frigate/docker/rocm/migraphx/targets/gpu/prefuse_ops.cpp

/*
 * 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/matcher.hpp>
#include <migraphx/permutation.hpp>
#include <migraphx/gpu/prefuse_ops.hpp>
#include <migraphx/gpu/gemm_softmax_gemm.hpp>
#include <migraphx/match/layernorm.hpp>
#include <migraphx/register_op.hpp>
#include <migraphx/pass_manager.hpp>
#include <migraphx/dead_code_elimination.hpp>
#include <migraphx/op/group_query_attention.hpp>
#ifdef MIGRAPHX_USE_COMPOSABLEKERNEL
#include <migraphx/gpu/ck.hpp>
#endif
#include <migraphx/gpu/fuse_mlir.hpp>

namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {

MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DISABLE_LAYERNORM_FUSION);

namespace {

template <class Derived, std::size_t N>
struct layernorm_base
{
    float epsilon = 1e-12f;
    template <class Self, class F>
    static auto reflect(Self& self, F f)
    {
        return pack(f(self.epsilon, "epsilon"));
    }
    shape compute_shape(std::vector<shape> inputs, std::vector<module_ref> mods) const
    {
        std::size_t nargs = N;
        if(not mods.empty())
        {
            auto* pm = mods.front();
            nargs += pm->get_parameter_names().size() - 1;
        }
        check_shapes{inputs, static_cast<const Derived&>(*this)}.has(nargs);
        auto s = inputs.front();
        auto t = s.type();
        if(not mods.empty())
            t = mods.front()->get_output_shapes().front().type();

        // Scalar output if all inputs are scalar
        if(inputs.front().elements() == 1 and
           all_of(inputs, [](const auto& ss) { return ss.scalar(); }))
            return inputs.front();
        auto l_s = shape::from_permutation(
            t, s.lens(), find_permutation(std::vector<shape>(inputs.begin(), inputs.begin() + N)));
        // just prelayernorm or preadd_layernorm
        if(nargs <= N)
            return l_s;
        // else, layernorm + pointwise fusion, preserve layout of fused op
        std::vector<shape> lp_s(inputs.begin() + N, inputs.end());
        lp_s.insert(lp_s.begin(), l_s);
        return shape::from_permutation(t, s.lens(), find_permutation(lp_s));
    }
};

struct layernorm : layernorm_base<layernorm, 1>
{

    std::string name() const { return "gpu::prelayernorm"; }
};
MIGRAPHX_REGISTER_OP(layernorm);

struct add_layernorm : layernorm_base<add_layernorm, 2>
{
    std::string name() const { return "gpu::preadd_layernorm"; }
};
MIGRAPHX_REGISTER_OP(add_layernorm);

struct find_layernorm
{
    auto matcher() const { return match::layernorm(); }

    void apply(module& m, const match::matcher_result& r) const
    {
        auto ins   = r.result;
        auto x_ins = r.instructions["x"];
        float eps  = 0;
        if(contains(r.instructions, "eps"))
            eps = r.instructions["eps"]->eval().at<float>();

        m.replace_instruction(ins, layernorm{eps}, x_ins);
    }
};

struct find_add_layernorm
{
    auto matcher() const
    {
        return match::name("gpu::prelayernorm")(
            match::args(match::name("add")(match::used_once()).bind("add")));
    }

    void apply(module& m, const match::matcher_result& r) const
    {
        auto ins     = r.result;
        auto add_ins = r.instructions["add"];
        auto op      = any_cast<layernorm>(ins->get_operator());

        m.replace_instruction(ins, add_layernorm{op.epsilon}, add_ins->inputs());
    }
};

struct pre_gemm_softmax_gemm : gemm_softmax_gemm
{
    std::string name() const { return "gpu::pre_gemm_softmax_gemm"; }
};
MIGRAPHX_REGISTER_OP(pre_gemm_softmax_gemm);

auto is_ck_gemm()
{
    return match::make_basic_pred_matcher([=](instruction_ref ins) {
#ifdef MIGRAPHX_USE_COMPOSABLEKERNEL
        if(not enabled(MIGRAPHX_ENABLE_CK{}))
            return false;
        if(ins->name() != "dot")
            return false;
        if(not pre_gemm_softmax_gemm::is_ck_supported_type(ins->get_shape().type()))
            return false;
        return true;
#else
        (void)ins;
        return false;
#endif
    });
}

auto is_test_gemm(bool enable_attention)
{
    return match::make_basic_pred_matcher([=](instruction_ref ins) {
        if(ins->name() != "dot")
            return false;
        return enable_attention;
    });
}

auto is_bias_supported()
{
    return match::make_basic_pred_matcher([=](instruction_ref) {
#ifdef MIGRAPHX_USE_COMPOSABLEKERNEL
        return not enabled(MIGRAPHX_ENABLE_CK{});
#else
        return true;
#endif
    });
}

struct find_gemm_softmax_gemm
{
    bool enable_attention = false;

    auto matcher() const
    {
        auto gemm1 = match::skip(match::name("contiguous"))(match::name("dot")(
            match::any_of(is_ck_gemm(), is_test_gemm(enable_attention)).bind("gemm1")));
        auto mul   = match::name("mul")(
            match::nargs(2), match::either_arg(0, 1)(match::is_constant().bind("scale"), gemm1));
        auto where = match::name("where")(match::arg(2)(match::is_constant().bind("select_const")),
                                          match::arg(1)(mul),
                                          match::arg(0)(match::any().bind("select_cond")));
        auto add =
            match::name("add")(is_bias_supported(),
                               match::nargs(2),
                               match::either_arg(0, 1)(match::none_of(mul).bind("bias"), mul));
        auto softmax = match::name("softmax")(match::arg(0)(match::any_of(mul, add, gemm1, where)))
                           .bind("softmax");

        return match::name("dot")(
            match::any_of(is_ck_gemm(), is_test_gemm(enable_attention)).bind("gemm2"))(
            match::arg(0)(softmax));
    }

    void apply(module_pass_manager& mpm, const match::matcher_result& r) const
    {
        auto ins       = r.result;
        auto gemm2_ins = r.instructions["gemm2"];
        auto gemm1_ins = r.instructions["gemm1"];

        float scale = 1.0;
        if(contains(r.instructions, "scale"))
        {
            auto scale_lit = r.instructions["scale"];
            // CK only supports single-valued scale
            scale_lit->eval().visit([&](const auto s) {
                // CK only supports single-valued scale
                if(not std::all_of(
                       s.begin() + 1, s.end(), [&](auto v) { return float_equal(v, s.front()); }))
                    return;
                scale = s.front();
            });
        }

        auto inputs = gemm1_ins->inputs(); // A, B
        if(contains(r.instructions, "select_cond"))
        {
            inputs.push_back(r.instructions["select_cond"]);
            inputs.push_back(r.instructions["select_const"]);
        }
        if(contains(r.instructions, "bias"))
        {
            inputs.push_back(r.instructions["bias"]);
        }

        inputs.push_back(gemm2_ins->inputs().back()); // B1

        mpm.get_module().replace_instruction(
            ins, pre_gemm_softmax_gemm{gemm2_ins->get_operator(), scale}, inputs);
    }
};

struct gpu_compute_attention_probabilities : op::group_query_attention
{
    std::string name() const { return "gpu::compute_attention_probabilities"; }

    shape compute_shape(std::vector<shape> inputs) const
    {
        auto query_lens        = inputs.front().lens();
        auto present_kv_seqlen = inputs.at(1).lens().at(2);
        std::vector<std::size_t> output_lens{
            query_lens.at(0), num_heads, query_lens.at(2), present_kv_seqlen};
        shape output_shape{inputs.front().type(), output_lens};
        return output_shape;
    }
};
MIGRAPHX_REGISTER_OP(gpu_compute_attention_probabilities);

struct gpu_compute_attention_scores : op::group_query_attention
{
    std::string name() const { return "gpu::compute_attention_scores"; }

    shape compute_shape(std::vector<shape> inputs) const
    {
        auto query_lens = inputs.front().lens();
        std::size_t q_hidden_size =
            (query_lens[1] * query_lens[3] * num_heads) / (num_heads + 2 * kv_num_heads);
        std::vector<std::size_t> output_lens{query_lens.at(0), query_lens.at(2), q_hidden_size};
        shape output_shape{inputs.front().type(), output_lens};
        return output_shape;
    }
};
MIGRAPHX_REGISTER_OP(gpu_compute_attention_scores);

struct gpu_gqa_rotary_embedding : op::group_query_attention
{
    std::string name() const { return "gpu::gqa_rotary_embedding"; }

    shape compute_shape(std::vector<shape> inputs) const { return inputs.front(); }
};
MIGRAPHX_REGISTER_OP(gpu_gqa_rotary_embedding);

struct gpu_gqa_softmax : op::group_query_attention
{
    std::string name() const { return "gpu::gqa_softmax"; }

    shape compute_shape(std::vector<shape> inputs) const { return inputs.at(2); }
};
MIGRAPHX_REGISTER_OP(gpu_gqa_softmax);

struct gpu_concat_past_present : op::group_query_attention
{
    std::string name() const { return "gpu::concat_past_present"; }

    shape compute_shape(std::vector<shape> inputs) const { return inputs[0]; }
};
MIGRAPHX_REGISTER_OP(gpu_concat_past_present);

struct find_group_query_attention
{
    auto matcher() const { return match::name("group_query_attention"); }

    void apply(module_pass_manager& mpm, const match::matcher_result& r) const
    {
        auto ins    = r.result;
        auto inputs = ins->inputs();
        auto v      = ins->get_operator().to_value();

        auto num_heads          = v.at("num_heads").to<std::size_t>();
        auto kv_num_heads       = v.at("kv_num_heads").to<std::size_t>();
        auto do_rotary          = v.at("do_rotary").to<bool>();
        auto local_window_size  = v.at("local_window_size").to<int>();
        auto rotary_interleaved = v.at("rotary_interleaved").to<bool>();
        auto scale              = v.at("scale").to<float>();

        auto q_shape                      = inputs[0]->get_shape();
        auto q_lens                       = q_shape.lens();
        const std::size_t batch_size      = q_lens[0];
        const std::size_t sequence_length = q_lens[1];
        std::size_t q_hidden_size         = q_lens[2];
        std::size_t head_size             = q_hidden_size / (num_heads + 2 * kv_num_heads);

        std::vector<std::size_t> bsnh{
            batch_size, sequence_length, num_heads + 2 * kv_num_heads, head_size};

        auto transposed_qkv = mpm.get_module().insert_instruction(
            ins, make_op("reshape", {{"dims", bsnh}}), inputs.at(0));

        transposed_qkv = mpm.get_module().insert_instruction(
            ins, make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), transposed_qkv);

        auto rotary_qkv = transposed_qkv;
        if(do_rotary)
        {
            std::vector<instruction_ref> rotary_inputs{
                transposed_qkv, inputs.at(5), inputs.at(7), inputs.at(8)};
            rotary_qkv =
                mpm.get_module().insert_instruction(ins,
                                                    gpu_gqa_rotary_embedding{do_rotary,
                                                                             kv_num_heads,
                                                                             local_window_size,
                                                                             num_heads,
                                                                             rotary_interleaved,
                                                                             scale},
                                                    rotary_inputs);
        }

        auto pres_k = inputs.at(3);
        auto pres_v = inputs.at(4);
        std::vector<instruction_ref> concat_inputs{rotary_qkv, pres_k, pres_v, inputs.at(5)};

        auto concat = mpm.get_module().insert_instruction(
            ins,
            gpu_concat_past_present{
                do_rotary, kv_num_heads, local_window_size, num_heads, rotary_interleaved, scale},
            concat_inputs);
        auto id =
            mpm.get_module().insert_instruction(ins, make_op("identity"), concat, pres_k, pres_v);

        std::vector<instruction_ref> attn_probs_inputs{id, pres_k, pres_v, inputs.at(5)};
        auto attn_probs = mpm.get_module().insert_instruction(
            ins,
            gpu_compute_attention_probabilities{
                do_rotary, kv_num_heads, local_window_size, num_heads, rotary_interleaved, scale},
            attn_probs_inputs);

        std::vector<instruction_ref> softmax_inputs{rotary_qkv, pres_k, attn_probs, inputs.at(5)};
        auto softmax = mpm.get_module().insert_instruction(
            ins,
            gpu_gqa_softmax{
                do_rotary, kv_num_heads, local_window_size, num_heads, rotary_interleaved, scale},
            softmax_inputs);
        std::vector<instruction_ref> new_inputs{rotary_qkv, pres_k, pres_v, inputs.at(5), softmax};

        auto get_tuple_elm_0 = std::next(ins);
        auto get_tuple_elm_1 = std::next(get_tuple_elm_0);
        auto get_tuple_elm_2 = std::next(get_tuple_elm_1);
        mpm.get_module().replace_instruction(get_tuple_elm_2, pres_v);
        mpm.get_module().replace_instruction(get_tuple_elm_1, pres_k);
        mpm.get_module().replace_instruction(
            get_tuple_elm_0,
            gpu_compute_attention_scores{
                do_rotary, kv_num_heads, local_window_size, num_heads, rotary_interleaved, scale},
            new_inputs);
    }
};

} // namespace

void prefuse_ops::apply(module_pass_manager& mpm) const
{
    if(not enabled(MIGRAPHX_DISABLE_LAYERNORM_FUSION{}))
    {
        match::find_matches(mpm.get_module(), find_layernorm{});
        mpm.run_pass(dead_code_elimination{});
        match::find_matches(mpm.get_module(), find_add_layernorm{});
    }
    match::find_matches(mpm, find_gemm_softmax_gemm{enable_attention});
    match::find_matches(mpm, find_group_query_attention{});
}

} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx