Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Use optimize_module pass for the quantization to fp16 #1974

Merged
merged 9 commits into from
Jul 21, 2023
Merged

Use optimize_module pass for the quantization to fp16 #1974

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
55d61f9
Use optimize_module pass for the quantization
umangyadav Jul 18, 2023
dea2a09
formatting
umangyadav Jul 18, 2023
0da1e71
change order of multiplication
umangyadav Jul 18, 2023
e5803ab
Fix test
umangyadav Jul 19, 2023
b50b5ca
Formatting
umangyadav Jul 19, 2023
0a14e61
Merge branch 'develop' into batchnorm_fp16
umangyadav Jul 19, 2023
b5fa13a
remove trace for run_passes()
umangyadav Jul 19, 2023
d4b3397
Merge branch 'develop' into batchnorm_fp16
umangyadav Jul 19, 2023
4147411
Merge branch 'develop' into batchnorm_fp16
causten Jul 21, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 11 additions & 13 deletions src/onnx/parse_batchnorm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,21 +57,19 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
auto x_rank = x_lens.size();
if(x_rank == 1 or x_rank == 2)
{
auto rt = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
auto numer = info.add_broadcastable_binary_op("sub", args[0], args[3]);
auto var_eps = info.add_broadcastable_binary_op("add", args[4], eps);
auto denom = info.add_broadcastable_binary_op("pow", var_eps, rt);
auto div0 = info.add_broadcastable_binary_op("div", numer, denom);
auto r0 = info.add_broadcastable_binary_op("mul", div0, args[1]);
auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
auto x_sub_mean = info.add_broadcastable_binary_op("sub", args[0], args[3]);
auto var_eps = info.add_broadcastable_binary_op("add", args[4], eps);
auto rsqrt = info.add_instruction(make_op("rsqrt"), var_eps);
auto mul0 = info.add_broadcastable_binary_op("mul", args[1], rsqrt);
auto r0 = info.add_broadcastable_binary_op("mul", x_sub_mean, mul0);
return info.add_broadcastable_binary_op("add", r0, args[2]);
}
else if(x_rank > 2)
{
// unsqueeze tensors of shape (C) to broadcast correctly
std::vector<int64_t> unsqueeze_axes(x_lens.size() - 2);
std::iota(unsqueeze_axes.begin(), unsqueeze_axes.end(), 1);
auto rt = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
auto scale_unsqueeze = info.add_instruction(
migraphx::make_op("unsqueeze", {{"axes", unsqueeze_axes}}), args[1]);
Expand All @@ -81,11 +79,11 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
migraphx::make_op("unsqueeze", {{"axes", unsqueeze_axes}}), args[3]);
auto var_unsqueeze = info.add_instruction(
migraphx::make_op("unsqueeze", {{"axes", unsqueeze_axes}}), args[4]);
auto numer = info.add_broadcastable_binary_op("sub", args[0], mean_unsqueeze);
auto var_eps = info.add_broadcastable_binary_op("add", var_unsqueeze, eps);
auto denom = info.add_broadcastable_binary_op("pow", var_eps, rt);
auto div0 = info.add_broadcastable_binary_op("div", numer, denom);
auto r0 = info.add_broadcastable_binary_op("mul", div0, scale_unsqueeze);
auto x_sub_mean = info.add_broadcastable_binary_op("sub", args[0], mean_unsqueeze);
auto var_eps = info.add_broadcastable_binary_op("add", var_unsqueeze, eps);
auto rsqrt = info.add_instruction(make_op("rsqrt"), var_eps);
auto mul0 = info.add_broadcastable_binary_op("mul", scale_unsqueeze, rsqrt);
auto r0 = info.add_broadcastable_binary_op("mul", x_sub_mean, mul0);
return info.add_broadcastable_binary_op("add", r0, bias_unsqueeze);
}
else
Expand Down
16 changes: 5 additions & 11 deletions src/quantization.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,7 @@
#include <migraphx/simplify_reshapes.hpp>
#include <migraphx/simplify_qdq.hpp>
#include <migraphx/eliminate_common_subexpression.hpp>
#include <migraphx/optimize_module.hpp>
#include <migraphx/dead_code_elimination.hpp>
#include <migraphx/program.hpp>
#include <migraphx/instruction.hpp>
Expand All @@ -48,19 +49,12 @@ MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_INT8_QUANTIZATION_PARAMS)

