From: Konstantinos Margaritis <konstantinos.margaritis@vectorcamp.gr>
Date: Sun, 25 Sep 2022 16:55:15 +0000 (+0000)
Subject: clean up, convert from uint64 for python due to rounding in Python, fix copying functions
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=ff1cb1d9b81b8e5a606ca93653efa75242844969;p=openpower-isa.git

clean up, convert from uint64 for python due to rounding in Python, fix copying functions
---

diff --git a/media/video/libvpx/variance_svp64_wrappers.c b/media/video/libvpx/variance_svp64_wrappers.c
index 38828e10..cd41e2f5 100644
--- a/media/video/libvpx/variance_svp64_wrappers.c
+++ b/media/video/libvpx/variance_svp64_wrappers.c
@@ -15,21 +15,20 @@ uint32_t vpx_get_mb_ss_svp64(const int16_t *src_ptr) {
     pypowersim_state_t *state = pypowersim_prepare();
 
     // Change the relevant elements, mandatory: body
-    //
     state->binary = PyBytes_FromStringAndSize((const char *)&vpx_get_mb_ss_svp64_real, 1000);
     // Set GPR #3 to the pointer
-    PyObject *address = PyLong_FromLongLong(src_ptr_svp64);
+    PyObject *address = PyLong_FromUnsignedLongLong(src_ptr_svp64);
     PyList_SetItem(state->initial_regs, 3, address);
     // Load data into buffer from real memory
     for (int i=0; i < 256; i += 4) {
-      PyObject *address = PyLong_FromLongLong(src_ptr_svp64);
-      uint64_t val = src_ptr[0];
-      val |= (uint64_t)(src_ptr[1]) << 16;
-      val |= (uint64_t)(src_ptr[2]) << 32;
-      val |= (uint64_t)(src_ptr[3]) << 48;
-      //printf("src: %p -> %04x %04x %04x %04x, val: %016x -> %p\n", src_ptr, src_ptr[0], src_ptr[1], src_ptr[2], src_ptr[3], val, src_ptr_svp64);
-      PyObject *word = PyLong_FromLongLong(val);
-      PyDict_SetItem(state->initial_mem, address, word);
+      PyObject *svp64_address = PyLong_FromUnsignedLongLong(src_ptr_svp64);
+      uint64_t val = (uint64_t)(src_ptr[0]) & 0xffff;
+      val |= ((uint64_t)(src_ptr[1]) & 0xffff) << 16;
+      val |= ((uint64_t)(src_ptr[2]) & 0xffff) << 32;
+      val |= ((uint64_t)(src_ptr[3]) & 0xffff) << 48;
+      // printf("src: %p -> %04x %04x %04x %04x\t val: %016lx -> %p\n", src_ptr, (uint16_t)src_ptr[0], (uint16_t)src_ptr[1], (uint16_t)src_ptr[2], (uint16_t)src_ptr[3], val, src_ptr_svp64);
+      PyObject *word = PyLong_FromUnsignedLongLong(val);
+      PyDict_SetItem(state->initial_mem, svp64_address, word);
       src_ptr += 4;
       src_ptr_svp64 += 8;
     }
@@ -39,26 +38,39 @@ uint32_t vpx_get_mb_ss_svp64(const int16_t *src_ptr) {
 
     // Call the function and get the resulting object
     state->result_obj = PyObject_CallObject(state->simulator, state->args);
-    Py_DECREF(state->simulator);
-    Py_DECREF(state->args);
     if (!state->result_obj) {
         PyErr_Print();
         printf("Error invoking 'run_a_simulation'\n");
+        pypowersim_finalize(state);
+	exit(1);
     }
 
     // Get the GPRs from the result_obj
     PyObject *final_regs = PyObject_GetAttrString(state->result_obj, "gpr");
     if (!final_regs) {
         PyErr_Print();
-        Py_DECREF(state->result_obj);
         printf("Error getting final GPRs\n");
+        pypowersim_finalize(state);
+	exit(1);
     }
 
     // GPR #3 holds the return value as an integer
     PyObject *key = PyLong_FromLongLong(3);
     PyObject *itm = PyDict_GetItem(final_regs, key);
+    if (!itm) {
+        PyErr_Print();
+        printf("Error getting GPR #3\n");
+        pypowersim_finalize(state);
+	exit(1);
+    }
     PyObject *value = PyObject_GetAttrString(itm, "value");
-    uint64_t val = PyLong_AsLongLong(value);
+    if (!value) {
+        PyErr_Print();
+        printf("Error getting value of GPR #3\n");
+        pypowersim_finalize(state);
+	exit(1);
+    }
+    uint64_t val = PyLong_AsUnsignedLongLong(value);
 
     // Return value
     return (uint32_t) val;
@@ -78,16 +90,17 @@ uint32_t vpx_get4x4sse_cs_svp64(const uint8_t *src_ptr, int src_stride,
     // Change the relevant elements, mandatory: body
     state->binary = PyBytes_FromStringAndSize((const char *)&vpx_get4x4sse_cs_svp64_real, 1000);
     // Set GPR #3 to the src_ptr
-    PyObject *src_address = PyLong_FromLongLong(src_ptr_svp64);
+    PyObject *src_address = PyLong_FromUnsignedLongLong(src_ptr_svp64);
     PyList_SetItem(state->initial_regs, 3, src_address);
     // Load data into buffer from real memory
     for (int r=0; r < 4; r++) {
-      PyObject *address = PyLong_FromLongLong(src_ptr_svp64);
-      uint64_t val = src_ptr[0];
-      val |= (uint64_t)(src_ptr[1]) << 16;
-      val |= (uint64_t)(src_ptr[2]) << 32;
-      val |= (uint64_t)(src_ptr[3]) << 48;
-      PyObject *word = PyLong_FromLongLong(val);
+      PyObject *address = PyLong_FromUnsignedLongLong(src_ptr_svp64);
+      uint64_t val = (uint64_t)(src_ptr[0]) & 0xffff;
+      val |= ((uint64_t)(src_ptr[1]) & 0xffff) << 16;
+      val |= ((uint64_t)(src_ptr[2]) & 0xffff) << 32;
+      val |= ((uint64_t)(src_ptr[3]) & 0xffff) << 48;
+      //printf("src: %p -> %04x %04x %04x %04x\t val: %016lx -> %p\n", src_ptr, (uint16_t)src_ptr[0], (uint16_t)src_ptr[1], (uint16_t)src_ptr[2], (uint16_t)src_ptr[3], val, src_ptr_svp64);
+      PyObject *word = PyLong_FromUnsignedLongLong(val);
       PyDict_SetItem(state->initial_mem, address, word);
       src_ptr += src_stride;
       src_ptr_svp64 += 8;
@@ -97,17 +110,17 @@ uint32_t vpx_get4x4sse_cs_svp64(const uint8_t *src_ptr, int src_stride,
     PyList_SetItem(state->initial_regs, 4, PyLong_FromLongLong(src_stride));
 
     // Set GPR #5 to the ref_ptr
-    PyObject *ref_address = PyLong_FromLongLong(ref_ptr_svp64);
+    PyObject *ref_address = PyLong_FromUnsignedLongLong(ref_ptr_svp64);
     PyList_SetItem(state->initial_regs, 5, ref_address);
     // Load data into buffer from real memory
     for (int r=0; r < 4; r++) {
-      PyObject *address = PyLong_FromLongLong(ref_ptr_svp64);
-      uint64_t val = ref_ptr[0];
-      val |= (uint64_t)(ref_ptr[1]) << 16;
-      val |= (uint64_t)(ref_ptr[2]) << 32;
-      val |= (uint64_t)(ref_ptr[3]) << 48;
+      PyObject *address = PyLong_FromUnsignedLongLong(ref_ptr_svp64);
+      uint64_t val = (uint64_t)(src_ptr[0]) & 0xffff;
+      val |= ((uint64_t)(src_ptr[1]) & 0xffff) << 16;
+      val |= ((uint64_t)(src_ptr[2]) & 0xffff) << 32;
+      val |= ((uint64_t)(src_ptr[3]) & 0xffff) << 48;
       //printf("ref: %p -> %04x %04x %04x %04x, val: %016lx -> %p\n", ref_ptr, ref_ptr[0], ref_ptr[1], ref_ptr[2], ref_ptr[3], val, ref_ptr_svp64);
-      PyObject *word = PyLong_FromLongLong(val);
+      PyObject *word = PyLong_FromUnsignedLongLong(val);
       PyDict_SetItem(state->initial_mem, address, word);
       ref_ptr += ref_stride;
       ref_ptr_svp64 += 8;
@@ -140,7 +153,7 @@ uint32_t vpx_get4x4sse_cs_svp64(const uint8_t *src_ptr, int src_stride,
     PyObject *key = PyLong_FromLongLong(3);
     PyObject *itm = PyDict_GetItem(final_regs, key);
     PyObject *value = PyObject_GetAttrString(itm, "value");
-    uint64_t val = PyLong_AsLongLong(value);
+    uint64_t val = PyLong_AsUnsignedLongLong(value);
 
