tests: lib: cmsis_dsp: matrix: Add Binary F16 tests for 1.9.0

This commit adds the matrix Binary F16 test patterns and
implementations for the CMSIS-DSP 1.9.0.

Signed-off-by: Stephanos Ioannidis <root@stephanos.io>

tests/lib/cmsis_dsp/matrix/CMakeLists.txt

@@ -43,6 +43,11 @@ target_sources_ifdef(
   app PRIVATE src/binary_q31.c
   )
 
+target_sources_ifdef(
+  CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_F16
+  app PRIVATE src/binary_f16.c
+  )
+
 target_sources_ifdef(
   CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_F32
   app PRIVATE src/binary_f32.c

tests/lib/cmsis_dsp/matrix/Kconfig

@@ -22,6 +22,10 @@ config CMSIS_DSP_TEST_MATRIX_BINARY_Q15
 config CMSIS_DSP_TEST_MATRIX_BINARY_Q31
 	bool "Test: Matrix Binary Q31"
 
+config CMSIS_DSP_TEST_MATRIX_BINARY_F16
+	bool "Test: Matrix Binary F16"
+	depends on CMSIS_DSP_FLOAT16
+
 config CMSIS_DSP_TEST_MATRIX_BINARY_F32
 	bool "Test: Matrix Binary F32"
 

tests/lib/cmsis_dsp/matrix/prj.conf

@@ -11,4 +11,5 @@ CONFIG_CMSIS_DSP_TEST_MATRIX_UNARY_F64=y
 CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_Q7=y
 CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_Q15=y
 CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_Q31=y
+CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_F16=y
 CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_F32=y

tests/lib/cmsis_dsp/matrix/src/binary_f16.c

@@ -0,0 +1,216 @@
+/*
+ * Copyright (c) 2021 Stephanos Ioannidis <root@stephanos.io>
+ * Copyright (C) 2010-2021 ARM Limited or its affiliates. All rights reserved.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include <ztest.h>
+#include <zephyr.h>
+#include <stdlib.h>
+#include <arm_math_f16.h>
+#include "../../common/test_common.h"
+
+#include "binary_f16.pat"
+
+#define SNR_ERROR_THRESH	((float32_t)60)
+#define REL_ERROR_THRESH	(2.0e-3)
+#define ABS_ERROR_THRESH	(2.0e-3)
+
+#define NUM_MATRICES		(ARRAY_SIZE(in_dims) / 3)
+#define MAX_MATRIX_DIM		(40)
+
+#define OP2_MULT		(0)
+#define OP2C_CMPLX_MULT		(0)
+
+static void test_op2(int op, const uint16_t *input1, const uint16_t *input2,
+	const uint16_t *ref, size_t length)
+{
+	size_t index;
+	uint16_t *dims = (uint16_t *)in_dims;
+	float16_t *tmp1, *tmp2, *output;
+	uint16_t rows, internal, columns;
+	arm_status status;
+
+	arm_matrix_instance_f16 mat_in1;
+	arm_matrix_instance_f16 mat_in2;
+	arm_matrix_instance_f16 mat_out;
+
+	/* Allocate buffers */
+	tmp1 = malloc(MAX_MATRIX_DIM * MAX_MATRIX_DIM * sizeof(float16_t));
+	zassert_not_null(tmp1, ASSERT_MSG_BUFFER_ALLOC_FAILED);
+
+	tmp2 = malloc(MAX_MATRIX_DIM * MAX_MATRIX_DIM * sizeof(float16_t));
+	zassert_not_null(tmp2, ASSERT_MSG_BUFFER_ALLOC_FAILED);
+
+	output = malloc(length * sizeof(float16_t));
+	zassert_not_null(output, ASSERT_MSG_BUFFER_ALLOC_FAILED);
+
+	/* Initialise contexts */
+	mat_in1.pData = tmp1;
+	mat_in2.pData = tmp2;
+	mat_out.pData = output;
+
+	/* Iterate matrices */
+	for (index = 0; index < NUM_MATRICES; index++) {
+		rows = *dims++;
+		internal = *dims++;
+		columns = *dims++;
+
+		/* Initialise matrix dimensions */
+		mat_in1.numRows = rows;
+		mat_in1.numCols = internal;
+
+		mat_in2.numRows = internal;
+		mat_in2.numCols = columns;
+
+		mat_out.numRows = rows;
+		mat_out.numCols = columns;
+
+		/* Load matrix data */
+		memcpy(mat_in1.pData, input1,
+		       rows * internal * sizeof(float16_t));
+
+		memcpy(mat_in2.pData, input2,
+		       internal * columns * sizeof(float16_t));
+
+		/* Run test function */
+		switch (op) {
+		case OP2_MULT:
+			status = arm_mat_mult_f16(&mat_in1, &mat_in2,
+						  &mat_out);
+			break;
+		default:
+			zassert_unreachable("invalid operation");
+		}
+
+		/* Validate status */
+		zassert_equal(status, ARM_MATH_SUCCESS,
+			      ASSERT_MSG_INCORRECT_COMP_RESULT);
+
+		/* Increment output pointer */
+		mat_out.pData += (rows * columns);
+	}
+
+	/* Validate output */
+	zassert_true(
+		test_snr_error_f16(length, output, (float16_t *)ref,
+			SNR_ERROR_THRESH),
+		ASSERT_MSG_SNR_LIMIT_EXCEED);
+
+	zassert_true(
+		test_close_error_f16(length, output, (float16_t *)ref,
+			ABS_ERROR_THRESH, REL_ERROR_THRESH),
+		ASSERT_MSG_ERROR_LIMIT_EXCEED);
+
+	/* Free buffers */
+	free(tmp1);
+	free(tmp2);
+	free(output);
+}
+
+DEFINE_TEST_VARIANT5(
+	op2, arm_mat_mult_f16, OP2_MULT,
+	in_mult1, in_mult2, ref_mult,
+	ARRAY_SIZE(ref_mult));
+
+static void test_op2c(int op, const uint16_t *input1, const uint16_t *input2,
+	const uint16_t *ref, size_t length)
+{
+	size_t index;
+	uint16_t *dims = (uint16_t *)in_dims;
+	float16_t *tmp1, *tmp2, *output;
+	uint16_t rows, internal, columns;
+	arm_status status;
+
+	arm_matrix_instance_f16 mat_in1;
+	arm_matrix_instance_f16 mat_in2;
+	arm_matrix_instance_f16 mat_out;
+
+	/* Allocate buffers */
+	tmp1 = malloc(2 * MAX_MATRIX_DIM * MAX_MATRIX_DIM * sizeof(float16_t));
+	zassert_not_null(tmp1, ASSERT_MSG_BUFFER_ALLOC_FAILED);
+
+	tmp2 = malloc(2 * MAX_MATRIX_DIM * MAX_MATRIX_DIM * sizeof(float16_t));
+	zassert_not_null(tmp2, ASSERT_MSG_BUFFER_ALLOC_FAILED);
+
+	output = malloc(2 * length * sizeof(float16_t));
+	zassert_not_null(output, ASSERT_MSG_BUFFER_ALLOC_FAILED);
+
+	/* Initialise contexts */
+	mat_in1.pData = tmp1;
+	mat_in2.pData = tmp2;
+	mat_out.pData = output;
+
+	/* Iterate matrices */
+	for (index = 0; index < NUM_MATRICES; index++) {
+		rows = *dims++;
+		internal = *dims++;
+		columns = *dims++;
+
+		/* Initialise matrix dimensions */
+		mat_in1.numRows = rows;
+		mat_in1.numCols = internal;
+
+		mat_in2.numRows = internal;
+		mat_in2.numCols = columns;
+
+		mat_out.numRows = rows;
+		mat_out.numCols = columns;
+
+		/* Load matrix data */
+		memcpy(mat_in1.pData, input1,
+		       2 * rows * internal * sizeof(float16_t));
+
+		memcpy(mat_in2.pData, input2,
+		       2 * internal * columns * sizeof(float16_t));
+
+		/* Run test function */
+		switch (op) {
+		case OP2C_CMPLX_MULT:
+			status = arm_mat_cmplx_mult_f16(&mat_in1, &mat_in2,
+							&mat_out);
+			break;
+		default:
+			zassert_unreachable("invalid operation");
+		}
+
+		/* Validate status */
+		zassert_equal(status, ARM_MATH_SUCCESS,
+			      ASSERT_MSG_INCORRECT_COMP_RESULT);
+
+		/* Increment output pointer */
+		mat_out.pData += (2 * rows * columns);
+	}
+
+	/* Validate output */
+	zassert_true(
+		test_snr_error_f16(2 * length, output, (float16_t *)ref,
+			SNR_ERROR_THRESH),
+		ASSERT_MSG_SNR_LIMIT_EXCEED);
+
+	zassert_true(
+		test_close_error_f16(length, output, (float16_t *)ref,
+			ABS_ERROR_THRESH, REL_ERROR_THRESH),
+		ASSERT_MSG_ERROR_LIMIT_EXCEED);
+
+	/* Free buffers */
+	free(tmp1);
+	free(tmp2);
+	free(output);
+}
+
+DEFINE_TEST_VARIANT5(
+	op2c, arm_mat_cmplx_mult_f16, OP2C_CMPLX_MULT,
+	in_cmplx_mult1, in_cmplx_mult2, ref_cmplx_mult,
+	ARRAY_SIZE(ref_cmplx_mult) / 2);
+
+void test_matrix_binary_f16(void)
+{
+	ztest_test_suite(matrix_binary_f16,
+		ztest_unit_test(test_op2_arm_mat_mult_f16),
+		ztest_unit_test(test_op2c_arm_mat_cmplx_mult_f16)
+		);
+
+	ztest_run_test_suite(matrix_binary_f16);
+}

