Merge pull request #10 from techpro-studio/dev

Dev

Author: alexmay23

CMakeLists.txt

@@ -76,14 +76,18 @@ endif()
 
 if (USE_SIGNAL_API)
     list(APPEND SOURCES
+        nntoolkitcore/signal/mel_filterbank.c
         nntoolkitcore/signal/dft.c
         nntoolkitcore/signal/dft.h
         nntoolkitcore/signal/spectrogram.c
+        nntoolkitcore/signal/log_mel_spectrogram.c
         nntoolkitcore/signal/window.c
     )
     list(APPEND PUBLIC_HEADERS
+        nntoolkitcore/signal/mel_filterbank.h
         nntoolkitcore/signal/window.h
         nntoolkitcore/signal/spectrogram.h
+        nntoolkitcore/signal/log_mel_spectrogram.h
     )
     if(NOT APPLE)
         list(APPEND SOURCES

NNToolkitCore.podspec

@@ -18,6 +18,8 @@ Pod::Spec.new do |s|
   'nntoolkitcore/layers/*.{h}',
   'nntoolkitcore/train/*.{h}',
   'nntoolkitcore/signal/spectrogram.h',
+  'nntoolkitcore/signal/mel_filterbank.h',
+  'nntoolkitcore/signal/log_mel_spectrogram.h',
   'nntoolkitcore/signal/window.h',
   'nntoolkitcore/core/debug.h',
   'nntoolkitcore/core/ops.h'

nntoolkitcore/core/apple_ops.c

@@ -15,23 +15,23 @@
 
 #define simd_float_16_init(var, value) \
 float var##arr[16] = { value, value, value, value, value, value, value, value, value, value, value, value, value, value, value, value };\
-simd_float16 var = simd_make_float16(*(simd_float16 *)(var##arr));\
+simd_float16 var = simd_make_float16(*(simd_packed_float16 *)(var##arr));\
 
 #define simd_float_8_init(var, value) \
 float var##arr[8] = { value, value, value, value, value, value, value, value };\
-simd_float8 var = simd_make_float8(*(simd_float8 *)(var##arr));\
+simd_float8 var = simd_make_float8(*(simd_packed_float8 *)(var##arr));\
 
 #define simd_float_4_init(var, value) \
 float var##arr[4] = { value, value, value, value };\
-simd_float4 var = simd_make_float4(*(simd_float4 *)(var##arr));\
+simd_float4 var = simd_make_float4(*(simd_packed_float4 *)(var##arr));\
 
 #define simd_float_3_init(var, value) \
 float var##arr[3] = { value, value, value };\
 simd_float3 var = simd_make_float3(*(simd_float3 *)(var##arr));\
 
 #define simd_float_2_init(var, value) \
 float var##arr[2] = { value, value };\
-simd_float2 var = simd_make_float2(*(simd_float2 *)(var##arr));\
+simd_float2 var = simd_make_float2(*(simd_packed_float2 *)(var##arr));\
 
 
 #define vector_dot_(NUM)  float op_vec_dot_##NUM(const float* a, const float *b, int size)\
@@ -83,7 +83,7 @@ float op_vec_dot_3(const float* a, const float *b, int size)
     {\
     simd_float_##NUM##_init(s_min, min)\
     simd_float_##NUM##_init(s_max, max)\
-        ((simd_float##NUM*) c)[i] = simd_clamp(((simd_float##NUM*) a)[i], s_min, s_max);\
+        ((simd_float##NUM*) c)[i] = simd_clamp(((simd_packed_float##NUM*) a)[i], s_min, s_max);\
     }\
     int left = size % NUM;\
     for (int i = 0; i < left; ++i)\
@@ -93,7 +93,6 @@ float op_vec_dot_3(const float* a, const float *b, int size)
 }
 
 op_vec_clamp_(2)
-op_vec_clamp_(3)
 op_vec_clamp_(4)
 op_vec_clamp_(8)
 op_vec_clamp_(16)
@@ -104,8 +103,8 @@ op_vec_clamp_(16)
     int iterations = size / NUM;\
     for (int i = 0; i < iterations; ++i)\
     {\
-    simd_float_##NUM##_init(s_b, b)\
-        ((simd_float##NUM*) c)[i] = simd_max(((simd_float##NUM*) a)[i], s_b);\
+        simd_float_##NUM##_init(s_b, b)\
+        ((simd_float##NUM*) c)[i] = simd_max(((simd_packed_float##NUM*) a)[i], s_b);\
     }\
     int left = size % NUM;\
     for (int i = 0; i < left; ++i)\
@@ -114,14 +113,23 @@ op_vec_clamp_(16)
     }\
 }
 
