Merge pull request #883 from xmos/feature/audio-example-with-state

examples/audio_network/Makefile

@@ -5,13 +5,22 @@
 TARGET = xk-audio-316-mc.xn
 APP_NAME =
 
+ifeq ($(XMOS_AITOOLSLIB_PATH),)
+$(error Path to XMOS AI Tools library and headers not set correctly!)
+endif
+include ${XMOS_AITOOLSLIB_PATH}/buildfiles/aitoolslib.make
+
 # The USED_MODULES variable lists other module used by the application. These
 # modules will extend the SOURCE_DIRS, INCLUDE_DIRS and LIB_DIRS variables. 
 # Modules are expected to be in the directory above the BASE_DIR directory.
 USED_MODULES = lib_xua lib_i2c lib_src lib_xcore_math
 
 # Audio Class 2, Async, I2S Master, 8xInput, 8xOutput
-XCC_FLAGS= -fcomment-asm -O3 -report -lquadflash -g -DUSB_TILE=tile[0] -DADAT_TX_USE_SHARED_BUFF=1 -DQUAD_SPI_FLASH=1 -DMIXER=0 #  -fxscope
+XCC_FLAGS= -fcomment-asm -O3 -report -lquadflash -g -DUSB_TILE=tile[0] -DADAT_TX_USE_SHARED_BUFF=1 -DQUAD_SPI_FLASH=1 -DMIXER=0 -fxscope
+
+XCC_FLAGS += $(APP_FLAGS) $(XMOS_AITOOLSLIB_DEFINITIONS) $(XMOS_AITOOLSLIB_INCLUDES)
+XCC_CPP_FLAGS += $(APP_FLAGS) -std=c++14 $(XMOS_AITOOLSLIB_DEFINITIONS) $(XMOS_AITOOLSLIB_INCLUDES)
+XCC_MAP_FLAGS += $(APP_FLAGS) $(XMOS_AITOOLSLIB_LIBRARIES)
 
 #=============================================================================
 # The following part of the Makefile includes the common build infrastructure

examples/audio_network/src/config.xscope

@@ -0,0 +1,3 @@
+<?xml version="1.0" encoding="UTF-8"?>
+
+<xSCOPEconfig ioMode="basic" enabled="true"/>

examples/audio_network/src/data_path.c

@@ -106,6 +106,8 @@ void nn_data_transport_thread(chanend_t c_data, chanend_t c_children) {
     }
 }
 
+extern void nn_predict_masks(int *masks, int button_state, int *mels);
+
 void nn_dsp_thread(uint32_t thread_number,
                    chanend_t c_parent,
                    chanend_t c_button_state) {
@@ -114,13 +116,6 @@ void nn_dsp_thread(uint32_t thread_number,
     int masks[MEL_BINS];
     int mels[MEL_BINS];
     int button_state = 0xF;
-    for(int i = 0; i < MEL_BINS; i++) {
-        if (i < 20) {
-            masks[i] = MEL_ONE_VALUE * (int64_t) (20 - i) / 20;
-        } else {
-            masks[i] = 0;
-        }
-    }
     SELECT_RES(
         CASE_THEN(c_parent,        go),
         CASE_THEN(c_button_state,  buttons)
@@ -132,7 +127,8 @@ void nn_dsp_thread(uint32_t thread_number,
         int gain = dsp_time_to_freq(fft_data, (int *)data_to_be_processed, &fft_state);
         dsp_calculate_mels(mels, fft_data, gain, MEL_BINS,
                            mel_coefficients, mel_bins_in_overlap);
-        dsp_apply_masks(fft_data, masks, button_state & 1, MEL_BINS,
+        nn_predict_masks(masks, button_state, mels);
+        dsp_apply_masks(fft_data, masks, 1, MEL_BINS,
                         mel_coefficients, mel_bins_in_overlap);
         dsp_freq_to_time((int *)data_processed, fft_data, &fft_state);
         continue;      // TODO: CONTINUE_NO_RESET

examples/audio_network/src/dsp.c

@@ -8,6 +8,7 @@
 #include "dsp.h"
 #include "log.h"
 #include "raised_cosine.h"
+#include "hypot.h"
 
 int dsp_time_to_freq(int64_t fft_output[], int data_to_be_processed[], fft_state_t *state) {
     static int input_frame[WINDOW_SIZE];    // State for overlap-add; first sample is oldest
@@ -22,8 +23,9 @@ int dsp_time_to_freq(int64_t fft_output[], int data_to_be_processed[], fft_state
     }
 
     // Use a raised cosing to window it, create two bits of headroom.
+    // Todo: divide the raised cosine table by 2
     for(int i = 0; i < WINDOW_SIZE; i++) {  // 42 us
-        ((int *)fft_output)[i] = (raised_cosine_512[i] * (int64_t)input_frame[i]) >> 32;
+        ((int *)fft_output)[i] = ((raised_cosine_512[i]>>1) * (int64_t)input_frame[i]) >> 32;
     }
 
     // 41 us
@@ -71,18 +73,20 @@ void dsp_calculate_mels(int mels[], int64_t fft_input[], int gain, int mel_bins,
                         int *mel_coefficients, int *mel_bins_in_overlap) {
     int magnitudes[WINDOW_SIZE/2+1];
 
-    magnitudes[0            ] = log2_16(((int *)fft_input)[0]) + (gain << LOG2_16_Q_VALUE);
-    magnitudes[WINDOW_SIZE/2] = log2_16(((int *)fft_input)[1]) + (gain << LOG2_16_Q_VALUE);
+    magnitudes[0            ] = abs(((int *)fft_input)[0]);
+    magnitudes[WINDOW_SIZE/2] = abs(((int *)fft_input)[1]);
+
     for(int i = 2; i < WINDOW_SIZE; i+=2) {           // 82 us
-        int64_t mag = (((int *)fft_input)[i] * (int64_t) ((int *)fft_input)[i] +
-                       ((int *)fft_input)[i] * (int64_t) ((int *)fft_input)[i]);
-        magnitudes[i/2] = log2_16_64(mag) + (gain << (1+LOG2_16_Q_VALUE));
+        magnitudes[i/2] = hypot_i(((int *)fft_input)[i], ((int *)fft_input)[i+1]);
     }
-
+    
     mel_compress(mels, magnitudes,
                  mel_coefficients,
                  mel_bins_in_overlap,
                  WINDOW_SIZE/2+1, mel_bins);         // 47 us
+    for(int i = 0; i < mel_bins; i++) {
+        mels[i] = (((log2_16(mels[i]) + (gain << LOG2_16_Q_VALUE) - (20 << LOG2_16_Q_VALUE)) * 21) >> LOG2_16_Q_VALUE) - 128;
+    }
 }
 
 void dsp_apply_masks(int64_t fft_output[], int masks_mel[], int enabled, int mel_bins,
@@ -140,6 +144,12 @@ int main(void) {
         dsp_calculate_mels(mels, fft_data, gain, MEL_BINS,                mel_coefficients,
                mel_bins_in_overlap
 );
+        printf("\n");
+        for(int i = 0; i < 64; i++) {
+            printf("%d ", mels[i]);
+        }
+        printf("\n");
+
         asm volatile("gettime %0" : "=r" (t1));
         for(int i = 0; i < MEL_BINS; i++) {
             masks[i] = MEL_ONE_VALUE;

