Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(510)

Issues Search
    My Issues | Starred     Open | Closed | All

Unified Diff: webrtc/modules/audio_processing/utility/ooura_fft.cc

Issue 2348213002: Move the aec_rdft* files to a more proper place beneath APM and make them thread-safe. (Closed)
Patch Set: Rebase Created 4 years, 2 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View side-by-side diff with in-line comments
Download patch
« no previous file with comments | « webrtc/modules/audio_processing/utility/ooura_fft.h ('k') | webrtc/modules/audio_processing/utility/ooura_fft_mips.cc » ('j') | no next file with comments »
Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
Index: webrtc/modules/audio_processing/utility/ooura_fft.cc
diff --git a/webrtc/modules/audio_processing/aec/aec_rdft.cc b/webrtc/modules/audio_processing/utility/ooura_fft.cc
similarity index 65%
rename from webrtc/modules/audio_processing/aec/aec_rdft.cc
rename to webrtc/modules/audio_processing/utility/ooura_fft.cc
index 690fe9f34f9c51ebf3f2c91e5553f2b2d9d46f1f..4ba88d7f17f0905426f86fd3a37421b7bdbe706b 100644
--- a/webrtc/modules/audio_processing/aec/aec_rdft.cc
+++ b/webrtc/modules/audio_processing/utility/ooura_fft.cc
@@ -10,6 +10,8 @@
* - Trivial type modifications.
* - Minimal code subset to do rdft of length 128.
* - Optimizations because of known length.
+ * - Removed the global variables by moving the code in to a class in order
+ * to make it thread safe.
*
* All changes are covered by the WebRTC license and IP grant:
* Use of this source code is governed by a BSD-style license
@@ -19,184 +21,19 @@
* be found in the AUTHORS file in the root of the source tree.
*/
-#include "webrtc/modules/audio_processing/aec/aec_rdft.h"
+#include "webrtc/modules/audio_processing//utility/ooura_fft.h"
#include <math.h>
+#include "webrtc/modules/audio_processing/utility/ooura_fft_tables_common.h"
#include "webrtc/system_wrappers/include/cpu_features_wrapper.h"
#include "webrtc/typedefs.h"
-// These tables used to be computed at run-time. For example, refer to:
-// https://code.google.com/p/webrtc/source/browse/trunk/webrtc/modules/audio_processing/aec/aec_rdft.c?r=6564
-// to see the initialization code.
-const float rdft_w[64] = {
- 1.0000000000f, 0.0000000000f, 0.7071067691f, 0.7071067691f,
- 0.9238795638f, 0.3826834559f, 0.3826834559f, 0.9238795638f,
- 0.9807852507f, 0.1950903237f, 0.5555702448f, 0.8314695954f,
- 0.8314695954f, 0.5555702448f, 0.1950903237f, 0.9807852507f,
- 0.9951847196f, 0.0980171412f, 0.6343933344f, 0.7730104327f,
- 0.8819212914f, 0.4713967443f, 0.2902846634f, 0.9569403529f,
