Fix the bug for AllocationRead() that cannot handle 3d Allocation

correctly.
Add the following functions to make it symmetric for read and write.

 - AllocationElementData
 - AllocationElementRead
 - Allocation3DRead

Change-Id: I8c89b5e3474c4c3c70e9795843a76e61f1f90f6a

9 files changed, 134 insertions, 54 deletions

diff --git a/cpp/rsDispatch.cpp b/cpp/rsDispatch.cpp
index e0d67883..6f379d09 100644
--- a/cpp/rsDispatch.cpp
+++ b/cpp/rsDispatch.cpp
@@ -389,6 +389,21 @@ bool loadSymbols(void* handle, dispatchTable& dispatchTab, int device_api) {
             LOG_API("Couldn't initialize dispatchTab.ScriptGroup2Create");
             return false;
         }
+        dispatchTab.AllocationElementData = (AllocationElementDataFnPtr)dlsym(handle, "rsAllocationElementData");
+        if (dispatchTab.AllocationElementData == NULL) {
+            LOG_API("Couldn't initialize dispatchTab.AllocationElementData");
+            return false;
+        }
+        dispatchTab.AllocationElementRead = (AllocationElementReadFnPtr)dlsym(handle, "rsAllocationElementRead");
+        if (dispatchTab.AllocationElementRead == NULL) {
+            LOG_API("Couldn't initialize dispatchTab.AllocationElementRead");
+            return false;
+        }
+        dispatchTab.Allocation3DRead = (Allocation3DReadFnPtr)dlsym(handle, "rsAllocation3DRead");
+        if (dispatchTab.Allocation3DRead == NULL) {
+            LOG_API("Couldn't initialize dispatchTab.Allocation3DRead");
+            return false;
+        }
     }
 
     return true;
diff --git a/cpp/rsDispatch.h b/cpp/rsDispatch.h
index c1c8d77f..5bda7c01 100644
--- a/cpp/rsDispatch.h
+++ b/cpp/rsDispatch.h
@@ -55,12 +55,15 @@ typedef RsElement (*ElementCreate2FnPtr) (RsContext, const RsElement*, size_t, c
 typedef void (*AllocationCopyToBitmapFnPtr) (RsContext, RsAllocation, void*, size_t);
 typedef void (*Allocation1DDataFnPtr) (RsContext, RsAllocation, uint32_t, uint32_t, uint32_t, const void*, size_t);
 typedef void (*Allocation1DElementDataFnPtr) (RsContext, RsAllocation, uint32_t, uint32_t, const void*, size_t, size_t);
+typedef void (*AllocationElementDataFnPtr) (RsContext, RsAllocation, uint32_t, uint32_t, uint32_t, uint32_t, const void*, size_t, size_t);
 typedef void (*Allocation2DDataFnPtr) (RsContext, RsAllocation, uint32_t, uint32_t, uint32_t, RsAllocationCubemapFace, uint32_t, uint32_t, const void*, size_t, size_t);
 typedef void (*Allocation3DDataFnPtr) (RsContext, RsAllocation, uint32_t, uint32_t, uint32_t, uint32_t, uint32_t, uint32_t, uint32_t, const void*, size_t, size_t);
 typedef void (*AllocationGenerateMipmapsFnPtr) (RsContext, RsAllocation);
 typedef void (*AllocationReadFnPtr) (RsContext, RsAllocation, void*, size_t);
 typedef void (*Allocation1DReadFnPtr) (RsContext, RsAllocation, uint32_t, uint32_t, uint32_t, void*, size_t);
+typedef void (*AllocationElementReadFnPtr) (RsContext, RsAllocation, uint32_t, uint32_t, uint32_t, uint32_t, void*, size_t, size_t);
 typedef void (*Allocation2DReadFnPtr) (RsContext, RsAllocation, uint32_t, uint32_t, uint32_t, RsAllocationCubemapFace, uint32_t, uint32_t, void*, size_t, size_t);
+typedef void (*Allocation3DReadFnPtr) (RsContext, RsAllocation, uint32_t, uint32_t, uint32_t, uint32_t, uint32_t, uint32_t, uint32_t, void*, size_t, size_t);
 typedef void (*AllocationSyncAllFnPtr) (RsContext, RsAllocation, RsAllocationUsageType);
 typedef void (*AllocationResize1DFnPtr) (RsContext, RsAllocation, uint32_t);
 typedef void (*AllocationCopy2DRangeFnPtr) (RsContext, RsAllocation, uint32_t, uint32_t, uint32_t, uint32_t, uint32_t, uint32_t, RsAllocation, uint32_t, uint32_t, uint32_t, uint32_t);
@@ -134,12 +137,15 @@ struct dispatchTable {
     AllocationCopyToBitmapFnPtr AllocationCopyToBitmap;
     Allocation1DDataFnPtr Allocation1DData;
     Allocation1DElementDataFnPtr Allocation1DElementData;
+    AllocationElementDataFnPtr AllocationElementData;
     Allocation2DDataFnPtr Allocation2DData;
     Allocation3DDataFnPtr Allocation3DData;
     AllocationGenerateMipmapsFnPtr AllocationGenerateMipmaps;
     AllocationReadFnPtr AllocationRead;
     Allocation1DReadFnPtr Allocation1DRead;
+    AllocationElementReadFnPtr AllocationElementRead;
     Allocation2DReadFnPtr Allocation2DRead;
+    Allocation3DReadFnPtr Allocation3DRead;
     AllocationSyncAllFnPtr AllocationSyncAll;
     AllocationResize1DFnPtr AllocationResize1D;
     AllocationCopy2DRangeFnPtr AllocationCopy2DRange;
diff --git a/driver/rsdAllocation.cpp b/driver/rsdAllocation.cpp
index f56ab56c..bbf73fb5 100644
--- a/driver/rsdAllocation.cpp
+++ b/driver/rsdAllocation.cpp
@@ -1114,13 +1114,12 @@ void rsdAllocationData3D_alloc(const android::renderscript::Context *rsc,
                                      srcXoff, srcYoff, srcZoff, srcLod);
 }
 
-void rsdAllocationElementData1D(const Context *rsc, const Allocation *alloc,
-                                uint32_t x,
-                                const void *data, uint32_t cIdx, size_t sizeBytes) {
+void rsdAllocationElementData(const Context *rsc, const Allocation *alloc,
+                              uint32_t x, uint32_t y, uint32_t z,
+                              const void *data, uint32_t cIdx, size_t sizeBytes) {
     DrvAllocation *drv = (DrvAllocation *)alloc->mHal.drv;
 
-    size_t eSize = alloc->mHal.state.elementSizeBytes;
-    uint8_t * ptr = GetOffsetPtr(alloc, x, 0, 0, 0, RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
+    uint8_t * ptr = GetOffsetPtr(alloc, x, y, z, 0, RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
 
     const Element * e = alloc->mHal.state.type->getElement()->getField(cIdx);
     ptr += alloc->mHal.state.type->getElement()->getFieldOffsetBytes(cIdx);
@@ -1134,24 +1133,17 @@ void rsdAllocationElementData1D(const Context *rsc, const Allocation *alloc,
     drv->uploadDeferred = true;
 }
 
-void rsdAllocationElementData2D(const Context *rsc, const Allocation *alloc,
-                                uint32_t x, uint32_t y,
-                                const void *data, uint32_t cIdx, size_t sizeBytes) {
+void rsdAllocationElementRead(const Context *rsc, const Allocation *alloc,
+                              uint32_t x, uint32_t y, uint32_t z,
+                              void *data, uint32_t cIdx, size_t sizeBytes) {
     DrvAllocation *drv = (DrvAllocation *)alloc->mHal.drv;
 
-    size_t eSize = alloc->mHal.state.elementSizeBytes;
-    uint8_t * ptr = GetOffsetPtr(alloc, x, y, 0, 0, RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
+    uint8_t * ptr = GetOffsetPtr(alloc, x, y, z, 0, RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
 
     const Element * e = alloc->mHal.state.type->getElement()->getField(cIdx);
     ptr += alloc->mHal.state.type->getElement()->getFieldOffsetBytes(cIdx);
 
-    if (alloc->mHal.state.hasReferences) {
-        e->incRefs(data);
-        e->decRefs(ptr);
-    }
-
-    memcpy(ptr, data, sizeBytes);
-    drv->uploadDeferred = true;
+    memcpy(data, ptr, sizeBytes);
 }
 
 static void mip565(const Allocation *alloc, int lod, RsAllocationCubemapFace face) {
diff --git a/driver/rsdAllocation.h b/driver/rsdAllocation.h
index 0b6d8bf1..eff5e302 100644
--- a/driver/rsdAllocation.h
+++ b/driver/rsdAllocation.h
@@ -161,14 +161,15 @@ void rsdAllocationData3D_alloc(const android::renderscript::Context *rsc,
                                uint32_t srcXoff, uint32_t srcYoff, uint32_t srcZoff,
                                uint32_t srcLod);
 
-void rsdAllocationElementData1D(const android::renderscript::Context *rsc,
-                                const android::renderscript::Allocation *alloc,
-                                uint32_t x,
-                                const void *data, uint32_t elementOff, size_t sizeBytes);
-void rsdAllocationElementData2D(const android::renderscript::Context *rsc,
-                                const android::renderscript::Allocation *alloc,
-                                uint32_t x, uint32_t y,
-                                const void *data, uint32_t elementOff, size_t sizeBytes);
+void rsdAllocationElementData(const android::renderscript::Context *rsc,
+                              const android::renderscript::Allocation *alloc,
+                              uint32_t x, uint32_t y, uint32_t z,
+                              const void *data, uint32_t elementOff, size_t sizeBytes);
+
+void rsdAllocationElementRead(const android::renderscript::Context *rsc,
+                              const android::renderscript::Allocation *alloc,
+                              uint32_t x, uint32_t y, uint32_t z,
+                              void *data, uint32_t elementOff, size_t sizeBytes);
 
 void rsdAllocationGenerateMipmaps(const android::renderscript::Context *rsc,
                                   const android::renderscript::Allocation *alloc);
diff --git a/driver/rsdCore.cpp b/driver/rsdCore.cpp
index 8812446e..5390faf2 100644
--- a/driver/rsdCore.cpp
+++ b/driver/rsdCore.cpp
@@ -108,8 +108,8 @@ static RsdHalFunctions FunctionTable = {
         rsdAllocationData1D_alloc,
         rsdAllocationData2D_alloc,
         rsdAllocationData3D_alloc,
-        rsdAllocationElementData1D,
-        rsdAllocationElementData2D,
+        rsdAllocationElementData,
+        rsdAllocationElementRead,
         rsdAllocationGenerateMipmaps,
         rsdAllocationUpdateCachedObject,
         rsdAllocationAdapterOffset
diff --git a/rs.spec b/rs.spec
index 841b89f3..8054faa6 100644
--- a/rs.spec
+++ b/rs.spec
@@ -174,6 +174,16 @@ Allocation1DElementData {
     param size_t comp_offset
     }
 
+AllocationElementData {
+    param RsAllocation va
+    param uint32_t x
+    param uint32_t y
+    param uint32_t z
+    param uint32_t lod
+    param const void *data
+    param size_t comp_offset
+    }
+
 Allocation2DData {
     param RsAllocation va
     param uint32_t xoff
@@ -216,6 +226,16 @@ Allocation1DRead {
     param void *data
     }
 
+AllocationElementRead {
+    param RsAllocation va
+    param uint32_t x
+    param uint32_t y
+    param uint32_t z
+    param uint32_t lod
+    param void *data
+    param size_t comp_offset
+    }
+
 Allocation2DRead {
     param RsAllocation va
     param uint32_t xoff
@@ -228,6 +248,18 @@ Allocation2DRead {
     param size_t stride
 }
 
+Allocation3DRead {
+    param RsAllocation va
+    param uint32_t xoff
+    param uint32_t yoff
+    param uint32_t zoff
+    param uint32_t lod
+    param uint32_t w
+    param uint32_t h
+    param uint32_t d
+    param void *data
+    param size_t stride
+    }
 
 AllocationSyncAll {
     param RsAllocation va
diff --git a/rsAllocation.cpp b/rsAllocation.cpp
index 386d86fb..1d3daaf6 100644
--- a/rsAllocation.cpp
+++ b/rsAllocation.cpp
@@ -256,17 +256,27 @@ void Allocation::read(Context *rsc, uint32_t xoff, uint32_t yoff, uint32_t zoff,
 
