ConvTranspose2d Operator (#91)

Add ConvTranspose2d operator

src/operators/ai.onnx/ConvTranspose/11/execute_operator__ai_onnx__convtranspose__11__T_tensor_float.c

@@ -0,0 +1,143 @@
+//this file was generated by ../../../../../../scripts/onnx_generator/OperatorTemplate.py
+#include "operator__ai_onnx__convtranspose__11.h"
+#include "tracing.h"
+#include "utils.h"
+
+//transformes a 3d pos into a 1d flat array pos
+static inline int calcArrayPos3D(int x, int y, int outputChannel, int width, int height) {
+    return outputChannel * height * width + y * width + x;
+}
+
+//transformes a 4d pos into a 1d flat array pos
+static inline int calcArrayPos4D(int x, int y, int outputChannel, int intputChannel, int width, int height, int nOfOutputChannels) {
+    return intputChannel * nOfOutputChannels *height * width 
+           + outputChannel * height * width 
+           + y * width
+           + x;
+}
+
+operator_status
+execute_operator__ai_onnx__convtranspose__11__T_tensor_float(
+    node_context *ctx
+)
+{
+    TRACE_ENTRY(1);
+
+    TRACE_NODE(2, true, ctx->onnx_node);
+
+    /* UNCOMMENT AS NEEDED */
+
+    Onnx__TensorProto *i_X = searchInputByName(ctx, 0);
+    Onnx__TensorProto *i_W = searchInputByName(ctx, 1);
+    Onnx__TensorProto *i_B = searchInputByName(ctx, 2);
+
+    TRACE_TENSOR(2, true, i_X);
+    TRACE_TENSOR(2, true, i_W);
+    TRACE_TENSOR(2, B, i_B);
+
+    context_operator__ai_onnx__convtranspose__11 *op_ctx = ctx->executer_context;
+
+    // size is not needed, because this operator works for one fixed size only
+
+    // char* auto_pad = op_ctx->auto_pad;
+    // size_t n_dilations = op_ctx->n_dilations;
+    int64_t* dilations = op_ctx->dilations;
+    // int64_t group = op_ctx->group;
+    // size_t n_kernel_shape = op_ctx->n_kernel_shape;
+    // int64_t* kernel_shape = op_ctx->kernel_shape;
+    // size_t n_output_padding = op_ctx->n_output_padding;
+    // int64_t* output_padding = op_ctx->output_padding;
+    // size_t n_output_shape = op_ctx->n_output_shape;
+    // int64_t* output_shape = op_ctx->output_shape;
+    // size_t n_pads = op_ctx->n_pads;
+    int64_t* pads = op_ctx->pads;
+    // size_t n_strides = op_ctx->n_strides;
+    int64_t* strides = op_ctx->strides;
+
+    // TRACE_VAR(2, true, auto_pad, "\"%s\"");
+    TRACE_ARRAY(2, true, dilations, , n_dilations, "%" PRId64);
+    // TRACE_VAR(2, true, group, "%" PRId64);
+    // TRACE_ARRAY(2, true, kernel_shape, , n_kernel_shape, "%" PRId64);
+    // TRACE_ARRAY(2, true, output_padding, , n_output_padding, "%" PRId64);
+    // TRACE_ARRAY(2, true, output_shape, , n_output_shape, "%" PRId64);
+    TRACE_ARRAY(2, true, pads, , n_pads, "%" PRId64);
+    TRACE_ARRAY(2, true, strides, , n_strides, "%" PRId64);
