path: root/cras/src/dsp/drc_kernel.c

/* Copyright (c) 2013 The Chromium OS Authors. All rights reserved.
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */

/* Copyright (C) 2011 Google Inc. All rights reserved.
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE.WEBKIT file.
 */

#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include "drc_math.h"
#include "drc_kernel.h"

#define MAX_PRE_DELAY_FRAMES 1024
#define MAX_PRE_DELAY_FRAMES_MASK (MAX_PRE_DELAY_FRAMES - 1)
#define DEFAULT_PRE_DELAY_FRAMES 256
#define DIVISION_FRAMES 32
#define DIVISION_FRAMES_MASK (DIVISION_FRAMES - 1)

#define assert_on_compile(e) ((void)sizeof(char[1 - 2 * !(e)]))
#define assert_on_compile_is_power_of_2(n)                                     \
	assert_on_compile((n) != 0 && (((n) & ((n)-1)) == 0))

const float uninitialized_value = -1;
static int drc_math_initialized;

void dk_init(struct drc_kernel *dk, float sample_rate)
{
	int i;

	if (!drc_math_initialized) {
		drc_math_initialized = 1;
		drc_math_init();
	}

	dk->sample_rate = sample_rate;
	dk->detector_average = 0;
	dk->compressor_gain = 1;
	dk->enabled = 0;
	dk->processed = 0;
	dk->last_pre_delay_frames = DEFAULT_PRE_DELAY_FRAMES;
	dk->pre_delay_read_index = 0;
	dk->pre_delay_write_index = DEFAULT_PRE_DELAY_FRAMES;
	dk->max_attack_compression_diff_db = -INFINITY;
	dk->ratio = uninitialized_value;
	dk->slope = uninitialized_value;
	dk->linear_threshold = uninitialized_value;
	dk->db_threshold = uninitialized_value;
	dk->db_knee = uninitialized_value;
	dk->knee_threshold = uninitialized_value;
	dk->ratio_base = uninitialized_value;
	dk->K = uninitialized_value;

	assert_on_compile_is_power_of_2(DIVISION_FRAMES);
	assert_on_compile(DIVISION_FRAMES % 4 == 0);
	/* Allocate predelay buffers */
	assert_on_compile_is_power_of_2(MAX_PRE_DELAY_FRAMES);
	for (i = 0; i < DRC_NUM_CHANNELS; i++) {
		size_t size = sizeof(float) * MAX_PRE_DELAY_FRAMES;
		dk->pre_delay_buffers[i] = (float *)calloc(1, size);
	}
}

void dk_free(struct drc_kernel *dk)
{
	int i;
	for (i = 0; i < DRC_NUM_CHANNELS; ++i)
		free(dk->pre_delay_buffers[i]);
}

/* Sets the pre-delay (lookahead) buffer size */
static void set_pre_delay_time(struct drc_kernel *dk, float pre_delay_time)
{
	int i;
	/* Re-configure look-ahead section pre-delay if delay time has
	 * changed. */
	unsigned pre_delay_frames = pre_delay_time * dk->sample_rate;
	pre_delay_frames = min(pre_delay_frames, MAX_PRE_DELAY_FRAMES - 1);

	/* Make pre_delay_frames multiplies of DIVISION_FRAMES. This way we
	 * won't split a division of samples into two blocks of memory, so it is
	 * easier to process. This may make the actual delay time slightly less
	 * than the specified value, but the difference is less than 1ms. */
	pre_delay_frames &= ~DIVISION_FRAMES_MASK;

	/* We need at least one division buffer, so the incoming data won't
	 * overwrite the output data */
	pre_delay_frames = max(pre_delay_frames, DIVISION_FRAMES);

	if (dk->last_pre_delay_frames != pre_delay_frames) {
		dk->last_pre_delay_frames = pre_delay_frames;
		for (i = 0; i < DRC_NUM_CHANNELS; ++i) {
			size_t size = sizeof(float) * MAX_PRE_DELAY_FRAMES;
			memset(dk->pre_delay_buffers[i], 0, size);
		}

		dk->pre_delay_read_index = 0;
		dk->pre_delay_write_index = pre_delay_frames;
	}
}

/* Exponential curve for the knee.  It is 1st derivative matched at
 * dk->linear_threshold and asymptotically approaches the value
 * dk->linear_threshold + 1 / k.
 *
 * This is used only when calculating the static curve, not used when actually
 * compress the input data (knee_curveK below is used instead).
 */
static float knee_curve(struct drc_kernel *dk, float x, float k)
{
	/* Linear up to threshold. */
	if (x < dk->linear_threshold)
		return x;

	return dk->linear_threshold +
	       (1 - knee_expf(-k * (x - dk->linear_threshold))) / k;
}

/* Approximate 1st derivative with input and output expressed in dB.  This slope
 * is equal to the inverse of the compression "ratio".  In other words, a
 * compression ratio of 20 would be a slope of 1/20.
 */
static float slope_at(struct drc_kernel *dk, float x, float k)
{
	if (x < dk->linear_threshold)
		return 1;

	float x2 = x * 1.001;

	float x_db = linear_to_decibels(x);
	float x2Db = linear_to_decibels(x2);

	float y_db = linear_to_decibels(knee_curve(dk, x, k));
	float y2Db = linear_to_decibels(knee_curve(dk, x2, k));

	float m = (y2Db - y_db) / (x2Db - x_db);

	return m;
}

static float k_at_slope(struct drc_kernel *dk, float desired_slope)
{
	float x_db = dk->db_threshold + dk->db_knee;
	float x = decibels_to_linear(x_db);

