def test_perf_gemm(m, k, n, xclbin_opts, post_scale=[1,0], A_range=32764, B_range=32764, bias_range=32764): ddrWidth = int(xclbin_opts["GEMX_ddrWidth"]) m = test.get_padded_size(m, int(xclbin_opts["GEMX_gemmMBlocks"]) * ddrWidth) k = test.get_padded_size(k, int(xclbin_opts["GEMX_gemmKBlocks"]) * ddrWidth) n = test.get_padded_size(n, int(xclbin_opts["GEMX_gemmNBlocks"]) * ddrWidth) mat_A = np.random.randint(low=-A_range, high=A_range, size=(m, k), dtype=np.int16) mat_B = np.random.randint(low=-B_range, high=B_range, size=(k, n), dtype=np.int16) bias = [] if bias_range != 0: bias = np.random.randint(low=-bias_range, high=bias_range, size=(m, n), dtype=np.int32) else: bias = np.zeros ((m, n), dtype=np.int32, order='C'); C_fpga = np.zeros( (m, n), dtype=np.int16) start_time = time.time() gemx.sendMat(mat_A) gemx.sendMat(mat_B) gemx.sendMat(C_fpga) gemx.sendMat(bias) gemx.addGEMMOp ( mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1]) gemx.execute() gemx.clearInstrBuf() gemx.getMat(C_fpga) end_time = time.time() total_operations = 2 * m * n * k + m * n * 3 test.test_perf(end_time-start_time,total_operations,m,k,n,ddrWidth) test.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def test_basic(self, PE, xclbin_opts, mat_A, mat_B, bias, post_scale=[1, 0]): m = mat_A.shape[0] k = mat_A.shape[1] n = mat_B.shape[1] print("test_basic(PE=%d): %d %d %d %d %d" % (PE, m, k, n, post_scale[0], post_scale[1])) print("A: ", np.amax(mat_A), np.amin(mat_A), np.average(mat_A)) print("B: ", np.amax(mat_B), np.amin(mat_B), np.average(mat_B)) print("bias: ", np.amax(bias), np.amin(bias), np.average(bias)) if xclbin_opts["GEMX_dataType"] == "short": C_fpga = np.zeros((m, n), dtype=np.int16, order='C') else: #float C_fpga = np.zeros((m, n), dtype=np.float32, order='C') gemx.sendMat(mat_A, PE) gemx.sendMat(mat_B, PE) gemx.sendMat(C_fpga, PE) gemx.sendMat(bias, PE) gemx.addGEMMOp(mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1], PE) # default test_basic will call addGEMMOp gemx.execute(PE) gemx.clearInstrBuf(PE) gemx.getMat(C_fpga, PE) self.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def loadInstr(self): gemx.clearInstrBuf() for i,l in enumerate(self.kmodel.layers): act = l.get_config()['activation'] if act == 'relu': gemx.addFCNOp( self.fpga_buf[i], self._qw[i], self.fpga_buf[i+1], self._qb[i], self.post_scale[i][0], self.post_scale[i][1], 0, 0) else: gemx.addGEMMOp( self.fpga_buf[i], self._qw[i], self.fpga_buf[i+1], self._qb[i], self.post_scale[i][0], self.post_scale[i][1])
def test_textfiles(self, path_to_a, path_to_b, path_to_bias, post_scale): mat_A = np.loadtxt(path_to_a, dtype=np.int16) mat_B = np.loadtxt(path_to_b, dtype=np.int16) bias = np.loadtxt(path_to_bias, dtype=np.int32) m = mat_A.shape[0] n = mat_B.shape[1] C_fpga = np.zeros((m, n), dtype=np.int16, order='C') gemx.sendMat(mat_A) gemx.sendMat(mat_B) gemx.sendMat(C_fpga) gemx.sendMat(bias) gemx.addGEMMOp (mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1]) gemx.execute() gemx.clearInstrBuf() gemx.getMat(C_fpga) self.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def test_multiInstrv1(int_range, m, k, n, add_bias=False): print("test_multiInstrv1: %d %d %d %d" % (int_range, m, k, n)) A = np.random.randint(low=-int_range, high=int_range, size=(m, k), dtype=np.int16) B = np.random.randint(low=-int_range, high=int_range, size=(k, n), dtype=np.int16) C = np.zeros((m, n), dtype=np.int16) D = np.random.randint(low=-int_range, high=int_range, size=(m, k), dtype=np.int16) E = np.zeros((m, n), dtype=np.int16) b0 = np.zeros((m, n), dtype=np.int32) b1 = np.zeros((m, n), dtype=np.int32) if add_bias == True: b0 = np.random.randint(low=-int_range, high=int_range, size=(m, n), dtype=np.int32) b1 = np.random.randint(low=-int_range, high=int_range, size=(m, n), dtype=np.int32) gemx.sendMat(A) gemx.sendMat(B) gemx.sendMat(b0) gemx.sendMat(C) gemx.sendMat(D) gemx.sendMat(E) gemx.sendMat(b1) gemx.addGEMMOp(A, B, C, b0, 1, 0) gemx.addGEMMOp(D, C, E, b1, 1, 0) gemx.execute() gemx.clearInstrBuf() gemx.getMat(C) gemx.getMat(E) print("test C") test.multiply_and_cmp(C, A, B, b0, m, n, [1, 0]) print("test E") test.multiply_and_cmp(E, D, C, b1, m, n, [1, 0])
def test_basic(self,PE, mat_A, mat_B, bias, post_scale = [1,1]): m = mat_A.shape[0] k = mat_A.shape[1] n = mat_B.shape[1] print ("test_basic(PE=%d): %d %d %d %d %d" % (PE,m, k, n, post_scale[0], post_scale[1] )) print ("A: ", np.amax(mat_A), np.amin(mat_A), np.average(mat_A)) print ("B: ", np.amax(mat_B), np.amin(mat_B), np.average(mat_B)) print ("bias: ", np.amax(bias), np.amin(bias), np.average(bias)) C_fpga = np.zeros( (m, n), dtype=np.int16) gemx.sendMat(mat_A,PE) gemx.sendMat(mat_B,PE) gemx.sendMat(C_fpga,PE) gemx.sendMat(bias, PE) gemx.addGEMMOp ( mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1], PE) # default test_basic will call addGEMMOp gemx.execute(PE) gemx.getMat(C_fpga,PE) self.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def test_perf_multi_gemm(ins_count, m_size, k_size, n_size, A_range, B_range, post_scale): total_operations = 0 total_parallel_operations = 0 mat_A = [] mat_C = [] mat_bias = [] for i in range(ins_count): total_operations += 2 * m_size[i] * n_size[i] * k_size[i] + m_size[ i] * n_size[i] * 3 total_parallel_operations += 2 * m_size[i] * n_size[i] * k_size[i] mat_A.append( np.random.randint(low=-A_range, high=A_range, size=(m_size[i], k_size[i]), dtype=np.int16)) mat_bias.append(np.zeros((m_size[i], n_size[i]), dtype=np.int32)) mat_C.append( np.zeros((m_size[i], n_size[i]), dtype=np.int16, order='C')) mat_B0 = np.random.randint(low=-B_range, high=B_range, size=(k_size[0], n_size[0]), dtype=np.int16) timePointKernel = [] timePointKernel.append(time.time()) # current time for i in range(ins_count): gemx.sendMat(mat_A[i]) gemx.sendMat(mat_C[i]) gemx.sendMat(mat_bias[i]) gemx.sendMat(mat_B0) gemx.addGEMMOp(mat_A[0], mat_B0, mat_C[0], mat_bias[0], post_scale[0], post_scale[1]) gemx.addGEMMOp(mat_A[1], mat_C[0], mat_C[1], mat_bias[1], post_scale[0], post_scale[1]) gemx.addGEMMOp(mat_A[2], mat_C[1], mat_C[2], mat_bias[2], post_scale[0], post_scale[1]) gemx.addGEMMOp(mat_A[3], mat_C[2], mat_C[3], mat_bias[3], post_scale[0], post_scale[1]) timePointKernel.append(time.time()) # send to FPGA gemx.execute() timePointKernel.append(time.time()) # call kernel gemx.getMat(mat_C[0]) gemx.getMat(mat_C[1]) gemx.getMat(mat_C[2]) gemx.getMat(mat_C[3]) timePointKernel.append(time.time()) # copy from FPGA freq = gemx.getFreq() test.test_perf(timePointKernel, total_operations, total_parallel_operations, freq, 0, 0, 0) if np.max(m_size) > 4096 and np.max(k_size) > 4096 and np.max( n_size) > 4096: print("Skip golden comparision because large matrix size") else: test.multiply_and_cmp(mat_C[3], mat_A[3], mat_C[2], mat_bias[3], m_size[3], n_size[3], post_scale)
def predict ( self, inp, in_scale, post_scale): row_padded, col_padded = self.get_padded_shape( inp.shape, self.min_m, self.min_k) padded_arr = np.zeros ( (row_padded, col_padded), dtype=inp.dtype, order='C') padded_arr[0:inp.shape[0], 0:inp.shape[1]] = inp print ("input shape", padded_arr.shape) np.copyto(self.fpga_buf[0], np.int16( padded_arr * in_scale ), casting='same_kind', where=True) gemx.sendMat(self.fpga_buf[0]) for i,l in enumerate(self.kmodel.layers): act = l.get_config()['activation'] if act == 'relu': gemx.addFCNOp( self.fpga_buf[i], self.w[i], self.fpga_buf[i+1], self.b[i], post_scale[i][0], post_scale[i][1], 0, 0) else: gemx.addGEMMOp( self.fpga_buf[i], self.w[i], self.fpga_buf[i+1], self.b[i], post_scale[i][0], post_scale[i][1]) gemx.execute() gemx.getMat (self.fpga_buf[-1]) return self.fpga_buf[-1][:self.out_dim[0],:self.out_dim[1]]
def test_perf_gemm(m, k, n, A_range=32764, B_range=32764, bias_range=32764, post_scale=[1, 0]): mat_A = np.random.randint(low=-A_range, high=A_range, size=(m, k), dtype=np.int16) mat_B = np.random.randint(low=-B_range, high=B_range, size=(k, n), dtype=np.int16) bias = [] if bias_range != 0: bias = np.random.randint(low=-bias_range, high=bias_range, size=(m, n), dtype=np.int32) else: bias = np.zeros((m, n), dtype=np.int32, order='C') C_fpga = np.zeros((m, n), dtype=np.int16) timePointKernel = [] timePointKernel.append(time.time()) # current time gemx.sendMat(mat_A) gemx.sendMat(mat_B) gemx.sendMat(C_fpga) gemx.sendMat(bias) gemx.addGEMMOp(mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1]) timePointKernel.append(time.time()) # send to FPGA gemx.execute() gemx.clearInstrBuf() timePointKernel.append(time.time()) # call kernel gemx.getMat(C_fpga) timePointKernel.append(time.time()) # copy from FPGA total_operations = 2 * m * n * k + m * n * 3 total_parallel_operations = 2 * m * n * k freq = gemx.getFreq() test.test_perf(timePointKernel, total_operations, total_parallel_operations, freq, m, k, n) test.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def test_perf_gemm_gemm(A_range, B_range, bias_range, m, k, n, post_scale): mat_A = np.random.randint(low=-A_range, high=A_range, size=(m, k), dtype=np.int16) mat_B = np.random.randint(low=-B_range, high=B_range, size=(k, n), dtype=np.int16) bias = [] if bias_range != 0: bias = np.random.randint(low=-bias_range, high=bias_range, size=(m, n), dtype=np.int32) else: bias = np.zeros((m, n), dtype=np.int32, order='C') C_fpga = np.zeros((m, n), dtype=np.int16) timePointKernel = [] timePointKernel.append(time.time()) # current time gemx.sendMat(mat_A) gemx.sendMat(mat_B) gemx.sendMat(C_fpga) gemx.sendMat(bias) gemx.addGEMMOp(mat_A, mat_B, C_fpga, bias, post_scale[0], post_scale[1]) timePointKernel.append(time.time()) # send to FPGA gemx.execute() timePointKernel.append(time.time()) # call kernel gemx.getMat(C_fpga) timePointKernel.append(time.time()) # copy from FPGA total_operations = 2 * m * n * k + m * n * 3 total_parallel_operations = 2 * m * n * k freq = gemx.getFreq() test.test_perf(timePointKernel, total_operations, total_parallel_operations, freq, m, k, n) if m > 4096 and n > 4096 and k > 4096: print("Skip golden comparision because large matrix size") else: test.multiply_and_cmp(C_fpga, mat_A, mat_B, bias, m, n, post_scale)
def loadInstr(self): gemx.clearInstrBuf() for i, (w_i, b_i) in enumerate(zip(self._qw, self._qb)): gemx.addGEMMOp(w_i, self.fpga_buf[i], self.fpga_buf[i + 1], b_i, self.post_scale[i][0], self.post_scale[i][1])