- 0.9569403529f, 0.2902846634f, 0.4713967443f, 0.8819212914f,
- 0.7730104327f, 0.6343933344f, 0.0980171412f, 0.9951847196f,
- 0.7071067691f, 0.4993977249f, 0.4975923598f, 0.4945882559f,
- 0.4903926253f, 0.4850156307f, 0.4784701765f, 0.4707720280f,
- 0.4619397819f, 0.4519946277f, 0.4409606457f, 0.4288643003f,
- 0.4157347977f, 0.4016037583f, 0.3865052164f, 0.3704755902f,
- 0.3535533845f, 0.3357794881f, 0.3171966672f, 0.2978496552f,
- 0.2777851224f, 0.2570513785f, 0.2356983721f, 0.2137775421f,
- 0.1913417280f, 0.1684449315f, 0.1451423317f, 0.1214900985f,
- 0.0975451618f, 0.0733652338f, 0.0490085706f, 0.0245338380f,
-};
-const float rdft_wk3ri_first[16] = {
- 1.000000000f, 0.000000000f, 0.382683456f, 0.923879564f,
- 0.831469536f, 0.555570245f, -0.195090353f, 0.980785251f,
- 0.956940353f, 0.290284693f, 0.098017156f, 0.995184720f,
- 0.634393334f, 0.773010492f, -0.471396863f, 0.881921172f,
-};
-const float rdft_wk3ri_second[16] = {
- -0.707106769f, 0.707106769f, -0.923879564f, -0.382683456f,
- -0.980785251f, 0.195090353f, -0.555570245f, -0.831469536f,
- -0.881921172f, 0.471396863f, -0.773010492f, -0.634393334f,
- -0.995184720f, -0.098017156f, -0.290284693f, -0.956940353f,
-};
-ALIGN16_BEG const float ALIGN16_END rdft_wk1r[32] = {
- 1.000000000f, 1.000000000f, 0.707106769f, 0.707106769f,
- 0.923879564f, 0.923879564f, 0.382683456f, 0.382683456f,
- 0.980785251f, 0.980785251f, 0.555570245f, 0.555570245f,
- 0.831469595f, 0.831469595f, 0.195090324f, 0.195090324f,
- 0.995184720f, 0.995184720f, 0.634393334f, 0.634393334f,
- 0.881921291f, 0.881921291f, 0.290284663f, 0.290284663f,
- 0.956940353f, 0.956940353f, 0.471396744f, 0.471396744f,
- 0.773010433f, 0.773010433f, 0.098017141f, 0.098017141f,
-};
-ALIGN16_BEG const float ALIGN16_END rdft_wk2r[32] = {
- 1.000000000f, 1.000000000f, -0.000000000f, -0.000000000f,
- 0.707106769f, 0.707106769f, -0.707106769f, -0.707106769f,
- 0.923879564f, 0.923879564f, -0.382683456f, -0.382683456f,
- 0.382683456f, 0.382683456f, -0.923879564f, -0.923879564f,
- 0.980785251f, 0.980785251f, -0.195090324f, -0.195090324f,
- 0.555570245f, 0.555570245f, -0.831469595f, -0.831469595f,
- 0.831469595f, 0.831469595f, -0.555570245f, -0.555570245f,
- 0.195090324f, 0.195090324f, -0.980785251f, -0.980785251f,
-};
-ALIGN16_BEG const float ALIGN16_END rdft_wk3r[32] = {
- 1.000000000f, 1.000000000f, -0.707106769f, -0.707106769f,
- 0.382683456f, 0.382683456f, -0.923879564f, -0.923879564f,
- 0.831469536f, 0.831469536f, -0.980785251f, -0.980785251f,
- -0.195090353f, -0.195090353f, -0.555570245f, -0.555570245f,
- 0.956940353f, 0.956940353f, -0.881921172f, -0.881921172f,
- 0.098017156f, 0.098017156f, -0.773010492f, -0.773010492f,