// This function is to convert any instructions specified in the input
// from double or float to float16 by inserting a convert operator.
// For the conversion, there could be cases of overflowing, but it
// is very rare in the area of deeping learning, so we just do a
// truncate of the input to get the fp16.
// For the conversion, there could be cases of overflowing or underflowing, but it
// is uncommon. Run optimize_module() before converting to fp16 to const eval and fold in FP32 to
// avoid loss of precision.
void quantize_fp16(program& prog, const std::vector<std::string>& ins_names)
{
run_passes(prog,
{quantize_fp16_pass{ins_names},
eliminate_common_subexpression{},
dead_code_elimination{},
simplify_reshapes{},
dead_code_elimination{},
simplify_qdq{},
dead_code_elimination{}});
run_passes(prog, {optimize_module{}, quantize_fp16_pass{ins_names}, optimize_module{}});
}

void quantize_int8(program& prog,
Expand Down
11 changes: 5 additions & 6 deletions src/tf/parse_batchnorm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,7 +52,6 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
auto x_type = args[0]->get_shape().type();

// unsqueeze tensors of shape (C) to broadcast correctly
auto rt = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});

auto scale_unsqueeze =
Expand All @@ -64,11 +63,11 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
auto var_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[4]);

auto numer = info.add_broadcastable_binary_op("sub", args[0], mean_unsqueeze);
auto var_eps = info.add_broadcastable_binary_op("add", var_unsqueeze, eps);
auto denom = info.add_broadcastable_binary_op("pow", var_eps, rt);
auto div0 = info.add_broadcastable_binary_op("div", numer, denom);
auto r0 = info.add_broadcastable_binary_op("mul", div0, scale_unsqueeze);
auto x_sub_mean = info.add_broadcastable_binary_op("sub", args[0], mean_unsqueeze);
auto var_eps = info.add_broadcastable_binary_op("add", var_unsqueeze, eps);
auto rsqrt = info.add_instruction(make_op("rsqrt"), var_eps);
auto mul0 = info.add_broadcastable_binary_op("mul", scale_unsqueeze, rsqrt);
auto r0 = info.add_broadcastable_binary_op("mul", x_sub_mean, mul0);
return info.add_broadcastable_binary_op("add", r0, bias_unsqueeze);
}
};
Expand Down
86 changes: 34 additions & 52 deletions test/onnx/onnx_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -440,14 +440,13 @@ TEST_CASE(batch_norm_flat_test)
auto mean = mm->add_parameter("mean", {migraphx::shape::float_type, {1}});
auto var = mm->add_parameter("variance", {migraphx::shape::float_type, {1}});

auto rt = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {1e-6f}});

auto numer = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {var, eps});
auto denom = add_common_op(*mm, migraphx::make_op("pow"), {var_eps, rt});
auto div0 = add_common_op(*mm, migraphx::make_op("div"), {numer, denom});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {div0, scale});
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {var, eps});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), {var_eps});
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
add_common_op(*mm, migraphx::make_op("add"), {r0, bias});

auto prog = optimize_onnx("batch_norm_flat_test.onnx");
Expand All @@ -465,14 +464,13 @@ TEST_CASE(batch_norm_rank_2_test)
auto mean = mm->add_parameter("mean", {migraphx::shape::float_type, {5}});
auto var = mm->add_parameter("variance", {migraphx::shape::float_type, {5}});

auto rt = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {1e-6f}});

auto numer = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {var, eps});
auto denom = add_common_op(*mm, migraphx::make_op("pow"), {var_eps, rt});
auto div0 = add_common_op(*mm, migraphx::make_op("div"), {numer, denom});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {div0, scale});
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {var, eps});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), {var_eps});
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
add_common_op(*mm, migraphx::make_op("add"), {r0, bias});

auto prog = optimize_onnx("batch_norm_rank_2_test.onnx");
Expand All @@ -490,19 +488,18 @@ TEST_CASE(batch_norm_1d_test)
auto mean = mm->add_parameter("mean", {migraphx::shape::float_type, {3}});
auto var = mm->add_parameter("variance", {migraphx::shape::float_type, {3}});

