FP8 OCP to FP8 FNUZ on hardware with only FP8 FNUZ support (#3684)

NANOO is short for NAN On Overflow, the data type comes from this paper: https://arxiv.org/pdf/2206.02915
Implements the method written about in Convert OCP FP8 model to FNUZ model inside MIGraphX #2717
This pass must run before simplify_qdq so that the adjusted scales and zero points are propagated to after the quantized operator.
The test in test/fp8_ocp_to_nanoo.cpp checks the pass works with simplify_qdq and does the expected operations
The test in test/ref/fp8_ocp_to_nanoo.cpp checks the pass produces the same result before and after
I will make a separate PR that removes the gpu context changes to get the gfx number
Fixed the cpp_generator that was using __builtin_nan incorrectly

src/CMakeLists.txt

@@ -58,6 +58,7 @@ add_library(migraphx
     file_buffer.cpp
     fileutils.cpp
     fp_to_double.cpp
+    fp8_ocp_to_fnuz.cpp
     fuse_concat.cpp
     fuse_pointwise.cpp
     fuse_pointwise_reduce.cpp

src/cpp_generator.cpp

@@ -221,7 +221,7 @@ cpp_generator::function cpp_generator::generate_module(const module& m,
                             string_literal = "-__builtin_huge_val()";
                     }
                     else if(std::isnan(static_cast<double>(x)))
-                        string_literal = "__builtin_nan()";
+                        string_literal = "__builtin_nan(\"0\")";
                     else
                         string_literal = ins->get_literal().to_string();
                 });

