Database/GEMM: updated profile

include/isaac/driver/buffer.h

@@ -38,6 +38,7 @@ class Buffer: public HandleInterface<Buffer, CUdeviceptr>
 {
 public:
   Buffer(Context const & context, size_t size);
+  void set_zero(Stream const & queue, size_t size);
   void set_zero(Stream const & queue);
   Handle<CUdeviceptr> const & cu() const;
 

lib/driver/buffer.cpp

@@ -40,12 +40,15 @@ Buffer::Buffer(Context const & context, size_t size) : context_(context), size_(
   dispatch::cuMemAlloc(&*cu_, size);
 }
 
-void Buffer::set_zero(Stream const & queue)
+void Buffer::set_zero(Stream const & queue, size_t size)
 {
   ContextSwitcher ctx_switch(context_);
-  dispatch::cuMemsetD8Async(*cu_, 0, size_, queue);
+  dispatch::cuMemsetD8Async(*cu_, 0, size, queue);
 }
 
+void Buffer::set_zero(Stream const & queue)
+{ set_zero(queue, size_); }
+
 Handle<CUdeviceptr> const & Buffer::cu() const
 { return cu_; }
 

lib/templates/gemm.cpp

@@ -100,9 +100,9 @@ void GEMM::check_valid(driver::Device const & device, size_t nkernels, uint32_t*
     DType dtype = (DType)(x[0]);
     IsaacOperation_t AT = (IsaacOperation_t)x[1];
     IsaacOperation_t BT = (IsaacOperation_t)x[2];
-    param_t vec = x[6], bm = x[7], kl = x[8], bn = x[9],
+    param_t M = x[3], N = x[4], vec = x[6], bm = x[7], kl = x[8], bn = x[9],
            ms = x[10], ks = x[11], ns = x[12], a_bf0 = x[13], a_bf1 = x[14], b_bf0 = x[15], b_bf1 = x[16],
-           rs = x[17], br = x[18];
+           rs = x[17], br = x[18], gridr = x[19];
     //Features
     param_t dtsize = size_of(dtype);
     param_t dtvec = (dtype==HALF_TYPE)?2:1;
@@ -111,6 +111,7 @@ void GEMM::check_valid(driver::Device const & device, size_t nkernels, uint32_t*
     param_t rl = rs*br;
     param_t ml = bm*ms;
     param_t nl = bn*ns;
+    param_t gridM = ceil(M, ml), gridN = ceil(N, nl);
     param_t nthreads = bm*bn*br;
     param_t cd_shareda = dtsize*(ml+(A_outer_contig?0:(vec*dtvec)));
     param_t cd_sharedb = dtsize*(nl+(B_outer_contig?0:(vec*dtvec)));
@@ -140,7 +141,7 @@ void GEMM::check_valid(driver::Device const & device, size_t nkernels, uint32_t*
                     &&  ns % (dtvec*vec) == 0
                     &&  kl % ks == 0
                     &&  size_shmem <= device.max_shared_memory()
-
+                    &&  (gridr == 1 || gridM*gridN < 64*64)
                     &&  n_instructions <= 1024 //Doesn't allow more than 1024 instructions in the inner loop
                     &&  bm <= device.max_block_dim()[0]
                     &&  bn <= device.max_block_dim()[1]
@@ -208,7 +209,7 @@ std::string GEMM::dump(drv::Device const & device, std::string const & name){
   if(vec_==1)
     vs[0] = "";
   //Load-Store alignments
-  io_conf Cio(ldc_, vec_, dtvec, dtsize, kg_>1?false:true);
+  io_conf Cio(ldc_, vec_, dtvec, dtsize, false);
   io_conf Aio(lda_, vec_, dtvec, dtsize, false);
   io_conf Bio(ldb_, vec_, dtvec, dtsize, false);
 
@@ -889,10 +890,10 @@ void GEMM::enqueue(driver::Kernel &gemm, driver::Stream &queue, const scalar& al
   gemm.setArg(14, bound);
   gemm.setArg(15, locks);
 
-//  std::cout << gridM << " " << gridN << " " << kg_ << std::endl;
+//  std::cout << gridM << " " << gridN << " " << std::endl;
   //Launch
   if(kg_ > 1)
-    locks.set_zero(queue);
+    locks.set_zero(queue, gridM*gridN*4);
   queue.enqueue(gemm, {gridM, gridN, kg_}, {bm_, bn_, bk_});
 }
 

python/setup.py

@@ -29,8 +29,8 @@ def recursive_glob(rootdir='.', suffix=''):
 def main():
     
     #Source
-    include = [os.path.join('src', 'include')]
+    include = [os.path.join(os.pardir, 'include')]
-    src =  recursive_glob(os.path.join('src','lib'), 'cpp')
+    src =  recursive_glob(os.path.join(os.pardir,'lib'), 'cpp')
     
     #Bindings
     include += [os.path.join('src', 'bind')]

python/src/bind/driver.cpp

@@ -24,6 +24,7 @@
 #include <pybind11/stl.h>
 #include "isaac/driver/backend.h"
 #include "isaac/driver/buffer.h"
+#include "isaac/driver/error.h"
 #include "isaac/driver/cublas.h"
 #include "isaac/driver/stream.h"
 #include "isaac/driver/context.h"
@@ -87,5 +88,9 @@ void export_driver(py::module&& m)
   py::class_<drv::Stream>(m, "Stream")
       .def(py::init<drv::Context>())
       .def("synchronize", &drv::Stream::synchronize);
+      
+      
+  py::register_exception<drv::exception::cuda::misaligned_address>(m, "MisalignedAddress");
+
 
 }

tune/dataset.py

@@ -34,7 +34,7 @@ def benchmarks(prefix, OpType, device, nsamples):
             op = OpType(params)
             try:
                 y = op.benchmark(ctx, stream)            
-            except RuntimeError:
+            except:
                 continue
             #Update
             bufX[nvalid % step, :] = params

tune/regression.py

@@ -50,7 +50,8 @@ def maximize(OpType, device, model, shapes, V):
     X[:, OpType.nshape_params:] = V
     X = OpType.get_valid(device, X)
     #Model predictions
-    predictions = model.predict(np.log2(X), batch_size=8192, verbose=0)
+    with tf.device('/cpu:0'):
+        predictions = model.predict(np.log2(X), batch_size=8192, verbose=0)
     pred_perfs = np.sort(predictions, axis=0)[::-1]
     pred_idxs = np.argsort(predictions, axis=0)[::-1]
     #Evaluate best predicted models