update hip build

driver/driver.hip.cpp

@@ -47,7 +47,7 @@ struct GeneratorTensor_3
         std::initializer_list<std::size_t> ids = {static_cast<std::size_t>(is)...};
         std::vector<std::size_t> lens(sizeof...(Is), 100);
         std::vector<std::size_t> strides(sizeof...(Is), 1);
-        std::partial_sum(lens.rbegin(), lens.rbegin() + (sizeof...(Is) - 1), strides.rbegin() + 1);
+        std::partial_sum(lens.rbegin(), lens.rbegin() + (sizeof...(Is)-1), strides.rbegin() + 1);
         return std::inner_product(ids.begin(), ids.end(), strides.begin(), std::size_t(0)) + 1;
 #endif
     }

src/include/blockwise_4d_tensor_op.hip.hpp

@@ -245,7 +245,8 @@ struct BlockwiseChwnTensorCopyPadded
         constexpr unsigned NLoop = ref_desc.GetElementSize() / BlockSize;
 
         const Float* p_src_tmp =
-            p_src + src_desc.Get1dIndex(c_block_data_begin,
+            p_src +
+            src_desc.Get1dIndex(c_block_data_begin,
                                 (ho_block_data_begin + h_block_pad_low) - h_global_pad_low,
                                 (wo_block_data_begin + w_block_pad_low) - w_global_pad_low,
                                 n_block_data_begin);

src/include/blockwise_direct_convolution.hip.hpp

@@ -93,7 +93,8 @@ __device__ void blockwise_direct_convolution(InBlockDesc,
         Float p_out_thread[out_thread_desc.GetElementSpace()];
 
         threadwise_4d_tensor_copy(out_block_desc,
-                                  p_out_block + out_block_desc.Get1dIndex(n_thread_data_begin,
+                                  p_out_block +
+                                      out_block_desc.Get1dIndex(n_thread_data_begin,
                                                                 k_thread_data_begin,
                                                                 ho_thread_data_begin,
                                                                 wo_thread_data_begin),
@@ -107,7 +108,8 @@ __device__ void blockwise_direct_convolution(InBlockDesc,
             // threadwise convolution
             threadwise_direct_convolution_2(
                 in_thread_block_desc,
-                p_in_block + in_block_desc.Get1dIndex(n_thread_data_begin,
+                p_in_block +
+                    in_block_desc.Get1dIndex(n_thread_data_begin,
                                              c_thread_data_begin,
                                              hi_thread_data_begin,
                                              wi_thread_data_begin),
@@ -122,7 +124,8 @@ __device__ void blockwise_direct_convolution(InBlockDesc,
         threadwise_4d_tensor_copy(out_thread_desc,
                                   p_out_thread,
                                   out_block_desc,
-                                  p_out_block + out_block_desc.Get1dIndex(n_thread_data_begin,
+                                  p_out_block +
+                                      out_block_desc.Get1dIndex(n_thread_data_begin,
                                                                 k_thread_data_begin,
                                                                 ho_thread_data_begin,
                                                                 wo_thread_data_begin),

src/include/blockwise_gemm.hip.hpp

@@ -431,12 +431,12 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
         constexpr unsigned MRepeat = MPerThread / MPerThreadSubC;
         constexpr unsigned NRepeat = NPerThread / NPerThreadSubC;
 
-        // loop over k
+// loop over k
 #pragma unroll
         for(unsigned k_begin = 0; k_begin < KPerBlock; k_begin += KPerThreadLoop)
         {
-            // read first batch of A, B
-            //   copy A-sub to form A
+// read first batch of A, B
+//   copy A-sub to form A
 #pragma unroll
             for(unsigned m_repeat = 0; m_repeat < MRepeat; ++m_repeat)
             {
@@ -449,7 +449,7 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
                     a_thread_sub_mtx.GetLengths());
             }
 
-            //   copy B-sub to form B
+//   copy B-sub to form B
 #pragma unroll
             for(unsigned n_repeat = 0; n_repeat < NRepeat; ++n_repeat)
             {
@@ -462,7 +462,7 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
                     b_thread_sub_mtx.GetLengths());
             }
 
-            // loop over batch
+// loop over batch
 #pragma unroll
             for(unsigned ib = 0; ib + 1 < BatchPerThread; ++ib)
             {
@@ -557,7 +557,8 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
             {
                 threadwise_matrix_copy(
                     c_thread_sub_mtx,
-                    p_c_thread + c_thread_sub_mtx.Get1dIndex(m_repeat * MPerLevel1Cluster,
+                    p_c_thread +
+                        c_thread_sub_mtx.Get1dIndex(m_repeat * MPerLevel1Cluster,
                                                     n_repeat * NPerLevel1Cluster),
                     c_block_mtx,
                     p_c_block +
@@ -656,7 +657,8 @@ struct BlockwiseGemmBlockABlockBThreadC
             constexpr unsigned NClusterWork =
                 (NPerBlock + NPerThread * NThreadPerCluster - 1) / (NPerThread * NThreadPerCluster);
 
-            static_assert(BlockSize == (MClusterWork * MThreadPerCluster) *
+            static_assert(BlockSize ==
+                              (MClusterWork * MThreadPerCluster) *
                                   (NClusterWork * NThreadPerCluster),
                           "wrong! wrong BlockSize");
 
@@ -1256,7 +1258,8 @@ struct BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2
                             p_b_thread + b_thread_mtx.Get1dIndex(0, n_repeat * NPerThreadSubC),
                             c_thread_sub_mtx,
                             False,
-                            p_c_thread + c_thread_mtx.Get1dIndex(m_repeat * MPerThreadSubC,
+                            p_c_thread +
+                                c_thread_mtx.Get1dIndex(m_repeat * MPerThreadSubC,
                                                         n_repeat * NPerThreadSubC),
                             f_accum);
                     }

src/include/common.hip.hpp

@@ -65,7 +65,7 @@ struct vector_type<half_float::half, 8>
 };
 #endif
 
-#if 1
+#if 0
 template <>
 struct vector_type<half, 1>
 {
@@ -139,6 +139,7 @@ struct Sequence
     }
 };
 
+#if DEVICE_BACKEND_CUDA
 template <typename T>
 __host__ __device__ constexpr T max(T a, T b)
 {
@@ -150,6 +151,7 @@ __host__ __device__ constexpr T min(T a, T b)
 {
     return a < b ? a : b;
 }
+#endif
 
 __host__ __device__ constexpr unsigned integer_divide_ceil(unsigned a, unsigned b)
 {

src/include/config.h.in

@@ -4,7 +4,6 @@
 
 #if DEVICE_BACKEND_HIP
 #include "hip/hip_runtime.h"
-#include "half.hpp"
 #elif DEVICE_BACKEND_CUDA
 #include "cuda_runtime.h"
 #include "nvToolsExt.h"

src/include/gridwise_direct_convolution_1.hip.hpp

@@ -113,7 +113,8 @@ __global__ void gridwise_direct_convolution_1(const Float* const __restrict__ p_
         c_block_work_begin += CPerBlock)
     {
         // copy input tensor to LDS
-        blockwise_in_copy.Run(p_in_global + in_global_desc.Get1dIndex(n_block_work_begin,
+        blockwise_in_copy.Run(p_in_global +
+                                  in_global_desc.Get1dIndex(n_block_work_begin,
                                                             c_block_work_begin,
                                                             hi_block_work_begin,
                                                             wi_block_work_begin),
@@ -143,9 +144,9 @@ __global__ void gridwise_direct_convolution_1(const Float* const __restrict__ p_
     }
 
     // copy output tensor from LDS to device mem
-    blockwise_out_copy.Run(p_out_block,
-                           p_out_global + out_global_desc.Get1dIndex(n_block_work_begin,
-                                                                     k_block_work_begin,
-                                                                     ho_block_work_begin,
-                                                                     wo_block_work_begin));
+    blockwise_out_copy.Run(
+        p_out_block,
+        p_out_global +
+            out_global_desc.Get1dIndex(
+                n_block_work_begin, k_block_work_begin, ho_block_work_begin, wo_block_work_begin));
 }

src/include/gridwise_direct_convolution_2.hip.hpp

@@ -139,7 +139,8 @@ __global__ void gridwise_direct_convolution_2(const Float* const __restrict__ p_
         c_block_data_begin += CPerBlock, __syncthreads())
     {
         // copy input tensor to LDS
-        blockwise_in_copy.Run(p_in_global + in_global_desc.Get1dIndex(n_block_data_begin,
+        blockwise_in_copy.Run(p_in_global +
+                                  in_global_desc.Get1dIndex(n_block_data_begin,
                                                             c_block_data_begin,
                                                             hi_block_data_begin,
                                                             wi_block_data_begin),
@@ -158,7 +159,8 @@ __global__ void gridwise_direct_convolution_2(const Float* const __restrict__ p_
 #if 1
             threadwise_direct_convolution_2(
                 in_thread_block_desc,
-                p_in_block + in_block_desc.Get1dIndex(n_thread_data_begin,
+                p_in_block +
+                    in_block_desc.Get1dIndex(n_thread_data_begin,
                                              c_thread_data,
                                              hi_thread_data_begin,
                                              wi_thread_data_begin),
@@ -169,7 +171,8 @@ __global__ void gridwise_direct_convolution_2(const Float* const __restrict__ p_
 #elif 0
             threadwise_direct_convolution_3(
                 in_thread_block_desc,
-                p_in_block + in_block_desc.Get1dIndex(n_thread_data_begin,
+                p_in_block +
+                    in_block_desc.Get1dIndex(n_thread_data_begin,
                                              c_thread_data,
                                              hi_thread_data_begin,
                                              wi_thread_data_begin),
@@ -186,7 +189,8 @@ __global__ void gridwise_direct_convolution_2(const Float* const __restrict__ p_
         out_thread_desc,
         p_out_thread,
         out_global_desc,
-        p_out_global + out_global_desc.Get1dIndex(n_block_data_begin + n_thread_data_begin,
+        p_out_global +
+            out_global_desc.Get1dIndex(n_block_data_begin + n_thread_data_begin,
                                        k_block_data_begin + k_thread_data_begin,
                                        ho_block_data_begin + ho_thread_data_begin,
                                        wo_block_data_begin + wo_thread_data_begin),

src/include/gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn.hip.hpp

@@ -184,7 +184,8 @@ gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn(const Float* const __restric
     threadwise_4d_tensor_set_zero(out_khwn_thread_desc, p_out_thread);
 
     const Float* p_in_global_block_begin =
-        p_in_global + in_chwn_global_desc.Get1dIndex(
+        p_in_global +
+        in_chwn_global_desc.Get1dIndex(
             0, hi_block_data_begin, wi_block_data_begin, n_block_data_begin);
 
     const Float* p_wei_global_block_begin =
@@ -216,7 +217,7 @@ gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn(const Float* const __restric
         }
     }
 
-    // output: register to global mem,
+// output: register to global mem,
 #if 0
     const auto c_thread_mtx_begin =
         blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
@@ -286,11 +287,12 @@ gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn(const Float* const __restric
         }
 #endif
 
-        threadwise_8d_tensor_copy(out_8d_thread_desc,
+        threadwise_8d_tensor_copy(
+            out_8d_thread_desc,
             p_out_thread,
             out_8d_global_desc,
-                                  p_out_global + out_khwn_global_desc.Get1dIndex(
-                                                     k_block_data_begin + k_thread_data_begin,
+            p_out_global +
+                out_khwn_global_desc.Get1dIndex(k_block_data_begin + k_thread_data_begin,
                                                 ho_block_data_begin + ho_thread_data_begin,
                                                 wo_block_data_begin + wo_thread_data_begin,
                                                 n_block_data_begin + n_thread_data_begin),

src/include/gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded.hip.hpp

@@ -283,7 +283,8 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded(
         out_hkwn_thread_desc,
         p_out_thread,
         out_khwn_global_desc,
-        p_out_global + out_khwn_global_desc.Get1dIndex(k_block_data_begin + k_thread_data_begin,
+        p_out_global +
+            out_khwn_global_desc.Get1dIndex(k_block_data_begin + k_thread_data_begin,
                                             ho_block_data_begin + ho_thread_data_begin,
                                             wo_block_data_begin + wo_thread_data_begin,
                                             n_block_data_begin + n_thread_data_begin),

src/include/tensor.hpp

@@ -22,8 +22,7 @@ std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim)
     return os;
 }
 
-typedef enum
-{
+typedef enum {
     Half  = 0,
     Float = 1,
 } DataType_t;