Merge branch 'develop' into eliminate-more-contiguous

src/targets/cpu/gemm.cpp

@@ -44,13 +44,9 @@ struct is_fast_gemm_type<float> : std::true_type
 {
 };
 
-template <class T>
-void migemm_impl(tensor_view<T> cmat,
-                 tensor_view<T> amat,
-                 tensor_view<T> bmat,
-                 float alpha,
-                 float beta,
-                 std::true_type)
+template <class T, class F>
+void migemm_impl(
+    tensor_view<T> cmat, tensor_view<T> amat, tensor_view<T> bmat, F alpha, F beta, std::true_type)
 {
     visit_mat(amat, [&](const auto& a) {
         visit_mat(bmat, [&](const auto& b) {
@@ -66,13 +62,9 @@ void migemm_impl(tensor_view<T> cmat,
     });
 }
 
-template <class T>
-void migemm_impl(tensor_view<T> cmat,
-                 tensor_view<T> amat,
-                 tensor_view<T> bmat,
-                 float alpha,
-                 float beta,
-                 std::false_type)
+template <class T, class F>
+void migemm_impl(
+    tensor_view<T> cmat, tensor_view<T> amat, tensor_view<T> bmat, F alpha, F beta, std::false_type)
 {
     std::size_t n_dims = cmat.get_shape().lens().size();
     std::size_t dim_0  = n_dims - 2;
@@ -95,9 +87,8 @@ void migemm_impl(tensor_view<T> cmat,
     });
 }
 
-template <class T>
-void migemm_impl(
-    tensor_view<T> cmat, tensor_view<T> amat, tensor_view<T> bmat, float alpha, float beta)
+template <class T, class F>
+void migemm_impl(tensor_view<T> cmat, tensor_view<T> amat, tensor_view<T> bmat, F alpha, F beta)
 {
     auto lens = amat.get_shape().lens();
     bool batch_mul =
@@ -113,13 +104,29 @@ void migemm_impl(
     }
 }
 
-void migemm(
-    const argument& c_arg, const argument& a_arg, const argument& b_arg, float alpha, float beta)
+template <class F>
+void migemm_tpl(
+    const argument& c_arg, const argument& a_arg, const argument& b_arg, F alpha, F beta)
 {
     visit_all(c_arg, a_arg, b_arg)(
         [&](auto cmat, auto amat, auto bmat) { migemm_impl(cmat, amat, bmat, alpha, beta); });
 }
 
+void migemm(
+    const argument& c_arg, const argument& a_arg, const argument& b_arg, float alpha, float beta)
+{
+    migemm_tpl(c_arg, a_arg, b_arg, alpha, beta);
+}
+
+void migemm(const argument& c_arg,
+            const argument& a_arg,
+            const argument& b_arg,
+            int32_t alpha,
+            int32_t beta)
+{
+    migemm_tpl(c_arg, a_arg, b_arg, alpha, beta);
+}
+
 } // namespace cpu
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/targets/cpu/include/migraphx/cpu/gemm.hpp

@@ -10,6 +10,11 @@ namespace cpu {
 
 void migemm(
     const argument& c_arg, const argument& a_arg, const argument& b_arg, float alpha, float beta);
+void migemm(const argument& c_arg,
+            const argument& a_arg,
+            const argument& b_arg,
+            int32_t alpha,
+            int32_t beta);
 
 } // namespace cpu
 } // namespace MIGRAPHX_INLINE_NS

src/targets/cpu/lowering.cpp

@@ -4,7 +4,9 @@
 #include <migraphx/dfor.hpp>
 #include <migraphx/op/batch_norm.hpp>
 #include <migraphx/op/convolution.hpp>
+#include <migraphx/op/quant_convolution.hpp>
 #include <migraphx/op/dot.hpp>
+#include <migraphx/op/quant_dot.hpp>
 #include <migraphx/op/elu.hpp>
 #include <migraphx/op/im2col.hpp>
 #include <migraphx/op/leaky_relu.hpp>
@@ -216,6 +218,61 @@ struct cpu_convolution
     }
 };
 
