AlexeyAB
2018-02-25 cad4d1618fee74471d335314cb77070fee951a42
Added support for Tensor Cores CC >= 7.0 (V100). For FP16/32 (mixed precision) define CUDNN_HALF should be used.
8 files modified
141 ■■■■ changed files
build/darknet/darknet.vcxproj 2 ●●● patch | view | raw | blame | history
src/convolutional_kernels.cu 122 ●●●● patch | view | raw | blame | history
src/convolutional_layer.c 8 ●●●● patch | view | raw | blame | history
src/convolutional_layer.h 1 ●●●● patch | view | raw | blame | history
src/layer.c 2 ●●●●● patch | view | raw | blame | history
src/layer.h 1 ●●●● patch | view | raw | blame | history
src/network.c 3 ●●●●● patch | view | raw | blame | history
src/network_kernels.cu 2 ●●● patch | view | raw | blame | history
build/darknet/darknet.vcxproj
@@ -145,7 +145,7 @@
      <GenerateDebugInformation>true</GenerateDebugInformation>
      <EnableCOMDATFolding>true</EnableCOMDATFolding>
      <OptimizeReferences>true</OptimizeReferences>
      <AdditionalLibraryDirectories>C:\opencv_3.0\opencv\build\x64\vc14\lib;C:\opencv_2.4.13\opencv\build\x64\vc12\lib;C:\opencv_2.4.13\opencv\build\x64\vc14\lib;$(CUDA_PATH)lib\$(PlatformName);$(cudnn)\lib\x64;%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
      <AdditionalLibraryDirectories>C:\opencv_3.0\opencv\build\x64\vc14\lib;C:\opencv_2.4.13\opencv\build\x64\vc14\lib;$(CUDA_PATH)lib\$(PlatformName);$(cudnn)\lib\x64;%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
      <AdditionalDependencies>..\..\3rdparty\lib\x64\pthreadVC2.lib;cublas.lib;curand.lib;cudart.lib;%(AdditionalDependencies)</AdditionalDependencies>
      <OutputFile>$(OutDir)\$(TargetName)$(TargetExt)</OutputFile>
    </Link>
src/convolutional_kernels.cu
@@ -81,8 +81,8 @@
    //if (idx < size) *((unsigned short *)output_f16 + idx) = __float2half(input_f32[idx]);
}
void cuda_convert_f32_to_f16(float* input_f32, size_t size, half *output_f16) {
    cuda_f32_to_f16 <<< size / BLOCK + 1, BLOCK, 0, get_cuda_stream() >>> (input_f32, size, output_f16);
void cuda_convert_f32_to_f16(float* input_f32, size_t size, float *output_f16) {
    cuda_f32_to_f16 <<< size / BLOCK + 1, BLOCK, 0, get_cuda_stream() >>> (input_f32, size, (half *)output_f16);
}
__global__ void cuda_f16_to_f32(half* input_f16, size_t size, float *output_f32)
@@ -92,8 +92,8 @@
    //if (idx < size) output_f32[idx] = __half2float(*((unsigned short *)input_f16 + idx));
}
void cuda_convert_f16_to_f32(half* input_f16, size_t size, float *output_f32) {
    cuda_f16_to_f32 <<< size / BLOCK + 1, BLOCK, 0, get_cuda_stream() >>> (input_f16, size, output_f32);
void cuda_convert_f16_to_f32(float* input_f16, size_t size, float *output_f32) {
    cuda_f16_to_f32 <<< size / BLOCK + 1, BLOCK, 0, get_cuda_stream() >>> ((half *)input_f16, size, output_f32);
}
half *cuda_make_f16_from_f32_array(float *src, size_t n)
@@ -102,7 +102,7 @@
    size_t size = sizeof(half)*n;
    check_error(cudaMalloc((void **)&dst16, size));
    if (src) {
        cuda_convert_f32_to_f16(src, n, dst16);
        cuda_convert_f32_to_f16(src, n, (float *)dst16);
    }
    if (!dst16) error("Cuda malloc failed\n");
    return dst16;
@@ -124,8 +124,8 @@
    }
#ifdef CUDNN
    //float one = 1;    // alpha[0], beta[0] is float for HALF and FLOAT
    float alpha = 1, beta = 0;
    float one = 1;  // alpha[0], beta[0] is float for HALF and FLOAT
    float alpha = 1, beta = 0;
#ifdef CUDNN_HALF
    // Note: For improved performance it is advised to use beta[0] = 0.0. 