tests/lib/cmsis_dsp/matrix/src/main.c

@@ -15,6 +15,7 @@ extern void test_matrix_unary_f64(void);
 extern void test_matrix_binary_q7(void);
 extern void test_matrix_binary_q15(void);
 extern void test_matrix_binary_q31(void);
+extern void test_matrix_binary_f16(void);
 extern void test_matrix_binary_f32(void);
 
 void test_main(void)
@@ -41,6 +42,9 @@ void test_main(void)
 #ifdef CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_Q31
 	test_matrix_binary_q31();
 #endif
+#ifdef CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_F16
+	test_matrix_binary_f16();
+#endif
 #ifdef CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_F32
 	test_matrix_binary_f32();
 #endif

tests/lib/cmsis_dsp/matrix/testcase.yaml

@@ -177,6 +177,32 @@ tests:
     extra_configs:
       - CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_Q31=y
       - CONFIG_FPU=y
+  libraries.cmsis_dsp.matrix.binary_f16:
+    filter: ((CONFIG_CPU_CORTEX_R or CONFIG_CPU_CORTEX_M) and TOOLCHAIN_HAS_NEWLIB == 1) or CONFIG_ARCH_POSIX
+    integration_platforms:
+      - frdm_k64f
+      - sam_e70_xplained
+      - mps2_an521
+      - native_posix
+    tags: cmsis_dsp
+    platform_exclude: frdm_kw41z
+    min_flash: 128
+    min_ram: 128
+    extra_args: CONF_FILE=prj_base.conf
+    extra_configs:
+      - CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_F16=y
+  libraries.cmsis_dsp.matrix.binary_f16.fpu:
+    filter: ((CONFIG_CPU_CORTEX_R or CONFIG_CPU_CORTEX_M) and CONFIG_CPU_HAS_FPU and TOOLCHAIN_HAS_NEWLIB == 1) or CONFIG_ARCH_POSIX
+    integration_platforms:
+      - mps2_an521_remote
+    tags: cmsis_dsp fpu
+    platform_exclude: frdm_kw41z
+    min_flash: 128
+    min_ram: 128
+    extra_args: CONF_FILE=prj_base.conf
+    extra_configs:
+      - CONFIG_CMSIS_DSP_TEST_MATRIX_BINARY_F16=y
+      - CONFIG_FPU=y
   libraries.cmsis_dsp.matrix.binary_f32:
     filter: ((CONFIG_CPU_CORTEX_R or CONFIG_CPU_CORTEX_M) and TOOLCHAIN_HAS_NEWLIB == 1) or CONFIG_ARCH_POSIX
     integration_platforms: