CUDA: add conv_2d_transpose (ggml-org#14287)

* CUDA: add conv_2d_transpose

* remove direct include of cuda_fp16

* Review: add brackets for readability, remove ggml_set_param and add asserts

Author: am17an

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -12,6 +12,7 @@
 #include "ggml-cuda/concat.cuh"
 #include "ggml-cuda/conv-transpose-1d.cuh"
 #include "ggml-cuda/conv2d-dw.cuh"
+#include "ggml-cuda/conv2d-transpose.cuh"
 #include "ggml-cuda/convert.cuh"
 #include "ggml-cuda/count-equal.cuh"
 #include "ggml-cuda/cpy.cuh"
@@ -2381,6 +2382,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_CONV_2D_DW:
             ggml_cuda_op_conv2d_dw(ctx, dst);
             break;
+        case GGML_OP_CONV_TRANSPOSE_2D:
+            ggml_cuda_conv_2d_transpose_p0(ctx, dst);
+            break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             ggml_cuda_op_conv_transpose_1d(ctx,dst);
             break;
@@ -3307,6 +3311,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         }
         case GGML_OP_IM2COL:
         case GGML_OP_CONV_2D_DW:
+        case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:

tests/test-backend-ops.cpp

@@ -2984,6 +2984,35 @@ struct test_conv_transpose_2d : public test_case {
     }
 };
 
+// GGML_OP_CONV_TRANSPOSE_2D
+struct test_conv_transpose_2d : public test_case {
+    const std::array<int64_t, 4> ne_input;
+    const std::array<int64_t, 4> ne_kernel;
+    const int stride;
+
+    std::string vars() override {
+        return VARS_TO_STR3(ne_input, ne_kernel, stride);
+    }
+
+    test_conv_transpose_2d(std::array<int64_t, 4> ne_input = {10, 10, 3, 1}, // [input_width, input_height, input_channels, 1]
+                           std::array<int64_t, 4> ne_kernel = {3, 3, 3, 1}, // [kernel_width, kernel_height, input_channels, 1]
+                           int stride = 1)
+        : ne_input(ne_input), ne_kernel(ne_kernel), stride(stride){}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data());
+        ggml_set_name(input, "input");
+
+        ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne_kernel.data());
+        ggml_set_name(kernel, "kernel");
+
+        ggml_tensor * out = ggml_conv_transpose_2d_p0(ctx, kernel, input, stride);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+};
+
 // GGML_OP_IM2COL
 struct test_im2col : public test_case {
     const ggml_type type_input;
@@ -4938,8 +4967,6 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
 
     test_cases.emplace_back(new test_conv_transpose_2d({256, 256, 256, 1}, {3, 3, 16, 256}, 1));
 
-    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, {256, 256, 3, 1}));
-
     return test_cases;
 }