@@ -154,8 +154,9 @@
        output16 = cuda_make_f16_from_f32_array(NULL, max_output16_size);
    }
    cuda_convert_f32_to_f16(state.input, input16_size, input16);
    cuda_convert_f32_to_f16(state.input, input16_size, (float *)input16);
    //fill_ongpu(output16_size / 2, 0, (float *)output16, 1);
    cudnnConvolutionForward(cudnn_handle(),
        &alpha,
        l.srcTensorDesc,
@@ -170,11 +171,12 @@
        l.dstTensorDesc,
        output16);
    
    cuda_convert_f16_to_f32(output16, output16_size, l.output_gpu);
    cuda_convert_f16_to_f32((float *)output16, output16_size, l.output_gpu);
#else
    cudnnConvolutionForward(cudnn_handle(),
                &alpha,
                &one,
                l.srcTensorDesc,
                state.input,
                l.weightDesc,
@@ -183,7 +185,7 @@
                l.fw_algo,
                state.workspace,
                l.workspace_size,
                &beta,
                &one,
                l.dstTensorDesc,
                l.output_gpu);
#endif
@@ -230,7 +232,88 @@
    if(l.xnor) state.input = l.binary_input_gpu;
#ifdef CUDNN
    float one = 1;
    float one = 1;
    float alpha = 1, beta = 0;
#ifdef CUDNN_HALF
    const size_t input16_size = l.batch*l.c*l.w*l.h;
    static size_t max_input16_size = input16_size;
    static half* input16 = cuda_make_f16_from_f32_array(NULL, max_input16_size);
    const size_t delta16_size = l.batch*l.n*l.out_w*l.out_h;
    static size_t max_delta16_size = delta16_size;
    static half* delta16 = cuda_make_f16_from_f32_array(NULL, max_delta16_size);
    if (max_input16_size < input16_size) {
        max_input16_size = input16_size;
        cuda_free((float *)input16);
        input16 = cuda_make_f16_from_f32_array(state.input, max_input16_size);
    }
    if (max_delta16_size < delta16_size) {
        max_delta16_size = delta16_size;
        cuda_free((float *)delta16);
        delta16 = cuda_make_f16_from_f32_array(NULL, max_delta16_size);
    }
    cuda_convert_f32_to_f16(state.input, input16_size, (float *)input16);
    cuda_convert_f32_to_f16(l.delta_gpu, delta16_size, (float *)delta16);
    // convert input: state.input (x), l.delta_gpu (y) from fp32 to fp16
    // get output: l.weight_updates_gpu (dw) and convert it to fp32 (ONLY if it is fp16)
    // calculate conv weight updates
    // Already: l.weight_updates_gpu = (l.weight_updates_gpu - l.weight*decay*batch*subdivision)*momentum
    //   so we should copy f32 to f16, or compute: f16=(w_up - w*d*b*s)*m
    cuda_convert_f32_to_f16(l.weight_updates_gpu, l.c*l.n*l.size*l.size, l.weight_updates_gpu16);
    cudnnConvolutionBackwardFilter(cudnn_handle(),
        &one,
        l.srcTensorDesc,
        input16, //state.input,
        l.ddstTensorDesc,
        delta16, //l.delta_gpu,
        l.convDesc,
        l.bf_algo,
        state.workspace,
        l.workspace_size,
        &one,
        l.dweightDesc,
        l.weight_updates_gpu16);    // l.weight_updates_gpu);
    cuda_convert_f16_to_f32(l.weight_updates_gpu16, l.c*l.n*l.size*l.size, l.weight_updates_gpu);
    if (state.delta) {
        if (l.binary || l.xnor) swap_binary(&l);
        // http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#cudnnConvolutionBackwardData
        // calculate delta for the next layer
        // convert input: l.weights_gpu (w), l.delta_gpu (dy) from fp32 to fp16
        // get output: state.delta (dx) and convert it to fp32 (ONLY if it is fp16)
