FP8 GPU implementation (#2455)

test/verify/test_literal_limits.cpp

@@ -26,6 +26,7 @@
 #include <migraphx/program.hpp>
 #include <migraphx/make_op.hpp>
 #include <limits>
+#include <type_traits>
 
 template <migraphx::shape::type_t Q, typename T>
 struct test_literal_limits : verify_program<test_literal_limits<Q, T>>
@@ -33,9 +34,13 @@ struct test_literal_limits : verify_program<test_literal_limits<Q, T>>
     migraphx::program create_program() const
     {
         migraphx::program p;
-        auto* mm          = p.get_main_module();
-        auto input_s      = migraphx::shape(Q, {3, 1});
-        auto infinity_val = std::numeric_limits<T>::infinity();
+        auto* mm     = p.get_main_module();
+        auto input_s = migraphx::shape(Q, {3, 1});
+        T infinity_val{0};
+        if constexpr(std::numeric_limits<T>::has_infinity and std::is_floating_point<T>{})
+        {
+            infinity_val = std::numeric_limits<T>::infinity();
+        }
         std::vector<T> s_data{
             infinity_val, static_cast<T>(-infinity_val), std::numeric_limits<T>::quiet_NaN()};
 
@@ -52,3 +57,4 @@ template struct test_literal_limits<migraphx::shape::double_type, double>;
 template struct test_literal_limits<migraphx::shape::half_type, migraphx::half>;
 template struct test_literal_limits<migraphx::shape::int32_type, int32_t>;
 template struct test_literal_limits<migraphx::shape::int8_type, int8_t>;
+template struct test_literal_limits<migraphx::shape::fp8e4m3fnuz_type, migraphx::fp8::fp8e4m3fnuz>;

test/verify/test_log.cpp

@@ -27,15 +27,20 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_log : verify_program<test_log>
+template <migraphx::shape::type_t DType>
+struct test_log : verify_program<test_log<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        migraphx::shape s{migraphx::shape::float_type, {6}};
+        migraphx::shape s{DType, {6}};
         auto x = mm->add_instruction(migraphx::make_op("abs"), mm->add_parameter("x", s));
         mm->add_instruction(migraphx::make_op("log"), x);
         return p;
     }
 };
+
+template struct test_log<migraphx::shape::float_type>;
+template struct test_log<migraphx::shape::half_type>;
+template struct test_log<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_min_max.cpp

@@ -46,7 +46,9 @@ struct test_min_max : verify_program<test_min_max<Op, T>>
 template struct test_min_max<migraphx::op::max, migraphx::shape::float_type>;
 template struct test_min_max<migraphx::op::max, migraphx::shape::half_type>;
 template struct test_min_max<migraphx::op::max, migraphx::shape::double_type>;
+template struct test_min_max<migraphx::op::max, migraphx::shape::fp8e4m3fnuz_type>;
 
 template struct test_min_max<migraphx::op::min, migraphx::shape::float_type>;
 template struct test_min_max<migraphx::op::min, migraphx::shape::half_type>;
 template struct test_min_max<migraphx::op::min, migraphx::shape::double_type>;
+template struct test_min_max<migraphx::op::min, migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_nearbyint.cpp

@@ -26,6 +26,7 @@
 #include <migraphx/program.hpp>
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/float8.hpp>
 
 template <class T>
 struct test_nearbyint : verify_program<test_nearbyint<T>>
@@ -45,3 +46,4 @@ struct test_nearbyint : verify_program<test_nearbyint<T>>
 
 template struct test_nearbyint<migraphx::half>;
 template struct test_nearbyint<float>;
+template struct test_nearbyint<migraphx::fp8::fp8e4m3fnuz>;

test/verify/test_pad.cpp

@@ -27,13 +27,14 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_pad : verify_program<test_pad>
+template <migraphx::shape::type_t DType>
+struct test_pad : verify_program<test_pad<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        migraphx::shape s0{migraphx::shape::int32_type, {1, 96, 165, 165}};
+        migraphx::shape s0{DType, {1, 96, 165, 165}};
         std::vector<int64_t> pads0 = {0, 0, 0, 0, 0, 0, 1, 1};
         std::vector<int64_t> pads1 = {0, 0, 0, 0, 1, 1, 1, 1};
         std::vector<int64_t> pads2 = {1, 1, 1, 1, 0, 0, 0, 0};
@@ -46,3 +47,8 @@ struct test_pad : verify_program<test_pad>
         return p;
     }
 };
+
+template struct test_pad<migraphx::shape::int32_type>;
+template struct test_pad<migraphx::shape::float_type>;
+template struct test_pad<migraphx::shape::half_type>;
+template struct test_pad<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_pow.cpp

