From 41a841f3e0992a578cf5c8f82ae44a552a6f8b2f Mon Sep 17 00:00:00 2001 From: txh18 Date: Wed, 2 Dec 2015 15:23:54 +0800 Subject: changed thres_mask function of matrix to a more standard api --- nerv/examples/lmptb/tnn/init.lua | 1 + nerv/layer/init.lua | 6 +++--- nerv/lib/matrix/cukernel.h | 2 +- nerv/lib/matrix/generic/cukernel.cu | 11 ++++++----- nerv/lib/matrix/generic/cumatrix.c | 5 +++-- nerv/matrix/generic/cumatrix.c | 9 +++++---- 6 files changed, 19 insertions(+), 15 deletions(-) diff --git a/nerv/examples/lmptb/tnn/init.lua b/nerv/examples/lmptb/tnn/init.lua index d45a2fa..ddaa6b8 100644 --- a/nerv/examples/lmptb/tnn/init.lua +++ b/nerv/examples/lmptb/tnn/init.lua @@ -41,6 +41,7 @@ function LayerT:get_dim() return self.dim_in, self.dim_out end +nerv.include('sutil.lua') nerv.include('tnn.lua') nerv.include('layersT/softmax_ce_t.lua') nerv.include('layersT/lstm_t.lua') diff --git a/nerv/layer/init.lua b/nerv/layer/init.lua index b8b7ea1..32b82d8 100644 --- a/nerv/layer/init.lua +++ b/nerv/layer/init.lua @@ -72,18 +72,18 @@ end function Layer:find_param(pid, l_conf, gconf, p_type, p_dim) if l_conf[pid] ~= nil then - nerv.info("Param [%s] of layer [%s] found in layer_conf.\n", pid, self.id) + nerv.info("Param [%s] of layer [%s] found in layer_conf.", pid, self.id) return l_conf[pid] end local pid_g = self.id .. '_' .. pid --global identifier local pr = l_conf.pr local p if pr ~= nil and pr:has_param(pid_g) == true then - nerv.info("Param [%s] of layer [%s] found in layer_conf.paramRepo.\n", pid, self.id) + nerv.info("Param [%s] of layer [%s] found in layer_conf.paramRepo.", pid, self.id) p = pr:get_param(pid_g) return p end - nerv.info("Param [%s] of layer [%s] is not found in layer_conf or layer_conf.paramRepo, switch to auto-generate.\n", pid, self.id) + nerv.info("Param [%s] of layer [%s] is not found in layer_conf or layer_conf.paramRepo, switch to auto-generate.", pid, self.id) p = p_type(pid_g, gconf) p.trans = gconf.cumat_type(unpack(p_dim)) p.trans:generate(gconf.param_random) diff --git a/nerv/lib/matrix/cukernel.h b/nerv/lib/matrix/cukernel.h index 47dc0a8..fe682d3 100644 --- a/nerv/lib/matrix/cukernel.h +++ b/nerv/lib/matrix/cukernel.h @@ -4,7 +4,7 @@ void cudak_(cuda_log_elem)(const Matrix *a, Matrix *b); void cudak_(cuda_sigmoid)(const Matrix *a, Matrix *b); void cudak_(cuda_sigmoid_grad)(const Matrix *output, const Matrix *err, Matrix *nerr); void cudak_(cuda_rand_uniform)(const Matrix *a); //a's curand_gen may be modified -void cudak_(cuda_thres_mask)(const Matrix *a, double thres, double low, double high); +void cudak_(cuda_thres_mask)(const Matrix *a, const Matrix *b, double thres, double low, double high); void cudak_(cuda_tanh)(const Matrix *a, Matrix *b); void cudak_(cuda_tanh_grad)(const Matrix *output, const Matrix *err, Matrix *nerr); void cudak_(cuda_rowsum)(const Matrix *a, Matrix *b); diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu index b092e4a..aa830b5 100644 --- a/nerv/lib/matrix/generic/cukernel.cu +++ b/nerv/lib/matrix/generic/cukernel.cu @@ -20,14 +20,14 @@ __global__ void cudak_(log_elem)(const MATRIX_ELEM *a, MATRIX_ELEM *b, b[idx] = log(tmp); } -__global__ void cudak_(thres_mask)(MATRIX_ELEM *a, double thres, double low, double high, +__global__ void cudak_(thres_mask)(MATRIX_ELEM *a, MATRIX_ELEM *b, double thres, double low, double high, int nrow, int ncol, int stride) { int j = blockIdx.x * blockDim.x + threadIdx.x; int i = blockIdx.y * blockDim.y + threadIdx.y; long idx; if (i >= nrow || j >= ncol) return; idx = j + i * stride; - if (a[idx] < thres) + if (b[idx] < thres) a[idx] = low; else a[idx] = high; @@ -389,7 +389,7 @@ extern "C" { cudaStreamSynchronize(0); } - void cudak_(cuda_rand_uniform)(Matrix *a) { + void cudak_(cuda_rand_uniform)(const Matrix *a) { #ifdef MATRIX_USE_FLOAT curandGenerateUniform(*(a->curand_gen), MATRIX_ELEM_PTR(a), a->nrow * a->stride / sizeof(MATRIX_ELEM)); #endif @@ -398,12 +398,13 @@ extern "C" { #endif } - void cudak_(cuda_thres_mask)(const Matrix *a, double thres, double low, double high) { + void cudak_(cuda_thres_mask)(const Matrix *a, const Matrix *b, double thres, double low, double high) { dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N); dim3 numBlocks(CEIL_DIV(a->ncol, threadsPerBlock.x), CEIL_DIV(a->nrow, threadsPerBlock.y)); cudak_(thres_mask)<<>> \ - (MATRIX_ELEM_PTR(a), thres, low, high, a->nrow, a->ncol, a->stride / sizeof(MATRIX_ELEM)); + (MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(b), + thres, low, high, a->nrow, a->ncol, a->stride / sizeof(MATRIX_ELEM)); cudaStreamSynchronize(0); } diff --git a/nerv/lib/matrix/generic/cumatrix.c b/nerv/lib/matrix/generic/cumatrix.c index cbb0481..65e0788 100644 --- a/nerv/lib/matrix/generic/cumatrix.c +++ b/nerv/lib/matrix/generic/cumatrix.c @@ -83,9 +83,10 @@ void nerv_matrix_(rand_uniform)(Matrix *a, Status *status) { NERV_SET_STATUS(status, NERV_NORMAL, 0); } -void nerv_matrix_(thres_mask)(Matrix *a, double thres, double low, double high, Status *status) { +void nerv_matrix_(thres_mask)(Matrix *a, Matrix *b, double thres, double low, double high, Status *status) { + CHECK_SAME_DIMENSION(a, b, status); PROFILE_START - cudak_(cuda_thres_mask)(a, thres, low, high); + cudak_(cuda_thres_mask)(a, b, thres, low, high); PROFILE_STOP NERV_SET_STATUS(status, NERV_NORMAL, 0); } diff --git a/nerv/matrix/generic/cumatrix.c b/nerv/matrix/generic/cumatrix.c index d1f763b..fb36033 100644 --- a/nerv/matrix/generic/cumatrix.c +++ b/nerv/matrix/generic/cumatrix.c @@ -65,10 +65,11 @@ static int nerv_matrix_(lua_sigmoid_grad)(lua_State *L) { static int nerv_matrix_(lua_thres_mask)(lua_State *L) { Status status; Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname)); - MATRIX_ELEM thres = luaL_checknumber(L, 2); - MATRIX_ELEM low = luaL_checknumber(L, 3); - MATRIX_ELEM high = luaL_checknumber(L, 4); - nerv_matrix_(thres_mask)(a, thres, low, high, &status); + Matrix *b = luaT_checkudata(L, 2, nerv_matrix_(tname)); + MATRIX_ELEM thres = luaL_checknumber(L, 3); + MATRIX_ELEM low = luaL_checknumber(L, 4); + MATRIX_ELEM high = luaL_checknumber(L, 5); + nerv_matrix_(thres_mask)(a, b, thres, low, high, &status); NERV_LUA_CHECK_STATUS(L, status); return 0; } -- cgit v1.2.3