        cudnnConvolutionBackwardData(cudnn_handle(),
            &alpha,
            l.weightDesc,
            l.weights_gpu16, //l.weights_gpu,
            l.ddstTensorDesc,
            delta16, //l.delta_gpu,
            l.convDesc,
            l.bd_algo,
            state.workspace,
            l.workspace_size,
            &beta,
            l.dsrcTensorDesc,
            input16);   // state.delta);
        cuda_convert_f16_to_f32((float *)input16, input16_size, state.delta);
        if (l.binary || l.xnor) swap_binary(&l);
        if (l.xnor) gradient_array_ongpu(original_input, l.batch*l.c*l.h*l.w, HARDTAN, state.delta);
    }
#else   // CUDNN_HALF
    // calculate conv weight updates
    // if used: beta=1 then loss decreases faster
    cudnnConvolutionBackwardFilter(cudnn_handle(),
            &one,
            l.srcTensorDesc,
@@ -248,6 +331,7 @@
    if(state.delta){
        if(l.binary || l.xnor) swap_binary(&l);
        // http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#cudnnConvolutionBackwardData
        // calculate delta for the next layer
        cudnnConvolutionBackwardData(cudnn_handle(),
                &one,
                l.weightDesc,
@@ -265,7 +349,9 @@
        if(l.xnor) gradient_array_ongpu(original_input, l.batch*l.c*l.h*l.w, HARDTAN, state.delta);
    }
#else
#endif  // CUDNN_HALF
#else   // CUDNN
    int m = l.n;
    int n = l.size*l.size*l.c;
    int k = l.out_w*l.out_h;
@@ -318,7 +404,7 @@
{
    cuda_push_array(layer.weights_gpu, layer.weights, layer.c*layer.n*layer.size*layer.size);
#ifdef CUDNN_HALF
    cuda_convert_f32_to_f16(layer.weights_gpu, layer.c*layer.n*layer.size*layer.size, (half *)layer.weights_gpu16);
    cuda_convert_f32_to_f16(layer.weights_gpu, layer.c*layer.n*layer.size*layer.size, layer.weights_gpu16);
#endif
    cuda_push_array(layer.biases_gpu, layer.biases, layer.n);
    cuda_push_array(layer.weight_updates_gpu, layer.weight_updates, layer.c*layer.n*layer.size*layer.size);
@@ -358,6 +444,14 @@
        adam_gpu(size, layer.weights_gpu, layer.m_gpu, layer.v_gpu, layer.B1, layer.B2, learning_rate/batch, layer.eps, layer.t+1);
        fill_ongpu(size, 0, layer.weight_updates_gpu, 1);
    }else{
        // update weights:
        // weights_gpu = weights_gpu*(1 - decay*lr) + weight_updates_gpu*lr / (batch*subdivision) =
        //  weights_gpu*(1 - 0.0005*0.001) + weight_updates_gpu*0.001/(64*8) =
        //  weights_gpu * 0.999 999 5 + weight_updates_gpu * 0.000 001 953125
        //
        // weight_updates_gpu = (weight_updates_gpu - weights_gpu*decay*batch*subdivision)*momentum =
        //  (weight_updates_gpu - weights_gpu * 0.0005 * 64 * 8) * 0.9 =
        //  weight_updates_gpu*0.9 - weights_gpu*0.2304
        axpy_ongpu(size, -decay*batch, layer.weights_gpu, 1, layer.weight_updates_gpu, 1);
        axpy_ongpu(size, learning_rate/batch, layer.weight_updates_gpu, 1, layer.weights_gpu, 1);
        scal_ongpu(size, momentum, layer.weight_updates_gpu, 1);
src/convolutional_layer.c
@@ -141,7 +141,8 @@
{
#ifdef CUDNN_HALF
    // TRUE_HALF_CONFIG is only supported on architectures with true fp16 support (compute capability 5.3 and 6.0): Tegra X1, Jetson TX1, DRIVE CX, DRIVE PX, Quadro GP100, Tesla P100
    // TRUE_HALF_CONFIG is only supported on architectures with true fp16 support (compute capability 5.3 and 6.0):
    //   Tegra X1, Jetson TX1, DRIVE CX, DRIVE PX, Quadro GP100, Tesla P100
    // PSEUDO_HALF_CONFIG is required for Tensor Cores - our case!
    const cudnnDataType_t data_type = CUDNN_DATA_HALF;
#else
@@ -164,10 +165,12 @@
    //   on architectures with DP4A support (compute capability 6.1 and later).
    //cudnnDataType_t data_type = CUDNN_DATA_INT8;
    // backward delta
    cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->c, l->h, l->w);
    cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
    cudnnSetFilter4dDescriptor(l->dweightDesc, data_type, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
    // forward
    cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->c, l->h, l->w);
    cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
    cudnnSetFilter4dDescriptor(l->weightDesc, data_type, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
@@ -302,7 +305,8 @@
        l.weights_gpu = cuda_make_array(l.weights, c*n*size*size);
#ifdef CUDNN_HALF
        l.weights_gpu16 = cuda_make_array(l.weights, c*n*size*size/2);
        l.weights_gpu16 = cuda_make_array(l.weights, c*n*size*size / 2);
        l.weight_updates_gpu16 = cuda_make_array(l.weight_updates, c*n*size*size / 2);
#endif
        l.weight_updates_gpu = cuda_make_array(l.weight_updates, c*n*size*size);
src/convolutional_layer.h
@@ -21,6 +21,7 @@
void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int size);
#ifdef CUDNN
void cudnn_convolutional_setup(layer *l, int cudnn_preference);
void cuda_convert_f32_to_f16(float* input_f32, size_t size, float *output_f16);
#endif
#endif
src/layer.c
@@ -83,6 +83,8 @@
    if (l.x_norm_gpu)              cuda_free(l.x_norm_gpu);
    if (l.weights_gpu)             cuda_free(l.weights_gpu);
    if (l.weight_updates_gpu)      cuda_free(l.weight_updates_gpu);
    if (l.weights_gpu16)           cuda_free(l.weights_gpu16);
    if (l.weight_updates_gpu16)    cuda_free(l.weight_updates_gpu16);
    if (l.biases_gpu)              cuda_free(l.biases_gpu);
    if (l.bias_updates_gpu)        cuda_free(l.bias_updates_gpu);
    if (l.scales_gpu)              cuda_free(l.scales_gpu);
src/layer.h
@@ -243,6 +243,7 @@
    float * weight_updates_gpu;
    float * weights_gpu16;
    float * weight_updates_gpu16;
    float * biases_gpu;
    float * bias_updates_gpu;
src/network.c
@@ -316,6 +316,8 @@
        net->layers[i].batch = b;
#ifdef CUDNN
        if(net->layers[i].type == CONVOLUTIONAL){
            cudnn_convolutional_setup(net->layers + i, cudnn_fastest);
            /*
            layer *l = net->layers + i;
            cudnn_convolutional_setup(l, cudnn_fastest);
            // check for excessive memory consumption 
@@ -327,6 +329,7 @@
                cudnn_convolutional_setup(l, cudnn_smallest);
                l->workspace_size = get_workspace_size(*l);
            }
            */
        }
#endif
    }
src/network_kernels.cu
@@ -117,7 +117,7 @@
    int i;
    for (i = 0; i < net.n; ++i) {
        layer l = net.layers[i];
        cuda_convert_f32_to_f16(l.weights_gpu, l.c*l.n*l.size*l.size, (half *)l.weights_gpu16);
        cuda_convert_f32_to_f16(l.weights_gpu, l.c*l.n*l.size*l.size, l.weights_gpu16);
    }
#endif
    forward_network_gpu(net, state);