Add build flag for fast math (#639)

* add flag

* formatting

* remove env variable

* fix api expression

* add api test

* add api test

* add op test

* formatting

* fix function name

* fix syntax

* formatting

* modify test

* remove test and update doc

* move test to new file

* formatting

* revert test files

* rewrite check

* New
Co-authored-by: Paul Fultz II <pfultz2@yahoo.com>

doc/src/reference/py.rst

@@ -151,12 +151,13 @@ program
 
     :rtype: shape
 
-.. py:method:: compile(t, offload_copy=True)
+.. py:method:: compile(t, offload_copy=True, fast_math=True)
 
     Compiles the program for the target and optimizes it.
 
     :param target t: This is the target to compile the program for.
     :param bool offload_copy: For targets with offloaded memory(such as the gpu), this will insert instructions during compilation to copy the input parameters to the offloaded memory and to copy the final result from the offloaded memory back to main memory.
+    :param bool fast_math: Optimize math functions to use faster approximate versions. There may be slight accuracy degredation when enabled.
 
 .. py:method:: run(params)
 

src/api/api.cpp

@@ -70,6 +70,7 @@ migraphx::compile_options to_compile_options(const migraphx_compile_options& opt
 {
     migraphx::compile_options result{};
     result.offload_copy = options.offload_copy;
+    result.fast_math    = options.fast_math;
     return result;
 }
 

src/api/include/migraphx/migraphx.h

@@ -42,6 +42,7 @@ typedef enum {
 typedef struct
 {
     bool offload_copy;
+    bool fast_math;
 } migraphx_compile_options;
 
 typedef struct

src/driver/main.cpp

@@ -221,6 +221,7 @@ struct compiler
     program_params parameters;
     bool gpu          = true;
     bool offload_copy = false;
+    bool fast_math    = true;
     int quantize      = 0;
 
     std::vector<std::string> fill0;
@@ -235,6 +236,10 @@ struct compiler
            {"--enable-offload-copy"},
            ap.help("Enable implicit offload copying"),
            ap.set_value(true));
+        ap(fast_math,
+           {"--disable-fast-math"},
+           ap.help("Disable fast math optimization"),
+           ap.set_value(false));
         ap(quantize, {"--fp16"}, ap.help("Quantize for fp16"), ap.set_value(q_fp16));
         ap(quantize, {"--int8"}, ap.help("Quantize for int8"), ap.set_value(q_int8));
     }
@@ -261,6 +266,7 @@ struct compiler
         }
         compile_options options;
         options.offload_copy = offload_copy;
+        options.fast_math    = fast_math;
         p.compile(t, options);
         l.save(p);
         return p;
@@ -300,6 +306,7 @@ struct verify : command<verify>
     bool per_instruction = false;
     bool reduce          = false;
     bool offload_copy    = false;
+    bool fast_math       = true;
     void parse(argument_parser& ap)
     {
         l.parse(ap);
@@ -308,6 +315,10 @@ struct verify : command<verify>
            {"--enable-offload-copy"},
            ap.help("Enable implicit offload copying"),
            ap.set_value(true));
+        ap(fast_math,
+           {"--disable-fast-math"},
+           ap.help("Disable fast math optimization"),
+           ap.set_value(false));
         ap(tolerance, {"--tolerance"}, ap.help("Tolerance for errors"));
         ap(per_instruction,
            {"-i", "--per-instruction"},
@@ -324,7 +335,7 @@ struct verify : command<verify>
 
         compile_options options;
         options.offload_copy = offload_copy;
-
+        options.fast_math    = fast_math;
         auto m               = parameters.generate(p, false);
 
         if(per_instruction)

src/include/migraphx/compile_options.hpp

@@ -10,6 +10,7 @@ inline namespace MIGRAPHX_INLINE_NS {
 struct compile_options
 {
     bool offload_copy = false;
+    bool fast_math    = true;
     tracer trace{};
 };
 

src/py/migraphx_py.cpp

@@ -172,14 +172,17 @@ PYBIND11_MODULE(migraphx, m)
         .def("get_parameter_names", &migraphx::program::get_parameter_names)
         .def("get_parameter_shapes", &migraphx::program::get_parameter_shapes)
         .def("get_output_shapes", &migraphx::program::get_output_shapes)
-        .def("compile",
-             [](migraphx::program& p, const migraphx::target& t, bool offload_copy) {
+        .def(
+            "compile",
+            [](migraphx::program& p, const migraphx::target& t, bool offload_copy, bool fast_math) {
                 migraphx::compile_options options;
                 options.offload_copy = offload_copy;
+                options.fast_math    = fast_math;
                 p.compile(t, options);
             },
             py::arg("t"),
-             py::arg("offload_copy") = true)
+            py::arg("offload_copy") = true,
+            py::arg("fast_math")    = true)
         .def("run",
              [](migraphx::program& p, py::dict params) {
                  migraphx::program::parameter_map pm;

src/targets/gpu/fuse_ops.cpp

@@ -29,7 +29,6 @@ inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DISABLE_MIOPEN_FUSION)
-MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DISABLE_FAST_GELU)
 
 struct fusion
 {
@@ -396,6 +395,7 @@ struct find_add_gelu
 
 struct find_gelu_new
 {
+    bool fast_math = true;
 
     static auto pow_fn()
     {
@@ -430,7 +430,7 @@ struct find_gelu_new
         auto x_ins = r.instructions["x"];
         auto args  = ins->inputs();
 