	/* Approximate k given initial values. */
	float minK = 0.1;
	float maxK = 10000;
	float k = 5;
	int i;

	for (i = 0; i < 15; ++i) {
		/* A high value for k will more quickly asymptotically approach
		 * a slope of 0. */
		float slope = slope_at(dk, x, k);

		if (slope < desired_slope) {
			/* k is too high. */
			maxK = k;
		} else {
			/* k is too low. */
			minK = k;
		}

		/* Re-calculate based on geometric mean. */
		k = sqrtf(minK * maxK);
	}

	return k;
}

static void update_static_curve_parameters(struct drc_kernel *dk,
					   float db_threshold, float db_knee,
					   float ratio)
{
	if (db_threshold != dk->db_threshold || db_knee != dk->db_knee ||
	    ratio != dk->ratio) {
		/* Threshold and knee. */
		dk->db_threshold = db_threshold;
		dk->linear_threshold = decibels_to_linear(db_threshold);
		dk->db_knee = db_knee;

		/* Compute knee parameters. */
		dk->ratio = ratio;
		dk->slope = 1 / dk->ratio;

		float k = k_at_slope(dk, 1 / dk->ratio);
		dk->K = k;
		/* See knee_curveK() for details */
		dk->knee_alpha = dk->linear_threshold + 1 / k;
		dk->knee_beta = -expf(k * dk->linear_threshold) / k;

		dk->knee_threshold = decibels_to_linear(db_threshold + db_knee);
		/* See volume_gain() for details */
		float y0 = knee_curve(dk, dk->knee_threshold, k);
		dk->ratio_base = y0 * powf(dk->knee_threshold, -dk->slope);
	}
}

/* This is the knee part of the compression curve. Returns the output level
 * given the input level x. */
static float knee_curveK(struct drc_kernel *dk, float x)
{
	/* The formula in knee_curveK is dk->linear_threshold +
	 * (1 - expf(-k * (x - dk->linear_threshold))) / k
	 * which simplifies to (alpha + beta * expf(gamma))
	 * where alpha = dk->linear_threshold + 1 / k
	 *	 beta = -expf(k * dk->linear_threshold) / k
	 *	 gamma = -k * x
	 */
	return dk->knee_alpha + dk->knee_beta * knee_expf(-dk->K * x);
}

/* Full compression curve with constant ratio after knee. Returns the ratio of
 * output and input signal. */
static float volume_gain(struct drc_kernel *dk, float x)
{
	float y;

	if (x < dk->knee_threshold) {
		if (x < dk->linear_threshold)
			return 1;
		y = knee_curveK(dk, x) / x;
	} else {
		/* Constant ratio after knee.
		 * log(y/y0) = s * log(x/x0)
		 * => y = y0 * (x/x0)^s
		 * => y = [y0 * (1/x0)^s] * x^s
		 * => y = dk->ratio_base * x^s
		 * => y/x = dk->ratio_base * x^(s - 1)
		 * => y/x = dk->ratio_base * e^(log(x) * (s - 1))
		 */
		y = dk->ratio_base * knee_expf(logf(x) * (dk->slope - 1));
	}

	return y;
}

void dk_set_parameters(struct drc_kernel *dk, float db_threshold, float db_knee,
		       float ratio, float attack_time, float release_time,
		       float pre_delay_time, float db_post_gain,
		       float releaseZone1, float releaseZone2,
		       float releaseZone3, float releaseZone4)
{
	float sample_rate = dk->sample_rate;

	update_static_curve_parameters(dk, db_threshold, db_knee, ratio);

	/* Makeup gain. */
	float full_range_gain = volume_gain(dk, 1);
	float full_range_makeup_gain = 1 / full_range_gain;

	/* Empirical/perceptual tuning. */
	full_range_makeup_gain = powf(full_range_makeup_gain, 0.6f);

	dk->main_linear_gain =
		decibels_to_linear(db_post_gain) * full_range_makeup_gain;

	/* Attack parameters. */
	attack_time = max(0.001f, attack_time);
	dk->attack_frames = attack_time * sample_rate;

	/* Release parameters. */
	float release_frames = sample_rate * release_time;

	/* Detector release time. */
	float sat_release_time = 0.0025f;
	float sat_release_frames = sat_release_time * sample_rate;
	dk->sat_release_frames_inv_neg = -1 / sat_release_frames;
	dk->sat_release_rate_at_neg_two_db =
		decibels_to_linear(-2 * dk->sat_release_frames_inv_neg) - 1;

	/* Create a smooth function which passes through four points.
	 * Polynomial of the form y = a + b*x + c*x^2 + d*x^3 + e*x^4
	 */
	float y1 = release_frames * releaseZone1;
	float y2 = release_frames * releaseZone2;
	float y3 = release_frames * releaseZone3;
	float y4 = release_frames * releaseZone4;

	/* All of these coefficients were derived for 4th order polynomial curve
	 * fitting where the y values match the evenly spaced x values as
	 * follows: (y1 : x == 0, y2 : x == 1, y3 : x == 2, y4 : x == 3)
	 */
	dk->kA = 0.9999999999999998f * y1 + 1.8432219684323923e-16f * y2 -
		 1.9373394351676423e-16f * y3 + 8.824516011816245e-18f * y4;
	dk->kB = -1.5788320352845888f * y1 + 2.3305837032074286f * y2 -
		 0.9141194204840429f * y3 + 0.1623677525612032f * y4;
	dk->kC = 0.5334142869106424f * y1 - 1.272736789213631f * y2 +
		 0.9258856042207512f * y3 - 0.18656310191776226f * y4;
	dk->kD = 0.08783463138207234f * y1 - 0.1694162967925622f * y2 +
		 0.08588057951595272f * y3 - 0.00429891410546283f * y4;
	dk->kE = -0.042416883008123074f * y1 + 0.1115693827987602f * y2 -
		 0.09764676325265872f * y3 + 0.028494263462021576f * y4;

	/* x ranges from 0 -> 3	      0	   1	2   3
	 *			     -15  -10  -5   0db
	 *
	 * y calculates adaptive release frames depending on the amount of
	 * compression.
	 */
	set_pre_delay_time(dk, pre_delay_time);
}

void dk_set_enabled(struct drc_kernel *dk, int enabled)
{
	dk->enabled = enabled;
}

/* Updates the envelope_rate used for the next division */
static void dk_update_envelope(struct drc_kernel *dk)
{
	const float kA = dk->kA;
	const float kB = dk->kB;
	const float kC = dk->kC;
	const float kD = dk->kD;
	const float kE = dk->kE;
	const float attack_frames = dk->attack_frames;

	/* Calculate desired gain */
	float desired_gain = dk->detector_average;

	/* Pre-warp so we get desired_gain after sin() warp below. */
	float scaled_desired_gain = warp_asinf(desired_gain);

	/* Deal with envelopes */

	/* envelope_rate is the rate we slew from current compressor level to
	 * the desired level.  The exact rate depends on if we're attacking or
	 * releasing and by how much.
	 */
	float envelope_rate;

	int is_releasing = scaled_desired_gain > dk->compressor_gain;

	/* compression_diff_db is the difference between current compression
	 * level and the desired level. */
	float compression_diff_db =
		linear_to_decibels(dk->compressor_gain / scaled_desired_gain);

	if (is_releasing) {
		/* Release mode - compression_diff_db should be negative dB */
		dk->max_attack_compression_diff_db = -INFINITY;

		/* Fix gremlins. */
		if (isbadf(compression_diff_db))
			compression_diff_db = -1;

		/* Adaptive release - higher compression (lower
		 * compression_diff_db) releases faster. Contain within range:
		 * -12 -> 0 then scale to go from 0 -> 3
		 */
		float x = compression_diff_db;
		x = max(-12.0f, x);
		x = min(0.0f, x);
		x = 0.25f * (x + 12);

		/* Compute adaptive release curve using 4th order polynomial.
		 * Normal values for the polynomial coefficients would create a
		 * monotonically increasing function.
		 */
		float x2 = x * x;
		float x3 = x2 * x;
		float x4 = x2 * x2;
		float release_frames =
			kA + kB * x + kC * x2 + kD * x3 + kE * x4;

#define kSpacingDb 5
		float db_per_frame = kSpacingDb / release_frames;
		envelope_rate = decibels_to_linear(db_per_frame);
	} else {
		/* Attack mode - compression_diff_db should be positive dB */

		/* Fix gremlins. */
		if (isbadf(compression_diff_db))
			compression_diff_db = 1;

		/* As long as we're still in attack mode, use a rate based off
		 * the largest compression_diff_db we've encountered so far.
		 */
		dk->max_attack_compression_diff_db =
			max(dk->max_attack_compression_diff_db,
			    compression_diff_db);

		float eff_atten_diff_db =
			max(0.5f, dk->max_attack_compression_diff_db);

		float x = 0.25f / eff_atten_diff_db;
		envelope_rate = 1 - powf(x, 1 / attack_frames);
	}

	dk->envelope_rate = envelope_rate;
	dk->scaled_desired_gain = scaled_desired_gain;
}

/* For a division of frames, take the absolute values of left channel and right
 * channel, store the maximum of them in output. */
#if defined(__aarch64__)
static inline void max_abs_division(float *output, const float *data0,
				    const float *data1)
{
	int count = DIVISION_FRAMES / 4;

	// clang-format off
	__asm__ __volatile__(
		"1:                                     \n"
		"ld1 {v0.4s}, [%[data0]], #16           \n"
		"ld1 {v1.4s}, [%[data1]], #16           \n"
		"fabs v0.4s, v0.4s                      \n"
		"fabs v1.4s, v1.4s                      \n"
		"fmax v0.4s, v0.4s, v1.4s               \n"
		"st1 {v0.4s}, [%[output]], #16          \n"
		"subs %w[count], %w[count], #1          \n"
		"b.ne 1b                                \n"
		: /* output */
		  [data0]"+r"(data0),
		  [data1]"+r"(data1),
		  [output]"+r"(output),
		  [count]"+r"(count)
		: /* input */
		: /* clobber */
		  "v0", "v1", "memory", "cc");
	// clang-format on
}
#elif defined(__ARM_NEON__)
static inline void max_abs_division(float *output, const float *data0,
				    const float *data1)
{
	int count = DIVISION_FRAMES / 4;