auto rt = mm->add_literal(migraphx::literal{migraphx::shape::half_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::half_type, {1e-5f}});

auto usq_scale = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), scale);
auto usq_bias = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), bias);
auto usq_mean = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), mean);
auto usq_var = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), var);

auto numer = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto denom = add_common_op(*mm, migraphx::make_op("pow"), {var_eps, rt});
auto div0 = add_common_op(*mm, migraphx::make_op("div"), {numer, denom});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {div0, usq_scale});
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), var_eps);
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {usq_scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
add_common_op(*mm, migraphx::make_op("add"), {r0, usq_bias});

auto prog = optimize_onnx("batch_norm_1d_test.onnx");
Expand All @@ -520,19 +517,18 @@ TEST_CASE(batch_norm_2d_test)
auto mean = mm->add_parameter("mean", {migraphx::shape::float_type, {3}});
auto var = mm->add_parameter("variance", {migraphx::shape::float_type, {3}});

auto rt = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {1e-5f}});

auto usq_scale = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), scale);
auto usq_bias = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), bias);
auto usq_mean = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), mean);
auto usq_var = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), var);

auto numer = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto denom = add_common_op(*mm, migraphx::make_op("pow"), {var_eps, rt});
auto div0 = add_common_op(*mm, migraphx::make_op("div"), {numer, denom});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {div0, usq_scale});
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), var_eps);
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {usq_scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
add_common_op(*mm, migraphx::make_op("add"), {r0, usq_bias});

auto prog = optimize_onnx("batch_norm_2d_test.onnx");
Expand All @@ -550,7 +546,6 @@ TEST_CASE(batch_norm_3d_test)
auto mean = mm->add_parameter("mean", {migraphx::shape::half_type, {2}});
auto var = mm->add_parameter("variance", {migraphx::shape::half_type, {2}});

auto rt = mm->add_literal(migraphx::literal{migraphx::shape::half_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::half_type, {1e-6f}});

auto usq_scale =
Expand All @@ -561,12 +556,13 @@ TEST_CASE(batch_norm_3d_test)
mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2, 3}}}), mean);
auto usq_var = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2, 3}}}), var);

auto numer = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto denom = add_common_op(*mm, migraphx::make_op("pow"), {var_eps, rt});
auto div0 = add_common_op(*mm, migraphx::make_op("div"), {numer, denom});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {div0, usq_scale});
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {x, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), var_eps);
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {usq_scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
add_common_op(*mm, migraphx::make_op("add"), {r0, usq_bias});

auto prog = optimize_onnx("batch_norm_3d_test.onnx");

EXPECT(p == prog);
Expand Down Expand Up @@ -908,7 +904,6 @@ TEST_CASE(constant_test)

TEST_CASE(constant_fill_test)
{

migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape s{migraphx::shape::float_type, {2, 3}};
Expand Down Expand Up @@ -1105,7 +1100,6 @@ TEST_CASE(conv_bn_relu_maxpool_test)
auto p5 = mm->add_parameter("5", {migraphx::shape::float_type, {1}});
auto p6 = mm->add_parameter("6", {migraphx::shape::float_type, {1}});

auto rt = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.5}});
auto eps = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {1e-5f}});

uint64_t axis = 1;
Expand All @@ -1120,25 +1114,12 @@ TEST_CASE(conv_bn_relu_maxpool_test)
auto usq_mean = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), p5);
auto usq_var = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), p6);

auto mb_mean = mm->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 28, 28}}}), usq_mean);
auto numer = mm->add_instruction(migraphx::make_op("sub"), l5, mb_mean);
auto mb_eps =
mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 1}}}), eps);
auto var_eps = mm->add_instruction(migraphx::make_op("add"), usq_var, mb_eps);
auto mb_rt =
mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 1}}}), rt);
auto denom = mm->add_instruction(migraphx::make_op("pow"), var_eps, mb_rt);
auto mb_denom = mm->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 28, 28}}}), denom);
auto div0 = mm->add_instruction(migraphx::make_op("div"), numer, mb_denom);
auto mb_scale = mm->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 28, 28}}}), usq_scale);
auto r0 = mm->add_instruction(migraphx::make_op("mul"), div0, mb_scale);
auto mb_bias = mm->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 28, 28}}}), usq_bias);

