ltpf: add Arm Helium (MVE) SIMD backend

src/ltpf.c

@@ -19,6 +19,7 @@
 #include "ltpf.h"
 #include "tables.h"
 
+#include "ltpf_mve.h"
 #include "ltpf_neon.h"
 #include "ltpf_arm.h"
 

src/ltpf_mve.h

@@ -0,0 +1,96 @@
+/******************************************************************************
+ *
+ *  Copyright 2026 Google LLC
+ *
+ *  Licensed under the Apache License, Version 2.0 (the "License");
+ *  you may not use this file except in compliance with the License.
+ *  You may obtain a copy of the License at:
+ *
+ *  http://www.apache.org/licenses/LICENSE-2.0
+ *
+ *  Unless required by applicable law or agreed to in writing, software
+ *  distributed under the License is distributed on an "AS IS" BASIS,
+ *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ *  See the License for the specific language governing permissions and
+ *  limitations under the License.
+ *
+ ******************************************************************************/
+
+/*
+ * Arm Helium (M-Profile Vector Extension, MVE) backend for the LTPF
+ * pitch-detection hotspot, targeting Armv8.1-M cores (e.g. Cortex-M55).
+ *
+ * `dot()` / `correlate()` accumulate int16 x int16 products into a 64-bit
+ * integer and apply the rounding `(v + 32) >> 6`. MVE `vmlaldavaq_s16`
+ * performs an 8-lane int16 x int16 -> int64 multiply-accumulate across the
+ * vector; the integer sum and the subsequent rounding are therefore
+ * identical to the scalar reference, i.e. the result is strictly bit-exact.
+ *
+ * This mirrors the NEON (`ltpf_neon.h`) and Arm DSP (`ltpf_arm.h`) backends
+ * and hooks the same `dot` / `correlate` override points. Include this file
+ * before the NEON/DSP backends in `ltpf.c` so its definitions take
+ * precedence on MVE targets.
+ */
+
+#if (__ARM_FEATURE_MVE & 1) && !defined(TEST_ARM) && !defined(TEST_NEON) \
+        || defined(TEST_MVE)
+
+#ifndef TEST_MVE
+#include <arm_mve.h>
+#endif /* TEST_MVE */
+
+
+/**
+ * Return dot product of 2 vectors
+ * a, b, n         The 2 vectors of size `n` (> 0 and <= 128)
+ * return          round( sum( a[i] * b[i] ) / 64 ), as float
+ *
+ * The size `n` of vectors is a multiple of 16 (so also of 8).
+ */
+#ifndef dot
+
+LC3_HOT static inline float mve_dot(const int16_t *a, const int16_t *b, int n)
+{
+    int64_t v = 0;
+
+    for (int i = 0; i < (n >> 3); i++) {
+        int16x8_t va = vldrhq_s16(a); a += 8;
+        int16x8_t vb = vldrhq_s16(b); b += 8;
+        v = vmlaldavaq_s16(v, va, vb);
+    }
+
+    int32_t v32 = (v + (1 << 5)) >> 6;
+    return (float)v32;
+}
+
+#ifndef TEST_MVE
+#define dot mve_dot
+#endif
+
+#endif /* dot */
+
+/**
+ * Return vector of correlations
+ * a, b, n         The 2 vectors of size `n` (> 0 and <= 128)
+ * y, nc           Output the correlation vector of size `nc`
+ *
+ * Sliding dot product: for each output the `b` window is shifted by one
+ * sample. Each lag reuses the 8-lane MVE `dot`, matching the scalar
+ * reference exactly.
+ */
+#ifndef correlate
+
+LC3_HOT static void mve_correlate(
+    const int16_t *a, const int16_t *b, int n, float *y, int nc)
+{
+    for (const float *ye = y + nc; y < ye; )
+        *(y++) = mve_dot(a, b--, n);
+}
+
+#ifndef TEST_MVE
+#define correlate mve_correlate
+#endif
+
+#endif /* correlate */
+
+#endif /* __ARM_FEATURE_MVE & 1 */

test/makefile.mk

@@ -26,5 +26,6 @@ test-clean:
 
 -include $(TEST_DIR)/arm/makefile.mk
 -include $(TEST_DIR)/neon/makefile.mk
+-include $(TEST_DIR)/mve/makefile.mk
 
 clean-all: test-clean

test/mve/ltpf_mve.c

@@ -0,0 +1,86 @@
+/******************************************************************************
+ *
+ *  Copyright 2026 Google LLC
+ *
+ *  Licensed under the Apache License, Version 2.0 (the "License");
+ *  you may not use this file except in compliance with the License.
+ *  You may obtain a copy of the License at:
+ *
+ *  http://www.apache.org/licenses/LICENSE-2.0
+ *
+ *  Unless required by applicable law or agreed to in writing, software
+ *  distributed under the License is distributed on an "AS IS" BASIS,
+ *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ *  See the License for the specific language governing permissions and
+ *  limitations under the License.
+ *
+ ******************************************************************************/
+
+#include "mve.h"
+
+#include <stdio.h>
+#include <stdint.h>
+#include <stdlib.h>
+
+/* -------------------------------------------------------------------------- */
+
+#define TEST_MVE
+#include <ltpf.c>
+
+void lc3_put_bits_generic(lc3_bits_t *a, unsigned b, int c)
+{ (void)a, (void)b, (void)c; }
+
+unsigned lc3_get_bits_generic(struct lc3_bits *a, int b)
+{ return (void)a, (void)b, 0; }
+
+/* -------------------------------------------------------------------------- */
+
+static int check_dot()
+{
+    int16_t x[200];
+    for (int i = 0; i < 200; i++)
+        x[i] = rand() & 0xffff;
+
+    float y = dot(x, x+3, 128);
+    float y_mve = mve_dot(x, x+3, 128);
+    if (y != y_mve)
+        return -1;
+
+    return 0;
+}
+
+static int check_correlate()
+{
+    int16_t alignas(4) a[500], b[500];
+    float y[100], y_mve[100];
+
+    for (int i = 0; i < 500; i++) {
+        a[i] = rand() & 0xffff;
+        b[i] = rand() & 0xffff;
+    }
+
+    correlate(a, b+200, 128, y, 100);
+    mve_correlate(a, b+200, 128, y_mve, 100);
+    if (memcmp(y, y_mve, 100 * sizeof(*y)) != 0)
+        return -1;
+
+    correlate(a, b+199, 128, y, 99);
+    mve_correlate(a, b+199, 128, y_mve, 99);
+    if (memcmp(y, y_mve, 99 * sizeof(*y)) != 0)
+        return -1;
+
+    return 0;
+}
+
+int check_ltpf(void)
+{
+    int ret;
+
+    if ((ret = check_dot()) < 0)
+        return ret;
+
+    if ((ret = check_correlate()) < 0)
+        return ret;
+
+    return 0;
+}

test/mve/makefile.mk

@@ -0,0 +1,31 @@
+#
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+test_mve_src += \
+    $(TEST_DIR)/mve/test_mve.c \
+    $(TEST_DIR)/mve/ltpf_mve.c \
+    $(SRC_DIR)/tables.c
+
+test_mve_include += $(SRC_DIR)
+test_mve_ldlibs += m
+
+$(eval $(call add-bin,test_mve))
+
+test_mve: $(test_mve_bin)
+	@echo "  RUN     $(notdir $<)"
+	$(V)$<
+
+test: test_mve

test/mve/mve.h

@@ -0,0 +1,61 @@
+/******************************************************************************
+ *
+ *  Copyright 2026 Google LLC
+ *
+ *  Licensed under the Apache License, Version 2.0 (the "License");
+ *  you may not use this file except in compliance with the License.
+ *  You may obtain a copy of the License at:
+ *
+ *  http://www.apache.org/licenses/LICENSE-2.0
+ *
+ *  Unless required by applicable law or agreed to in writing, software
+ *  distributed under the License is distributed on an "AS IS" BASIS,
+ *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ *  See the License for the specific language governing permissions and
+ *  limitations under the License.
+ *
+ ******************************************************************************/
+
+#if (__ARM_FEATURE_MVE & 1)
+
+#include <arm_mve.h>
+
+#else
+
+#include <stdint.h>
+
+
+/* ----------------------------------------------------------------------------
+ *  Integer
+ * -------------------------------------------------------------------------- */
+
+typedef struct { int16_t e[8]; } int16x8_t;
+
+
+/**
+ * Load
+ */
+
+__attribute__((unused))
+static int16x8_t vldrhq_s16(const int16_t *p)
+{
+    return (int16x8_t){ { p[0], p[1], p[2], p[3],
+                          p[4], p[5], p[6], p[7] } };
+}
+
+
+/**
+ * Multiply-accumulate across vector, into a 64-bit accumulator
+ */
+
+__attribute__((unused))
+static int64_t vmlaldavaq_s16(int64_t a, int16x8_t b, int16x8_t c)
+{
+    for (int i = 0; i < 8; i++)
+        a += (int32_t)b.e[i] * c.e[i];
+
+    return a;
+}
+
+
+#endif /* __ARM_FEATURE_MVE */

test/mve/test_mve.c

@@ -0,0 +1,32 @@
+/******************************************************************************
+ *
+ *  Copyright 2026 Google LLC
+ *
+ *  Licensed under the Apache License, Version 2.0 (the "License");
+ *  you may not use this file except in compliance with the License.
+ *  You may obtain a copy of the License at:
+ *
+ *  http://www.apache.org/licenses/LICENSE-2.0
+ *
+ *  Unless required by applicable law or agreed to in writing, software
+ *  distributed under the License is distributed on an "AS IS" BASIS,
+ *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ *  See the License for the specific language governing permissions and
+ *  limitations under the License.
+ *
+ ******************************************************************************/
+
+#include <stdio.h>
+
+int check_ltpf(void);
+
+int main()
+{
+    int r, ret = 0;
+
+    printf("Checking LTPF MVE... "); fflush(stdout);
+    printf("%s\n", (r = check_ltpf()) == 0 ? "OK" : "Failed");
+    ret = ret || r;
+
+    return ret;
+}