+
+    Onnx__TensorProto *o_Y = searchOutputByName(ctx, 0);
+
+    // TRACE_TENSOR(2, true, o_Y);
+
+    /* DO CALCULATION HERE */
+
+    const int inputSizeX = i_X->dims[3];
+    const int inputSizeY = i_X->dims[2];
+    const int inputChannels = i_X->dims[1];
+
+    const float *input = i_X->float_data;
+    const float *weights = i_W->float_data;
+
+    const int kernelSizeX = i_W->dims[3];
+    const int kernelSizeY = i_W->dims[2];
+    const int outputChannels = i_W->dims[1];
+
+    const int strideX = strides[1];
+    const int strideY = strides[0];
+
+    const int dilationsX = dilations[1];
+    const int dilationsY = dilations[0];
+
+    const int padStartY = pads[0];
+    const int padStartX = pads[1];
+
+    //not used because outputSize is used to test for the padEnd
+    //const int padEndY = pads[2];
+    //const int padEndX = pads[3];
+
+    const int outputSizeX = o_Y->dims[2];
+    const int outputSizeY = o_Y->dims[3];
+
+    float* output = o_Y->float_data;
+
+    //fill with bias
+    for(int c=0; c<outputChannels; c++) {
+        float bias = i_B?i_B->float_data[c]:0;
+        for(int y=0; y<outputSizeY; y++) {
+            for(int x=0; x<outputSizeX; x++) {
+                output[calcArrayPos3D(x,y,c,outputSizeX, outputSizeY)] = bias;
+            }
+        }
+    }
+
+    //actual transpose convolution
+    for(int i=0; i < inputChannels; i++) {
+        for(int c=0; c<outputChannels; c++) {
+            for(int inputPosY=0; inputPosY<inputSizeY; inputPosY++) {
+                for(int inputPosX=0; inputPosX<inputSizeX; inputPosX++) {
+                    float _input = input[calcArrayPos3D(inputPosX, inputPosY, i, inputSizeX, inputSizeY)];
+
+                    for(int kernelPosX=0; kernelPosX<kernelSizeX; kernelPosX++) {
+                        int x = inputPosX*strideX+kernelPosX*dilationsX - padStartX;
+                        if(x < 0 || x >= outputSizeX) {
+                            continue;
+                        }
+
+                        for(int kernelPosY=0; kernelPosY<kernelSizeY; kernelPosY++) {
+                            int y = inputPosY*strideY+kernelPosY*dilationsY - padStartY;
+                            if(y < 0 || y >= outputSizeY) {
+                                continue;
+                            }
+
+                            const float _weight = weights[calcArrayPos4D(kernelPosX, kernelPosY, c, i, kernelSizeX, kernelSizeY, outputChannels)];
+                            output[calcArrayPos3D(x, y, c, outputSizeX, outputSizeY)] += _input * _weight;
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    TRACE_EXIT(1);
+
+    /* CHANGE RETURN CODE IF THIS EXECUTER IS VALID */
+    return OP_OK;
+}

