T&) -> decltype(has_finalize_op(rank<1>{},
      virtual('is_context_free', returns = 'bool', const = True, default = 'is_context_free_op'),
      virtual('has_finalize', returns = 'bool', const = True, default = 'has_finalize_op'),
      virtual('output_alias',
-             returns = 'int',
+             returns = 'std::ptrdiff_t',
              input   = 'const std::vector<shape>&',
              const   = True,
              default = 'output_alias_op'),

tools/include/schedule_model.hpp

+#ifndef MIGRAPHX_GUARD_SCHEDULE_MODEL_HPP
+#define MIGRAPHX_GUARD_SCHEDULE_MODEL_HPP
+
+#include <cassert>
+#include <string>
+#include <functional>
+#include <memory>
+#include <type_traits>
+#include <utility>
+
+#include <migraphx/config.hpp>
+#include <migraphx/instruction_ref.hpp>
+#include <vector>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct program;
+struct operation;
+
+#ifdef DOXYGEN
+
+/// An interface for target-dependent model for the scheduler
+struct schedule_model
+{
+    /// Get the number of concurrent instruction allowed
+    std::size_t concurrency() const;
+    /// Schedule a concurrent instruction
+    void sched(program& p, instruction_ref ins, std::size_t n) const;
+    // Insert necessary waits before an instruction
+    void wait(program& p, instruction_ref ins, std::size_t wait_id) const;
+    // Insert necessary records after an instruction
+    void record(program& p, instruction_ref ins, std::size_t wait_id) const;
+    /// Compute weights for an operation
+    std::size_t weight(const operation& op) const;
+};
+
+#else
+
+<%
+interface('schedule_model',
+    virtual('concurrency', returns='std::size_t', const=True),
+    virtual('sched', p='program&', ins='instruction_ref', n='std::size_t', const=True),
+    virtual('wait', p='program&', ins='instruction_ref', wait_id='std::size_t', const=True),
+    virtual('record', p='program&', ins='instruction_ref', wait_id='std::size_t', const=True),
+    virtual('weight', returns='std::size_t', op='const operation&', const=True)
+)
+%>
+
+#endif
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

tools/include/target.hpp

@@ -22,10 +22,8 @@ struct target
 {
     /// A unique name used to identify the target
     std::string name() const;
-    /// The transformation passes to be run
     /**
      * @brief The transformation pass to be run during compilation.
-     * @details [long description]
      *
      * @param ctx This is the target-dependent context that is created by `get_context`
      * @return The passes to be ran

tools/te.py

 import string, sys, re, os
 
+
+trivial = [
+    'std::size_t',
+    'instruction_ref'
+]
+
 headers = '''
 #include <algorithm>
 #include <cassert>
@@ -286,7 +292,7 @@ def convert_member(d, struct_name):
                     member['this'] = x
                     if 'const' in t:
                         member['member_const'] = 'const'
-                if t.endswith(('&', '*')):
+                if t.endswith(('&', '*')) or t in trivial:
                     if use_member: member_args.append(x)
                     args.append(arg_name)
                 else: