resolve conflict

driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp

@@ -423,7 +423,7 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
         printf("Elapsed time : %f ms, %f TFlop/s\n",
                time,
                (float)calculate_convolution_flops(InDesc{}, WeiDesc{}, OutDesc{}) /
-                   (std::size_t(1024) * 1024 * 1024 * 1024) / (time / 1000));
+                   (std::size_t(1000) * 1000 * 1000) / time);
         usleep(std::min(time * 1000, float(10000)));
     }
 

src/include/blockwise_batched_gemm.hip.hpp

@@ -211,52 +211,12 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
 #pragma unroll
         for(index_t k_begin = 0; k_begin < KPerBlock; k_begin += KPerThreadLoop)
         {
-// read first batch of A, B
-//   copy A-sub to form A
-#pragma unroll
-            for(index_t m_repeat = 0; m_repeat < MRepeat; ++m_repeat)
-            {
-                threadwise_matrix_copy(
-                    a_block_mtx,
-                    p_a_block + a_block_mtx.Get1dIndex(k_begin, m_repeat * MPerLevel1Cluster) +
-                        mMyThreadOffsetA,
-                    a_thread_mtx,
-                    p_a_thread + a_thread_mtx.Get1dIndex(0, m_repeat * MPerThreadSubC),
-                    a_thread_sub_mtx.GetLengths(),
-                    Number<DataPerReadA>{});
-            }
-
-//   copy B-sub to form B
-#pragma unroll
-            for(index_t n_repeat = 0; n_repeat < NRepeat; ++n_repeat)
-            {
-                threadwise_matrix_copy(
-                    b_block_mtx,
-                    p_b_block + b_block_mtx.Get1dIndex(k_begin, n_repeat * NPerLevel1Cluster) +
-                        mMyThreadOffsetB,
-                    b_thread_mtx,
-                    p_b_thread + b_thread_mtx.Get1dIndex(0, n_repeat * NPerThreadSubC),
-                    b_thread_sub_mtx.GetLengths(),
-                    Number<DataPerReadB>{});
-            }
-
 // loop over batch
 #pragma unroll
-            for(index_t ib = 0; ib + 1 < BatchPerThread; ++ib)
+            for(index_t ib = 0; ib < BatchPerThread; ++ib)
             {
-                // do current batch of gemm
-                threadwise_gemm(a_thread_mtx,
-                                True,
-                                p_a_thread,
-                                b_thread_mtx,
-                                False,
-                                p_b_thread,
-                                c_thread_mtx,
-                                False,
-                                p_c_thread + ib * ThreadMatrixStrideC);
-
                 // read next batch of a, b
-                if(BlockMatrixStrideA != 0)
+                if(BlockMatrixStrideA != 0 or ib == 0)
                 {
 #pragma unroll
                     for(index_t m_repeat = 0; m_repeat < MRepeat; ++m_repeat)
@@ -265,7 +225,7 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
                             a_block_mtx,
                             p_a_block +
                                 a_block_mtx.Get1dIndex(k_begin, m_repeat * MPerLevel1Cluster) +
-                                (ib + 1) * BlockMatrixStrideA + mMyThreadOffsetA,
+                                ib * BlockMatrixStrideA + mMyThreadOffsetA,
                             a_thread_mtx,
                             p_a_thread + a_thread_mtx.Get1dIndex(0, m_repeat * MPerThreadSubC),
                             a_thread_sub_mtx.GetLengths(),
@@ -273,7 +233,7 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
                     }
                 }
 
-                if(BlockMatrixStrideB != 0)
+                if(BlockMatrixStrideB != 0 or ib == 0)
                 {
 #pragma unroll
                     for(index_t n_repeat = 0; n_repeat < NRepeat; ++n_repeat)
@@ -282,25 +242,24 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
                             b_block_mtx,
                             p_b_block +
                                 b_block_mtx.Get1dIndex(k_begin, n_repeat * NPerLevel1Cluster) +
-                                (ib + 1) * BlockMatrixStrideB + mMyThreadOffsetB,
+                                ib * BlockMatrixStrideB + mMyThreadOffsetB,
                             b_thread_mtx,
                             p_b_thread + b_thread_mtx.Get1dIndex(0, n_repeat * NPerThreadSubC),
                             b_thread_sub_mtx.GetLengths(),
                             Number<DataPerReadB>{});
                     }
                 }
-            }
 
-            // do last batch of gemm
-            threadwise_gemm(a_thread_mtx,
-                            True,
-                            p_a_thread,
-                            b_thread_mtx,
-                            False,
-                            p_b_thread,
-                            c_thread_mtx,
-                            False,
-                            p_c_thread + (BatchPerThread - 1) * ThreadMatrixStrideC);
+                threadwise_gemm(a_thread_mtx,
+                                True,
+                                p_a_thread,
+                                b_thread_mtx,
+                                False,
+                                p_b_thread,
+                                c_thread_mtx,
+                                False,
+                                p_c_thread + ib * ThreadMatrixStrideC);
+            }
         }
     }
 

src/include/gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn_lds_double_buffer.hip.hpp

@@ -276,9 +276,9 @@ struct GridwiseConvolutionImplicitGemm_v2_chwn_cyxk_khwn_lds_double_buffer
 #elif 1
                         blockwise_gemm.Run_asm
 #endif
-                        (p_wei_block_now + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
-                         p_in_block_now + y * Wi + x,
-                         p_out_thread);
+                            (p_wei_block_now + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
+                             p_in_block_now + y * Wi + x,
+                             p_out_thread);
                     }
                 }
 
@@ -320,9 +320,9 @@ struct GridwiseConvolutionImplicitGemm_v2_chwn_cyxk_khwn_lds_double_buffer
 #elif 1
                     blockwise_gemm.Run_asm
 #endif
-                    (p_wei_block_double + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
-                     p_in_block_double + y * Wi + x,
-                     p_out_thread);
+                        (p_wei_block_double + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
+                         p_in_block_double + y * Wi + x,
+                         p_out_thread);
                 }
             }
 
@@ -345,10 +345,10 @@ struct GridwiseConvolutionImplicitGemm_v2_chwn_cyxk_khwn_lds_double_buffer
 #elif 1
                     blockwise_gemm.Run_asm
 #endif
-                    (p_wei_block_double + wei_block_space +
-                         wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
-                     p_in_block_double + in_block_space + y * Wi + x,
-                     p_out_thread);
+                        (p_wei_block_double + wei_block_space +
+                             wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
+                         p_in_block_double + in_block_space + y * Wi + x,
+                         p_out_thread);
                 }
             }
         }