+struct cpu_quant_convolution
+{
+    op::quant_convolution op;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
+    std::string name() const { return "cpu::quant_convolution"; }
+    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
+    argument compute(context&, shape output_shape, std::vector<argument> args) const
+    {
+        argument result{output_shape};
+        auto output = result.get<int32_t>();
+        visit_all(args[0], args[1])([&](auto input, auto weights) {
+            auto in   = input.get_shape().lens();
+            auto in_h = in[2];
+            auto in_w = in[3];
+
+            auto wei   = weights.get_shape().lens();
+            auto wei_n = wei[0];
+            auto wei_c = wei[1];
+            auto wei_h = wei[2];
+            auto wei_w = wei[3];
+
+            par_dfor(output_shape.lens()[0],
+                     output_shape.lens()[1],
+                     output_shape.lens()[2],
+                     output_shape.lens()[3])(
+                [&](std::size_t o, std::size_t w, std::size_t i, std::size_t j) {
+                    const auto start_x  = i * op.stride[0] - op.padding[0];
+                    const auto start_y  = j * op.stride[1] - op.padding[1];
+                    const auto group_id = w / (wei_n / op.group);
+
+                    int32_t acc = 0;
+                    dfor(wei_c, wei_h, wei_w)([&](std::size_t k, std::size_t x, std::size_t y) {
+                        const auto in_x  = start_x + x;
+                        const auto in_y  = start_y + y;
+                        const auto in_ch = group_id * wei_c + k;
+                        if(in_x >= 0 && in_x < in_h && in_y >= 0 && in_y < in_w)
+                        {
+                            acc += static_cast<int32_t>(input(o, in_ch, in_x, in_y)) *
+                                   weights(w, k, x, y);
+                        }
+                    });
+                    output(o, w, i, j) = acc;
+                });
+        });
+
+        return result;
+    }
+};
+
 struct cpu_im2col
 {
     op::im2col op;
@@ -245,17 +302,17 @@ struct cpu_im2col
             const std::size_t& stride_h = op.stride[0];
             const std::size_t& stride_w = op.stride[1];
 
-            auto kdiv2_h = kernel_h / 2;
-            auto kdiv2_w = kernel_w / 2;
+            long kdiv2_h = long(kernel_h) / 2;
+            long kdiv2_w = long(kernel_w) / 2;
             // calculate output sizes
             const std::size_t col_height = (height - kernel_h + 2 * pad_h) / stride_h + 1;
             const std::size_t col_width  = (width - kernel_w + 2 * pad_w) / stride_w + 1;
             // account for padding for the starting position of the input pixels
-            std::size_t iinput = kdiv2_h - pad_h;
+            long iinput = kdiv2_h - long(pad_h);
             // loop over output pixels (ioutput, joutput)
             for(std::size_t ioutput = 0; ioutput < col_height; ioutput++, iinput += stride_h)
             {
-                std::size_t jinput = kdiv2_w - pad_w;
+                long jinput = kdiv2_w - long(pad_w);
                 for(std::size_t joutput = 0; joutput < col_width; joutput++, jinput += stride_w)
                 {
                     // compute linear index for output
@@ -264,8 +321,8 @@ struct cpu_im2col
                     dfor(channels,
                          kernel_h,
                          kernel_w)([&](std::size_t c, std::size_t koffset, std::size_t loffset) {
-                        auto idx    = iinput + koffset - kdiv2_h;
-                        auto jdx    = jinput + loffset - kdiv2_w;
+                        auto idx    = iinput + long(koffset) - kdiv2_h;
+                        auto jdx    = jinput + long(loffset) - kdiv2_w;
                         col(ldx, p) = ((idx >= 0) && (idx < height) && (jdx >= 0) && (jdx < width))
                                           ? input(0, c, idx, jdx)
                                           : 0;
@@ -433,7 +490,7 @@ struct cpu_gemm
     {
         argument result{output_shape};
         // 3 inputs, it is alpha * A * B + beta * C, then
-        // A and B are matrics, and C is broadcastable to A * B
+        // A and B are matrices, and C is of the same shape as A * B
         if(args.size() == 3)
         {
             // no need to consider the value of args[2]
@@ -460,13 +517,79 @@ struct cpu_gemm
     }
 };
 
