add vpx_get4x4sse_cs_svp64_real() and wrapper

diff --git a/media/video/libvpx/Makefile b/media/video/libvpx/Makefile
index 06b82323e008522a79ce4807be0903b9cf58759b..c4f129f85caf1280eedf169eb6ff8547716e3189 100644 (file)
--- a/media/video/libvpx/Makefile
+++ b/media/video/libvpx/Makefile
@@ -10,15 +10,15 @@ CXXFLAGS= -Iinclude -O -g3
 ASFLAGS= -mlibresoc
 LDFLAGS=-lgtest -pthread -lpython3.7m
 
-BINFILES = variancefuncs_svp64.bin
-ASFILES  = variancefuncs_svp64.s
+BINFILES = vpx_get_mb_ss_svp64_real.bin vpx_get4x4sse_cs_svp64_real.bin 
+ASFILES  = vpx_get_mb_ss_svp64_real.s vpx_get4x4sse_cs_svp64_real.s
 CFILES   = variance_ref.c  variance_svp64.c  variance_svp64_wrappers.c  vpx_mem.c
 CPPFILES = test_libvpx.cc  variance_test.cc
 EXAMPLEC = pypowersim_wrapper_example.c
 EXAMPLEOBJ= ${EXAMPLEC:.c=.o}
 OBJFILES = $(CFILES:.c=.o) $(CPPFILES:.cc=.o) $(ASFILES:.s=.o)
 
-variancefuncs_svp64.bin: variancefuncs_svp64.o
+%.bin: %.o
        ${OBJCOPY} -I elf64-little -O binary $< $@ 
 
 ${TARGET}: ${OBJFILES} ${BINFILES}

diff --git a/media/video/libvpx/variance_svp64.c b/media/video/libvpx/variance_svp64.c
index f0fc3f5544d70f60edd683fe9b63155f2aa07c17..cdfaa052496e6685807ee7fc04c4d238df2a40e1 100644 (file)
--- a/media/video/libvpx/variance_svp64.c
+++ b/media/video/libvpx/variance_svp64.c
@@ -21,23 +21,6 @@ static const uint8_t bilinear_filters[8][2] = {
   { 64, 64 }, { 48, 80 },  { 32, 96 }, { 16, 112 },
 };
 
-uint32_t vpx_get4x4sse_cs_svp64(const uint8_t *src_ptr, int src_stride,
-                            const uint8_t *ref_ptr, int ref_stride) {
-  int distortion = 0;
-  int r, c;
-
-  for (r = 0; r < 4; ++r) {
-    for (c = 0; c < 4; ++c) {
-      int diff = src_ptr[c] - ref_ptr[c];
-      distortion += diff * diff;
-    }
-
-    src_ptr += src_stride;
-    ref_ptr += ref_stride;
-  }
-
-  return distortion;
-}
 
 static void variance_svp64(const uint8_t *src_ptr, int src_stride,
                      const uint8_t *ref_ptr, int ref_stride, int w, int h,

diff --git a/media/video/libvpx/variance_svp64_wrappers.c b/media/video/libvpx/variance_svp64_wrappers.c
index c6e8431b250cb79366936d3bed3d46fbd364d04b..da70fdf3f84bcf11bca38bd65e791f42b9e54541 100644 (file)
--- a/media/video/libvpx/variance_svp64_wrappers.c
+++ b/media/video/libvpx/variance_svp64_wrappers.c
@@ -49,7 +49,7 @@ uint32_t vpx_get_mb_ss_svp64(const int16_t *src_ptr) {
     }
 
     // GPR #3 holds the return value as an integer
-    PyObject *key = PyLong_FromLong(3);
+    PyObject *key = PyLong_FromLongLong(3);
     PyObject *itm = PyDict_GetItem(final_regs, key);
     PyObject *value = PyObject_GetAttrString(itm, "value");
     uint64_t val = PyLong_AsLongLong(value);
@@ -57,3 +57,91 @@ uint32_t vpx_get_mb_ss_svp64(const int16_t *src_ptr) {
     // Return value
     return (uint32_t) val;
 }
+
+uint32_t vpx_get4x4sse_cs_svp64(const uint8_t *src_ptr, int src_stride,
+                                const uint8_t *ref_ptr, int ref_stride) {
+    // It cannot be the same pointer as the original function, as it is really a separate CPU/RAM
+    // we have to memcpy from src_ptr to this pointer, the address was chosen arbitrarily
+    const uint64_t src_ptr_svp64 = 0x100000;
+    const uint64_t ref_ptr_svp64 = 0x200000;
+    const uint64_t *src_ptr64 = (const uint64_t *) src_ptr;
+    const uint64_t *ref_ptr64 = (const uint64_t *) ref_ptr;
+
+    // Create the pypowersim_state
+    pypowersim_state_t *state = pypowersim_prepare();
+
+    // 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);
+    PyList_SetItem(state->initial_regs, 3, src_address);
+    // Load data into buffer from real memory
+    size_t size = 16*sizeof(uint16_t)/sizeof(uint64_t);
+    for (int i=0; i < size; i++) {
+      PyObject *address = PyLong_FromLongLong(src_ptr_svp64 + i*8);
+      PyObject *word = PyLong_FromLongLong(*(src_ptr64 + i));
+      PyDict_SetItem(state->initial_mem, address, word);
+    }
+
+    // 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);
+    PyList_SetItem(state->initial_regs, 5, ref_address);
+    // Load data into buffer from real memory
+    for (int i=0; i < size; i++) {
+      PyObject *address = PyLong_FromLongLong(ref_ptr_svp64 + i*8);
+      PyObject *word = PyLong_FromLongLong(*(ref_ptr64 + i));
+      PyDict_SetItem(state->initial_mem, address, word);
+    }
+
+    // Set GPR #6 to the ref_stride 
+    PyList_SetItem(state->initial_regs, 4, PyLong_FromLongLong(ref_stride));
+
+    // Prepare the arguments object for the call
+    pypowersim_prepareargs(state);
+
+    // 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");
+    }
+
+    // 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");
+    }
+
+    // GPR #3 holds the return value as an integer
+    PyObject *key = PyLong_FromLongLong(3);
+    PyObject *itm = PyDict_GetItem(final_regs, key);
+    PyObject *value = PyObject_GetAttrString(itm, "value");
+    uint64_t val = PyLong_AsLongLong(value);
+
+    // Return value
+    return (uint32_t) val;
+
+  int distortion = 0;
+  int r, c;
+
+  for (r = 0; r < 4; ++r) {
+    for (c = 0; c < 4; ++c) {
+      int diff = src_ptr[c] - ref_ptr[c];
+      distortion += diff * diff;
+    }
+
+    src_ptr += src_stride;
+    ref_ptr += ref_stride;
+  }
+
+  return distortion;
+}
+

diff --git a/media/video/libvpx/variance_svp64_wrappers.h b/media/video/libvpx/variance_svp64_wrappers.h
index d38ecff03ea5e91489d3e9b4b7f39fbfc30619a8..4441dcceafda96c5b9b120e49500ceaf7efa977c 100644 (file)
--- a/media/video/libvpx/variance_svp64_wrappers.h
+++ b/media/video/libvpx/variance_svp64_wrappers.h
@@ -1,3 +1,7 @@
 #include <stdint.h>
 
 uint32_t vpx_get_mb_ss_svp64_real(const int16_t *src_ptr);
+
+uint32_t vpx_get4x4sse_cs_svp64_real(const uint8_t *src_ptr, int src_stride,
+                                     const uint8_t *ref_ptr, int ref_stride);
+

diff --git a/media/video/libvpx/vpx_get4x4sse_cs_svp64_real.c.in b/media/video/libvpx/vpx_get4x4sse_cs_svp64_real.c.in
new file mode 100644 (file)
index 0000000..279f3fe
--- /dev/null
+++ b/media/video/libvpx/vpx_get4x4sse_cs_svp64_real.c.in
@@ -0,0 +1,20 @@
+#include <stdint.h>
+
+uint32_t vpx_get4x4sse_cs_svp64_real(const uint8_t *src_ptr, int src_stride,
+                                     const uint8_t *ref_ptr, int ref_stride) {
+  int distortion = 0;
+  int r, c;
+
+  for (r = 0; r < 4; ++r) {
+    for (c = 0; c < 4; ++c) {
+      int diff = src_ptr[c] - ref_ptr[c];
+      distortion += diff * diff;
+    }
+
+    src_ptr += src_stride;
+    ref_ptr += ref_stride;
+  }
+
+  return distortion;
+}
+

diff --git a/media/video/libvpx/vpx_get4x4sse_cs_svp64_real.s b/media/video/libvpx/vpx_get4x4sse_cs_svp64_real.s
new file mode 100644 (file)
index 0000000..1d7c6e5
--- /dev/null
+++ b/media/video/libvpx/vpx_get4x4sse_cs_svp64_real.s
@@ -0,0 +1,40 @@
+       .file   "vpx_get4x4sse_cs_svp64_real.c"
+       .abiversion 2
+       .section        ".text"
+       .align 2
+       .globl vpx_get4x4sse_cs_svp64_real
+       .type   vpx_get4x4sse_cs_svp64_real, @function
+vpx_get4x4sse_cs_svp64_real:
+.LFB0:
+       .cfi_startproc
+       addi 5,5,-1
+       addi 3,3,3
+       li 12,4
+       li 8,0
+.L2:
+       addi 7,3,-4
+       mr 11,5
+       subf 9,7,3
+       mtctr 9
+.L3:
+       lbzu 9,1(7)
+       lbzu 10,1(11)
+       subf 9,10,9
+       mullw 9,9,9
+       add 9,9,8
+       extsw 8,9
+       bdnz .L3
+       addi 9,12,-1
+       add 5,5,6
+       add 3,3,4
+       rldicl. 12,9,0,32
+       bne 0,.L2
+       rldicl 3,8,0,32
+       blr
+       .long 0
+       .byte 0,0,0,0,0,0,0,0
+       .cfi_endproc
+.LFE0:
+       .size   vpx_get4x4sse_cs_svp64_real,.-vpx_get4x4sse_cs_svp64_real
+       .ident  "GCC: (Debian 8.3.0-6) 8.3.0"
+       .section        .note.GNU-stack,"",@progbits