public static void fpbinary_gemm_cpu(IntPtr a, IntPtr b, IntPtr c, int m, int nn, int k, int transb, int beta, int alpha, IntPtr alphas) { if (THWrapper.THFloatTensor_nDimension(c) != 2 || THWrapper.THFloatTensor_size(c, 0) * THWrapper.THFloatTensor_size(c, 1) < m * k) { THWrapper.THFloatTensor_resize2d(c, m, k); //THFloatTensor_resize2d(c, m, k); } /* * uint32_t* A = (uint32_t*)THIntTensor_data(a); * uint32_t* B = (uint32_t*)THIntTensor_data(b); * float* C = THFloatTensor_data(c); * float* D = THFloatTensor_data(alphas); */ var A = THWrapper.THIntTensor_data(a); var B = THWrapper.THIntTensor_data(b); var C = THWrapper.THFloatTensor_data(c); var D = THWrapper.THFloatTensor_data(alphas); var aa = InternalArray.FromTHIntTensor(a); var bb = InternalArray.FromTHIntTensor(b); var cc = InternalArray.FromTHFloatTensor(c); var dd = InternalArray.FromTHFloatTensor(alphas); int n = 1 + (nn - 1) / matmul.ENCODE_BIT; int brow = transb != 0 ? 1 : k; int bcol = transb != 0 ? n : 1; //matmul.dgemm_nn(m, k, nn, A, n, 1, B, brow, bcol, C, k, 1, beta, alpha, D); //matmul.dgemm_nn(m, k, nn, A, n, 1, B, brow, bcol, C, k, 1, beta, alpha, D); matmul.fpdgemm_nn(m, k, nn, A, n, 1, B, brow, bcol, C, k, 1, beta); if (alpha != 0) { for (int i = 0; i < m; i++) { for (int j = 0; j < k; j++) { //C[i * n + j] *= alphas[i]; var aa1 = matmul.GetFloat(C, i * k + j); short aq1 = (short)(aa1 * 256); var aa2 = matmul.GetFloat(D, i); var aq2 = (short)(aa2 * 256); var val4 = (short)((int)(aq1 * aq2) >> 8); var quant_res = val4 / 256f; var orig = aa1 * aa2; //matmul.SetFloat(C, i * k + j, aa1 * aa2); matmul.SetFloat(C, i * k + j, val4); //C[i * n + j] = (float)(C[i * n + j] * alphas[i]); } } } }
public static InternalArray FromTHFloatTensor(IntPtr tensor) { var dims = THWrapper.THFloatTensor_nDimension(tensor); int[] d = new int[dims]; for (int i = 0; i < dims; i++) { d[i] = (int)THWrapper.THFloatTensor_size(tensor, i); } InternalArray ret = new InternalArray(d); if (dims == 2) { for (int i = 0; i < d[0]; i++) { for (int j = 0; j < d[1]; j++) { ret.Set2D(i, j, THWrapper.THFloatTensor_get2d(tensor, i, j)); } } } if (dims == 3) { for (int k = 0; k < d[0]; k++) { for (int i = 0; i < d[1]; i++) { for (int j = 0; j < d[2]; j++) { ret.Set3D(k, i, j, THWrapper.THFloatTensor_get3d(tensor, k, i, j)); } } } } if (dims == 4) { for (int k1 = 0; k1 < d[0]; k1++) { for (int k = 0; k < d[1]; k++) { for (int i = 0; i < d[2]; i++) { for (int j = 0; j < d[3]; j++) { ret.Set4D(k1, k, i, j, THWrapper.THFloatTensor_get4d(tensor, k1, k, i, j)); } } } } } ret.UpdateOffsets(); return(ret); }
public static void THNN_Bin_SpatialConvolutionMM_updateOutput_frame( IntPtr output, //float IntPtr weight, //int IntPtr bias, //float IntPtr ones, //float IntPtr bin_col, //int IntPtr alphas, //float int kW, int kH, int dW, int dH, int padW, int padH, Int64 nInputPlane, Int64 inputWidth, Int64 inputHeight, Int64 nOutputPlane, Int64 outputWidth, Int64 outputHeight, bool quantOutput = false) { IntPtr output2d; //var output2d = THFloatTensor_newWithStorage2d(output->storage, output->storageOffset, nOutputPlane, -1, outputHeight * outputWidth, -1); //var output2d = THFloatTensor_newWithStorage2d(output, (int)nOutputPlane, -1, (int)(outputHeight * outputWidth), -1); var strg = THWrapper.THFloatTensor_storage(output); var offset = THWrapper.THFloatTensor_storageOffset(output); output2d = THWrapper.THFloatTensor_newWithStorage2d(strg, offset, nOutputPlane, (long)-1, outputWidth * outputHeight, (long)-1); //InternalArray output2d = new InternalArray(new int[] { }); THWrapper.THFloatTensor_zero(output2d); binary_gemm_cpu(weight, bin_col, output2d, (int)nOutputPlane, (int)(kW * kH * nInputPlane), (int)(outputHeight * outputWidth), 0, 1, 1, alphas, quantOutput); if (bias != null && THWrapper.THFloatTensor_nDimension(bias) != 0) { THWrapper.THFloatTensor_addmm(output2d, 1, output2d, 1, bias, ones); //THFloatTensor_addmm(output2d, 1, output2d, 1, bias, ones); } THWrapper.THFloatTensor_free(output2d); //THWrapper.THFloatTensor_free(_ones); //THFloatTensor_free(output2d); }
/*public static void memcpy(float[] a1, long dst, float[] a2, long dst2, long size) * { * for (int i = 0; i < size / 4; i++) * { * a1[i + dst] = a2[i + dst2]; * } * }*/ /* * THFloatTensor *input, * THFloatTensor *output, * THIntTensor *weight, * THFloatTensor *bias, * THFloatTensor *columns, * THFloatTensor *alphas, * int kH, int kW, * int dH, int dW, * int padH, int padW)*/ public static void THNN_Bin_SpatialConvolutionMM_updateOutput( IntPtr input, IntPtr output, IntPtr weight, IntPtr bias, IntPtr columns, IntPtr alphas, int kH, int kW, int dH, int dW, int padH, int padW , bool quantOutput = false) { int ndim = THWrapper.THFloatTensor_nDimension(input); int dimf = 0; int dimh = 1; int dimw = 2; if (ndim == 4) { dimf++; dimh++; dimw++; } var nInputPlane = THWrapper.THFloatTensor_size(input, dimf); var inputHeight = THWrapper.THFloatTensor_size(input, dimh); var inputWidth = THWrapper.THFloatTensor_size(input, dimw); var nOutputPlane = THWrapper.THFloatTensor_size(weight, 0); var outputHeight = (inputHeight + 2 * padH - kH) / dH + 1; var outputWidth = (inputWidth + 2 * padW - kW) / dW + 1; //InternalArray ones = new InternalArray(new int[] { 1 }); IntPtr ones = THWrapper.THFloatTensor_new(); if (bias != null && THWrapper.THFloatTensor_nDimension(bias) == 1) { THWrapper.THFloatTensor_resize2d(bias, THWrapper.THFloatTensor_size(bias, 0), 1); //THFloatTensor_resize2d(bias, bias.Shape[0], 1); } THWrapper.THFloatTensor_resize2d(ones, 1, outputHeight * outputWidth); //THFloatTensor_resize2d(ones, 1, outputHeight * outputWidth); THWrapper.THFloatTensor_fill(ones, 1); //THFloatTensor_fill(ones, 1); var T = THWrapper.THFloatTensor_size(input, 0); //InternalArray bin_col = new InternalArray(new int[] { 1 }); var bin_col = THWrapper.THIntTensor_new(); THWrapper.THFloatTensor_resize4d(output, T, (int)nOutputPlane, outputHeight, outputWidth); //THFloatTensor_resize4d(output, T, (int)nOutputPlane, outputHeight, outputWidth); THWrapper.THFloatTensor_resize3d(columns, T, kW * kH * nInputPlane, outputHeight * outputWidth); //THFloatTensor_resize3d(columns, T, kW * kH * nInputPlane, outputHeight * outputWidth); THWrapper.THIntTensor_resize3d(bin_col, T, (int)nOutputPlane, outputHeight * outputWidth); //THIntTensor_resize3d(bin_col, T, (int)nOutputPlane, outputHeight * outputWidth); for (int t = 0; t < T; t++) { /*var input_t = input.Get2DImageFrom4DArray(0, t); * var columns_t = columns.Get2DImageFrom4DArray(0, t); * var bin_col_t = bin_col.Get2DImageFrom4DArray(0, t);*/ //var _bin_col = bin_col.ToTHTensor(); var input_t = THWrapper.THFloatTensor_newSelect(input, 0, t); var columns_t = THWrapper.THFloatTensor_newSelect(columns, 0, t); var bin_col_t = THWrapper.THIntTensor_newSelect(bin_col, 0, t); /* var bbb = InternalArray.FromTHFloatTensor(columns_t); * for (int i = 0; i < bbb.Data.Length; i++) * { * var res = bbb.Data[i]; * if (res != 0) * { * * } * * }*/ THNN_unfolded_copy( columns_t, input_t, kW, kH, dW, dH, padW, padH, (int)nInputPlane, (int)inputWidth, (int)inputHeight, (int)outputWidth, (int)outputHeight ); // bbb = InternalArray.FromTHFloatTensor(columns_t); //Debug.Assert(bbb.Data[1800] == -0.053401); //Debug.Assert(bbb.Data[4200] == -0.852360); /* for (int i = 0; i < bbb.Data.Length; i++) * { * var res = bbb.Data[i]; * if (res != 0) * { * * } * }*/ encode_cols_cpu(columns_t, bin_col_t); // var bb = InternalArray.FromTHIntTensor(bin_col_t); /* for (int i = 0; i < bb.IntData.Length; i++) * { * var res = bb.IntData[i]; * if (res != 0) * { * * } * }*/ } for (int t = 0; t < T; t++) { /* THFloatTensor* output_t = THFloatTensor_newSelect(output, 0, t); * THIntTensor* bin_col_t = THIntTensor_newSelect(bin_col, 0, t); */ /*var output_t = output.Get2DImageFrom4DArray(0, t); * var bin_col_t = bin_col.Get2DImageFrom4DArray(0, t);*/ //var _output = output.ToTHTensor(); //var _bin_col = bin_col.ToTHTensor(); var output_t = THWrapper.THFloatTensor_newSelect(output, 0, t); var bin_col_t = THWrapper.THIntTensor_newSelect(bin_col, 0, t); THNN_Bin_SpatialConvolutionMM_updateOutput_frame( output_t, weight, bias, ones, bin_col_t, alphas, kW, kH, dW, dH, padW, padH, nInputPlane, inputWidth, inputHeight, nOutputPlane, outputWidth, outputHeight, quantOutput ); } }