src/operators/ai.onnx/ConvTranspose/11/free_operator__ai_onnx__convtranspose__11.c

@@ -0,0 +1,78 @@
+//this file was generated by ../../../../../../scripts/onnx_generator/OperatorTemplate.py
+#include "operator__ai_onnx__convtranspose__11.h"
+#include "tracing.h"
+#include "utils.h"
+
+void
+free_operator__ai_onnx__convtranspose__11(
+    node_context *ctx
+)
+{
+    TRACE_ENTRY(1);
+
+    TRACE_NODE(2, true, ctx->onnx_node);
+
+    /* UNCOMMENT AS NEEDED */
+
+    // Onnx__TensorProto *i_X = searchInputByName(ctx, 0);
+    // Onnx__TensorProto *i_W = searchInputByName(ctx, 1);
+    // Onnx__TensorProto *i_B = searchInputByName(ctx, 2);
+
+    // TRACE_TENSOR(2, true, i_X);
+    // TRACE_TENSOR(2, true, i_W);
+    // TRACE_TENSOR(2, B, i_B);
+
+    // Onnx__AttributeProto *a_auto_pad = searchAttributeNyName(ctx->onnx_node->n_attribute,ctx->onnx_node->attribute,"auto_pad");
+    // Onnx__AttributeProto *a_dilations = searchAttributeNyName(ctx->onnx_node->n_attribute,ctx->onnx_node->attribute,"dilations");
+    // Onnx__AttributeProto *a_group = searchAttributeNyName(ctx->onnx_node->n_attribute,ctx->onnx_node->attribute,"group");
+    // Onnx__AttributeProto *a_kernel_shape = searchAttributeNyName(ctx->onnx_node->n_attribute,ctx->onnx_node->attribute,"kernel_shape");
+    // Onnx__AttributeProto *a_output_padding = searchAttributeNyName(ctx->onnx_node->n_attribute,ctx->onnx_node->attribute,"output_padding");
+    // Onnx__AttributeProto *a_output_shape = searchAttributeNyName(ctx->onnx_node->n_attribute,ctx->onnx_node->attribute,"output_shape");
+    // Onnx__AttributeProto *a_pads = searchAttributeNyName(ctx->onnx_node->n_attribute,ctx->onnx_node->attribute,"pads");
+    // Onnx__AttributeProto *a_strides = searchAttributeNyName(ctx->onnx_node->n_attribute,ctx->onnx_node->attribute,"strides");
+
+    // TRACE_ATTRIBUTE(2, a_auto_pad, a_auto_pad);
+    // TRACE_ATTRIBUTE(2, a_dilations, a_dilations);
+    // TRACE_ATTRIBUTE(2, a_group, a_group);
+    // TRACE_ATTRIBUTE(2, a_kernel_shape, a_kernel_shape);
+    // TRACE_ATTRIBUTE(2, a_output_padding, a_output_padding);
+    // TRACE_ATTRIBUTE(2, a_output_shape, a_output_shape);
+    // TRACE_ATTRIBUTE(2, a_pads, a_pads);
+    // TRACE_ATTRIBUTE(2, a_strides, a_strides);
+
+    Onnx__TensorProto *o_Y = searchOutputByName(ctx, 0);
+
+    // TRACE_TENSOR(2, true, o_Y);
+
+    /* FREE CONTEXT HERE IF NEEDED */
+
+    context_operator__ai_onnx__convtranspose__11 *op_ctx = ctx->executer_context;
+
+    // TRACE_VAR(2, true, op_ctx->auto_pad, "\"%s\"");
+    TRACE_ARRAY(2, true, op_ctx->dilations, , op_ctx->n_dilations, "%" PRId64);
+    // TRACE_VAR(2, true, op_ctx->group, "%" PRId64);
+    // TRACE_ARRAY(2, true, op_ctx->kernel_shape, , op_ctx->n_kernel_shape, "%" PRId64);
+    // TRACE_ARRAY(2, true, op_ctx->output_padding, , op_ctx->n_output_padding, "%" PRId64);
+    // TRACE_ARRAY(2, true, op_ctx->output_shape, , op_ctx->n_output_shape, "%" PRId64);
+    TRACE_ARRAY(2, true, op_ctx->pads, , op_ctx->n_pads, "%" PRId64);
+    TRACE_ARRAY(2, true, op_ctx->strides, , op_ctx->n_strides, "%" PRId64);
+
+    // free(op_ctx->auto_pad);
+    free(op_ctx->dilations);
+    // free(op_ctx->kernel_shape);
+    // free(op_ctx->output_padding);
+    // free(op_ctx->output_shape);
+    free(op_ctx->pads);
+    free(op_ctx->strides);
+
+    free(op_ctx);
+
+
+    /* FREE OUTPUT DATA_TYPE AND SHAPE HERE */
+    /* DO NOT FREE THE TENSOR ITSELF */
+
+    // freeTensorData(o_Y);
+    free(o_Y->dims);
+
+    TRACE_EXIT(1);
+}

src/operators/ai.onnx/ConvTranspose/11/info_operator__ai_onnx__convtranspose__11.c