     // Return value
     return (uint32_t) val;
@@ -167,39 +180,43 @@ void variance_svp64(const uint8_t *src_ptr, int src_stride,
     // Change the relevant elements, mandatory: body
     state->binary = PyBytes_FromStringAndSize((const char *)&variance_svp64_real, 1000);
     // Set GPR #3 to the src_ptr
-    PyObject *src_address = PyLong_FromLongLong(src_ptr_svp64);
+    PyObject *src_address = PyLong_FromUnsignedLongLong(src_ptr_svp64);
     PyList_SetItem(state->initial_regs, 3, src_address);
     // Load data into buffer from real memory
-    for (int r=0; r < 4; r++) {
-      PyObject *address = PyLong_FromLongLong(src_ptr_svp64);
-      uint64_t val = src_ptr[0];
-      val |= (uint64_t)(src_ptr[1]) << 16;
-      val |= (uint64_t)(src_ptr[2]) << 32;
-      val |= (uint64_t)(src_ptr[3]) << 48;
-      PyObject *word = PyLong_FromLongLong(val);
-      PyDict_SetItem(state->initial_mem, address, word);
+    for (int r=0; r < h; r++) {
+      for (int c=0; c < w; c += 4) {
+        PyObject *address = PyLong_FromUnsignedLongLong(src_ptr_svp64 + c*2);
+        uint64_t val = (uint64_t)(src_ptr[c + 0]) & 0xffff;
+        val |= ((uint64_t)(src_ptr[c + 1]) & 0xffff) << 16;
+        val |= ((uint64_t)(src_ptr[c + 2]) & 0xffff) << 32;
+        val |= ((uint64_t)(src_ptr[c + 3]) & 0xffff) << 48;
+        PyObject *word = PyLong_FromUnsignedLongLong(val);
+        PyDict_SetItem(state->initial_mem, address, word);
+      }
       src_ptr += src_stride;
-      src_ptr_svp64 += 8;
+      src_ptr_svp64 += w*2;
     }
 
     // Set GPR #4 to the src_stride 
     PyList_SetItem(state->initial_regs, 4, PyLong_FromLongLong(src_stride));
 