-        if(enabled(MIGRAPHX_DISABLE_FAST_GELU{}))
+        if(not fast_math)
             p.replace_instruction(ins, hip_gelu_new{}, x_ins, args.back());
         else
             p.replace_instruction(ins, hip_gelu{}, x_ins, args.back());
@@ -807,7 +807,7 @@ struct find_commutative_broadcast
 
 void fuse_ops::apply(program& p) const
 {
-    match::find_matches(p, find_gelu{}, find_gelu_new{});
+    match::find_matches(p, find_gelu{}, find_gelu_new{fast_math});
     run_passes(p, {dead_code_elimination{}});
     match::find_matches(p, find_triadd{});
     match::find_matches(p,

src/targets/gpu/include/migraphx/gpu/fuse_ops.hpp

@@ -13,6 +13,7 @@ namespace gpu {
 struct fuse_ops
 {
     context* ctx   = nullptr;
+    bool fast_math = true;
     std::string name() const { return "gpu::fuse_ops"; }
     void apply(program& p) const;
 };

src/targets/gpu/target.cpp

@@ -73,7 +73,7 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         dead_code_elimination{},
         pack_int8_args{},
         dead_code_elimination{},
-        fuse_ops{&ctx},
+        fuse_ops{&ctx, options.fast_math},
         dead_code_elimination{},
         write_literals{&ctx},
         schedule{gpu::schedule_model{ctx.get_current_device().nstreams()}, not enabled(MIGRAPHX_DISABLE_SCHEDULE_PASS{})},

test/gpu/fast_math.cpp

+#include <test.hpp>
+#include <migraphx/quantization.hpp>
+#include <migraphx/iterator_for.hpp>
+#include "test_utils.hpp"
+#include "test.hpp"
+
+migraphx::program create_gelu()
+{
+    migraphx::program p;
+    std::vector<float> data0 = {0.044715};
+    std::vector<float> data1 = {0.797885};
+    std::vector<float> data2 = {3};
+    std::vector<float> data3 = {0.5};
+    migraphx::shape s0{migraphx::shape::float_type, {1}};
+
+    std::vector<size_t> x_dims{1, 1, 5};
+
+    auto x         = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, x_dims});
+    auto const_val = p.add_literal(migraphx::literal{s0, data0});
+    auto sqrt_2_pi = p.add_literal(migraphx::literal{s0, data1});
+    auto three_val = p.add_literal(migraphx::literal{s0, data2});
+    auto half_val  = p.add_literal(migraphx::literal{s0, data3});
+
+    auto mbcast_3         = p.add_instruction(migraphx::op::multibroadcast{x_dims}, three_val);
+    auto pow_op           = p.add_instruction(migraphx::op::pow{}, x, mbcast_3);
+    auto mbcast_const     = p.add_instruction(migraphx::op::multibroadcast{x_dims}, const_val);
+    auto mul_const        = p.add_instruction(migraphx::op::mul{}, mbcast_const, pow_op);
+    auto add_x            = p.add_instruction(migraphx::op::add{}, x, mul_const);
+    auto mbcast_sqrt_2_pi = p.add_instruction(migraphx::op::multibroadcast{x_dims}, sqrt_2_pi);
+    auto mul_add_x        = p.add_instruction(migraphx::op::mul{}, mbcast_sqrt_2_pi, add_x);
+    auto tanh_op          = p.add_instruction(migraphx::op::tanh{}, mul_add_x);
+    auto mbcast_half      = p.add_instruction(migraphx::op::multibroadcast{x_dims}, half_val);
+    auto mul_half         = p.add_instruction(migraphx::op::mul{}, mbcast_half, tanh_op);
+    auto add_mul_half     = p.add_instruction(migraphx::op::add{}, mul_half, mbcast_half);
+    auto mul_x            = p.add_instruction(migraphx::op::mul{}, x, add_mul_half);
+    p.add_return({mul_x});
+
+    return p;
+}
+
+TEST_CASE(enable_fast_gelu)
+{
+    migraphx::program p = create_gelu();
+    p.compile(migraphx::gpu::target{});
+    CHECK(any_of(p, [&](auto&& i) { return i.name() == "gpu::gelu"; }));
+}
+
+TEST_CASE(disable_fast_gelu)
+{
+    migraphx::program p = create_gelu();
+    migraphx::compile_options options;
+    options.fast_math = false;
+    p.compile(migraphx::gpu::target{}, options);
+    CHECK(any_of(p, [&](auto&& i) { return i.name() == "gpu::gelu_new"; }));
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

tools/api/api.cpp

@@ -70,6 +70,7 @@ migraphx::compile_options to_compile_options(const migraphx_compile_options& opt
 {
     migraphx::compile_options result{};
     result.offload_copy = options.offload_copy;
+    result.fast_math    = options.fast_math;
     return result;
 }
 

tools/api/migraphx.h

@@ -42,6 +42,7 @@ typedef enum {
 typedef struct
 {
     bool offload_copy;
+    bool fast_math;
 } migraphx_compile_options;
 
 typedef struct