@@ -0,0 +1,152 @@
+//this file was generated by ../../../../../../scripts/onnx_generator/OperatorInfo.py
+#include "operators/operator_info.h"
+#include "operator__ai_onnx__convtranspose__11.h"
+
+/* attributes */
+static
+operator_info_attribute
+attributes[] = {
+{
+    .name     = "auto_pad",
+    .optional = true,
+    .type     = ONNX__ATTRIBUTE_PROTO__ATTRIBUTE_TYPE__STRING
+},
+{
+    .name     = "dilations",
+    .optional = true,
+    .type     = ONNX__ATTRIBUTE_PROTO__ATTRIBUTE_TYPE__INTS
+},
+{
+    .name     = "group",
+    .optional = true,
+    .type     = ONNX__ATTRIBUTE_PROTO__ATTRIBUTE_TYPE__INT
+},
+{
+    .name     = "kernel_shape",
+    .optional = true,
+    .type     = ONNX__ATTRIBUTE_PROTO__ATTRIBUTE_TYPE__INTS
+},
+{
+    .name     = "output_padding",
+    .optional = true,
+    .type     = ONNX__ATTRIBUTE_PROTO__ATTRIBUTE_TYPE__INTS
+},
+{
+    .name     = "output_shape",
+    .optional = true,
+    .type     = ONNX__ATTRIBUTE_PROTO__ATTRIBUTE_TYPE__INTS
+},
+{
+    .name     = "pads",
+    .optional = true,
+    .type     = ONNX__ATTRIBUTE_PROTO__ATTRIBUTE_TYPE__INTS
+},
+{
+    .name     = "strides",
+    .optional = true,
+    .type     = ONNX__ATTRIBUTE_PROTO__ATTRIBUTE_TYPE__INTS
+}
+};
+
+/* input tensors */
+static
+uint32_t
+input_tensor_type_X[] = {
+ONNX__TENSOR_PROTO__DATA_TYPE__DOUBLE,
+ONNX__TENSOR_PROTO__DATA_TYPE__FLOAT,
+ONNX__TENSOR_PROTO__DATA_TYPE__FLOAT16
+};
+
+static
+uint32_t
+input_tensor_type_W[] = {
+ONNX__TENSOR_PROTO__DATA_TYPE__DOUBLE,
+ONNX__TENSOR_PROTO__DATA_TYPE__FLOAT,
+ONNX__TENSOR_PROTO__DATA_TYPE__FLOAT16
+};
+
+static
+uint32_t
+input_tensor_type_B[] = {
+ONNX__TENSOR_PROTO__DATA_TYPE__DOUBLE,
+ONNX__TENSOR_PROTO__DATA_TYPE__FLOAT,
+ONNX__TENSOR_PROTO__DATA_TYPE__FLOAT16
+};
+
+static
+operator_info_tensor
+inputs[] = {
+{
+    .name        = "X",
+    .optional    = false,
+    .variadic    = false,
+    .homogeneous = true,
+    .constraint  = "T",
+    .n_types     = 3,
+    .types       = input_tensor_type_X
+},
+{
+    .name        = "W",
+    .optional    = false,
+    .variadic    = false,
+    .homogeneous = true,
+    .constraint  = "T",
+    .n_types     = 3,
+    .types       = input_tensor_type_W
+},
+{
+    .name        = "B",
+    .optional    = true,
+    .variadic    = true,
+    .homogeneous = true,
+    .constraint  = "T",
+    .n_types     = 3,
+    .types       = input_tensor_type_B
+}
+};
+
+/* output tensors */
+static
+uint32_t
+output_tensor_type_Y[] = {
+ONNX__TENSOR_PROTO__DATA_TYPE__DOUBLE,
+ONNX__TENSOR_PROTO__DATA_TYPE__FLOAT,
+ONNX__TENSOR_PROTO__DATA_TYPE__FLOAT16
+};
+
+static
+operator_info_tensor
+outputs[] = {
+{
+    .name        = "Y",
+    .optional    = false,
+    .variadic    = false,
+    .homogeneous = true,
+    .constraint  = "T",
+    .n_types     = 3,
+    .types       = output_tensor_type_Y
+}
+};
+
+/* constraints */
+static
+operator_info_constraint
+constraints[] = {
+{ "T" }
+};
+
+/* operator info */
+operator_info
+info_operator__ai_onnx__convtranspose__11 = {
+    .name         = "ConvTranspose",
+    .range_input  = { 2, 3 },
+    .range_output = { 1, 1 },
+    .n_attribute  = 8,
+    .attribute    = attributes,
+    .n_input      = 3,
+    .input        = inputs,
+    .n_output     = 1,
+    .output       = outputs,
+    .n_constraint = 1,
+    .constraint   = constraints
+};