nn-testing/src/mat.cl

typedef struct __attribute__((packed)) {
	uint rows;
	uint cols;
	char transposed;
} cl_GPUMat;

#define matrixAt(mat, mat_values, r, c) mat_values[(!mat.transposed ? r * mat.cols + c : c * mat.rows + r)]

#define matrixGetIJ(mat, idx, i, j) \
	{                               \
		if(!mat.transposed) {       \
			i = idx / mat.cols;     \
			j = idx % mat.cols;     \
		} else {                    \
			i = idx % mat.rows;     \
			j = idx / mat.rows;     \
		}                           \
	}

void mat_multiply(cl_GPUMat matA, __global float *matA_values, cl_GPUMat matB, __global float *matB_values, cl_GPUMat matOut, __global float *matOut_values) {
	uint idx = get_global_id(0);
	if(idx >= matOut.rows * matOut.cols) return;

	// TODO: might not work with transposed matOut
	uint i = idx / matOut.cols;
	uint j = idx % matOut.cols;

	float sum = 0;
	for(unsigned int k = 0; k < matA.cols; k++) {
		sum += matrixAt(matA, matA_values, i, k) * matrixAt(matB, matB_values, k, j);
	}

	matOut_values[idx] = sum;
}

void mat_add(cl_GPUMat matA, __global float *matA_values, cl_GPUMat matB, __global float *matB_values) {
	uint idx = get_global_id(0);
	if(idx >= matA.rows * matA.cols) return;

	matA_values[idx] += matB_values[idx];
}

void mat_sigmoid(cl_GPUMat mat, __global float *mat_values) {
	uint idx = get_global_id(0);
	if(idx >= mat.rows * mat.cols) return;

	mat_values[idx] = 1 / (1 + exp(-mat_values[idx]));
}

void mat_dSigmoid(cl_GPUMat mat, __global float *mat_values) {
	uint idx = get_global_id(0);
	if(idx >= mat.rows * mat.cols) return;

	float v = mat_values[idx];
	mat_values[idx] = v * (1 - v);
}

void mat_copy(cl_GPUMat mat, __global float *mat_values, cl_GPUMat other, __global float *other_values) {
	uint idx = get_global_id(0);
	if(idx >= mat.rows * mat.cols) return;

	other_values[idx] = mat_values[idx];
}

void mat_multiply_elements(cl_GPUMat mat, __global float *mat_values, cl_GPUMat other, __global float *other_values) {
	uint idx = get_global_id(0);
	if(idx >= mat.rows * mat.cols) return;

	other_values[idx] *= mat_values[idx];
}

void mat_multiply_scalar(cl_GPUMat mat, __global float *mat_values, float scalar) {
	uint idx = get_global_id(0);
	if(idx >= mat.rows * mat.cols) return;

	mat_values[idx] *= scalar;
}

cl_GPUMat clm_matrixTranspose(cl_GPUMat mat) {
	cl_GPUMat tr = {0};
	tr.cols = mat.rows;
	tr.rows = mat.cols;
	tr.transposed = !mat.transposed;
	return tr;
}

__kernel void linear_forward(unsigned int batchSize,
	cl_GPUMat input, __global float *input_values,
	cl_GPUMat weights, __global float *weights_values,
	cl_GPUMat bias, __global float *bias_values,
	cl_GPUMat out, __global float *out_values) {
	// FIXME mat_multiply possibly doesn't index at idx when out is transposed, which is important to ensure it always accesses the same index for all operations!
	for(unsigned int b = 0; b < batchSize; b++) {
		__global float *batchInput_values = input_values + b * input.rows * input.cols;
		__global float *batchOut_values = out_values + b * out.rows * out.cols;
		mat_multiply(weights, weights_values, input, batchInput_values, out, batchOut_values);
		mat_add(out, batchOut_values, bias, bias_values);
		mat_sigmoid(out, batchOut_values);
	}
}

__kernel void linear_backprop_1(unsigned int batchSize, float learnRate,
	cl_GPUMat weights, __global float *weights_values,
	cl_GPUMat outputs, __global float *outputs_values,
	cl_GPUMat inputErrors, __global float *inputErrors_values,
	cl_GPUMat outputGradients, __global float *outputGradients_values) {

	for(unsigned int b = 0; b < batchSize; b++) {
		__global float *batchOut_values = outputs_values + b * outputs.rows * outputs.cols;
		__global float *batchInErrors_values = inputErrors_values + b * inputErrors.rows * inputErrors.cols;
		__global float *batchOutGradients_values = outputGradients_values + b * outputGradients.rows * outputGradients.cols;

		mat_copy(outputs, batchOut_values, outputGradients, batchOutGradients_values);
		mat_dSigmoid(outputGradients, batchOutGradients_values); // dsig(yhat)
		mat_multiply_elements(outputGradients, batchOutGradients_values, inputErrors, batchInErrors_values); // (yhat - y) . dsig(yhat)
		mat_multiply_scalar(outputGradients, batchOutGradients_values, learnRate);
	}
}

__kernel void linear_backprop_2(unsigned int batchSize,
	cl_GPUMat weights, __global float *weights_values,
	cl_GPUMat inputs, __global float *inputs_values,
	cl_GPUMat inputErrors, __global float *inputErrors_values,
	char updateErrors,
	cl_GPUMat outputErrors, __global float *outputErrors_values,
	cl_GPUMat outputWeightsErrors, __global float *outputWeightsErrors_values,
	cl_GPUMat outputGradients, __global float *outputGradients_values) {
	for(unsigned int b = 0; b < batchSize; b++) {
		__global float *batchInput_values = inputs_values + b * inputs.rows * inputs.cols;
		__global float *batchInErrors_values = inputErrors_values + b * inputErrors.rows * inputErrors.cols;
		__global float *batchOutErrors_values = outputErrors_values + b * outputErrors.rows * outputErrors.cols;
		__global float *batchOutWeightsErrors_values = outputWeightsErrors_values + b * outputWeightsErrors.rows * outputWeightsErrors.cols;
		__global float *batchOutGradients_values = outputGradients_values + b * outputGradients.rows * outputGradients.cols;

		cl_GPUMat inputsT = clm_matrixTranspose(inputs);
		mat_multiply(outputGradients, batchOutGradients_values, inputsT, batchInput_values, outputWeightsErrors, batchOutWeightsErrors_values);

		if(updateErrors) {
			cl_GPUMat weightsT = clm_matrixTranspose(weightsT);
			// mat_multiply(weightsT, weights_values, inputErrors, batchInErrors_values, outputErrors, batchOutErrors_values);
		}
	}
}