examples/audio_network/src/hypot.S

@@ -0,0 +1,115 @@
+// Copyright (c) 2016, XMOS Ltd, All rights reserved
+
+	.section    .cp.rodata, "ac", @progbits
+	g_hyp_scalar:
+	.word 0x9b74eda7 // (unsigned)((double)UINT_MAX*0.607252935008881);
+
+	.section	.dp.data,"awd",@progbits
+	.text
+	.cc_top hypot_i.function
+	.globl	hypot_i
+	.align	4
+	.type	hypot_i,@function
+
+#define ITT(N) \
+	{lss r3, r0, r11; ldc r2, N};\
+	{bt r3, neg_ ## N; shr r3, r1, r2};\
+		{shr r2, r0, r2;sub r0, r0, r3};\
+		{bu next_ ## N; add r1, r1, r2};\
+	neg_ ## N:;\
+		ashr r2, r0, r2;\
+		{add r0, r0, r3; sub r1, r1, r2};\
+	next_ ## N:
+
+#define ITT_FINAL(N) \
+	{lss r3, r0, r11; ldc r2, N};\
+	{bt r3, neg_ ## N;};\
+		{shr r2, r0, r2; ldc r3, 0};\
+		{bu next_ ## N; add r1, r1, r2};\
+	neg_ ## N:;\
+		ashr r2, r0, r2;\
+		{sub r1, r1, r2; ldc r3, 0};\
+	next_ ## N:
+
+hypot_i:
+.align 8
+.issue_mode dual
+	ashr r0, r0, 2
+	ldc r11, 0
+	lss r2, r1, r11
+	{bf r2, start; lss r2, r0, r11}
+		{neg r1, r1}
+	start:
+	{bf r2, neg_0;shr r1, r1, 2}
+		{add r0, r0, r1; sub r1, r1, r0}
+		bu next_0
+	neg_0:
+		{sub r0, r0, r1; add r1, r1, r0}
+	next_0:
+
+	ITT(1)
+	ITT(2)
+	ITT(3)
+	ITT(4)
+	ITT(5)
+	ITT(6)
+	ITT(7)
+	ITT(8)
+	ITT(9)
+	ITT(10)
+	ITT(11)
+	ITT(12)
+	ITT(13)
+	ITT(14)
+	ITT(15)
+	ITT(16)
+	ITT(17)
+	ITT_FINAL(18)
+
+	ldw r0, cp[g_hyp_scalar]
+	maccu r11, r3, r1, r0
+	{retsp 0;shl r0, r11, 2}
+
+
+.tmp_hypot_i:
+	.size	hypot_i, .tmp_hypot_i-hypot_i
+	.align	4
+	.cc_bottom hypot_i.function
+
+	.set	hypot_i.nstackwords,0
+	.globl	hypot_i.nstackwords
+	.set	hypot_i.maxcores,1
+	.globl	hypot_i.maxcores
+	.set	hypot_i.maxtimers,0
+	.globl	hypot_i.maxtimers
+	.set	hypot_i.maxchanends,0
+	.globl	hypot_i.maxchanends
+
+	.cc_top scale.function
+	.globl	scale
+	.align	4
+	.type	scale,@function
+scale:
+.align 8
+.issue_mode dual
+	ldc r3, 0
+	ldivu r1,r11,r3,r1,r2
+	ldivu r0,r2,r11,r0,r2
+	retsp 0
+
+.tmp_scale:
+	.size	scale, .tmp_scale-scale
+	.align	4
+	.cc_bottom scale.function
+
+	.set	scale.nstackwords,0
+	.globl	scale.nstackwords
+	.set	scale.maxcores,1
+	.globl	scale.maxcores
+	.set	scale.maxtimers,0
+	.globl	scale.maxtimers
+	.set	scale.maxchanends,0
+	.globl	scale.maxchanends
+
+
+

examples/audio_network/src/hypot.h

@@ -0,0 +1 @@
+extern int hypot_i(int x, int y);

examples/audio_network/src/log.c

@@ -144,11 +144,11 @@ struct {int adder, multiplier; } xxx[] = {
 
 int log2_16(uint32_t x) {
     int base;
-    int lookup;
+    uint32_t lookup;
     asm("clz %0, %1" : "=r" (base) : "r" (x));
     asm("shl %0, %1, %2" : "=r" (lookup) : "r" (x), "r" (base-25));
     asm("shl %0, %1, %2" : "=r" (x) : "r" (x), "r" (base-16));
-    asm("zext %0, 6" : "+r" (lookup));
+    asm("{zext %0, 6;zext %1, 16}" : "+r" (lookup), "+r" (x));
     int log_multiplier = xxx[lookup].multiplier;
     int log_adder = xxx[lookup].adder;
     return log_adder + ((log_multiplier * x) >> 16) + ((31-base)<<10);
@@ -175,17 +175,20 @@ int log2_16_64(uint64_t x) {
         return log2_16((uint32_t)x);
     }
     asm("lextract %0, %1, %2, %3, 32" : "=r" (y) : "r" ((int32_t)(x >> 32)), "r" ((uint32_t)x), "r" (base));
-    return log2_16(y) + base;
+    return log2_16(y) + (base<<10);
 }
 
 #ifdef LOCAL_LOG_MAIN
 
 int main(void) {
+    for(int i = 0; i < 64; i = i +1) {
+        printf("%d: %d %f\n", i, log2_16(i), log((float)i)/log(2.0)*8*128);
+    }
     for(int i = 1000; i < 0*(1<<30); i = i * 2 + 834) {
         printf("%d: %d %f\n", i, log2_8(i), log((float)i)/log(2.0)*8);
     }
-    for(int i = 1000; i < (1<<30); i = i * 2 + 834) {
-        printf("%d: %d %f\n", i, log2_16(i), log((float)i)/log(2.0)*8*128);
+    for(int64_t i = 100000000; i < (1LL<<60); i = i * 2 + 834) {
+        printf("%lld: %d %f\n", i, log2_16_64(i), log((float)i)/log(2.0)*8*128);
     }
 }