- 0.634393334f, 0.634393334f, -0.995184720f, -0.995184720f,
- -0.471396863f, -0.471396863f, -0.290284693f, -0.290284693f,
-};
-ALIGN16_BEG const float ALIGN16_END rdft_wk1i[32] = {
- -0.000000000f, 0.000000000f, -0.707106769f, 0.707106769f,
- -0.382683456f, 0.382683456f, -0.923879564f, 0.923879564f,
- -0.195090324f, 0.195090324f, -0.831469595f, 0.831469595f,
- -0.555570245f, 0.555570245f, -0.980785251f, 0.980785251f,
- -0.098017141f, 0.098017141f, -0.773010433f, 0.773010433f,
- -0.471396744f, 0.471396744f, -0.956940353f, 0.956940353f,
- -0.290284663f, 0.290284663f, -0.881921291f, 0.881921291f,
- -0.634393334f, 0.634393334f, -0.995184720f, 0.995184720f,
-};
-ALIGN16_BEG const float ALIGN16_END rdft_wk2i[32] = {
- -0.000000000f, 0.000000000f, -1.000000000f, 1.000000000f,
- -0.707106769f, 0.707106769f, -0.707106769f, 0.707106769f,
- -0.382683456f, 0.382683456f, -0.923879564f, 0.923879564f,
- -0.923879564f, 0.923879564f, -0.382683456f, 0.382683456f,
- -0.195090324f, 0.195090324f, -0.980785251f, 0.980785251f,
- -0.831469595f, 0.831469595f, -0.555570245f, 0.555570245f,
- -0.555570245f, 0.555570245f, -0.831469595f, 0.831469595f,
- -0.980785251f, 0.980785251f, -0.195090324f, 0.195090324f,
-};
-ALIGN16_BEG const float ALIGN16_END rdft_wk3i[32] = {
- -0.000000000f, 0.000000000f, -0.707106769f, 0.707106769f,
- -0.923879564f, 0.923879564f, 0.382683456f, -0.382683456f,
- -0.555570245f, 0.555570245f, -0.195090353f, 0.195090353f,
- -0.980785251f, 0.980785251f, 0.831469536f, -0.831469536f,
- -0.290284693f, 0.290284693f, -0.471396863f, 0.471396863f,
- -0.995184720f, 0.995184720f, 0.634393334f, -0.634393334f,
- -0.773010492f, 0.773010492f, 0.098017156f, -0.098017156f,
- -0.881921172f, 0.881921172f, 0.956940353f, -0.956940353f,
-};
-ALIGN16_BEG const float ALIGN16_END cftmdl_wk1r[4] = {
- 0.707106769f, 0.707106769f, 0.707106769f, -0.707106769f,
-};
+namespace webrtc {
-static void bitrv2_128_C(float* a) {
- /*
- Following things have been attempted but are no faster:
- (a) Storing the swap indexes in a LUT (index calculations are done
- for 'free' while waiting on memory/L1).
- (b) Consolidate the load/store of two consecutive floats by a 64 bit
- integer (execution is memory/L1 bound).
- (c) Do a mix of floats and 64 bit integer to maximize register
- utilization (execution is memory/L1 bound).
- (d) Replacing ip[i] by ((k<<31)>>25) + ((k >> 1)<<5).
- (e) Hard-coding of the offsets to completely eliminates index
- calculations.
- */
-
- unsigned int j, j1, k, k1;
- float xr, xi, yr, yi;
-
- static const int ip[4] = {0, 64, 32, 96};
- for (k = 0; k < 4; k++) {
- for (j = 0; j < k; j++) {
- j1 = 2 * j + ip[k];