+struct cpu_quant_gemm
+{
+    op::quant_dot op;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
+    std::string name() const { return "cpu::quant_dot"; }
+    shape compute_shape(const std::vector<shape>& inputs) const
+    {
+        if(inputs.size() == 3)
+        {
+            auto c_shape = inputs.at(2);
+            check_shapes{{c_shape}}.not_broadcasted();
+        }
+        return op.compute_shape(inputs);
+    }
+
+    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
+    {
+        argument result{output_shape};
+        // 3 inputs, it is alpha * A * B + beta * C, then
+        // A and B are matrices, and C is of the same shape to A * B
+
+        // first, convert the args[0] and args[1] from int8_t to int32_t
+        argument arg_0{{shape::int32_type, {args.at(0).get_shape().lens()}}};
+        argument arg_1{{shape::int32_type, {args.at(1).get_shape().lens()}}};
+        arg_0.visit([&](auto output) {
+            args.at(0).visit(
+                [&](auto input) { std::copy(input.begin(), input.end(), output.begin()); });
+        });
+
+        arg_1.visit([&](auto output) {
+            args.at(1).visit(
+                [&](auto input) { std::copy(input.begin(), input.end(), output.begin()); });
+        });
+
+        if(args.size() == 3)
+        {
+            // no need to consider the value of args[2]
+            if(op.beta == 0)
+            {
+                result.visit([&](auto output) { std::fill(output.begin(), output.end(), 0); });
+            }
+            else
+            {
+                visit_all(result, args[2])([&](auto output, auto input) {
+                    std::copy(input.begin(), input.end(), output.begin());
+                });
+            }
+
+            migemm(result, arg_0, arg_1, op.alpha, op.beta);
+
+            return result;
+        }
+
+        // 2 input arguments
+        migemm(result, arg_0, arg_1, op.alpha, int32_t{0});
+
+        return result;
+    }
+};
+
 struct leaky_relu_op
 {
     op::leaky_relu op;
     std::string name() const { return "cpu::leaky_relu"; }
     auto fcn() const
     {
-        auto& a = op.alpha;
+        auto a = op.alpha;
         return [a](auto x) { return x > 0 ? x : x * a; };
     }
 };
@@ -477,7 +600,7 @@ struct elu_op
     std::string name() const { return "cpu::elu"; }
     auto fcn() const
     {
-        auto& a = op.alpha;
+        auto a = op.alpha;
         return [a](auto x) { return x > 0 ? x : a * std::expm1(x); };
     }
 };
@@ -671,15 +794,17 @@ struct cpu_apply
     {
         apply_map["batch_norm_inference"] =
             extend_op<cpu_batch_norm_inference, op::batch_norm_inference>();
-        apply_map["convolution"] = extend_op<cpu_convolution, op::convolution>();
-        apply_map["dot"]         = extend_op<cpu_gemm, op::dot>();
-        apply_map["elu"]         = extend_op<cpu_unary<elu_op>, op::elu>();
-        apply_map["im2col"]      = extend_op<cpu_im2col, op::im2col>();
-        apply_map["leaky_relu"]  = extend_op<cpu_unary<leaky_relu_op>, op::leaky_relu>();
-        apply_map["logsoftmax"]  = extend_op<cpu_logsoftmax, op::logsoftmax>();
-        apply_map["lrn"]         = extend_op<cpu_lrn, op::lrn>();
-        apply_map["pad"]         = extend_op<cpu_pad, op::pad>();
-        apply_map["softmax"]     = extend_op<cpu_softmax, op::softmax>();
+        apply_map["convolution"]       = extend_op<cpu_convolution, op::convolution>();
+        apply_map["dot"]               = extend_op<cpu_gemm, op::dot>();
+        apply_map["quant_dot"]         = extend_op<cpu_quant_gemm, op::quant_dot>();
+        apply_map["quant_convolution"] = extend_op<cpu_quant_convolution, op::quant_convolution>();
+        apply_map["elu"]               = extend_op<cpu_unary<elu_op>, op::elu>();
+        apply_map["im2col"]            = extend_op<cpu_im2col, op::im2col>();
+        apply_map["leaky_relu"]        = extend_op<cpu_unary<leaky_relu_op>, op::leaky_relu>();
+        apply_map["logsoftmax"]        = extend_op<cpu_logsoftmax, op::logsoftmax>();
+        apply_map["lrn"]               = extend_op<cpu_lrn, op::lrn>();
+        apply_map["pad"]               = extend_op<cpu_pad, op::pad>();
+        apply_map["softmax"]           = extend_op<cpu_softmax, op::softmax>();
     }
 
     void apply()