@@ -27,13 +27,15 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_pow : verify_program<test_pow>
+template <typename CType>
+struct test_pow : verify_program<test_pow<CType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
-        auto* mm = p.get_main_module();
-        migraphx::shape s{migraphx::shape::float_type, {6}};
+        migraphx::shape::type_t dtype = migraphx::shape::get_type<CType>();
+        auto* mm                      = p.get_main_module();
+        migraphx::shape s{dtype, {6}};
         std::vector<float> vec_e(s.elements(), 2.0f);
         auto b = mm->add_parameter("x", s);
         auto e = mm->add_literal(migraphx::literal(s, vec_e));
@@ -41,3 +43,6 @@ struct test_pow : verify_program<test_pow>
         return p;
     }
 };
+template struct test_pow<float>;
+template struct test_pow<migraphx::half>;
+template struct test_pow<migraphx::fp8::fp8e4m3fnuz>;

test/verify/test_reduce_add.cpp

@@ -27,14 +27,16 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 #include <migraphx/instruction.hpp>
+#include <migraphx/shape.hpp>
 
-struct test_reduce_add : verify_program<test_reduce_add>
+template <migraphx::shape::type_t DType>
+struct test_reduce_add : verify_program<test_reduce_add<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        migraphx::shape s{migraphx::shape::float_type, {4, 1000, 2, 2}};
+        migraphx::shape s{DType, {4, 1000, 2, 2}};
         migraphx::shape bs{migraphx::shape::half_type, {1, 32, 128}};
         auto x = mm->add_parameter("x", s);
         auto reduce_mean =
@@ -46,3 +48,6 @@ struct test_reduce_add : verify_program<test_reduce_add>
         return p;
     };
 };
+
+template struct test_reduce_add<migraphx::shape::float_type>;
+template struct test_reduce_add<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_reduce_mean_nhwc.cpp

@@ -28,14 +28,14 @@
 #include <migraphx/make_op.hpp>
 #include <migraphx/instruction.hpp>
 
-struct test_reduce_mean_nhwc : verify_program<test_reduce_mean_nhwc>
+template <migraphx::shape::type_t DType>
+struct test_reduce_mean_nhwc : verify_program<test_reduce_mean_nhwc<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
-        auto* mm = p.get_main_module();
-        auto s   = migraphx::shape::from_permutation(
-            migraphx::shape::float_type, {4, 256, 2, 2}, {0, 2, 3, 1});
+        auto* mm    = p.get_main_module();
+        auto s      = migraphx::shape::from_permutation(DType, {4, 256, 2, 2}, {0, 2, 3, 1});
         auto x      = mm->add_parameter("x", s);
         auto reduce = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {1}}}), x);
         auto abs    = mm->add_instruction(migraphx::make_op("abs"), reduce);
@@ -44,3 +44,7 @@ struct test_reduce_mean_nhwc : verify_program<test_reduce_mean_nhwc>
         return p;
     };
 };
+
+template struct test_reduce_mean_nhwc<migraphx::shape::float_type>;
+template struct test_reduce_mean_nhwc<migraphx::shape::half_type>;
+template struct test_reduce_mean_nhwc<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_reduce_op_large.cpp

@@ -51,6 +51,22 @@ template struct test_reduce_op_large<migraphx::op::reduce_min, 1, migraphx::shap
 template struct test_reduce_op_large<migraphx::op::reduce_prod, 2, migraphx::shape::float_type>;
 template struct test_reduce_op_large<migraphx::op::reduce_sum, 1, migraphx::shape::float_type>;
 
+template struct test_reduce_op_large<migraphx::op::reduce_max,
+                                     1,
+                                     migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_reduce_op_large<migraphx::op::reduce_mean,
+                                     1,
+                                     migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_reduce_op_large<migraphx::op::reduce_min,
+                                     1,
+                                     migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_reduce_op_large<migraphx::op::reduce_prod,
+                                     2,
+                                     migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_reduce_op_large<migraphx::op::reduce_sum,
+                                     1,
+                                     migraphx::shape::fp8e4m3fnuz_type>;
+
 struct test_reduce_mean_1 : verify_program<test_reduce_mean_1>
 {
     migraphx::program create_program() const

test/verify/test_reduce_op_small.cpp

@@ -56,3 +56,19 @@ template struct test_reduce_op_small<migraphx::op::reduce_mean, 2, migraphx::sha
 template struct test_reduce_op_small<migraphx::op::reduce_max, 2, migraphx::shape::half_type>;
 template struct test_reduce_op_small<migraphx::op::reduce_min, 2, migraphx::shape::half_type>;
 template struct test_reduce_op_small<migraphx::op::reduce_prod, -2, migraphx::shape::half_type>;
+
+template struct test_reduce_op_small<migraphx::op::reduce_sum,
+                                     2,
+                                     migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_reduce_op_small<migraphx::op::reduce_mean,
+                                     2,
+                                     migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_reduce_op_small<migraphx::op::reduce_max,
+                                     2,
+                                     migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_reduce_op_small<migraphx::op::reduce_min,
+                                     2,
+                                     migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_reduce_op_small<migraphx::op::reduce_prod,
+                                     -2,
+                                     migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_roialign.cpp

@@ -27,15 +27,16 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_roialign : verify_program<test_roialign>
+template <migraphx::shape::type_t DType>
+struct test_roialign : verify_program<test_roialign<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        migraphx::shape x_s{migraphx::shape::float_type, {5, 4, 10, 10}};
+        migraphx::shape x_s{DType, {5, 4, 10, 10}};
 
-        migraphx::shape roi_s{migraphx::shape::float_type, {5, 4}};
+        migraphx::shape roi_s{DType, {5, 4}};
 
         migraphx::shape ind_s{migraphx::shape::int64_type, {5}};
         std::vector<int64_t> ind_vec = {0, 2, 3, 4, 1};
@@ -44,10 +45,10 @@ struct test_roialign : verify_program<test_roialign>
         auto roi = mm->add_parameter("roi", roi_s);
         auto ind = mm->add_literal(migraphx::literal(ind_s, ind_vec));
         auto r   = mm->add_instruction(migraphx::make_op("roialign",
-                                                       {{"spatial_scale", 1.0},
-                                                        {"output_height", 5},
-                                                        {"output_width", 5},
-                                                        {"sampling_ratio", 2}}),
+                                                         {{"spatial_scale", 1.0},
+                                                          {"output_height", 5},
+                                                          {"output_width", 5},
+                                                          {"sampling_ratio", 2}}),
                                      x,
                                      roi,
                                      ind);
@@ -56,3 +57,7 @@ struct test_roialign : verify_program<test_roialign>
         return p;
     }
 };
+
+template struct test_roialign<migraphx::shape::float_type>;
+template struct test_roialign<migraphx::shape::half_type>;
+template struct test_roialign<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_rsqrt.cpp

@@ -23,22 +23,26 @@
  */
 
 #include "verify_program.hpp"
+#include <migraphx/float8.hpp>
 #include <migraphx/program.hpp>
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_rsqrt : verify_program<test_rsqrt>
+template <typename CType>
+struct test_rsqrt : verify_program<test_rsqrt<CType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
-        auto* mm = p.get_main_module();
+        auto* mm                      = p.get_main_module();
+        migraphx::shape::type_t dtype = migraphx::shape::get_type<CType>();
         std::vector<size_t> input_lens{1, 3, 16, 16};
-        migraphx::shape s{migraphx::shape::float_type, input_lens};
+        migraphx::shape s{dtype, input_lens};
         auto x       = mm->add_parameter("x", s);
-        auto min_val = mm->add_literal(1.0f);
-        auto max_val = mm->add_literal(std::numeric_limits<float>::max());
-        min_val      = mm->add_instruction(
+        auto min_val = mm->add_literal(migraphx::literal{migraphx::shape{dtype}, {1.0}});
+        auto max_val = mm->add_literal(
+            migraphx::literal{migraphx::shape{dtype}, {std::numeric_limits<CType>::max()}});
+        min_val = mm->add_instruction(
             migraphx::make_op("multibroadcast", {{"out_lens", input_lens}}), min_val);
         max_val = mm->add_instruction(
             migraphx::make_op("multibroadcast", {{"out_lens", input_lens}}), max_val);
@@ -47,3 +51,7 @@ struct test_rsqrt : verify_program<test_rsqrt>
         return p;
     };
 };
+
+template struct test_rsqrt<float>;
+template struct test_rsqrt<migraphx::half>;
+template struct test_rsqrt<migraphx::fp8::fp8e4m3fnuz>;

test/verify/test_scatternd.cpp

@@ -25,18 +25,19 @@
 #include <migraphx/program.hpp>
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/shape.hpp>
 
-struct test_scatternd : verify_program<test_scatternd>
+template <migraphx::shape::type_t DType>
+struct test_scatternd : verify_program<test_scatternd<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
         auto* mm   = p.get_main_module();
-        auto dtype = migraphx::shape::float_type;
         auto itype = migraphx::shape::int64_type;
-        migraphx::shape ds{dtype, {1}};
+        migraphx::shape ds{DType, {1}};
         migraphx::shape is{itype, {4, 1}};
-        migraphx::shape us{dtype, {4}};
+        migraphx::shape us{DType, {4}};
         std::vector<int64_t> ind_vec{4, 3, 1, 7};
 
         auto ld = mm->add_literal(migraphx::literal{ds, {1}});
@@ -51,3 +52,7 @@ struct test_scatternd : verify_program<test_scatternd>
         return p;
     }
 };
+
+template struct test_scatternd<migraphx::shape::float_type>;
+template struct test_scatternd<migraphx::shape::half_type>;
+template struct test_scatternd<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_sin.cpp

