fix mdconv addmm bug for parrots (#450)

* fix mdconv addmm bug for parrots

* fix mdconv ctv save tensor
Co-authored-by: hanyachao <hanyachao@sensetime.com>

mmcv/ops/csrc/parrots/modulated_deform_conv_cuda.cu

@@ -219,14 +219,26 @@ void ModulatedDeformConvBackwardCUDAKernelLauncher(
                           weight.dim(2), weight.dim(3)});
 
     for (size_t g = 0; g < group; g++) {
-      auto columns_g = columns[g];
-      gemm(ctx, 1, true,
-           weight[g].view(
-               {weight.dim(1), weight.dim(2) * weight.dim(3) * weight.dim(4)}),
-           false,
-           grad_output[b][g].view(
-               {grad_output.dim(2), grad_output.dim(3) * grad_output.dim(4)}),
-           0, columns_g);
+      auto columns_g = ctx.createDArrayLite(
+          weight.elemType(), DArrayShape(columns.dim(1), columns.dim(2)));
+      copy(ctx, columns_g, columns[g]);
+      auto weight_g = weight[g].view(
+          {weight.dim(1), weight.dim(2) * weight.dim(3) * weight.dim(4)});
+      weight_g = transpose(ctx, weight_g, 0, 1);
+
+      auto grad_output_bg = ctx.createDArrayLite(
+          grad_output.elemType(),
+          DArrayShape(grad_output.dim(2), grad_output.dim(3),
+                      grad_output.dim(4)));
+      copy(ctx, grad_output_bg, grad_output[b][g]);
+      grad_output_bg =
+          grad_output_bg.view({grad_output_bg.dim(0),
+                               grad_output_bg.dim(1) * grad_output_bg.dim(2)});
+
+      columns_g =
+          parrots::op::addmm(ctx, columns[g], weight_g, grad_output_bg, 0, 1);
+      auto columns_out = columns[g];
+      copy(ctx, columns_out, columns_g);
     }
 
     columns = columns.view({columns.dim(0) * columns.dim(1), columns.dim(2)});

mmcv/ops/csrc/parrots_cuda_helper.hpp

@@ -5,6 +5,7 @@
 #include <float.h>
 
 #include <parrots/darray/darraymath.hpp>
+#include <parrots/darray/mathfunctions.hpp>
 #include <parrots/extension.hpp>
 #include <parrots/foundation/darrayutil.hpp>
 #include <parrots/foundation/exceptions.hpp>

mmcv/ops/modulated_deform_conv.py

@@ -57,9 +57,7 @@ class ModulatedDeformConv2dFunction(Function):
         ctx.with_bias = bias is not None
         if not ctx.with_bias:
             bias = input.new_empty(0)  # fake tensor
-        if weight.requires_grad or mask.requires_grad or offset.requires_grad \
-           or input.requires_grad:
-            ctx.save_for_backward(input, offset, mask, weight, bias)
+        ctx.save_for_backward(input, offset, mask, weight, bias)
         output = input.new_empty(
             ModulatedDeformConv2dFunction._output_size(ctx, input, weight))
         ctx._bufs = [input.new_empty(0), input.new_empty(0)]