	// clang-format off
	__asm__ __volatile__(
		"1:                                     \n"
		"vld1.32 {q0}, [%[data0]]!              \n"
		"vld1.32 {q1}, [%[data1]]!              \n"
		"vabs.f32 q0, q0                        \n"
		"vabs.f32 q1, q1                        \n"
		"vmax.f32 q0, q1                        \n"
		"vst1.32 {q0}, [%[output]]!             \n"
		"subs %[count], #1                      \n"
		"bne 1b                                 \n"
		: /* output */
		  [data0]"+r"(data0),
		  [data1]"+r"(data1),
		  [output]"+r"(output),
		  [count]"+r"(count)
		: /* input */
		: /* clobber */
		  "q0", "q1", "memory", "cc");
	// clang-format on
}
#elif defined(__SSE3__)
#include <emmintrin.h>
static inline void max_abs_division(float *output, const float *data0,
				    const float *data1)
{
	__m128 x, y;
	int count = DIVISION_FRAMES / 4;
	// clang-format off
	__asm__ __volatile__(
		"1:                                     \n"
		"lddqu (%[data0]), %[x]                 \n"
		"lddqu (%[data1]), %[y]                 \n"
		"andps %[mask], %[x]                    \n"
		"andps %[mask], %[y]                    \n"
		"maxps %[y], %[x]                       \n"
		"movdqu %[x], (%[output])               \n"
		"add $16, %[data0]                      \n"
		"add $16, %[data1]                      \n"
		"add $16, %[output]                     \n"
		"sub $1, %[count]                       \n"
		"jnz 1b                                 \n"
		: /* output */
		  [data0]"+r"(data0),
		  [data1]"+r"(data1),
		  [output]"+r"(output),
		  [count]"+r"(count),
		  [x]"=&x"(x),
		  [y]"=&x"(y)
		: /* input */
		  [mask]"x"(_mm_set1_epi32(0x7fffffff))
		: /* clobber */
		  "memory", "cc");
	// clang-format on
}
#else
static inline void max_abs_division(float *output, const float *data0,
				    const float *data1)
{
	int i;
	for (i = 0; i < DIVISION_FRAMES; i++)
		output[i] = fmaxf(fabsf(data0[i]), fabsf(data1[i]));
}
#endif

/* Update detector_average from the last input division. */
static void dk_update_detector_average(struct drc_kernel *dk)
{
	float abs_input_array[DIVISION_FRAMES];
	const float sat_release_frames_inv_neg = dk->sat_release_frames_inv_neg;
	const float sat_release_rate_at_neg_two_db =
		dk->sat_release_rate_at_neg_two_db;
	float detector_average = dk->detector_average;
	int div_start, i;

	/* Calculate the start index of the last input division */
	if (dk->pre_delay_write_index == 0) {
		div_start = MAX_PRE_DELAY_FRAMES - DIVISION_FRAMES;
	} else {
		div_start = dk->pre_delay_write_index - DIVISION_FRAMES;
	}

	/* The max abs value across all channels for this frame */
	max_abs_division(abs_input_array, &dk->pre_delay_buffers[0][div_start],
			 &dk->pre_delay_buffers[1][div_start]);

	for (i = 0; i < DIVISION_FRAMES; i++) {
		/* Compute compression amount from un-delayed signal */
		float abs_input = abs_input_array[i];

		/* Calculate shaped power on undelayed input.  Put through
		 * shaping curve. This is linear up to the threshold, then
		 * enters a "knee" portion followed by the "ratio" portion. The
		 * transition from the threshold to the knee is smooth (1st
		 * derivative matched). The transition from the knee to the
		 * ratio portion is smooth (1st derivative matched).
		 */
		float gain = volume_gain(dk, abs_input);
		int is_release = (gain > detector_average);
		if (is_release) {
			if (gain > NEG_TWO_DB) {
				detector_average +=
					(gain - detector_average) *
					sat_release_rate_at_neg_two_db;
			} else {
				float gain_db = linear_to_decibels(gain);
				float db_per_frame =
					gain_db * sat_release_frames_inv_neg;
				float sat_release_rate =
					decibels_to_linear(db_per_frame) - 1;
				detector_average += (gain - detector_average) *
						    sat_release_rate;
			}
		} else {
			detector_average = gain;
		}

		/* Fix gremlins. */
		if (isbadf(detector_average))
			detector_average = 1.0f;
		else
			detector_average = min(detector_average, 1.0f);
	}

	dk->detector_average = detector_average;
}

/* Calculate compress_gain from the envelope and apply total_gain to compress
 * the next output division. */
/* TODO(fbarchard): Port to aarch64 */
#if defined(__ARM_NEON__)
#include <arm_neon.h>
static void dk_compress_output(struct drc_kernel *dk)
{
	const float main_linear_gain = dk->main_linear_gain;
	const float envelope_rate = dk->envelope_rate;
	const float scaled_desired_gain = dk->scaled_desired_gain;
	const float compressor_gain = dk->compressor_gain;
	const int div_start = dk->pre_delay_read_index;
	float *ptr_left = &dk->pre_delay_buffers[0][div_start];
	float *ptr_right = &dk->pre_delay_buffers[1][div_start];
	int count = DIVISION_FRAMES / 4;

	/* See warp_sinf() for the details for the constants. */
	const float32x4_t A7 = vdupq_n_f32(-4.3330336920917034149169921875e-3f);
	const float32x4_t A5 = vdupq_n_f32(7.9434238374233245849609375e-2f);
	const float32x4_t A3 = vdupq_n_f32(-0.645892798900604248046875f);
	const float32x4_t A1 = vdupq_n_f32(1.5707910060882568359375f);

	/* Exponential approach to desired gain. */
	if (envelope_rate < 1) {
		float c = compressor_gain - scaled_desired_gain;
		float r = 1 - envelope_rate;
		float32x4_t x0 = { c * r, c * r * r, c * r * r * r,
				   c * r * r * r * r };
		float32x4_t x, x2, x4, left, right, tmp1, tmp2;

		// clang-format off
		__asm__ __volatile(
			"b 2f                                               \n"
			"1:                                                 \n"
			"vmul.f32 %q[x0], %q[r4]                            \n"
			"2:                                                 \n"
			"vld1.32 {%e[left],%f[left]}, [%[ptr_left]]         \n"
			"vld1.32 {%e[right],%f[right]}, [%[ptr_right]]      \n"
			"vadd.f32 %q[x], %q[x0], %q[base]                   \n"
			/* Calculate warp_sin() for four values in x. */
			"vmul.f32 %q[x2], %q[x], %q[x]                      \n"
			"vmov.f32 %q[tmp1], %q[A5]                          \n"
			"vmov.f32 %q[tmp2], %q[A1]                          \n"
			"vmul.f32 %q[x4], %q[x2], %q[x2]                    \n"
			"vmla.f32 %q[tmp1], %q[A7], %q[x2]                  \n"
			"vmla.f32 %q[tmp2], %q[A3], %q[x2]                  \n"
			"vmla.f32 %q[tmp2], %q[tmp1], %q[x4]                \n"
			"vmul.f32 %q[tmp2], %q[tmp2], %q[x]                 \n"
			/* Now tmp2 contains the result of warp_sin(). */
			"vmul.f32 %q[tmp2], %q[tmp2], %q[g]                 \n"
			"vmul.f32 %q[left], %q[tmp2]                        \n"
			"vmul.f32 %q[right], %q[tmp2]                       \n"
			"vst1.32 {%e[left],%f[left]}, [%[ptr_left]]!        \n"
			"vst1.32 {%e[right],%f[right]}, [%[ptr_right]]!     \n"
			"subs %[count], #1                                  \n"
			"bne 1b                                             \n"
			: /* output */
			  "=r"(count),
			  "=r"(ptr_left),
			  "=r"(ptr_right),
			  "=w"(x0),
			  [x]"=&w"(x),
			  [x2]"=&w"(x2),
			  [x4]"=&w"(x4),
			  [left]"=&w"(left),
			  [right]"=&w"(right),
			  [tmp1]"=&w"(tmp1),
			  [tmp2]"=&w"(tmp2)
			: /* input */
			  [count]"0"(count),
			  [ptr_left]"1"(ptr_left),
			  [ptr_right]"2"(ptr_right),
			  [x0]"3"(x0),
			  [A1]"w"(A1),
			  [A3]"w"(A3),
			  [A5]"w"(A5),
			  [A7]"w"(A7),
			  [base]"w"(vdupq_n_f32(scaled_desired_gain)),
			  [r4]"w"(vdupq_n_f32(r*r*r*r)),
			  [g]"w"(vdupq_n_f32(main_linear_gain))
			: /* clobber */
			  "memory", "cc");
		// clang-format on
		dk->compressor_gain = x[3];
	} else {
		float c = compressor_gain;
		float r = envelope_rate;
		float32x4_t x = { c * r, c * r * r, c * r * r * r,
				  c * r * r * r * r };
		float32x4_t x2, x4, left, right, tmp1, tmp2;

		// clang-format off
		__asm__ __volatile(
			"b 2f                                               \n"
			"1:                                                 \n"
			"vmul.f32 %q[x], %q[r4]                             \n"
			"2:                                                 \n"
			"vld1.32 {%e[left],%f[left]}, [%[ptr_left]]         \n"
			"vld1.32 {%e[right],%f[right]}, [%[ptr_right]]      \n"
			"vmin.f32 %q[x], %q[one]                            \n"
			/* Calculate warp_sin() for four values in x. */
			"vmul.f32 %q[x2], %q[x], %q[x]                      \n"
			"vmov.f32 %q[tmp1], %q[A5]                          \n"
			"vmov.f32 %q[tmp2], %q[A1]                          \n"
			"vmul.f32 %q[x4], %q[x2], %q[x2]                    \n"
			"vmla.f32 %q[tmp1], %q[A7], %q[x2]                  \n"
			"vmla.f32 %q[tmp2], %q[A3], %q[x2]                  \n"
			"vmla.f32 %q[tmp2], %q[tmp1], %q[x4]                \n"
			"vmul.f32 %q[tmp2], %q[tmp2], %q[x]                 \n"
			/* Now tmp2 contains the result of warp_sin(). */
			"vmul.f32 %q[tmp2], %q[tmp2], %q[g]                 \n"
			"vmul.f32 %q[left], %q[tmp2]                        \n"
			"vmul.f32 %q[right], %q[tmp2]                       \n"
			"vst1.32 {%e[left],%f[left]}, [%[ptr_left]]!        \n"
			"vst1.32 {%e[right],%f[right]}, [%[ptr_right]]!     \n"
			"subs %[count], #1                                  \n"
			"bne 1b                                             \n"
			: /* output */
			  "=r"(count),
			  "=r"(ptr_left),
			  "=r"(ptr_right),
			  "=w"(x),
			  [x2]"=&w"(x2),
			  [x4]"=&w"(x4),
			  [left]"=&w"(left),
			  [right]"=&w"(right),
			  [tmp1]"=&w"(tmp1),
			  [tmp2]"=&w"(tmp2)
			: /* input */
			  [count]"0"(count),
			  [ptr_left]"1"(ptr_left),
			  [ptr_right]"2"(ptr_right),
			  [x]"3"(x),
			  [A1]"w"(A1),
			  [A3]"w"(A3),
			  [A5]"w"(A5),
			  [A7]"w"(A7),
			  [one]"w"(vdupq_n_f32(1)),
			  [r4]"w"(vdupq_n_f32(r*r*r*r)),
			  [g]"w"(vdupq_n_f32(main_linear_gain))
			: /* clobber */
			  "memory", "cc");
		// clang-format on
		dk->compressor_gain = x[3];
	}
}
#elif defined(__SSE3__) && defined(__x86_64__)
#include <emmintrin.h>
static void dk_compress_output(struct drc_kernel *dk)
{
	const float main_linear_gain = dk->main_linear_gain;
	const float envelope_rate = dk->envelope_rate;
	const float scaled_desired_gain = dk->scaled_desired_gain;
	const float compressor_gain = dk->compressor_gain;
	const int div_start = dk->pre_delay_read_index;
	float *ptr_left = &dk->pre_delay_buffers[0][div_start];
	float *ptr_right = &dk->pre_delay_buffers[1][div_start];
	int count = DIVISION_FRAMES / 4;

	/* See warp_sinf() for the details for the constants. */
	const __m128 A7 = _mm_set1_ps(-4.3330336920917034149169921875e-3f);
	const __m128 A5 = _mm_set1_ps(7.9434238374233245849609375e-2f);
	const __m128 A3 = _mm_set1_ps(-0.645892798900604248046875f);
	const __m128 A1 = _mm_set1_ps(1.5707910060882568359375f);

	/* Exponential approach to desired gain. */
	if (envelope_rate < 1) {
		float c = compressor_gain - scaled_desired_gain;
		float r = 1 - envelope_rate;
		__m128 x0 = { c * r, c * r * r, c * r * r * r,
			      c * r * r * r * r };
		__m128 x, x2, x4, left, right, tmp1, tmp2;

		// clang-format off
		__asm__ __volatile(
			"jmp 2f                                     \n"
			"1:                                         \n"
			"mulps %[r4], %[x0]                         \n"
			"2:                                         \n"
			"lddqu (%[ptr_left]), %[left]               \n"
			"lddqu (%[ptr_right]), %[right]             \n"
			"movaps %[x0], %[x]                         \n"
			"addps %[base], %[x]                        \n"
			/* Calculate warp_sin() for four values in x. */
			"movaps %[x], %[x2]                         \n"
			"mulps %[x], %[x2]                          \n"
			"movaps %[x2], %[x4]                        \n"
			"movaps %[x2], %[tmp1]                      \n"
			"movaps %[x2], %[tmp2]                      \n"
			"mulps %[x2], %[x4]                         \n"
			"mulps %[A7], %[tmp1]                       \n"
			"mulps %[A3], %[tmp2]                       \n"
			"addps %[A5], %[tmp1]                       \n"
			"addps %[A1], %[tmp2]                       \n"
			"mulps %[x4], %[tmp1]                       \n"
			"addps %[tmp1], %[tmp2]                     \n"
			"mulps %[x], %[tmp2]                        \n"
			/* Now tmp2 contains the result of warp_sin(). */
			"mulps %[g], %[tmp2]                        \n"
			"mulps %[tmp2], %[left]                     \n"
			"mulps %[tmp2], %[right]                    \n"
			"movdqu %[left], (%[ptr_left])              \n"
			"movdqu %[right], (%[ptr_right])            \n"
			"add $16, %[ptr_left]                       \n"
			"add $16, %[ptr_right]                      \n"
			"sub $1, %[count]                           \n"
			"jne 1b                                     \n"
			: /* output */
			  "=r"(count),
			  "=r"(ptr_left),
			  "=r"(ptr_right),
			  "=x"(x0),
			  [x]"=&x"(x),
			  [x2]"=&x"(x2),
			  [x4]"=&x"(x4),
			  [left]"=&x"(left),
			  [right]"=&x"(right),
			  [tmp1]"=&x"(tmp1),
			  [tmp2]"=&x"(tmp2)
			: /* input */
			  [count]"0"(count),
			  [ptr_left]"1"(ptr_left),
			  [ptr_right]"2"(ptr_right),
			  [x0]"3"(x0),
			  [A1]"x"(A1),
			  [A3]"x"(A3),
			  [A5]"x"(A5),
			  [A7]"x"(A7),
			  [base]"x"(_mm_set1_ps(scaled_desired_gain)),
			  [r4]"x"(_mm_set1_ps(r*r*r*r)),
			  [g]"x"(_mm_set1_ps(main_linear_gain))
			: /* clobber */
			  "memory", "cc");
		// clang-format on
		dk->compressor_gain = x[3];
	} else {
		/* See warp_sinf() for the details for the constants. */
		__m128 A7 = _mm_set1_ps(-4.3330336920917034149169921875e-3f);
		__m128 A5 = _mm_set1_ps(7.9434238374233245849609375e-2f);
		__m128 A3 = _mm_set1_ps(-0.645892798900604248046875f);
		__m128 A1 = _mm_set1_ps(1.5707910060882568359375f);

		float c = compressor_gain;
		float r = envelope_rate;
		__m128 x = { c * r, c * r * r, c * r * r * r,
			     c * r * r * r * r };
		__m128 x2, x4, left, right, tmp1, tmp2;

		// clang-format off
		__asm__ __volatile(
			"jmp 2f                                     \n"
			"1:                                         \n"
			"mulps %[r4], %[x]                          \n"
			"2:                                         \n"
			"lddqu (%[ptr_left]), %[left]               \n"
			"lddqu (%[ptr_right]), %[right]             \n"
			"minps %[one], %[x]                         \n"
			/* Calculate warp_sin() for four values in x. */
			"movaps %[x], %[x2]                         \n"
			"mulps %[x], %[x2]                          \n"
			"movaps %[x2], %[x4]                        \n"
			"movaps %[x2], %[tmp1]                      \n"
			"movaps %[x2], %[tmp2]                      \n"
			"mulps %[x2], %[x4]                         \n"
			"mulps %[A7], %[tmp1]                       \n"
			"mulps %[A3], %[tmp2]                       \n"
			"addps %[A5], %[tmp1]                       \n"
			"addps %[A1], %[tmp2]                       \n"
			"mulps %[x4], %[tmp1]                       \n"
			"addps %[tmp1], %[tmp2]                     \n"
			"mulps %[x], %[tmp2]                        \n"
			/* Now tmp2 contains the result of warp_sin(). */
			"mulps %[g], %[tmp2]                        \n"
			"mulps %[tmp2], %[left]                     \n"
			"mulps %[tmp2], %[right]                    \n"
			"movdqu %[left], (%[ptr_left])              \n"
			"movdqu %[right], (%[ptr_right])            \n"
			"add $16, %[ptr_left]                       \n"
			"add $16, %[ptr_right]                      \n"
			"sub $1, %[count]                           \n"
			"jne 1b                                     \n"
			: /* output */
			  "=r"(count),
			  "=r"(ptr_left),
			  "=r"(ptr_right),
			  "=x"(x),
			  [x2]"=&x"(x2),
			  [x4]"=&x"(x4),
			  [left]"=&x"(left),
			  [right]"=&x"(right),
			  [tmp1]"=&x"(tmp1),
			  [tmp2]"=&x"(tmp2)
			: /* input */
			  [count]"0"(count),
			  [ptr_left]"1"(ptr_left),
			  [ptr_right]"2"(ptr_right),
			  [x]"3"(x),
			  [A1]"x"(A1),
			  [A3]"x"(A3),
			  [A5]"x"(A5),
			  [A7]"x"(A7),
			  [one]"x"(_mm_set1_ps(1)),
			  [r4]"x"(_mm_set1_ps(r*r*r*r)),
			  [g]"x"(_mm_set1_ps(main_linear_gain))
			: /* clobber */
			  "memory", "cc");
		// clang-format on
		dk->compressor_gain = x[3];
	}
}
#else
static void dk_compress_output(struct drc_kernel *dk)
{
	const float main_linear_gain = dk->main_linear_gain;
	const float envelope_rate = dk->envelope_rate;
	const float scaled_desired_gain = dk->scaled_desired_gain;
	const float compressor_gain = dk->compressor_gain;
	const int div_start = dk->pre_delay_read_index;
	float *ptr_left = &dk->pre_delay_buffers[0][div_start];
	float *ptr_right = &dk->pre_delay_buffers[1][div_start];
	int count = DIVISION_FRAMES / 4;

	int i, j;

	/* Exponential approach to desired gain. */
	if (envelope_rate < 1) {
		/* Attack - reduce gain to desired. */
		float c = compressor_gain - scaled_desired_gain;
		float base = scaled_desired_gain;
		float r = 1 - envelope_rate;
		float x[4] = { c * r, c * r * r, c * r * r * r,
			       c * r * r * r * r };
		float r4 = r * r * r * r;

		i = 0;
		while (1) {
			for (j = 0; j < 4; j++) {
				/* Warp pre-compression gain to smooth out sharp
				 * exponential transition points.
				 */
				float post_warp_compressor_gain =
					warp_sinf(x[j] + base);