src/targets/gpu/CMakeLists.txt

@@ -34,12 +34,14 @@ add_library(migraphx_device
     device/contiguous.cpp
     device/logsoftmax.cpp
     device/softmax.cpp
+    device/sigmoid.cpp
     device/convert.cpp
     device/mul.cpp
     device/concat.cpp
     device/pad.cpp
     device/gather.cpp
     device/sub.cpp
+    device/int8_gemm_pack.cpp
     device/div.cpp
     device/clip.cpp
     device/reduce_sum.cpp
@@ -65,8 +67,10 @@ add_library(migraphx_gpu
     target.cpp
     lowering.cpp
     gemm.cpp
+    quant_gemm.cpp
     pooling.cpp
     convolution.cpp
+    quant_convolution.cpp
     softmax.cpp
     logsoftmax.cpp
     contiguous.cpp
@@ -75,17 +79,20 @@ add_library(migraphx_gpu
     batchnorm.cpp
     write_literals.cpp
     rocblas.cpp
-    sigmoid.cpp
     abs.cpp
     elu.cpp
     pad.cpp
     gather.cpp
+    convert.cpp
     lrn.cpp
     schedule_model.cpp
     adjust_allocation.cpp
+    pack_int8_args.cpp
     clip.cpp
     reduce_sum.cpp
     reduce_mean.cpp
+    int8_gemm_pack.cpp
+    int8_conv_pack.cpp
 )
 set_target_properties(migraphx_gpu PROPERTIES EXPORT_NAME gpu)
 rocm_clang_tidy_check(migraphx_gpu)

src/targets/gpu/convert.cpp

+#include <migraphx/gpu/convert.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/device/convert.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+shape hip_convert::compute_shape(std::vector<shape> inputs) const
+{
+    inputs.pop_back();
+    check_shapes{inputs}.packed();
+    return op.compute_shape(inputs);
+}
+
+argument hip_convert::compute(context& ctx, const shape&, const std::vector<argument>& args) const
+{
+    device::convert(ctx.get_stream().get(), args[1], args[0]);
+    return args[1];
+}
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/include/migraphx/gpu/device/reduce.hpp

@@ -155,8 +155,8 @@ __device__ void dpp_reduce(T& in, Op op)
 
 __device__ inline void dpp_reduce(float& x, sum)
 {
-#ifdef MIGRAPHX_USE_CLANG_TIDY
-    (void)x;
+#if defined(MIGRAPHX_USE_CLANG_TIDY) || defined(CPPCHECK)
+    x = 1;
 #else
     __asm__ volatile("s_nop 4\n"
                      "v_add_f32 %0 %0 %0 row_shr:1\n"
@@ -245,8 +245,7 @@ void reduce_standard_impl(hipStream_t stream,
                           T init,
                           Input read_input,
                           Output read_output,
-                          std::size_t relements,
-                          std::size_t stride)
+                          std::size_t relements)
 {
     hip_visit_all(result, arg)([&](auto output, auto input) {
         auto nelements = result.get_shape().elements();
@@ -255,7 +254,7 @@ void reduce_standard_impl(hipStream_t stream,
         const std::size_t block_size     = compute_block_size(relements, max_block_size);
         gs_launch(stream, nelements * block_size, block_size)([=](auto i, auto idx) __device__ {
             const auto out_idx  = i / block_size;
-            const auto base_idx = out_idx * stride;
+            const auto base_idx = out_idx * relements;
             auto r = block_reduce<max_block_size>(idx, op, init, relements, [&](auto j) __device__ {
                 return read_input(input.data()[base_idx + j]);
             });
@@ -276,25 +275,15 @@ void reduce(hipStream_t stream,
 {
     auto&& output_shape = result.get_shape();
     auto&& input_shape  = arg.get_shape();
+    assert(output_shape.lens().size() == input_shape.lens().size());
     if(input_shape.standard() and output_shape.standard() and
        output_shape.lens().back() != input_shape.lens().back() and
        std::equal(output_shape.lens().begin(),
                   std::prev(output_shape.lens().end()),
                   input_shape.lens().begin()))
     {
-        std::size_t stride = std::accumulate(input_shape.strides().begin(),
-                                             input_shape.strides().end(),
-                                             1,
-                                             std::multiplies<size_t>());
-        reduce_standard_impl(stream,
-                             result,
-                             arg,
-                             op,
-                             init,
-                             read_input,
-                             read_output,
-                             input_shape.lens().back(),
-                             stride);
+        reduce_standard_impl(
+            stream, result, arg, op, init, read_input, read_output, input_shape.lens().back());
     }
     else
     {

src/targets/gpu/device/include/migraphx/gpu/device/tensor.hpp

@@ -31,6 +31,7 @@ struct hip_tensor_descriptor
             result[is] = tidx / strides[is];
             tidx       = tidx % strides[is];
         }
+
         return result;
     }
     __device__ __host__ std::size_t linear(hip_tensor_index<NDim> s) const

src/targets/gpu/device/int8_gemm_pack.cpp

+#include <migraphx/shape.hpp>
+#include <migraphx/argument.hpp>
+#include <migraphx/gpu/device/int8_gemm_pack.hpp>
+#include <migraphx/gpu/device/launch.hpp>
+#include <migraphx/gpu/device/types.hpp>
+#include <migraphx/gpu/device/tensor.hpp>
+#include <migraphx/gpu/hip.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void int8_gemm_pack_a(hipStream_t stream, const argument& result, const argument& arg)
+{
+    auto comp_shape    = arg.get_shape();
+    auto out_lens      = comp_shape.lens();
+    auto dim_0         = out_lens.size() - 2;
+    auto dim_1         = out_lens.size() - 1;
+    std::size_t lda    = comp_shape.strides()[dim_0];
+    std::size_t m_size = out_lens[dim_0] * out_lens[dim_1];
+    visit_all(result, arg)([&](auto output, auto input) {
+        std::size_t nelements = comp_shape.elements();
+        auto* out_ptr         = device_cast(output.data());
+        auto* in_ptr          = device_cast(input.data());
+        visit_tensor_size(out_lens.size(), [&](auto out_dim) {
+            hip_tensor_descriptor<out_dim> desc(comp_shape);
+            gs_launch(stream, nelements, 256)([=](auto ii) {
+                const size_t nb    = 4;
+                auto idx           = desc.multi(ii);
+                std::size_t i_m    = idx[dim_1];
+                std::size_t i_k    = idx[dim_0];
+                std::size_t offset = ii / m_size * m_size;
+                out_ptr[i_k % nb + (i_m + (i_k / nb) * lda) * nb + offset] =
+                    in_ptr[i_m + i_k * lda + offset];
+            });
+        });
+    });
+}
+
+void int8_gemm_pack_b(hipStream_t stream, const argument& result, const argument& arg)
+{
+    auto trans_shape = arg.get_shape();
+    auto out_lens    = trans_shape.lens();
+    auto dim_0       = trans_shape.lens().size() - 2;
+    auto dim_1       = trans_shape.lens().size() - 1;
+    std::size_t ldb  = trans_shape.strides()[dim_1];
+
+    auto wrap_lens = out_lens;
+    std::swap(wrap_lens[dim_0], wrap_lens[dim_1]);
+    shape comp_shape{trans_shape.type(), wrap_lens};
+    std::size_t m_size = out_lens[dim_0] * out_lens[dim_1];
+    visit_all(result, arg)([&](auto output, auto input) {
+        std::size_t nelements = comp_shape.elements();
+        auto* out_ptr         = device_cast(output.data());
+        auto* in_ptr          = device_cast(input.data());
+        visit_tensor_size(out_lens.size(), [&](auto out_dim) {
+            hip_tensor_descriptor<out_dim> desc(comp_shape);
+            gs_launch(stream, nelements, 256)([=](auto ii) {
+                const size_t nb    = 4;
+                auto idx           = desc.multi(ii);
+                std::size_t i_n    = idx[dim_1];
+                std::size_t i_k    = idx[dim_0];
+                std::size_t offset = ii / m_size * m_size;
+                out_ptr[i_k % nb + (i_n + (i_k / nb) * ldb) * nb + offset] =
+                    in_ptr[i_n + i_k * ldb + offset];
+            });
+        });
+    });
+}
+
+void sync_stream(hipStream_t stream) { hipStreamSynchronize(stream); }
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/sigmoid.cpp

+#include <migraphx/gpu/device/sigmoid.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+#include <migraphx/gpu/device/types.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void sigmoid(hipStream_t stream, const argument& result, const argument& arg)
+{
+    nary(stream, result, arg)([](auto x) { return 1.f / (1.f + ::exp(to_hip_type(-x))); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/gemm.cpp

@@ -233,6 +233,10 @@ argument miopen_gemm::compute(context& ctx,
         auto to_pointer = [&](auto&& arg) { return to_rocblas_type(as.from(arg.data())); };
         if(num_matrices == 1)
         {
+            // the rocblas_gemm API handles inputs and output matrices as
+            // column-major format. When doing a C = A * B, we actually do
+            // C^T = (B^T) * (A^T). That is the reason we input args[1] as
+            // A and args[0] as B in calling the rocblas_gemm.
             generic_rocblas_gemm(as,
                                  ctx.get_stream().get_rocblas(),
                                  transb ? rocblas_operation_transpose : rocblas_operation_none,

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

@@ -3,8 +3,6 @@
 
 #include <migraphx/shape.hpp>
 #include <migraphx/op/convert.hpp>
-#include <migraphx/gpu/oper.hpp>
-#include <migraphx/gpu/device/convert.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -12,7 +10,7 @@ namespace gpu {
 
 struct context;
 
-struct hip_convert : unary_device<hip_convert, device::convert>
+struct hip_convert
 {
     op::convert op;
 
@@ -22,13 +20,15 @@ struct hip_convert : unary_device<hip_convert, device::convert>
         return migraphx::reflect(self.op, f);
     }
 
-    hip_convert(op::convert oper) : op(oper) {}
+    std::string name() const { return "gpu::convert"; }
 
-    shape compute_shape(std::vector<shape> inputs) const
+    shape compute_shape(std::vector<shape> inputs) const;
+
+    argument compute(context& ctx, const shape&, const std::vector<argument>& args) const;
+
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
     {
-        inputs.pop_back();
-        check_shapes{inputs}.packed();
-        return op.compute_shape(inputs);
+        return shapes.size() - 1;
     }
 };
 

src/targets/gpu/include/migraphx/gpu/device/int8_gemm_pack.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_INT8_GEMM_PACK_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_INT8_GEMM_PACK_HPP
+
+#include <migraphx/argument.hpp>
+#include <migraphx/config.hpp>
+#include <hip/hip_runtime_api.h>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void int8_gemm_pack_a(hipStream_t stream, const argument& result, const argument& arg);
+
+void int8_gemm_pack_b(hipStream_t stream, const argument& result, const argument& arg);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/include/migraphx/gpu/device/sigmoid.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_SIGMOID_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_SIGMOID_HPP
+
+#include <migraphx/argument.hpp>
+#include <migraphx/config.hpp>
+#include <hip/hip_runtime_api.h>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void sigmoid(hipStream_t stream, const argument& result, const argument& arg);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/include/migraphx/gpu/int8_conv_pack.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_INT8_CONV_PACK_HPP
+#define MIGRAPHX_GUARD_RTGLIB_INT8_CONV_PACK_HPP
+
+#include <migraphx/op/quant_dot.hpp>
+#include <migraphx/config.hpp>
+#include <utility>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct context;
+
+struct miopen_int8_conv_pack
+{
+    std::string name() const { return "gpu::int8_conv_pack"; }
+    shape compute_shape(const std::vector<shape>& inputs) const;
+    argument compute(context& ctx, const shape&, const std::vector<argument>& args) const;
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/include/migraphx/gpu/int8_gemm_pack.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_INT8_GEMM_PACK_HPP
+#define MIGRAPHX_GUARD_RTGLIB_INT8_GEMM_PACK_HPP
+
+#include <migraphx/op/quant_dot.hpp>
+#include <migraphx/config.hpp>
+#include <utility>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct context;
+
+struct hip_int8_gemm_pack_a
+{
+    std::string name() const { return "gpu::int8_gemm_pack_a"; }
+    shape compute_shape(const std::vector<shape>& inputs) const;
+    argument compute(context& ctx, const shape&, const std::vector<argument>& args) const;
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
+};
+
+struct hip_int8_gemm_pack_b
+{
+    std::string name() const { return "gpu::int8_gemm_pack_b"; }
+    shape compute_shape(const std::vector<shape>& inputs) const;
+    argument compute(context& ctx, const shape&, const std::vector<argument>& args) const;
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

@@ -34,11 +34,11 @@ Result make_obj(F f, Ts... xs)
     auto status                = f(&x, xs...);
     Result r{x};
     if(status != miopenStatusSuccess)
-        MIGRAPHX_THROW("MIOpen call failed");
+        MIGRAPHX_THROW("MAKE_OBJ: MIOpen call failed");
     return r;
 }
 
-inline tensor_descriptor make_tensor(const migraphx::shape& s)
+inline tensor_descriptor make_tensor(const migraphx::shape& s, bool pack = false)
 {
     auto t = make_obj<tensor_descriptor>(&miopenCreateTensorDescriptor);
     // Convert to ints
@@ -49,13 +49,33 @@ inline tensor_descriptor make_tensor(const migraphx::shape& s)
         d = miopenFloat;
     else if(s.type() == shape::half_type)
         d = miopenHalf;
+    else if(s.type() == shape::int32_type)
+        d = miopenInt32;
+    else if(s.type() == shape::int8_type)
+    {
+        if(pack)
+        {
+            // update the lens and corresponding strides
+            d          = miopenInt8x4;
+            lens[1]    = ((lens[1] + 3) / 4) * 4;
+            strides[0] = strides[1] * lens[1];
+        }
+        else
+        {
+            d = miopenInt8;
+        }
+    }
     else
-        MIGRAPHX_THROW("Unsupported type");
+    {
+        MIGRAPHX_THROW("MAKE_TENSOR: unsupported type");
+    }
     miopenSetTensorDescriptor(t.get(), d, s.lens().size(), lens.data(), strides.data());
+
     return t;
 }
 
-inline convolution_descriptor make_conv(const migraphx::op::convolution& op)
+template <class T>
+inline convolution_descriptor make_conv(const T& op)
 {
     auto c = make_obj<convolution_descriptor>(&miopenCreateConvolutionDescriptor);
     miopenConvolutionMode_t c_mode = miopenConvolution;

src/targets/gpu/include/migraphx/gpu/pack_int8_args.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_PACK_INT8_ARGS_HPP
+#define MIGRAPHX_GUARD_RTGLIB_PACK_INT8_ARGS_HPP
+
+#include <migraphx/program.hpp>
+#include <migraphx/config.hpp>
+#include <migraphx/gpu/context.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+namespace gpu {
+
+struct pack_int8_args
+{
+    std::string name() const { return "gpu::pack_int8_args"; }
+    void apply(program& p) const;
+    shape pack_int8_shape(const shape& s) const;
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_QUANT_CONVOLUTION_HPP
+#define MIGRAPHX_GUARD_RTGLIB_QUANT_CONVOLUTION_HPP
+
+#include <migraphx/shape.hpp>
+#include <migraphx/op/quant_convolution.hpp>
+#include <migraphx/gpu/miopen.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct context;
+
+struct miopen_quant_convolution
+{
+    op::quant_convolution op;
+    shared<convolution_descriptor> cd;
+    miopenConvFwdAlgorithm_t algo{};
+    miopenHandle_t handle = nullptr;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        // TODO: Add algo
+        return op::quant_convolution::reflect(self.op, f);
+    }
+
+    std::string name() const { return "gpu::quant_convolution"; }
+    shape compute_shape(const std::vector<shape>& inputs) const;
+    argument
+    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
+    shape compile(context& ctx, const shape& output_shape, std::vector<shape> inputs);
+    void finalize(context& ctx, const shape& output_shape, std::vector<shape> inputs);
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
+
+    private:
+    shape pack_int8_shape(const shape& s) const;
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_QUANT_GEMM_HPP
+#define MIGRAPHX_GUARD_RTGLIB_QUANT_GEMM_HPP
+
+#include <migraphx/shape.hpp>
+#include <migraphx/op/quant_dot.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct context;
+
+struct rocblas_quant_gemm
+{
+    op::quant_dot op;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
+    std::string name() const { return "gpu::quant_gemm"; }
+    shape compute_shape(const std::vector<shape>& inputs) const;
+    argument
+    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
+    void batch_not_transposed(const std::vector<std::size_t>& strides) const;
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_SIGMOID_HPP
 #define MIGRAPHX_GUARD_RTGLIB_SIGMOID_HPP
 
-#include <migraphx/shape.hpp>
-#include <migraphx/gpu/miopen.hpp>
+#include <migraphx/gpu/oper.hpp>
+#include <migraphx/gpu/device/sigmoid.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-struct context;
-
-struct miopen_sigmoid
+struct hip_sigmoid : unary_device<hip_sigmoid, device::sigmoid>
 {
-    shared<activation_descriptor> ad;
-
-    template <class Self, class F>
-    static auto reflect(Self& self, F f)
-    {
-        return gpu::reflect(self.ad.get(), f);
-    }
-
-    std::string name() const { return "gpu::sigmoid"; }
-    shape compute_shape(const std::vector<shape>& inputs) const;
-    argument
-    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
-    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
-    {
-        return shapes.size() - 1;
-    }
 };
 
 } // namespace gpu