blocks

spmvm/spmvm.cuh

@@ -40,6 +40,7 @@ struct spmvm_context_t
     scalar_t *d_scalars;
     // output scalars
     scalar_t *d_out;
+
 };
 
 template <typename scalar_t>
@@ -88,6 +89,10 @@ struct spmvm_host_t
     size_t num_rows;
     size_t num_cols;
     size_t nnz;
+
+    const size_t *blocks;
+    size_t num_blocks;
+    size_t block_size;
 };
 
 template <typename scalar_t>
@@ -115,21 +120,21 @@ public:
     // scalar_t *d_out;
 
 public:
-    spmvm_t(size_t num_rows, size_t num_cols, size_t nnz, int device_id = -1)
+    spmvm_t(spmvm_host_t<scalar_t> *csr, int device_id = -1)
         : gpu(select_gpu(device_id))
     {
         this->context = reinterpret_cast<spmvm_context_t<scalar_t> *>(malloc(sizeof(spmvm_context_t<scalar_t>)));
 
-        context->d_data = reinterpret_cast<scalar_t *>(gpu.Dmalloc(nnz * sizeof(scalar_t)));
-        context->d_col_idx = reinterpret_cast<size_t *>(gpu.Dmalloc(nnz * sizeof(size_t)));
-        context->d_row_ptr = reinterpret_cast<size_t *>(gpu.Dmalloc((num_rows + 1) * sizeof(size_t)));
+        context->d_data = reinterpret_cast<scalar_t *>(gpu.Dmalloc(csr->block_size * sizeof(scalar_t)));
+        context->d_col_idx = reinterpret_cast<size_t *>(gpu.Dmalloc(csr->block_size * sizeof(size_t)));
+        context->d_row_ptr = reinterpret_cast<size_t *>(gpu.Dmalloc((csr->num_rows + 1) * sizeof(size_t)));
 
-        context->num_rows = num_rows;
-        context->num_cols = num_cols;
-        context->nnz = nnz;
+        context->num_rows = csr->num_rows;
+        context->num_cols = csr->num_cols;
+        context->nnz = csr->nnz;
 
-        context->d_scalars = reinterpret_cast<scalar_t *>(gpu.Dmalloc(num_cols * sizeof(scalar_t)));
-        context->d_out = reinterpret_cast<scalar_t *>(gpu.Dmalloc(num_rows * sizeof(scalar_t)));
+        context->d_scalars = reinterpret_cast<scalar_t *>(gpu.Dmalloc(csr->num_cols * sizeof(scalar_t)));
+        context->d_out = reinterpret_cast<scalar_t *>(gpu.Dmalloc(csr->num_rows * sizeof(scalar_t)));
 
         this->owned = true;
     }
@@ -202,21 +207,44 @@ public:
 
         try
         {
-            if (csr->data)
-                gpu[2].HtoD(&context->d_data[0], &csr->data[0], context->nnz);
-            if (csr->col_idx)
-                gpu[2].HtoD(&context->d_col_idx[0], &csr->col_idx[0], context->nnz);
-            if (csr->row_ptr)
-                gpu[2].HtoD(&context->d_row_ptr[0], &csr->row_ptr[0], context->num_rows + 1);
-
             if (scalars)
                 gpu[2].HtoD(&context->d_scalars[0], &scalars[0], context->num_cols);
-
-            spmvm_context_t<scalar_t> *d_context = reinterpret_cast<spmvm_context_t<scalar_t> *>(gpu[2].Dmalloc(sizeof(spmvm_context_t<scalar_t>)));
-            gpu[2].HtoD(d_context, context, 1);
             cudaMemsetAsync(&context->d_out[0], 0, context->num_rows * sizeof(scalar_t), gpu[2]);
-            csr_vector_mul<scalar_t><<<gpu.sm_count(), nthreads, 0, gpu[2]>>>(d_context);
-            CUDA_OK(cudaGetLastError());
+
+            size_t start_row = 0;
+            size_t end_row = 0;
+            size_t num_rows = 0;
+
+            size_t start_data = 0;
+            size_t end_data = 0;
+            size_t num_data = 0;
+
+            for (size_t i = 0; i < csr->num_blocks - 1; ++i) {
+                start_row = csr->blocks[i];
+                end_row = csr->blocks[i + 1];
+                num_rows = end_row - start_row;
+
+                start_data = csr->row_ptr[start_row];
+                end_data = csr->row_ptr[end_row];
+                num_data = end_data - start_data;
+
+                if (csr->data)
+                    gpu[i&1].HtoD(&context->d_data[0], &csr->data[start_data], num_data);
+                if (csr->col_idx)
+                    gpu[i&1].HtoD(&context->d_col_idx[0], &csr->col_idx[start_data], num_data);
+                if (csr->row_ptr) {
+                    gpu[i&1].HtoD(&context->d_row_ptr[0], &csr->row_ptr[start_row], num_rows + 1);
+                    csr->num_rows = num_rows;
+                }
+
+                gpu[i&1].sync();
+
+                spmvm_context_t<scalar_t> *d_context = reinterpret_cast<spmvm_context_t<scalar_t> *>(gpu[i&1].Dmalloc(sizeof(spmvm_context_t<scalar_t>)));
+                gpu[i&1].HtoD(d_context, context, 1);
+
+                csr_vector_mul<scalar_t><<<gpu.sm_count(), nthreads, 0, gpu[i&1]>>>(d_context);
+                CUDA_OK(cudaGetLastError());
+            }
 
             gpu[2].DtoH(&out[0], &context->d_out[0], context->num_rows);
             gpu.sync();