auto l6 = mm->add_instruction(migraphx::make_op("add"), r0, mb_bias);
auto x_sub_mean = add_common_op(*mm, migraphx::make_op("sub"), {l5, usq_mean});
auto var_eps = add_common_op(*mm, migraphx::make_op("add"), {usq_var, eps});
auto rsqrt = mm->add_instruction(migraphx::make_op("rsqrt"), var_eps);
auto mul0 = add_common_op(*mm, migraphx::make_op("mul"), {usq_scale, rsqrt});
auto r0 = add_common_op(*mm, migraphx::make_op("mul"), {x_sub_mean, mul0});
auto l6 = add_common_op(*mm, migraphx::make_op("add"), {r0, usq_bias});

auto l7 = mm->add_instruction(migraphx::make_op("relu"), l6);
mm->add_instruction(migraphx::make_op("pooling",
Expand Down Expand Up @@ -7150,7 +7131,8 @@ TEST_CASE(variable_batch_user_input_test6)

TEST_CASE(variable_batch_user_input_test7)
{
// if entry in map_dyn_input_dims is all fixed dynamic_dimensions, convert it to a static shape
// if entry in map_dyn_input_dims is all fixed dynamic_dimensions, convert it to a static
// shape
migraphx::program p;
auto* mm = p.get_main_module();
auto l0 = mm->add_parameter("0", migraphx::shape{migraphx::shape::float_type, {2, 3, 16, 16}});
Expand Down
18 changes: 6 additions & 12 deletions test/quantization.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -379,28 +379,23 @@ TEST_CASE(fp16_subgraph)
auto create_fp16_program = [] {
migraphx::program p;
auto* mm = p.get_main_module();
migraphx::shape sd{migraphx::shape::float_type, {1}};
auto l1 = mm->add_literal(migraphx::literal(sd, {1}));
auto l2 = mm->add_literal(migraphx::literal(sd, {2}));
auto l3 = mm->add_literal(migraphx::literal(sd, {3}));
migraphx::shape sd{migraphx::shape::half_type, {1}};
migraphx::shape sx{migraphx::shape::float_type, {1, 4}};
migraphx::shape sy{migraphx::shape::float_type, {3, 4}};
migraphx::shape sc{migraphx::shape::bool_type};
auto cond = mm->add_parameter("cond", sc);
auto x = mm->add_parameter("x", sx);
auto y = mm->add_parameter("y", sy);
auto* then_mod = p.create_module("If_6_if");
auto hl1 = then_mod->add_instruction(
migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), l1);
auto mhl1 = then_mod->add_instruction(
auto hl2 = then_mod->add_literal(migraphx::literal(sd, {2}));
auto hl1 = then_mod->add_literal(migraphx::literal(sd, {1}));
auto mhl1 = then_mod->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {1, 4}}}), hl1);
auto hx = then_mod->add_instruction(
migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), x);
auto ad = then_mod->add_instruction(migraphx::make_op("add"), hx, mhl1);
auto fad = then_mod->add_instruction(
migraphx::make_op("convert", {{"target_type", migraphx::shape::float_type}}), ad);
auto hl2 = then_mod->add_instruction(
migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), l2);
auto mhl2 = then_mod->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {3, 4}}}), hl2);
auto hy1 = then_mod->add_instruction(
Expand All @@ -411,9 +406,8 @@ TEST_CASE(fp16_subgraph)
then_mod->add_return({fad, fmu, mu});

auto* else_mod = p.create_module("If_6_else");
auto hl3 = else_mod->add_instruction(
migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), l3);
auto mhl3 = else_mod->add_instruction(
auto hl3 = else_mod->add_literal(migraphx::literal(sd, {3}));
auto mhl3 = else_mod->add_instruction(
migraphx::make_op("multibroadcast", {{"out_lens", {1, 4}}}), hl3);
auto hx2 = else_mod->add_instruction(
migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), x);
Expand Down
Loading
Loading