add some code

managed_components/espressif__esp-dsp/modules/math/sub/fixed/dsps_sub_s16_ae32.S

@@ -0,0 +1,74 @@
+/*
+ * SPDX-FileCopyrightText: 2024 Espressif Systems (Shanghai) CO LTD
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "dsps_sub_platform.h"
+#if (dsps_sub_s16_ae32_enabled == 1)
+
+    .text
+    .align  4
+    .global dsps_sub_s16_ae32
+    .type   dsps_sub_s16_ae32,@function
+// The function implements the following C code:
+// esp_err_t dsps_sub_s16_ansi(const int16_t *input1, const int16_t *input2, int16_t *output, int len, int step1, int step2, int step_out, int shift)
+// {
+//     for (int i = 0 ; i < len ; i++) {
+//         int32_t acc = (int32_t)input1[i * step1] + (int32_t)input2[i * step2];
+//         output[i * step_out] = acc >> shift;
+//     }
+//     return ESP_OK;
+// }
+dsps_sub_s16_ae32: 
+// input1   - a2
+// input2   - a3
+// output   - a4
+// len      - a5
+// step_in1 - a6
+// step_in2 - a7
+// step_out - stack (a10)
+// shift    - stack (a9)
+
+    entry	a1, 16
+
+
+    // l32i.n	a10, a1, 16
+    // s16i	a10, a4, 0
+    // l32i.n	a10, a1, 20
+    // s16i	a10, a4, 2       
+
+    l32i.n	a10, a1, 16     // Load step_out to the a10 register
+    l32i.n	a9, a1,  20     // Load shift to the a9 register
+    ssr     a9              // sar = a9
+
+    slli 	a6, a6, 1  	// a6 - step_in<<1
+    slli 	a7, a7, 1  	// a7 - step_in<<1
+    slli 	a10, a10, 1 // a8 - step_out<<1
+
+    // s16i	a10, a4, 0
+    // s16i	a6, a4, 2
+    // s16i	a7, a4, 4
+    // s16i	a5, a4, 6
+
+    l16si   a11, a2, 0
+    l16si   a8,  a3, 0
+    add     a8, a11, a8
+    srl     a9, a8          // a8 = a8>>sar    
+
+    loopnez a5, .loop_end_sub_s16_ae32
+        add.n   a2, a2, a6      // input1+=step_in1;
+        add.n   a3, a3, a7      // input2+=step_in2;
+
+        l16si   a11, a2, 0
+        l16si   a8,  a3, 0
+        s16i	a9,  a4, 0      // store result to the putput
+        sub     a8, a11, a8
+    	srl     a9, a8          // a8 = a8>>sar    
+
+        add.n   a4, a4, a10     // output+=step_out;
+.loop_end_sub_s16_ae32:
+    movi.n	a2, 0 // return status ESP_OK
+    retw.n
+
+#endif // dsps_sub_s16_ae32_enabled

managed_components/espressif__esp-dsp/modules/math/sub/fixed/dsps_sub_s16_aes3.S

@@ -0,0 +1,104 @@
+/*
+ * SPDX-FileCopyrightText: 2024 Espressif Systems (Shanghai) CO LTD
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "dsps_sub_platform.h"
+#if (dsps_sub_s16_aes3_enabled == 1)
+
+    .text
+    .align  4
+    .global dsps_sub_s16_aes3
+    .type   dsps_sub_s16_aes3,@function
+// The function implements the following C code:
+// esp_err_t dsps_sub_s16_ansi(const int16_t *input1, const int16_t *input2, int16_t *output, int len, int step1, int step2, int step_out, int shift)
+// {
+//     for (int i = 0 ; i < len ; i++) {
+//         int32_t acc = (int32_t)input1[i * step1] + (int32_t)input2[i * step2];
+//         output[i * step_out] = acc >> shift;
+//     }
+//     return ESP_OK;
+// }
+dsps_sub_s16_aes3: 
+// input1   - a2
+// input2   - a3
+// output   - a4
+// len      - a5
+// step_in1 - a6
+// step_in2 - a7
+// step_out - stack (a10)
+// shift    - stack (a9)
+
+    entry	a1, 16
+
+    l32i.n	a10, a1, 16     // Load step_out to the a10 register
+    l32i.n	a9, a1,  20     // Load shift to the a9 register
+    ssr     a9              // sar = a9
+
+    // Check if any of steps is not 0
+    addi    a15, a6, -1
+    bnez    a15, .sub_s16_ae32_mode // Branch if step !=0
+    addi    a15, a7, -1
+    bnez    a15, .sub_s16_ae32_mode // Branch if step !=0
+    addi    a15, a10, -1
+    bnez    a15,.sub_s16_ae32_mode  // Branch if step !=0
+
+    // Check addresses
+    movi        a15, 0xF                                              // modulo 16 mask
+    bany        a2, a15, .sub_s16_ae32_mode                          // jump if != 0
+    bany        a3, a15, .sub_s16_ae32_mode                          // jump if != 0
+
+    // Check length (should be divided to 8)
+    movi        a15, 0x7                                              // modulo 8 mask
+    bany        a5, a15, .sub_s16_ae32_mode                          // jump if != 0
+
+    // Process main function for S3
+    
+    slli 	a6, a6, 1  	// a6 - step_in<<1
+    slli 	a7, a7, 1  	// a7 - step_in<<1
+    slli 	a10, a10, 1 // a8 - step_out<<1
+	wsr.sar	a9          // load sar register
+
+    // Preload q1 from a3
+    //ee.vld.128.ip     q1, a3, 16
+    srli    a5, a5, 3
+    ee.vld.128.ip     q0, a2, 16
+    loopnez a5, .loop_end_sub_s16_aes3_main
+        ee.vld.128.ip     q1, a3, 16
+        ee.vsubs.s16.ld.incp q0, a2, q4, q0, q1
+        ee.vst.128.ip     q4, a4, 16
+.loop_end_sub_s16_aes3_main:
+
+    // Exit for Esp32s3 mode
+    movi.n	a2, 0 // return status ESP_OK
+    retw.n
+
+
+.sub_s16_ae32_mode:
+    slli 	a6, a6, 1  	// a6 - step_in<<1
+    slli 	a7, a7, 1  	// a7 - step_in<<1
+    slli 	a10, a10, 1 // a8 - step_out<<1
+
+    l16si   a11, a2, 0
+    l16si   a8,  a3, 0
+    sub     a8, a11, a8
+    srl     a9, a8          // a8 = a8>>sar    
+
+    loopnez a5, .loop_end_sub_s16_aes3
+        add.n   a2, a2, a6      // input1+=step_in1;
+        add.n   a3, a3, a7      // input2+=step_in2;
+
+        l16si   a11, a2, 0
+        l16si   a8,  a3, 0
+        s16i	a9,  a4, 0      // store result to the putput
+        sub     a8, a11, a8
+    	srl     a9, a8          // a8 = a8>>sar    
+
+        add.n   a4, a4, a10     // output+=step_out;
+.loop_end_sub_s16_aes3:
+    // Exit for Esp32 mode
+    movi.n	a2, 0 // return status ESP_OK
+    retw.n
+
+#endif // dsps_sub_s16_aes3_enabled

managed_components/espressif__esp-dsp/modules/math/sub/fixed/dsps_sub_s16_ansi.c

@@ -0,0 +1,26 @@
+/*
+ * SPDX-FileCopyrightText: 2024 Espressif Systems (Shanghai) CO LTD
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "dsps_add.h"
+
+esp_err_t dsps_sub_s16_ansi(const int16_t *input1, const int16_t *input2, int16_t *output, int len, int step1, int step2, int step_out, int shift)
+{
+    if (NULL == input1) {
+        return ESP_ERR_DSP_PARAM_OUTOFRANGE;
+    }
+    if (NULL == input2) {
+        return ESP_ERR_DSP_PARAM_OUTOFRANGE;
+    }
+    if (NULL == output) {
+        return ESP_ERR_DSP_PARAM_OUTOFRANGE;
+    }
+
+    for (int i = 0 ; i < len ; i++) {
+        int32_t acc = (int32_t)input1[i * step1] - (int32_t)input2[i * step2];
+        output[i * step_out] = acc >> shift;
+    }
+    return ESP_OK;
+}

managed_components/espressif__esp-dsp/modules/math/sub/fixed/dsps_sub_s8_aes3.S

@@ -0,0 +1,97 @@
+/*
+ * SPDX-FileCopyrightText: 2024 Espressif Systems (Shanghai) CO LTD
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "dsps_sub_platform.h"
+#if (dsps_sub_s16_aes3_enabled == 1)
+
+    .text
+    .align  4
+    .global dsps_sub_s8_aes3
+    .type   dsps_sub_s8_aes3,@function
+// The function implements the following C code:
+// esp_err_t dsps_sub_s8_ansi(const int16_t *input1, const int16_t *input2, int16_t *output, int len, int step1, int step2, int step_out, int shift)
+// {
+//     for (int i = 0 ; i < len ; i++) {
+//         int32_t acc = (int32_t)input1[i * step1] + (int32_t)input2[i * step2];
+//         output[i * step_out] = acc >> shift;
+//     }
+//     return ESP_OK;
+// }
+dsps_sub_s8_aes3: 
+// input1   - a2
+// input2   - a3
+// output   - a4
+// len      - a5
+// step_in1 - a6
+// step_in2 - a7
+// step_out - stack (a10)
+// shift    - stack (a9)
+
+    entry	a1, 16
+
+    l32i.n	a10, a1, 16     // Load step_out to the a10 register
+    l32i.n	a9, a1,  20     // Load shift to the a9 register
+    ssr     a9              // sar = a9
+
+    // Check if any of steps is not 0
+    addi    a15, a6, -1
+    bnez    a15, .sub_s8_ae32_mode // Branch if step !=0
+    addi    a15, a7, -1
+    bnez    a15, .sub_s8_ae32_mode // Branch if step !=0
+    addi    a15, a10, -1
+    bnez    a15,.sub_s8_ae32_mode  // Branch if step !=0
+
+    // Check addresses
+    movi        a15, 0xF                                              // modulo 16 mask
+    bany        a2, a15, .sub_s8_ae32_mode                          // jump if != 0
+    bany        a3, a15, .sub_s8_ae32_mode                          // jump if != 0
+
+    // Check length (should be divided to 8)
+    movi        a15, 0xf                                              // modulo 8 mask
+    bany        a5, a15, .sub_s8_ae32_mode                          // jump if != 0
+
+    // Process main function for S3
+    
+	wsr.sar	a9          // load sar register
+
+    // Preload q1 from a3
+    //ee.vld.128.ip     q1, a3, 16
+    srli    a5, a5, 4
+    ee.vld.128.ip     q0, a2, 16
+    loopnez a5, .loop_end_sub_s8_aes3_main
+        ee.vld.128.ip     q1, a3, 16
+        ee.vsubs.s8.ld.incp q0, a2, q4, q0, q1
+        ee.vst.128.ip     q4, a4, 16
+.loop_end_sub_s8_aes3_main:
+
+    // Exit for Esp32s3 mode
+    movi.n	a2, 0 // return status ESP_OK
+    retw.n
+
+
+.sub_s8_ae32_mode:
+    l8ui    a11, a2, 0
+    l8ui    a8,  a3, 0
+    sub     a8, a11, a8
+    srl     a9, a8          // a8 = a8>>sar    
+
+    loopnez a5, .loop_end_sub_s8_aes3
+        add.n   a2, a2, a6      // input1+=step_in1;
+        add.n   a3, a3, a7      // input2+=step_in2;
+
+        l8ui   a11, a2, 0
+        l8ui   a8,  a3, 0
+        s8i	a9,  a4, 0      // store result to the putput
+        sub     a8, a11, a8
+    	srl     a9, a8          // a8 = a8>>sar    
+
+        add.n   a4, a4, a10     // output+=step_out;
+.loop_end_sub_s8_aes3:
+    // Exit for Esp32 mode
+    movi.n	a2, 0 // return status ESP_OK
+    retw.n
+
+#endif // dsps_sub_s8_aes3_enabled

managed_components/espressif__esp-dsp/modules/math/sub/fixed/dsps_sub_s8_ansi.c

@@ -0,0 +1,26 @@
+/*
+ * SPDX-FileCopyrightText: 2024 Espressif Systems (Shanghai) CO LTD
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "dsps_sub.h"
+
+esp_err_t dsps_sub_s8_ansi(const int8_t *input1, const int8_t *input2, int8_t *output, int len, int step1, int step2, int step_out, int shift)
+{
+    if (NULL == input1) {
+        return ESP_ERR_DSP_PARAM_OUTOFRANGE;
+    }
+    if (NULL == input2) {
+        return ESP_ERR_DSP_PARAM_OUTOFRANGE;
+    }
+    if (NULL == output) {
+        return ESP_ERR_DSP_PARAM_OUTOFRANGE;
+    }
+
+    for (int i = 0 ; i < len ; i++) {
+        int32_t acc = (int32_t)input1[i * step1] - (int32_t)input2[i * step2];
+        output[i * step_out] = acc >> shift;
+    }
+    return ESP_OK;
+}