Merge pull request #7 from DavidDiazGuerra/mixed_precision

Mixed precision

.gitignore

@@ -41,4 +41,4 @@ Thumbs.db
 ###########
 build/
 dist/
-gpurRIR.egg-info/
+gpuRIR.egg-info/

CMakeLists.txt

@@ -1,22 +1,23 @@
 # https://stackoverflow.com/questions/51907755/building-a-pybind11-module-with-cpp-and-cuda-sources-using-cmake
 
 cmake_minimum_required(VERSION 3.12)
-#set(CMAKE_CXX_COMPILER "/usr/local/gcc55/bin/g++55")
+
 project(gpuRIR LANGUAGES CXX CUDA)
 find_package(CUDA)
 
-set(PYBIND11_CPP_STANDARD -std=c++11)
-add_subdirectory(third_party/pybind11)
+# Select several CUDA architectures
+include(FindCUDA)
+cuda_select_nvcc_arch_flags(CUDA_ARCH_FLAGS Common) # Change Common by Auto for autodetect
+string(REPLACE ";" " " CUDA_ARCH_FLAGS "${CUDA_ARCH_FLAGS}")
+set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${CUDA_ARCH_FLAGS}")
+
+set(CMAKE_CUDA_FLAGS  "${CMAKE_CUDA_FLAGS} -D_FORCE_INLINES")
+# set(CMAKE_CUDA_FLAGS  "${CMAKE_CUDA_FLAGS} -lineinfo")
 
 add_library(gpuRIRcu STATIC
     src/gpuRIR_cuda.cu
 )
 
-SET(CMAKE_CUDA_FLAGS  "${CMAKE_CUDA_FLAGS} -D_FORCE_INLINES")
-# SET(CMAKE_CUDA_FLAGS  "${CMAKE_CUDA_FLAGS} -lineinfo")
-
-#target_link_libraries(gpuRIRcu PRIVATE -lcurand)
-
 set_target_properties(gpuRIRcu PROPERTIES 
     POSITION_INDEPENDENT_CODE ON
     CUDA_VISIBILITY_PRESET "hidden"
@@ -27,6 +28,9 @@ add_library(gpuRIR_bind MODULE
     src/python_bind.cpp
 )
 
+set(PYBIND11_CPP_STANDARD -std=c++11)
+add_subdirectory(third_party/pybind11)
+
 set_target_properties(gpuRIR_bind PROPERTIES 
     CXX_VISIBILITY_PRESET "hidden"
     PREFIX "${PYTHON_MODULE_PREFIX}"

README.md

@@ -9,6 +9,7 @@
 - [Documentation](#documentation)
   * [`simulateRIR`](#simulaterir)
   * [`simulateTrajectory`](#simulatetrajectory)
+  * [`activateMixedPrecision`](#activateMixedPrecision)
   * [`beta_SabineEstimation`](#beta_sabineestimation)
   * [`att2t_SabineEstimator`](#att2t_sabineestimator)
   * [`t2n`](#t2n)
@@ -101,6 +102,15 @@ Filter an audio signal by the RIRs of a motion trajectory recorded with a microp
 *2D ndarray*
 	Matrix with the signals captured by each microphone in each column.
 
+### `activateMixedPrecision`
+
+Activate the mixed precision mode, only for Pascal GPU architecture or superior.
+
+#### Parameters
+
+* **activate** : *bool, optional.*
+        True for activate and Flase for deactivate. True by default.
+
 ### `beta_SabineEstimation`
 
 Estimation of the reflection coefficients needed to have the desired reverberation time.

examples/example.py

@@ -9,6 +9,7 @@
 from math import ceil
 
 import gpuRIR
+gpuRIR.activateMixedPrecision(False)
 
 room_sz = [3,3,2.5]  # Size of the room [m]
 nb_src = 2  # Number of sources
@@ -31,3 +32,4 @@
 
 t = np.arange(int(ceil(Tmax * fs))) / fs
 plt.plot(t, RIRs.reshape(nb_src*nb_rcv, -1).transpose())
+plt.show()

examples/simulate_trajectory.py

@@ -9,6 +9,7 @@
 from scipy.io import wavfile
 
 import gpuRIR
+gpuRIR.activateMixedPrecision(False)
 
 fs, source_signal = wavfile.read('source_signal.wav')
 
@@ -33,3 +34,4 @@
 filtered_signal = gpuRIR.simulateTrajectory(source_signal, RIRs)
 wavfile.write('filtered_signal.wav', fs, filtered_signal)
 plt.plot(filtered_signal)
+plt.show()

examples/time_vs_T60.py

@@ -6,9 +6,10 @@
 
 import numpy as np
 import matplotlib.pyplot as plt
-import  time
+import time
 
 import gpuRIR
+gpuRIR.activateMixedPrecision(False)
 
 T60_vec = np.arange(0.1, 2.2, 0.2) # Reverberation times to measure
 nb_test_per_point = 10 # Number of simulations per T60 to average the runtime
@@ -45,4 +46,5 @@
 
 plt.semilogy(T60_vec, times)
 plt.ylabel("time [s]")
-plt.xlabel("T60 [s]")
+plt.xlabel("T60 [s]")
+plt.show()

examples/time_vs_nbRIRs.py

@@ -6,9 +6,10 @@
 
 import numpy as np
 import matplotlib.pyplot as plt
-import  time
+import time
 
 import gpuRIR
+gpuRIR.activateMixedPrecision(False)
 
 nb_src_vec = np.concatenate([2**np.arange(12), [4094]]) # Number of RIRs to measure
 nb_test_per_point = 10 # Number of simulations per T60 to average the runtime
@@ -46,3 +47,4 @@
 plt.loglog(nb_src_vec, times)
 plt.ylabel("time [s]")
 plt.xlabel("number of RIRs")
+plt.show()

gpuRIR/__init__.py

@@ -6,9 +6,9 @@
 from scipy.optimize import minimize
 from scipy.signal import convolve
 
-from gpuRIR_bind import simulateRIR_bind, gpu_conv, cuda_warmup
+from gpuRIR_bind import gpuRIR_bind
 
-__all__ = ["mic_patterns", "beta_SabineEstimation", "att2t_SabineEstimator", "t2n", "simulateRIR", "simulateTrajectory"]
+__all__ = ["mic_patterns", "beta_SabineEstimation", "att2t_SabineEstimator", "t2n", "simulateRIR", "simulateTrajectory", "activate_mixed_precision"]
 
 mic_patterns =	{
   "omni": 0,
@@ -153,7 +153,7 @@ def simulateRIR(room_sz, beta, pos_src, pos_rcv, nb_img, Tmax, fs, Tdiff=None, m
 	if orV_rcv is None: orV_rcv = np.zeros_like(pos_rcv)
 	else: orV_rcv = orV_rcv.astype('float32', order='C', copy=False)
 
-	return simulateRIR_bind(room_sz, beta, pos_src, pos_rcv, orV_rcv, mic_patterns[mic_pattern], nb_img, Tdiff, Tmax, fs, c)
+	return gpuRIR_bind_simulator.simulateRIR_bind(room_sz, beta, pos_src, pos_rcv, orV_rcv, mic_patterns[mic_pattern], nb_img, Tdiff, Tmax, fs, c)
 
 def simulateTrajectory(source_signal, RIRs, timestamps=None, fs=None):
 	''' Filter an audio signal by the RIRs of a motion trajectory recorded with a microphone array.
@@ -192,7 +192,7 @@ def simulateTrajectory(source_signal, RIRs, timestamps=None, fs=None):
 	for n in range(nPts):
 		segments[n,0:w_len[n]] = source_signal[w_ini[n]:w_ini[n+1]]
 	segments = segments.astype('float32', order='C', copy=False)
-	convolution = gpu_conv(segments, RIRs)
+	convolution = gpuRIR_bind_simulator.gpu_conv(segments, RIRs)
 
 	filtered_signal = np.zeros((nSamples+lenRIR-1, nRcv))
 	for m in range(nRcv):
@@ -201,5 +201,16 @@ def simulateTrajectory(source_signal, RIRs, timestamps=None, fs=None):
 
 	return filtered_signal
 
-# Initialize the CUDA runtime API and the cuFFT library when the module is loaded
-cuda_warmup()
+def activateMixedPrecision(activate=True):
+	''' Activate the mixed precision mode, only for Pascal GPU architecture or superior.
+
+	Parameters
+	----------
+	activate : bool, optional
+		True for activate and Flase for deactivate. True by default.
+
+	'''
+	gpuRIR_bind_simulator.activate_mixed_precision_bind(activate)
+
+# Create the simulator object when the module is loaded
+gpuRIR_bind_simulator = gpuRIR_bind()

setup.py

@@ -66,7 +66,7 @@ def build_extension(self, ext):
 
 setup(
     name='gpuRIR',
-    version='1.1.0',
+    version='1.2.0',
     author='David Diaz-Guerra',
     author_email='[email protected]',
     url='',