				/* Calculate total gain using main gain. */
				float total_gain = main_linear_gain *
						   post_warp_compressor_gain;

				/* Apply final gain. */
				*ptr_left++ *= total_gain;
				*ptr_right++ *= total_gain;
			}

			if (++i == count)
				break;

			for (j = 0; j < 4; j++)
				x[j] = x[j] * r4;
		}

		dk->compressor_gain = x[3] + base;
	} else {
		/* Release - exponentially increase gain to 1.0 */
		float c = compressor_gain;
		float r = envelope_rate;
		float x[4] = { c * r, c * r * r, c * r * r * r,
			       c * r * r * r * r };
		float r4 = r * r * r * r;

		i = 0;
		while (1) {
			for (j = 0; j < 4; j++) {
				/* Warp pre-compression gain to smooth out sharp
				 * exponential transition points.
				 */
				float post_warp_compressor_gain =
					warp_sinf(x[j]);

				/* Calculate total gain using main gain. */
				float total_gain = main_linear_gain *
						   post_warp_compressor_gain;

				/* Apply final gain. */
				*ptr_left++ *= total_gain;
				*ptr_right++ *= total_gain;
			}

			if (++i == count)
				break;

			for (j = 0; j < 4; j++)
				x[j] = min(1.0f, x[j] * r4);
		}

		dk->compressor_gain = x[3];
	}
}
#endif

/* After one complete divison of samples have been received (and one divison of
 * samples have been output), we calculate shaped power average
 * (detector_average) from the input division, update envelope parameters from
 * detector_average, then prepare the next output division by applying the
 * envelope to compress the samples.
 */
static void dk_process_one_division(struct drc_kernel *dk)
{
	dk_update_detector_average(dk);
	dk_update_envelope(dk);
	dk_compress_output(dk);
}

/* Copy the input data to the pre-delay buffer, and copy the output data back to
 * the input buffer */
static void dk_copy_fragment(struct drc_kernel *dk, float *data_channels[],
			     unsigned frame_index, int frames_to_process)
{
	int write_index = dk->pre_delay_write_index;
	int read_index = dk->pre_delay_read_index;
	int j;

	for (j = 0; j < DRC_NUM_CHANNELS; ++j) {
		memcpy(&dk->pre_delay_buffers[j][write_index],
		       &data_channels[j][frame_index],
		       frames_to_process * sizeof(float));
		memcpy(&data_channels[j][frame_index],
		       &dk->pre_delay_buffers[j][read_index],
		       frames_to_process * sizeof(float));
	}

	dk->pre_delay_write_index =
		(write_index + frames_to_process) & MAX_PRE_DELAY_FRAMES_MASK;
	dk->pre_delay_read_index =
		(read_index + frames_to_process) & MAX_PRE_DELAY_FRAMES_MASK;
}

/* Delay the input sample only and don't do other processing. This is used when
 * the kernel is disabled. We want to do this to match the processing delay in
 * kernels of other bands.
 */
static void dk_process_delay_only(struct drc_kernel *dk, float *data_channels[],
				  unsigned count)
{
	int read_index = dk->pre_delay_read_index;
	int write_index = dk->pre_delay_write_index;
	int i = 0;

	while (i < count) {
		int j;
		int small = min(read_index, write_index);
		int large = max(read_index, write_index);
		/* chunk is the minimum of readable samples in contiguous
		 * buffer, writable samples in contiguous buffer, and the
		 * available input samples. */
		int chunk = min(large - small, MAX_PRE_DELAY_FRAMES - large);
		chunk = min(chunk, count - i);
		for (j = 0; j < DRC_NUM_CHANNELS; ++j) {
			memcpy(&dk->pre_delay_buffers[j][write_index],
			       &data_channels[j][i], chunk * sizeof(float));
			memcpy(&data_channels[j][i],
			       &dk->pre_delay_buffers[j][read_index],
			       chunk * sizeof(float));
		}
		read_index = (read_index + chunk) & MAX_PRE_DELAY_FRAMES_MASK;
		write_index = (write_index + chunk) & MAX_PRE_DELAY_FRAMES_MASK;
		i += chunk;
	}

	dk->pre_delay_read_index = read_index;
	dk->pre_delay_write_index = write_index;
}

void dk_process(struct drc_kernel *dk, float *data_channels[], unsigned count)
{
	int i = 0;
	int fragment;

	if (!dk->enabled) {
		dk_process_delay_only(dk, data_channels, count);
		return;
	}

	if (!dk->processed) {
		dk_update_envelope(dk);
		dk_compress_output(dk);
		dk->processed = 1;
	}

	int offset = dk->pre_delay_write_index & DIVISION_FRAMES_MASK;
	while (i < count) {
		fragment = min(DIVISION_FRAMES - offset, count - i);
		dk_copy_fragment(dk, data_channels, i, fragment);
		i += fragment;
		offset = (offset + fragment) & DIVISION_FRAMES_MASK;

		/* Process the input division (32 frames). */
		if (offset == 0)
			dk_process_one_division(dk);
	}
}