+void op_vec_max_sc_4(const float* a, float b, float *c, int size)\
+{\
+    int iterations = size / 4;
+    for (int i = 0; i < iterations; ++i)
+    {
+        simd_float_4_init(s_b, b)
+        simd_float4 s_a = ((simd_packed_float4 *) a)[i];
+        ((simd_packed_float4*) c)[i] = simd_max(s_a, s_b);
+    }
+    int left = size % 4;
+    for (int i = 0; i < left; ++i)
+    {
+        c[iterations * 4 + i] = simd_max(a[iterations * 4 + i], b);
+    }
+}
 
 
-op_vec_max_sc_(2)
-op_vec_max_sc_(3)
-op_vec_max_sc_(4)
-op_vec_max_sc_(8)
-op_vec_max_sc_(16)
-
 typedef enum {
     two = 2, three = 3, four = 4, eight = 8, sixteen = 16
 }optimal_vector_size;
@@ -146,24 +154,6 @@ optimal_vector_size get_optimal_vector_size(int size){
     return values[optimalIndex];
 }
 
-#define get_optimized(func) func##_fn func##_get_optimized(int size){\
-    optimal_vector_size value = get_optimal_vector_size(size);\
-    switch (value) {\
-        case two:\
-            return func##_2;\
-        case three:\
-            return func##_3;\
-        case four:\
-            return func##_4;\
-        case eight:\
-            return func##_8;\
-        case sixteen:\
-            return func##_16;\
-        default:\
-            return func##_4;\
-    }\
-}
-
 typedef float (*op_vec_dot_fn)(const float *a, const float *b, int size);
 
 typedef void (*op_vec_clamp_fn)(const float *a, float *c, float min, float max, int size);
@@ -176,9 +166,6 @@ op_vec_clamp_fn op_vec_clamp_get_optimized(int size);
 
 op_vec_max_sc_fn op_vec_max_sc_get_optimized(int size);
 
-get_optimized(op_vec_dot)
-get_optimized(op_vec_clamp)
-get_optimized(op_vec_max_sc)
 
 void op_vec_max(const float *a, const float *b, float *c, int size){
     vDSP_vmax(a, 1, b, 1, c, 1, size);
@@ -194,9 +181,6 @@ float op_vec_dot(const float *a, const float *b, int size) {
 #else
     return op_vec_dot_4(a, b, size);
 #endif
-//    float c;
-//    vDSP_dotpr(a, 1, b, 1, &c, size);
-//    return c;
 }
 
 void op_vec_clamp(const float *a, float *c, float min, float max, int size){
@@ -215,7 +199,7 @@ void op_vec_sub(const float *a, const float *b, float *result, int size){
     vDSP_vsub(b, 1, a, 1, result, 1, size);
 }
 
-void op_vec_sum(const float *a, float* result, int size){
+void op_vec_sum(const float *a, float* result, int size) {
     vDSP_sve(a, 1, result, size);
 }
 
@@ -247,6 +231,15 @@ void op_vec_exp(const float *a, float *c, int size) {
     vvexpf(c, a, &size);
 }
 
+void op_vec_pow(const float *a, const float *b, float *c, int size) {
+    vvpowf(c, b, a, &size);
+}
+
+void op_vec_pow_sc(const float *a, const float b, float *c, int size) {
+    vvpowsf(c, &b, a, &size);
+}
+
+
 void op_vec_log(const float *a, float *c, int size) {
     vvlogf(c, a, &size);
 }
@@ -263,7 +256,7 @@ void op_vec_div(const float *a, const float *b, float *c, int size) {
     vDSP_vdiv(b, 1, a, 1, c, 1, size);
 }
 