@@ -27,15 +27,20 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_sin : verify_program<test_sin>
+template <migraphx::shape::type_t DType>
+struct test_sin : verify_program<test_sin<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        migraphx::shape s{migraphx::shape::float_type, {10}};
+        migraphx::shape s{DType, {10}};
         auto x = mm->add_parameter("x", s);
         mm->add_instruction(migraphx::make_op("sin"), x);
         return p;
     }
 };
+
+template struct test_sin<migraphx::shape::float_type>;
+template struct test_sin<migraphx::shape::half_type>;
+template struct test_sin<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_sinh.cpp

@@ -27,15 +27,20 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_sinh : verify_program<test_sinh>
+template <migraphx::shape::type_t DType>
+struct test_sinh : verify_program<test_sinh<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        migraphx::shape s{migraphx::shape::float_type, {16}};
+        migraphx::shape s{DType, {16}};
         auto x = mm->add_parameter("x", s);
         mm->add_instruction(migraphx::make_op("sinh"), x);
         return p;
     }
 };
+
+template struct test_sinh<migraphx::shape::float_type>;
+template struct test_sinh<migraphx::shape::half_type>;
+template struct test_sinh<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_softmax.cpp

@@ -48,3 +48,7 @@ template struct test_softmax<0, migraphx::shape::half_type>;
 template struct test_softmax<1, migraphx::shape::half_type>;
 template struct test_softmax<2, migraphx::shape::half_type>;
 template struct test_softmax<3, migraphx::shape::half_type>;
+template struct test_softmax<0, migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_softmax<1, migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_softmax<2, migraphx::shape::fp8e4m3fnuz_type>;
+template struct test_softmax<3, migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_sqrt.cpp

@@ -27,16 +27,21 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_sqrt : verify_program<test_sqrt>
+template <migraphx::shape::type_t DType>
+struct test_sqrt : verify_program<test_sqrt<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        migraphx::shape s{migraphx::shape::float_type, {2, 3, 4, 6}};
+        migraphx::shape s{DType, {2, 3, 4, 6}};
         auto param     = mm->add_parameter("x", s);
         auto param_abs = mm->add_instruction(migraphx::make_op("abs"), param);
         mm->add_instruction(migraphx::make_op("sqrt"), param_abs);
         return p;
     }
 };
+
+template struct test_sqrt<migraphx::shape::float_type>;
+template struct test_sqrt<migraphx::shape::half_type>;
+template struct test_sqrt<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_tan.cpp

@@ -27,15 +27,20 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_tan : verify_program<test_tan>
+template <migraphx::shape::type_t DType>
+struct test_tan : verify_program<test_tan<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        migraphx::shape s{migraphx::shape::float_type, {16}};
+        migraphx::shape s{DType, {16}};
         auto x = mm->add_parameter("x", s);
         mm->add_instruction(migraphx::make_op("tan"), x);
         return p;
     }
 };
+
+template struct test_tan<migraphx::shape::float_type>;
+template struct test_tan<migraphx::shape::half_type>;
+template struct test_tan<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_tanh.cpp

@@ -27,14 +27,19 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_tanh : verify_program<test_tanh>
+template <migraphx::shape::type_t DType>
+struct test_tanh : verify_program<test_tanh<DType>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        auto x = mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, {4, 3, 3, 3}});
+        auto x   = mm->add_parameter("x", migraphx::shape{DType, {4, 3, 3, 3}});
         mm->add_instruction(migraphx::make_op("tanh"), x);
         return p;
     }
 };
+
+template struct test_tanh<migraphx::shape::float_type>;
+template struct test_tanh<migraphx::shape::half_type>;
+template struct test_tanh<migraphx::shape::fp8e4m3fnuz_type>;

test/verify/test_where.cpp

@@ -27,7 +27,8 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_where : verify_program<test_where>
+template <migraphx::shape::type_t DType>
+struct test_where : verify_program<test_where<DType>>
 {
     migraphx::program create_program() const
     {
@@ -44,3 +45,7 @@ struct test_where : verify_program<test_where>
         return p;
     };
 };
+
+template struct test_where<migraphx::shape::float_type>;
+template struct test_where<migraphx::shape::half_type>;
+template struct test_where<migraphx::shape::fp8e4m3fnuz_type>;