src/fp8_ocp_to_fnuz.cpp

@@ -0,0 +1,178 @@
+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/fp8_ocp_to_fnuz.hpp>
+#include <migraphx/matcher.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/match/dq_helpers.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace {
+
+using fp8::fp8e4m3fnuz;
+
+std::unordered_set<std::string> get_quantizable_op_names()
+{
+    static std::unordered_set<std::string> s = {"convolution", "dot"};
+    return s;
+}
+
+struct match_fp8ocp_convert_to_fp8fnuz
+{
+    auto matcher() const
+    {
+        auto dq1 = match::arg(0)(
+            skip_post_dq_ops(match::dequantizelinear_op("scale1", "zp1").bind("dq1")));
+        auto dq2 = match::arg(1)(
+            skip_post_dq_ops(match::dequantizelinear_op("scale2", "zp2").bind("dq2")));
+        return match::name(get_quantizable_op_names())(dq1, dq2);
+    }
+
+    static auto bit_cast_and_handle_specials(module& m,
+                                             const instruction_ref dq,
+                                             const instruction_ref x,
+                                             const instruction_ref bits_0x80_lit,
+                                             const instruction_ref bits_0x7f_lit,
+                                             const instruction_ref bits_0xff_lit,
+                                             const instruction_ref bits_0x00_lit)
+    {
+        auto x_lens     = x->get_shape().lens();
+        auto cast_input = m.insert_instruction(
+            dq, make_op("bit_cast", {{"target_type", shape::fp8e4m3fnuz_type}}), x);
+        auto mb_bits_0x80_lit = m.insert_instruction(
+            dq, make_op("multibroadcast", {{"out_lens", x_lens}}), bits_0x80_lit);
+        auto mb_bits_0x7f_lit = m.insert_instruction(
+            dq, make_op("multibroadcast", {{"out_lens", x_lens}}), bits_0x7f_lit);
+        auto mb_bits_0xff_lit = m.insert_instruction(
+            dq, make_op("multibroadcast", {{"out_lens", x_lens}}), bits_0xff_lit);
+        auto mb_zero_lit = m.insert_instruction(
+            dq, make_op("multibroadcast", {{"out_lens", x_lens}}), bits_0x00_lit);
+        // negative zero in fp8e4m3fn to zero in fp8e4m3fnuz
+        // a == 0x80 ? 0x0 : a
+        auto is_neg_zero = m.insert_instruction(dq, make_op("equal"), cast_input, mb_bits_0x80_lit);
+        auto ret = m.insert_instruction(dq, make_op("where"), is_neg_zero, mb_zero_lit, cast_input);
+
+        // positive and negative NaN in fp8e4m3fn to NaN in fp8e4m3fnuz
+        // (a == 0x7f or a == 0xff) ? 0x80 : a
+        auto eq_0x7f = m.insert_instruction(dq, make_op("equal"), ret, mb_bits_0x7f_lit);
+
+        auto eq_0xff = m.insert_instruction(dq, make_op("equal"), ret, mb_bits_0xff_lit);
+
+        auto cond = m.insert_instruction(dq, make_op("logical_or"), eq_0x7f, eq_0xff);
+        ret       = m.insert_instruction(dq, make_op("where"), cond, mb_bits_0x80_lit, ret);
+        return ret;
+    }
+
+    // Add the same broadcast instructions after adjusted scales or
+    // adjusted zero points from after the originals. Similar to
+    // propagate_quantized_ins in simplify_qdq.
+    static auto propagate_broadcasts(module& m,
+                                     const instruction_ref adj,
+                                     const instruction_ref ori,
+                                     const instruction_ref start,
+                                     const instruction_ref insert_pt)
+    {
+        auto prev_ins = start;
+        std::vector<instruction_ref> ins_between;
+        // matcher skips continguous, multi/broadcasts and transposes, collect all those
+        // instructions
+        while(prev_ins != ori)
+        {
+            ins_between.push_back(prev_ins);
+            prev_ins = prev_ins->inputs().front();
+        }
+        auto ret = adj;
+        for(auto ins : reverse_iterator_for(ins_between))
+        {
+            ret = m.insert_instruction(insert_pt, (*ins)->get_operator(), {ret});
+        }
+        return ret;
+    }
+
+    static auto cast_to_fnuz(module& m,
+                             const instruction_ref dq,
+                             const instruction_ref input,
+                             const instruction_ref dq_scale,
+                             const instruction_ref dq_zp)
+    {
+        auto x                             = input;
+        std::vector<fp8e4m3fnuz> bits_0x80 = {fp8e4m3fnuz(0x80, fp8e4m3fnuz::from_bits())};
+        auto bits_0x80_lit = m.add_literal(shape{shape::fp8e4m3fnuz_type, {1}, {0}}, bits_0x80);
+
+        std::vector<fp8e4m3fnuz> bits_0x7f = {fp8e4m3fnuz(0x7f, fp8e4m3fnuz::from_bits())};
+        auto bits_0x7f_lit = m.add_literal(shape{shape::fp8e4m3fnuz_type, {1}, {0}}, bits_0x7f);
+
+        std::vector<fp8e4m3fnuz> bits_0xff = {fp8e4m3fnuz(0xff, fp8e4m3fnuz::from_bits())};
+        auto bits_0xff_lit = m.add_literal(shape{shape::fp8e4m3fnuz_type, {1}, {0}}, bits_0xff);
+
+        std::vector<fp8e4m3fnuz> bits_0x00 = {fp8e4m3fnuz(0x00, fp8e4m3fnuz::from_bits())};
+        auto bits_0x00_lit = m.add_literal(shape{shape::fp8e4m3fnuz_type, {1}, {0}}, bits_0x00);
+
+        x = bit_cast_and_handle_specials(
+            m, dq, x, bits_0x80_lit, bits_0x7f_lit, bits_0xff_lit, bits_0x00_lit);
+        auto adj_dq_zp = bit_cast_and_handle_specials(
+            m, dq, dq_zp, bits_0x80_lit, bits_0x7f_lit, bits_0xff_lit, bits_0x00_lit);
+
+        // adj_scale = 2 * scale
+        auto two_lit = m.add_literal(literal{shape{dq_scale->get_shape().type()}, {2}});
+        two_lit      = m.insert_instruction(
+            dq, make_op("multibroadcast", {{"out_lens", dq_scale->get_shape().lens()}}), two_lit);
+        auto adj_dq_scale = m.insert_instruction(dq, make_op("mul"), dq_scale, two_lit);
+
+        adj_dq_scale = propagate_broadcasts(m, adj_dq_scale, dq_scale, dq->inputs().at(1), dq);
+        adj_dq_zp    = propagate_broadcasts(m, adj_dq_zp, dq_zp, dq->inputs().at(2), dq);
+        m.replace_instruction(dq, make_op("dequantizelinear"), x, adj_dq_scale, adj_dq_zp);
+    }
+
+    auto apply(module& m, const match::matcher_result& r) const
+    {
+        auto dq1    = r.instructions["dq1"];
+        auto dq2    = r.instructions["dq2"];
+        auto scale1 = r.instructions["scale1"];
+        auto scale2 = r.instructions["scale2"];
+        auto zp1    = r.instructions["zp1"];
+        auto zp2    = r.instructions["zp2"];
+
+        std::set<migraphx::shape::type_t> supported_types = {migraphx::shape::fp8e4m3fn_type};
+        if(not contains(supported_types, dq1->inputs().front()->get_shape().type()) or
+           not contains(supported_types, dq2->inputs().front()->get_shape().type()))
+            return;
+
+        cast_to_fnuz(m, dq1, dq1->inputs().front(), scale1, zp1);
+        cast_to_fnuz(m, dq2, dq2->inputs().front(), scale2, zp2);
+    }
+};
+
+} // namespace
+
+void fp8_ocp_to_fnuz::apply(module_pass_manager& mpm) const
+{
+    module_ref mm = &mpm.get_module();
+    match::find_matches(*mm, match_fp8ocp_convert_to_fp8fnuz{});
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/include/migraphx/fp8_ocp_to_fnuz.hpp