     // Set GPR #5 to the ref_ptr
-    PyObject *ref_address = PyLong_FromLongLong(ref_ptr_svp64);
+    PyObject *ref_address = PyLong_FromUnsignedLongLong(ref_ptr_svp64);
     PyList_SetItem(state->initial_regs, 5, ref_address);
     // Load data into buffer from real memory
-    for (int r=0; r < 4; r++) {
-      PyObject *address = PyLong_FromLongLong(ref_ptr_svp64);
-      uint64_t val = ref_ptr[0];
-      val |= (uint64_t)(ref_ptr[1]) << 16;
-      val |= (uint64_t)(ref_ptr[2]) << 32;
-      val |= (uint64_t)(ref_ptr[3]) << 48;
-      //printf("ref: %p -> %04x %04x %04x %04x, val: %016lx -> %p\n", ref_ptr, ref_ptr[0], ref_ptr[1], ref_ptr[2], ref_ptr[3], val, ref_ptr_svp64);
-      PyObject *word = PyLong_FromLongLong(val);
-      PyDict_SetItem(state->initial_mem, address, word);
+    for (int r=0; r < h; r++) {
+      for (int c=0; c < w; c += 4) {
+        PyObject *address = PyLong_FromUnsignedLongLong(ref_ptr_svp64 + c*2);
+        uint64_t val = (uint64_t)(src_ptr[c + 0]) & 0xffff;
+        val |= ((uint64_t)(src_ptr[c + 1]) & 0xffff) << 16;
+        val |= ((uint64_t)(src_ptr[c + 2]) & 0xffff) << 32;
+        val |= ((uint64_t)(src_ptr[c + 3]) & 0xffff) << 48;
+        //printf("ref: %p -> %04x %04x %04x %04x, val: %016lx -> %p\n", ref_ptr, ref_ptr[0], ref_ptr[1], ref_ptr[2], ref_ptr[3], val, ref_ptr_svp64);
+        PyObject *word = PyLong_FromUnsignedLongLong(val);
+        PyDict_SetItem(state->initial_mem, address, word);
+      }
       ref_ptr += ref_stride;
-      ref_ptr_svp64 += 8;
+      ref_ptr_svp64 += w*2;
     }
 
     // Set GPR #6 to the ref_stride 
@@ -209,12 +226,12 @@ void variance_svp64(const uint8_t *src_ptr, int src_stride,
     // Set GPR #8 to the height
     PyList_SetItem(state->initial_regs, 8, PyLong_FromLongLong(h));
     // Set GPR #9 to the sse pointer
-    PyList_SetItem(state->initial_regs, 9, PyLong_FromLongLong(sse_ptr_svp64));
+    PyList_SetItem(state->initial_regs, 9, PyLong_FromUnsignedLongLong(sse_ptr_svp64));
     // Set GPR #10 to the sum pointer
-    PyList_SetItem(state->initial_regs, 10, PyLong_FromLongLong(sum_ptr_svp64));
+    PyList_SetItem(state->initial_regs, 10, PyLong_FromUnsignedLongLong(sum_ptr_svp64));
 
-    PyObject *sse_address = PyLong_FromLongLong(sse_ptr_svp64);
-    PyObject *sum_address = PyLong_FromLongLong(sum_ptr_svp64);
+    PyObject *sse_address = PyLong_FromUnsignedLongLong(sse_ptr_svp64);
+    PyObject *sum_address = PyLong_FromUnsignedLongLong(sum_ptr_svp64);
     PyObject *word = PyLong_FromLongLong(0);
     PyDict_SetItem(state->initial_mem, sse_address, word);
     PyDict_SetItem(state->initial_mem, sum_address, word);
@@ -253,13 +270,16 @@ void variance_svp64(const uint8_t *src_ptr, int src_stride,
         printf("Error getting mem dict\n");
     }
 
-    sse_address = PyLong_FromLongLong(sse_ptr_svp64/8);
-    sum_address = PyLong_FromLongLong(sum_ptr_svp64/8);
+    sse_address = PyLong_FromUnsignedLongLong(sse_ptr_svp64/8);
+    sum_address = PyLong_FromUnsignedLongLong(sum_ptr_svp64/8);
 
     PyObject *sse_val = PyDict_GetItem(mem, sse_address);
-    *sse = (uint32_t) PyLong_AsLongLong(sse_val);
+    uint64_t val = PyLong_AsUnsignedLongLong(sse_val);
+    *sse = (uint32_t) val;
+    printf("val: %016lx, sse: %d/%08x\n", val, *sse, *sse);
 
     PyObject *sum_val = PyDict_GetItem(mem, sum_address);
-    *sum = (int32_t)PyLong_AsLongLong(sum_val);
-    printf("sse: %d/%08x, sum: %d/%08x\n", *sse, *sse, *sum, *sum);
+    val = PyLong_AsUnsignedLongLong(sum_val);
+    *sum = (int32_t) val;
+    printf("val: %016lx, sum: %d/%08x\n", val, *sum, *sum);
 }