-void op_vec_magnitudes(float *a, float *b, float *c, int size) {
+void op_vec_magn_sq(float *a, float *b, float *c, int size) {
     DSPSplitComplex split = {a, b};
     vDSP_zvmags(&split, 1, c, 1, size);
 }
@@ -281,3 +274,5 @@ void op_mat_transp(const float *a, float *b, int M, int N) {
 }
 
 
+
+

nntoolkitcore/core/default_ops.cc

@@ -2,6 +2,7 @@
 // Created by Alex on 07.11.2020.
 //
 
+
 #include "ops.h"
 #include "math.h"
 #include "third_party/eigen3/Eigen/Dense"
@@ -305,6 +306,44 @@ void op_vec_exp(const float *a, float *c, int size){
 #endif
 }
 
+static inline void op_vec_pow_c(const float *a, const float *b, float *c, int size) {
+    for (int i = 0; i < size; ++i){
+        c[i] = powf(a[i], b[i]);
+    }
+}
+// x ^^ m = exp(m * log(x))
+
+void op_vec_pow(const float *a, const float *b, float *c, int size){
+#if NEON
+    int parts = size / 4, remaining = size % 4;
+    for (int i = 0; i < parts; ++i){
+        vst1q_f32(c + i * 4, exp_neon(vmulq_f32(vld1q_f32(b + 4 * i), log_neon(vld1q_f32(a + 4 * i)))));
+    }
+    op_vec_pow_c(a + parts * 4, b + parts * 4, c + parts * 4, remaining);
+#else
+    op_vec_pow_c(a, b, c, size);
+#endif
+}
+
+static inline void op_vec_pow_sc_c(const float *a, const float b, float *c, int size) {
+    for (int i = 0; i < size; ++i){
+        c[i] = powf(a[i], b);
+    }
+}
+
+void op_vec_pow_sc(const float *a, const float b, float *c, int size) {
+#if NEON
+    int parts = size / 4, remaining = size % 4;
+    for (int i = 0; i < parts; ++i){
+        vst1q_f32(c + i * 4, exp_neon(vmulq_f32(vdupq_n_f32(b), log_neon(vld1q_f32(a + 4 * i)))));
+    }
+    op_vec_pow_sc_c(a + parts * 4, b, c + parts * 4, remaining);
+#else
+    op_vec_pow_sc_c(a, b, c, size);
+#endif
+}
+
+
 static inline void op_vec_log_c(const float *a, float *c, int size){
     for (int i = 0; i < size; ++i){
         c[i] = logf(a[i]);
@@ -465,7 +504,7 @@ void op_vec_add_sc(const float *a, float b, float *c, int size) {
 
 static void op_vec_sub_sc_c(const float *a, float b, float *c, int size){
     for (int i = 0; i < size; ++i){
-        c[i] = a[i] + b;
+        c[i] = a[i] - b;
     }
 }
 
@@ -624,23 +663,23 @@ void op_vec_sum(const float *a, float* c, int size) {
 }
 
 
-static void op_vec_magnitudes_c(const float *a, const float *b, float *c, int size){
+static void op_vec_magn_sq_c(const float *a, const float *b, float *c, int size){
     for (int i = 0; i < size; ++i){
         c[i] = a[i] * a[i] + b[i] * b[i];
     }
 }
 
-void op_vec_magnitudes(float *a, float *b, float *c, int size) {
+void op_vec_magn_sq(float *a, float *b, float *c, int size) {
 #if NEON
     int parts = size / 4, remaining = size % 4;
     for (int i = 0; i < parts; ++i){
         float32x4_t a_4 = vld1q_f32(a + 4 * i);
         float32x4_t b_4 = vld1q_f32(b + 4 * i);
         vst1q_f32(c + i * 4, vaddq_f32(vmulq_f32(a_4, a_4), vmulq_f32(b_4, b_4)));
     }
-    op_vec_magnitudes_c(a + parts * 4, b + parts * 4, c + parts * 4, remaining);
+    op_vec_magn_sq_c(a + parts * 4, b + parts * 4, c + parts * 4, remaining);
 #else
-    op_vec_magnitudes_c(a, b, c, size);
+    op_vec_magn_sq_c(a, b, c, size);
 #endif
 }
 
@@ -715,3 +754,4 @@ void op_mat_transp(const float *a, float *b, int M, int N) {
 
 
 
+

nntoolkitcore/core/memory.h

@@ -5,6 +5,8 @@
 #ifndef memory_h
 #define memory_h
 
+#include "stdlib.h"
+
 float *f_malloc(unsigned long size);
 
 void f_copy(float *dst, const float *src, unsigned long size);

nntoolkitcore/core/ops.h

@@ -44,13 +44,21 @@ void op_vec_sqrt(const float *a, float *c, int size);
 
 void op_vec_exp(const float *a, float *c, int size);
 
+void op_vec_pow(const float *a, const float *b, float *c, int size);
+
+void op_vec_pow_sc(const float *a, const float b, float *c, int size);
+
+void op_vec_min(const float *a, const float *b, float *c, int size);
+
+void op_vec_max(const float *a, const float *b, float *c, int size);
+
 void op_vec_log(const float *a, float *c, int size);
 
 void op_vec_tanh(const float *a, float *c, int size);
 
 void op_vec_reciprocal(const float *a, float *c, int size);
 
-void op_vec_magnitudes(float *a, float *b, float *c, int size);
+void op_vec_magn_sq(float *a, float *b, float *c, int size);
 
 void op_vec_db(float *a, float b, float *c, int size);
 

nntoolkitcore/signal/log_mel_spectrogram.c

@@ -0,0 +1,42 @@
+//
+// Created by Alex on 18.12.2020.
+//
+
+#include "log_mel_spectrogram.h"
+#include "stdlib.h"
+#include "nntoolkitcore/core/memory.h"
+#include "nntoolkitcore/core/ops.h"
+
+struct LogMelSpectrogramStruct {
+    MelFilterBankConfig mel_filter_bank_config;
+    Spectrogram spectrogram;
+    MelFilterBank bank;
+    int ts;
+    int mel_output_size;
+    float *buffer;
+};
+
+LogMelSpectrogram LogMelSpectrogramCreate(Spectrogram spectrogram, MelFilterBankConfig mel_filter_bank_config){
+    LogMelSpectrogram filter = malloc(sizeof(struct LogMelSpectrogramStruct));
+    filter->spectrogram = spectrogram;
+    SpectrogramConfig s_cfg = SpectrogramGetConfig(spectrogram);
+    filter->ts = s_cfg.ntime_series;
+    filter->mel_output_size = filter->ts * mel_filter_bank_config.n_mels;
+    filter->mel_filter_bank_config = mel_filter_bank_config;
+    filter->bank = MelFilterBankCreate(mel_filter_bank_config);
+    filter->buffer = f_malloc(s_cfg.ntime_series * s_cfg.nfreq);
+    return filter;
+}
+
+void LogMelSpectrogramApply(LogMelSpectrogram filter, const float *input, float* output){
+    SpectrogramApply(filter->spectrogram, input, filter->buffer);
+    MelFilterBankApply(filter->bank, filter->buffer, output, filter->ts);
+    op_vec_add_sc(output, 1.5849e-13, output, filter->mel_output_size);
+    op_vec_log(output, output, filter->mel_output_size);
+}
+
+void LogMelSpectrogramDestroy(LogMelSpectrogram filter){
+    MelFilterBankDestroy(filter->bank);
+    free(filter->buffer);
+    free(filter);
+}

nntoolkitcore/signal/log_mel_spectrogram.h

@@ -0,0 +1,20 @@
+//
+// Created by Alex on 18.12.2020.
+//
+
+#ifndef log_mel_spectrogram_h
+#define log_mel_spectrogram_h
+
+#include "spectrogram.h"
+#include "mel_filterbank.h"
+
+
+typedef struct LogMelSpectrogramStruct* LogMelSpectrogram;
+
+LogMelSpectrogram LogMelSpectrogramCreate(Spectrogram spectrogram, MelFilterBankConfig mel_filter_bank_config);
+
+void LogMelSpectrogramApply(LogMelSpectrogram filter, const float *input, float* output);
+
+void LogMelSpectrogramDestroy(LogMelSpectrogram filter);
+
+#endif //log_mel_spectrogram_h