@@ -0,0 +1,49 @@
+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#ifndef MIGRAPHX_GUARD_RTGLIB_FP8_OCP_TO_FNUZ_HPP
+#define MIGRAPHX_GUARD_RTGLIB_FP8_OCP_TO_FNUZ_HPP
+
+#include <migraphx/config.hpp>
+#include <migraphx/pass_manager.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+/**
+ * Convert fp8e4m3fn to fp8e4m3fnuz for hardware that only supports fp8e4m3fnuz data types
+ * intrinsically. Conversion uses the same bit representation and adjusts scaling factors at the
+ * dequantization. Using the same bit representation from fp8e4m3fn to fp8e4m3fnuz halves the
+ * floating point representation. This pass should run before simplify_qdq so that the scales and
+ * zero points calculated by simplify_qdq have the correct adjusted scaling factors
+ */
+struct MIGRAPHX_EXPORT fp8_ocp_to_fnuz
+{
+    std::string name() const { return "fp8_ocp_to_fnuz"; }
+    void apply(module_pass_manager& mpm) const;
+};
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/match/dq_helpers.hpp

@@ -0,0 +1,62 @@
+
+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#ifndef MIGRAPHX_GUARD_MATCH_DQ_HELPERS_HPP
+#define MIGRAPHX_GUARD_MATCH_DQ_HELPERS_HPP
+
+#include <migraphx/config.hpp>
+#include <migraphx/matcher.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace match {
+
+/**
+ * Find dequantizelinear (DQ) instruction with constant scale and zero point input
+ * while skipping broadcast instructions between DQ and scale/zero point. Used
+ * in simplify_qdq and fp8_ocp_to_fnuz.
+ */
+inline auto dequantizelinear_op(const std::string& scale, const std::string& zp)
+{
+    return match::name("dequantizelinear")(
+        match::arg(1)(match::skip_broadcasts(match::is_constant().bind(scale))),
+        match::arg(2)(match::skip_broadcasts(match::is_constant().bind(zp))));
+}
+
+/**
+ * Skip certain operators after DQ instruction.
+ * Used in simplify_qdq and fp8_ocp_to_fnuz.
+ */
+template <class... Ms>
+auto skip_post_dq_ops(Ms... ms)
+{
+    return match::skip(match::name(
+        "broadcast", "multibroadcast", "contiguous", "transpose", "reshape", "convert"))(ms...);
+}
+
+} // namespace match
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/op/bit_cast.hpp

@@ -80,6 +80,7 @@ struct bit_cast : unary<bit_cast>
             args[0].visit([&](auto input) {
                 using itype = typename decltype(input)::value_type;
                 if constexpr(sizeof(otype) == sizeof(itype))
+
                 {
                     par_transform(input.begin(), input.end(), output.begin(), [&](auto x) {
                         return migraphx::bit_cast<otype>(x);

src/simplify_qdq.cpp

@@ -36,18 +36,12 @@
 #include <migraphx/op/quant_dot.hpp>
 #include <migraphx/register_op.hpp>
 #include <migraphx/fp8_types.hpp>
+#include <migraphx/match/dq_helpers.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace {
 
-template <class... Ms>
-auto skip_post_dq_ops(Ms... ms)
-{
-    return match::skip(match::name(
-        "broadcast", "multibroadcast", "contiguous", "transpose", "reshape", "convert"))(ms...);
-}
-
 std::unordered_set<std::string> get_quantizable_op_names()
 {
     static std::unordered_set<std::string> s = {"convolution", "dot"};
@@ -117,20 +111,12 @@ struct match_find_quantizable_ops
         return qinp;
     }
 
-    static auto dequantizelinear_op(const std::string& scale, const std::string& zp)
-    {
-        return match::name("dequantizelinear")(
-            match::arg(0)(match::skip(match::name("quantizelinear"))(match::any())),
-            match::arg(1)(match::skip_broadcasts(match::is_constant().bind(scale))),
-            match::arg(2)(match::skip_broadcasts(match::is_constant().bind(zp))));
-    }
-
     auto matcher() const
     {
-        auto dq1 =
-            match::arg(0)(skip_post_dq_ops(dequantizelinear_op("scale1", "zp1").bind("dq1")));
-        auto dq2 =
-            match::arg(1)(skip_post_dq_ops(dequantizelinear_op("scale2", "zp2").bind("dq2")));
+        auto dq1 = match::arg(0)(
+            skip_post_dq_ops(match::dequantizelinear_op("scale1", "zp1").bind("dq1")));
+        auto dq2 = match::arg(1)(
+            skip_post_dq_ops(match::dequantizelinear_op("scale2", "zp2").bind("dq2")));
         return match::name(get_quantizable_op_names())(dq1, dq2);
     }
 
@@ -231,7 +217,9 @@ struct match_find_quantizable_ops
                    is_valid_qparam(zp1, out_lens, out_lens.size() - 2) and
                    is_valid_qparam(scale2, out_lens, out_lens.size() - 1) and
                    is_valid_qparam(zp2, out_lens, out_lens.size() - 1)))
+            {
                 return;
+            }
 
             // This implementation supports both arguments being per-axis affine quantized
             // In practice, inputs are per-tensor affine and weights are per-axis symmetric

src/targets/gpu/target.cpp

@@ -31,6 +31,7 @@
 #include <migraphx/eliminate_data_type.hpp>
 #include <migraphx/eliminate_identity.hpp>
 #include <migraphx/eliminate_pad.hpp>
+#include <migraphx/fp8_ocp_to_fnuz.hpp>
 #include <migraphx/fuse_concat.hpp>
 #include <migraphx/fuse_pointwise_reduce.hpp>
 #include <migraphx/inline_module.hpp>
@@ -179,6 +180,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         dead_code_elimination{},
         eliminate_identity{},
         dead_code_elimination{},
+        enable_pass(not gpu::gfx_has_fp8ocp_intrinsics() and gpu::gfx_has_fp8fnuz_intrinsics(), fp8_ocp_to_fnuz{}),
+        enable_pass(not gpu::gfx_has_fp8ocp_intrinsics() and gpu::gfx_has_fp8fnuz_intrinsics(), dead_code_elimination{}),
         simplify_qdq{},
         enable_pass(not mlir_enabled(), rewrite_quantization{}),
         dead_code_elimination{},