- k1 = 2 * k + ip[j];
- xr = a[j1 + 0];
- xi = a[j1 + 1];
- yr = a[k1 + 0];
- yi = a[k1 + 1];
- a[j1 + 0] = yr;
- a[j1 + 1] = yi;
- a[k1 + 0] = xr;
- a[k1 + 1] = xi;
- j1 += 8;
- k1 += 16;
- xr = a[j1 + 0];
- xi = a[j1 + 1];
- yr = a[k1 + 0];
- yi = a[k1 + 1];
- a[j1 + 0] = yr;
- a[j1 + 1] = yi;
- a[k1 + 0] = xr;
- a[k1 + 1] = xi;
- j1 += 8;
- k1 -= 8;
- xr = a[j1 + 0];
- xi = a[j1 + 1];
- yr = a[k1 + 0];
- yi = a[k1 + 1];
- a[j1 + 0] = yr;
- a[j1 + 1] = yi;
- a[k1 + 0] = xr;
- a[k1 + 1] = xi;
- j1 += 8;
- k1 += 16;
- xr = a[j1 + 0];
- xi = a[j1 + 1];
- yr = a[k1 + 0];
- yi = a[k1 + 1];
- a[j1 + 0] = yr;
- a[j1 + 1] = yi;
- a[k1 + 0] = xr;
- a[k1 + 1] = xi;
- }
- j1 = 2 * k + 8 + ip[k];
- k1 = j1 + 8;
- xr = a[j1 + 0];
- xi = a[j1 + 1];
- yr = a[k1 + 0];
- yi = a[k1 + 1];
- a[j1 + 0] = yr;
- a[j1 + 1] = yi;
- a[k1 + 0] = xr;
- a[k1 + 1] = xi;
- }
-}
+namespace {
+#if !(defined(MIPS_FPU_LE) || defined(WEBRTC_HAS_NEON))
static void cft1st_128_C(float* a) {
const int n = 128;
int j, k1, k2;
@@ -431,67 +268,6 @@ static void cftmdl_128_C(float* a) {
}
}
-static void cftfsub_128_C(float* a) {
- int j, j1, j2, j3, l;
- float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
-
- cft1st_128(a);
- cftmdl_128(a);
- l = 32;
- for (j = 0; j < l; j += 2) {
- j1 = j + l;
- j2 = j1 + l;
- j3 = j2 + l;
- x0r = a[j] + a[j1];
- x0i = a[j + 1] + a[j1 + 1];
- x1r = a[j] - a[j1];
- x1i = a[j + 1] - a[j1 + 1];
- x2r = a[j2] + a[j3];
- x2i = a[j2 + 1] + a[j3 + 1];
- x3r = a[j2] - a[j3];
- x3i = a[j2 + 1] - a[j3 + 1];
- a[j] = x0r + x2r;
- a[j + 1] = x0i + x2i;
- a[j2] = x0r - x2r;
- a[j2 + 1] = x0i - x2i;
- a[j1] = x1r - x3i;
- a[j1 + 1] = x1i + x3r;
- a[j3] = x1r + x3i;
- a[j3 + 1] = x1i - x3r;
- }
-}
-
-static void cftbsub_128_C(float* a) {
- int j, j1, j2, j3, l;
- float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
-
- cft1st_128(a);
- cftmdl_128(a);
- l = 32;
-
- for (j = 0; j < l; j += 2) {
- j1 = j + l;
- j2 = j1 + l;
- j3 = j2 + l;
- x0r = a[j] + a[j1];
- x0i = -a[j + 1] - a[j1 + 1];
- x1r = a[j] - a[j1];
- x1i = -a[j + 1] + a[j1 + 1];
- x2r = a[j2] + a[j3];
- x2i = a[j2 + 1] + a[j3 + 1];
- x3r = a[j2] - a[j3];
- x3i = a[j2 + 1] - a[j3 + 1];
- a[j] = x0r + x2r;
- a[j + 1] = x0i - x2i;
- a[j2] = x0r - x2r;
- a[j2 + 1] = x0i + x2i;
- a[j1] = x1r - x3i;
- a[j1 + 1] = x1i - x3r;
- a[j3] = x1r + x3i;
- a[j3 + 1] = x1i + x3r;
- }
-}
-
static void rftfsub_128_C(float* a) {
const float* c = rdft_w + 32;
int j1, j2, k1, k2;
@@ -535,8 +311,22 @@ static void rftbsub_128_C(float* a) {
}
a[65] = -a[65];
}
+#endif
+
-void aec_rdft_forward_128(float* a) {
+} // namespace
+
+OouraFft::OouraFft() {
