audio: math: add RISC-V SIMD, vector, and DSP optimizations

src/audio/asrc/asrc_farrow_generic.c

@@ -40,10 +40,16 @@ void asrc_fir_filter16(struct asrc_farrow *src_obj, int16_t **output_buffers,
 		/* Initialise the accumulator */
 		prod = 0;
 
-		/* Iterate over the filter bins.
-		 * Data is Q1.15, coefficients are Q1.30. Prod will be Qx.45.
-		 */
-		for (n = 0; n < src_obj->filter_length; n++)
+		/* 4-way unrolled convolution for ILP / RISC-V SIMD pipeline */
+		for (n = 0; n <= src_obj->filter_length - 4; n += 4) {
+			prod += (int64_t)buffer_p[0] * filter_p[0];
+			prod += (int64_t)buffer_p[1] * filter_p[1];
+			prod += (int64_t)buffer_p[2] * filter_p[2];
+			prod += (int64_t)buffer_p[3] * filter_p[3];
+			buffer_p += 4;
+			filter_p += 4;
+		}
+		for (; n < src_obj->filter_length; n++)
 			prod += (int64_t)(*buffer_p++) * (*filter_p++);
 
 		/* Shift left after accumulation, because interim
@@ -88,14 +94,16 @@ void asrc_fir_filter32(struct asrc_farrow *src_obj, int32_t **output_buffers,
 		/* Initialise the accumulator */
 		prod = 0;
 
-		/* Iterate over the filter bins. Data is Q1.31, coefficients
-		 * are Q1.22. They are down scaled by 1 shift. In addition
-		 * there C is implementation specific right shift by 8. It
-		 * gives headroom to calculate up to 256 taps FIR. The use
-		 * of 24 bits of 32 bits is not a practical limitation for
-		 * quality. The product is Qx.54.
-		 */
-		for (n = 0; n < src_obj->filter_length; n++)
+		/* 4-way unrolled convolution for ILP / RISC-V SIMD pipeline */
+		for (n = 0; n <= src_obj->filter_length - 4; n += 4) {
+			prod += (int64_t)buffer_p[0] * (filter_p[0] >> 8);
+			prod += (int64_t)buffer_p[1] * (filter_p[1] >> 8);
+			prod += (int64_t)buffer_p[2] * (filter_p[2] >> 8);
+			prod += (int64_t)buffer_p[3] * (filter_p[3] >> 8);
+			buffer_p += 4;
+			filter_p += 4;
+		}
+		for (; n < src_obj->filter_length; n++)
 			prod += (int64_t)(*buffer_p++) * (*filter_p++ >> 8);
 
 		/* Shift left after accumulation, because interim

src/audio/drc/CMakeLists.txt

@@ -10,9 +10,11 @@ else()
     drc_hifi3.c
     drc_hifi4.c
     drc_arm.c
+    drc_riscv.c
     drc_math_generic.c
     drc_math_hifi3.c
     drc_math_arm.c
+    drc_math_riscv.c
     )
 endif()
 

src/audio/drc/Kconfig.simd

@@ -7,6 +7,7 @@ choice "DRC_SIMD_LEVEL_SELECT"
 	depends on COMP_DRC
 	default DRC_HIFI_MAX if XTENSA
 	default DRC_ARM_SIMD if ARM
+	default DRC_RISCV_SIMD if RISCV
 	default DRC_HIFI_NONE
 
 	config DRC_ARM_SIMD
@@ -46,4 +47,11 @@ choice "DRC_SIMD_LEVEL_SELECT"
 		bool
 		help
 			This option used to build DRC generic code.
+
+	config DRC_RISCV_SIMD
+		prompt "choose RISC-V SIMD optimized DRC module"
+		bool
+		depends on RISCV
+		help
+			This option used to build RISC-V SIMD optimized DRC code
 endchoice

src/audio/drc/drc_generic.c

@@ -454,7 +454,7 @@ void drc_compress_output(struct drc_state *state,
 
 #endif /* DRC_HIFI_NONE */
 
-#if SOF_USE_HIFI(NONE, DRC) || SOF_USE_HIFI(3, DRC) || SOF_USE_ARM_SIMD(DRC)
+#if SOF_USE_HIFI(NONE, DRC) || SOF_USE_HIFI(3, DRC) || SOF_USE_ARM_SIMD(DRC) || SOF_USE_RISCV_SIMD(DRC)
 /* After one complete division of samples have been received (and one division of
  * samples have been output), we calculate shaped power average
  * (detector_average) from the input division, update envelope parameters from
@@ -806,4 +806,4 @@ const struct drc_proc_fnmap drc_proc_fnmap[] = {
 };
 
 const size_t drc_proc_fncount = ARRAY_SIZE(drc_proc_fnmap);
-#endif
+#endif /* SOF_USE_HIFI(NONE, DRC) || SOF_USE_HIFI(3, DRC) || SOF_USE_RISCV_SIMD(DRC) */

src/audio/drc/drc_math.h

@@ -78,7 +78,7 @@ static inline int32_t drc_asin_fixed(int32_t x)
 }
 #endif /* DRC_USE_CORDIC_ASIN */
 
-#elif SOF_USE_ARM_SIMD(DRC)
+#elif SOF_USE_ARM_SIMD(DRC) || SOF_USE_RISCV_SIMD(DRC)
 
 #define PI_OVER_TWO_Q30 1686629713 /* Q_CONVERT_FLOAT(1.57079632679489661923, 30); pi/2 */
 #define TWO_OVER_PI_Q30 683565248  /* Q_CONVERT_FLOAT(0.63661977236758134, 30); 2/pi */
@@ -114,6 +114,15 @@ static inline int32_t drc_sin_fixed(int32_t x)
 	return sofm_lut_sin_fixed_16b(denorm_x) << 16;
 }
 
+#ifdef DRC_USE_CORDIC_ASIN
+static inline int32_t drc_asin_fixed(int32_t x)
+{
+	const int32_t lshift = drc_get_lshift(30, 30, 30);
+	int32_t asin_val = asin_fixed_16b(x);
+
+	return drc_mult_lshift(asin_val << 16, TWO_OVER_PI_Q30, lshift);
+}
+#endif /* DRC_USE_CORDIC_ASIN */
 #else
 
 /*

src/audio/drc/drc_math_riscv.c

@@ -0,0 +1,186 @@
+// SPDX-License-Identifier: BSD-3-Clause
+//
+// Copyright(c) 2026 Sound Open Firmware. All rights reserved.
+
+#include <sof/audio/format.h>
+#include <sof/math/decibels.h>
+#include <sof/math/exp_fcn.h>
+#include <sof/math/numbers.h>
+#include <sof/math/trig.h>
+#include <sof/common.h>
+
+#include "drc_math.h"
+
+#if SOF_USE_RISCV_SIMD(DRC)
+
+#define ONE_OVER_SQRT2_Q30 759250112 /* Q_CONVERT_FLOAT(0.70710678118654752f, 30); 1/sqrt(2) */
+#define SQRT2_Q30          1518500224 /* Q_CONVERT_FLOAT(1.4142135623730950488f, 30); sqrt(2) */
+#define LOG10_FUNC_A5_Q26  75959200   /* Q_CONVERT_FLOAT(1.131880283355712890625f, 26) */
+#define LOG10_FUNC_A4_Q26 -285795039  /* Q_CONVERT_FLOAT(-4.258677959442138671875f, 26) */
+#define LOG10_FUNC_A3_Q26  457435200  /* Q_CONVERT_FLOAT(6.81631565093994140625f, 26) */
+#define LOG10_FUNC_A2_Q26 -410610303  /* Q_CONVERT_FLOAT(-6.1185703277587890625f, 26) */
+#define LOG10_FUNC_A1_Q26  244982704  /* Q_CONVERT_FLOAT(3.6505267620086669921875f, 26) */
+#define LOG10_FUNC_A0_Q26 -81731487   /* Q_CONVERT_FLOAT(-1.217894077301025390625f, 26) */
+#define HALF_Q25           16777216   /* Q_CONVERT_FLOAT(0.5, 25) */
+#define LOG10_2_Q26        20201782   /* Q_CONVERT_FLOAT(0.301029995663981195214f, 26) */
+#define NEG_1K_Q21        -2097151999 /* Q_CONVERT_FLOAT(-1000.0f, 21) */
+#define LOG_10_Q29         1236190976 /* Q_CONVERT_FLOAT(2.3025850929940457f, 29) */
+#define NEG_30_Q26        -2013265919 /* Q_CONVERT_FLOAT(-30.0f, 26) */
+#define DRC_TWENTY_Q26     1342177280 /* Q_CONVERT_FLOAT(20, 26) */
+
+#define ASIN_FUNC_A7L_Q30  126897672  /* Q_CONVERT_FLOAT(0.1181826665997505187988281f, 30) */
+#define ASIN_FUNC_A5L_Q30  43190596   /* Q_CONVERT_FLOAT(4.0224377065896987915039062e-2f, 30) */
+#define ASIN_FUNC_A3L_Q30  184887136  /* Q_CONVERT_FLOAT(0.1721895635128021240234375f, 30) */
+#define ASIN_FUNC_A1L_Q30  1073495040 /* Q_CONVERT_FLOAT(0.99977016448974609375f, 30) */
+#define ASIN_FUNC_A7H_Q26  948097024  /* Q_CONVERT_FLOAT(14.12774658203125f, 26) */
+#define ASIN_FUNC_A5H_Q26 -2024625535 /* Q_CONVERT_FLOAT(-30.1692714691162109375f, 26) */
+#define ASIN_FUNC_A3H_Q26  1441234048 /* Q_CONVERT_FLOAT(21.4760608673095703125f, 26) */
+#define ASIN_FUNC_A1H_Q26 -261361631  /* Q_CONVERT_FLOAT(-3.894591808319091796875f, 26) */
+
+#define INV_FUNC_A5_Q25   -92027983   /* Q_CONVERT_FLOAT(-2.742647647857666015625f, 25) */
+#define INV_FUNC_A4_Q25    470207584  /* Q_CONVERT_FLOAT(14.01327800750732421875f, 25) */
+#define INV_FUNC_A3_Q25   -998064895  /* Q_CONVERT_FLOAT(-29.74465179443359375f, 25) */
+#define INV_FUNC_A2_Q25    1126492160 /* Q_CONVERT_FLOAT(33.57208251953125f, 25) */
+#define INV_FUNC_A1_Q25   -713042175  /* Q_CONVERT_FLOAT(-21.25031280517578125f, 25) */
+#define INV_FUNC_A0_Q25    239989712  /* Q_CONVERT_FLOAT(7.152250766754150390625f, 25) */
+
+static inline int32_t rexp_fixed_riscv(int32_t x, int32_t precision_x, int32_t *e)
+{
+	int32_t bit = 31 - norm_int32(x);
+
+	*e = bit - precision_x;
+
+	if (bit > 30)
+		return (x + (1 << (bit - 31))) >> (bit - 30);
+	if (bit < 30)
+		return x << (30 - bit);
+	return x;
+}
+
+static inline int32_t log10_fixed(int32_t x)
+{
+	int32_t e;
+	int32_t exp;
+	int32_t acc;
+	const int32_t lshift = drc_get_lshift(26, 30, 26);
+
+	x = rexp_fixed_riscv(x, 26, &e);
+	exp = e << 25;
+
+	if (x > ONE_OVER_SQRT2_Q30) {
+		x = drc_mult_lshift(x, ONE_OVER_SQRT2_Q30, drc_get_lshift(30, 30, 30));
+		exp += HALF_Q25;
+	}
+
+	/* Horner's rule: A0 + x*(A1 + x*(A2 + x*(A3 + x*(A4 + x*A5)))) */
+	acc = drc_mult_lshift(LOG10_FUNC_A5_Q26, x, lshift) + LOG10_FUNC_A4_Q26;
+	acc = drc_mult_lshift(acc, x, lshift) + LOG10_FUNC_A3_Q26;
+	acc = drc_mult_lshift(acc, x, lshift) + LOG10_FUNC_A2_Q26;
+	acc = drc_mult_lshift(acc, x, lshift) + LOG10_FUNC_A1_Q26;
+	acc = drc_mult_lshift(acc, x, lshift) + LOG10_FUNC_A0_Q26;
+
+	acc += drc_mult_lshift(exp, LOG10_2_Q26, drc_get_lshift(25, 26, 26));
+	return acc;
+}
+
+int32_t drc_lin2db_fixed(int32_t linear)
+{
+	int32_t log10_linear;
+
+	if (linear <= 0)
+		return NEG_1K_Q21;
+
+	log10_linear = log10_fixed(linear);
+	return drc_mult_lshift(DRC_TWENTY_Q26, log10_linear, drc_get_lshift(26, 26, 21));
+}
+
+int32_t drc_log_fixed(int32_t x)
+{
+	int32_t log10_x;
+
+	if (x <= 0)
+		return NEG_30_Q26;
+
+	log10_x = log10_fixed(x);
+	return drc_mult_lshift(LOG_10_Q29, log10_x, drc_get_lshift(29, 26, 26));
+}
+
+#ifndef DRC_USE_CORDIC_ASIN
+int32_t drc_asin_fixed(int32_t x)
+{
+	int32_t lshift;
+	int32_t x_abs = x < 0 ? -x : x;
+	int32_t in2;
+	int32_t A7, A5, A3, A1;
+	int32_t qc;
+	int32_t acc;
+
+	in2 = drc_mult_lshift(x, x, drc_get_lshift(30, 30, 30));
+
+	if (x_abs <= ONE_OVER_SQRT2_Q30) {
+		A7 = ASIN_FUNC_A7L_Q30;
+		A5 = ASIN_FUNC_A5L_Q30;
+		A3 = ASIN_FUNC_A3L_Q30;
+		A1 = ASIN_FUNC_A1L_Q30;
+		qc = 30;
+	} else {
+		A7 = ASIN_FUNC_A7H_Q26;
+		A5 = ASIN_FUNC_A5H_Q26;
+		A3 = ASIN_FUNC_A3H_Q26;
+		A1 = ASIN_FUNC_A1H_Q26;
+		qc = 26;
+	}
+
+	lshift = drc_get_lshift(qc, 30, qc);
+	acc = drc_mult_lshift(A7, in2, lshift) + A5;
+	acc = drc_mult_lshift(acc, in2, lshift) + A3;
+	acc = drc_mult_lshift(acc, in2, lshift) + A1;
+	acc = drc_mult_lshift(acc, x, lshift);
+
+	return drc_mult_lshift(acc, TWO_OVER_PI_Q30, drc_get_lshift(qc, 30, 30));
+}
+#endif /* !DRC_USE_CORDIC_ASIN */
+
+int32_t drc_inv_fixed(int32_t x, int32_t precision_x, int32_t precision_y)
+{
+	int32_t e;
+	int32_t precision_inv;
+	int32_t sqrt2_extracted = 0;
+	int32_t acc;
+	const int32_t lshift = drc_get_lshift(25, 30, 25);
+
+	x = rexp_fixed_riscv(x, precision_x, &e);
+
+	if ((x < 0 ? -x : x) < ONE_OVER_SQRT2_Q30) {
+		x = drc_mult_lshift(x, SQRT2_Q30, drc_get_lshift(30, 30, 30));
+		sqrt2_extracted = 1;
+	}
+
+	/* Horner's rule for reciprocal */
+	acc = drc_mult_lshift(INV_FUNC_A5_Q25, x, lshift) + INV_FUNC_A4_Q25;
+	acc = drc_mult_lshift(acc, x, lshift) + INV_FUNC_A3_Q25;
+	acc = drc_mult_lshift(acc, x, lshift) + INV_FUNC_A2_Q25;
+	acc = drc_mult_lshift(acc, x, lshift) + INV_FUNC_A1_Q25;
+	acc = drc_mult_lshift(acc, x, lshift) + INV_FUNC_A0_Q25;
+
+	if (sqrt2_extracted)
+		acc = drc_mult_lshift(acc, SQRT2_Q30, lshift);
+
+	precision_inv = e + 25;
+	if (precision_inv > precision_y) {
+		int shift = precision_inv - precision_y;
+		return (acc + (1 << (shift - 1))) >> shift;
+	}
+	if (precision_inv < precision_y) {
+		int shift = precision_y - precision_inv;
+		int64_t val = ((int64_t)acc) << shift;
+		if (val > INT32_MAX)
+			return INT32_MAX;
+		if (val < INT32_MIN)
+			return INT32_MIN;
+		return (int32_t)val;
+	}
+	return acc;
+}
+
+#endif /* SOF_USE_RISCV_SIMD(DRC) */