+#if defined(WEBRTC_ARCH_X86_FAMILY)
+ use_sse2_ = (WebRtc_GetCPUInfo(kSSE2) != 0);
+#else
+ use_sse2_ = false;
+#endif
+}
+
+OouraFft::~OouraFft() = default;
+
+void OouraFft::Fft(float* a) const {
float xi;
bitrv2_128(a);
cftfsub_128(a);
@@ -545,8 +335,7 @@ void aec_rdft_forward_128(float* a) {
a[0] += a[1];
a[1] = xi;
}
-
-void aec_rdft_inverse_128(float* a) {
+void OouraFft::InverseFft(float* a) const {
a[1] = 0.5f * (a[0] - a[1]);
a[0] -= a[1];
rftbsub_128(a);
@@ -554,32 +343,193 @@ void aec_rdft_inverse_128(float* a) {
cftbsub_128(a);
}
-// code path selection
-RftSub128 cft1st_128;
-RftSub128 cftmdl_128;
-RftSub128 rftfsub_128;
-RftSub128 rftbsub_128;
-RftSub128 cftfsub_128;
-RftSub128 cftbsub_128;
-RftSub128 bitrv2_128;
-
-void aec_rdft_init(void) {
- cft1st_128 = cft1st_128_C;
- cftmdl_128 = cftmdl_128_C;
- rftfsub_128 = rftfsub_128_C;
- rftbsub_128 = rftbsub_128_C;
- cftfsub_128 = cftfsub_128_C;
- cftbsub_128 = cftbsub_128_C;
- bitrv2_128 = bitrv2_128_C;
-#if defined(WEBRTC_ARCH_X86_FAMILY)
- if (WebRtc_GetCPUInfo(kSSE2)) {
- aec_rdft_init_sse2();
+void OouraFft::cft1st_128(float* a) const {
+#if defined(MIPS_FPU_LE)
+ cft1st_128_mips(a);
+#elif defined(WEBRTC_HAS_NEON)
+ cft1st_128_neon(a);
+#else
+ if (use_sse2_) {
+ cft1st_128_SSE2(a);
+ } else {
+ cft1st_128_C(a);
}
#endif
+}
+void OouraFft::cftmdl_128(float* a) const {
#if defined(MIPS_FPU_LE)
- aec_rdft_init_mips();
+ cftmdl_128_mips(a);
+#elif defined(WEBRTC_HAS_NEON)
+ cftmdl_128_neon(a);
+#else
+ if (use_sse2_) {
+ cftmdl_128_SSE2(a);
+ } else {
+ cftmdl_128_C(a);
+ }
#endif
-#if defined(WEBRTC_HAS_NEON)
- aec_rdft_init_neon();
+}
+void OouraFft::rftfsub_128(float* a) const {
+#if defined(MIPS_FPU_LE)
+ rftfsub_128_mips(a);
+#elif defined(WEBRTC_HAS_NEON)
+ rftfsub_128_neon(a);
+#else
+ if (use_sse2_) {
+ rftfsub_128_SSE2(a);
+ } else {
+ rftfsub_128_C(a);
+ }
+#endif
+}
+
+void OouraFft::rftbsub_128(float* a) const {
+#if defined(MIPS_FPU_LE)
+ rftbsub_128_mips(a);
+#elif defined(WEBRTC_HAS_NEON)
+ rftbsub_128_neon(a);
+#else
+ if (use_sse2_) {
+ rftbsub_128_SSE2(a);
+ } else {
+ rftbsub_128_C(a);
+ }
#endif
}
+
+void OouraFft::cftbsub_128(float* a) const {
+ int j, j1, j2, j3, l;
+ float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
+
+ cft1st_128(a);
+ cftmdl_128(a);
+ l = 32;
+
+ for (j = 0; j < l; j += 2) {
+ j1 = j + l;
+ j2 = j1 + l;
+ j3 = j2 + l;
+ x0r = a[j] + a[j1];
+ x0i = -a[j + 1] - a[j1 + 1];
+ x1r = a[j] - a[j1];
+ x1i = -a[j + 1] + a[j1 + 1];
+ x2r = a[j2] + a[j3];
+ x2i = a[j2 + 1] + a[j3 + 1];
+ x3r = a[j2] - a[j3];
+ x3i = a[j2 + 1] - a[j3 + 1];
+ a[j] = x0r + x2r;
+ a[j + 1] = x0i - x2i;
+ a[j2] = x0r - x2r;
+ a[j2 + 1] = x0i + x2i;
+ a[j1] = x1r - x3i;
+ a[j1 + 1] = x1i - x3r;
+ a[j3] = x1r + x3i;
+ a[j3 + 1] = x1i + x3r;
+ }
+}
+
+void OouraFft::cftfsub_128(float* a) const {
+ int j, j1, j2, j3, l;
+ float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
+
+ cft1st_128(a);
+ cftmdl_128(a);
+ l = 32;
+ for (j = 0; j < l; j += 2) {
+ j1 = j + l;
+ j2 = j1 + l;
+ j3 = j2 + l;
+ x0r = a[j] + a[j1];
+ x0i = a[j + 1] + a[j1 + 1];
+ x1r = a[j] - a[j1];
+ x1i = a[j + 1] - a[j1 + 1];
+ x2r = a[j2] + a[j3];
+ x2i = a[j2 + 1] + a[j3 + 1];
+ x3r = a[j2] - a[j3];
+ x3i = a[j2 + 1] - a[j3 + 1];
+ a[j] = x0r + x2r;
+ a[j + 1] = x0i + x2i;
+ a[j2] = x0r - x2r;
+ a[j2 + 1] = x0i - x2i;
+ a[j1] = x1r - x3i;
+ a[j1 + 1] = x1i + x3r;
+ a[j3] = x1r + x3i;
+ a[j3 + 1] = x1i - x3r;
+ }
+}
+
+void OouraFft::bitrv2_128(float* a) const {
+ /*
+ Following things have been attempted but are no faster:
+ (a) Storing the swap indexes in a LUT (index calculations are done
+ for 'free' while waiting on memory/L1).
+ (b) Consolidate the load/store of two consecutive floats by a 64 bit
+ integer (execution is memory/L1 bound).
+ (c) Do a mix of floats and 64 bit integer to maximize register
+ utilization (execution is memory/L1 bound).
+ (d) Replacing ip[i] by ((k<<31)>>25) + ((k >> 1)<<5).
+ (e) Hard-coding of the offsets to completely eliminates index
+ calculations.
+ */
+
+ unsigned int j, j1, k, k1;
+ float xr, xi, yr, yi;
+
+ const int ip[4] = {0, 64, 32, 96};
+ for (k = 0; k < 4; k++) {
+ for (j = 0; j < k; j++) {
+ j1 = 2 * j + ip[k];
+ k1 = 2 * k + ip[j];
+ xr = a[j1 + 0];
+ xi = a[j1 + 1];
+ yr = a[k1 + 0];
+ yi = a[k1 + 1];
+ a[j1 + 0] = yr;
+ a[j1 + 1] = yi;
+ a[k1 + 0] = xr;
+ a[k1 + 1] = xi;
+ j1 += 8;
+ k1 += 16;
+ xr = a[j1 + 0];
+ xi = a[j1 + 1];
+ yr = a[k1 + 0];
+ yi = a[k1 + 1];
+ a[j1 + 0] = yr;
+ a[j1 + 1] = yi;
+ a[k1 + 0] = xr;
+ a[k1 + 1] = xi;
+ j1 += 8;
+ k1 -= 8;
+ xr = a[j1 + 0];
+ xi = a[j1 + 1];
+ yr = a[k1 + 0];
+ yi = a[k1 + 1];
+ a[j1 + 0] = yr;
+ a[j1 + 1] = yi;
+ a[k1 + 0] = xr;
+ a[k1 + 1] = xi;
+ j1 += 8;
+ k1 += 16;
+ xr = a[j1 + 0];
+ xi = a[j1 + 1];
+ yr = a[k1 + 0];
+ yi = a[k1 + 1];
+ a[j1 + 0] = yr;
+ a[j1 + 1] = yi;
+ a[k1 + 0] = xr;
+ a[k1 + 1] = xi;
+ }
+ j1 = 2 * k + 8 + ip[k];
+ k1 = j1 + 8;
+ xr = a[j1 + 0];
+ xi = a[j1 + 1];
+ yr = a[k1 + 0];
+ yi = a[k1 + 1];
+ a[j1 + 0] = yr;
+ a[j1 + 1] = yi;
+ a[k1 + 0] = xr;
+ a[k1 + 1] = xi;
+ }
+}
+
+} // namespace webrtc
« no previous file with comments | « webrtc/modules/audio_processing/utility/ooura_fft.h ('k') | webrtc/modules/audio_processing/utility/ooura_fft_mips.cc » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698