 }
 
-void Allocation::elementData(Context *rsc, uint32_t x, const void *data,
-                                uint32_t cIdx, size_t sizeBytes) {
+void Allocation::elementData(Context *rsc, uint32_t x, uint32_t y, uint32_t z,
+                             const void *data, uint32_t cIdx, size_t sizeBytes) {
     size_t eSize = mHal.state.elementSizeBytes;
 
-    if (cIdx >= mHal.state.type->getElement()->getFieldCount()) {
-        rsc->setError(RS_ERROR_BAD_VALUE, "subElementData component out of range.");
+    if (x >= mHal.drvState.lod[0].dimX) {
+        rsc->setError(RS_ERROR_BAD_VALUE, "subElementData X offset out of range.");
         return;
     }
 
-    if (x >= mHal.drvState.lod[0].dimX) {
-        rsc->setError(RS_ERROR_BAD_VALUE, "subElementData X offset out of range.");
+    if (y > 0 && y >= mHal.drvState.lod[0].dimY) {
+        rsc->setError(RS_ERROR_BAD_VALUE, "subElementData Y offset out of range.");
+        return;
+    }
+
+    if (z > 0 && z >= mHal.drvState.lod[0].dimZ) {
+        rsc->setError(RS_ERROR_BAD_VALUE, "subElementData Z offset out of range.");
+        return;
+    }
+
+    if (cIdx >= mHal.state.type->getElement()->getFieldCount()) {
+        rsc->setError(RS_ERROR_BAD_VALUE, "subElementData component out of range.");
         return;
     }
 
@@ -277,12 +287,12 @@ void Allocation::elementData(Context *rsc, uint32_t x, const void *data,
         return;
     }
 
-    rsc->mHal.funcs.allocation.elementData1D(rsc, this, x, data, cIdx, sizeBytes);
+    rsc->mHal.funcs.allocation.elementData(rsc, this, x, y, z, data, cIdx, sizeBytes);
     sendDirty(rsc);
 }
 
-void Allocation::elementData(Context *rsc, uint32_t x, uint32_t y,
-                                const void *data, uint32_t cIdx, size_t sizeBytes) {
+void Allocation::elementRead(Context *rsc, uint32_t x, uint32_t y, uint32_t z,
+                             void *data, uint32_t cIdx, size_t sizeBytes) {
     size_t eSize = mHal.state.elementSizeBytes;
 
     if (x >= mHal.drvState.lod[0].dimX) {
@@ -290,8 +300,13 @@ void Allocation::elementData(Context *rsc, uint32_t x, uint32_t y,
         return;
     }
 
-    if (y >= mHal.drvState.lod[0].dimY) {
-        rsc->setError(RS_ERROR_BAD_VALUE, "subElementData X offset out of range.");
+    if (y > 0 && y >= mHal.drvState.lod[0].dimY) {
+        rsc->setError(RS_ERROR_BAD_VALUE, "subElementData Y offset out of range.");
+        return;
+    }
+
+    if (z > 0 && z >= mHal.drvState.lod[0].dimZ) {
+        rsc->setError(RS_ERROR_BAD_VALUE, "subElementData Z offset out of range.");
         return;
     }
 
@@ -307,8 +322,7 @@ void Allocation::elementData(Context *rsc, uint32_t x, uint32_t y,
         return;
     }
 
-    rsc->mHal.funcs.allocation.elementData2D(rsc, this, x, y, data, cIdx, sizeBytes);
-    sendDirty(rsc);
+    rsc->mHal.funcs.allocation.elementRead(rsc, this, x, y, z, data, cIdx, sizeBytes);
 }
 
 void Allocation::addProgramToDirty(const Program *p) {
@@ -653,16 +667,16 @@ void rsi_Allocation1DData(Context *rsc, RsAllocation va, uint32_t xoff, uint32_t
     a->data(rsc, xoff, lod, count, data, sizeBytes);
 }
 
-void rsi_Allocation2DElementData(Context *rsc, RsAllocation va, uint32_t x, uint32_t y, uint32_t lod, RsAllocationCubemapFace face,
-                                 const void *data, size_t sizeBytes, size_t eoff) {
+void rsi_Allocation1DElementData(Context *rsc, RsAllocation va, uint32_t x,
+                                 uint32_t lod, const void *data, size_t sizeBytes, size_t eoff) {
     Allocation *a = static_cast<Allocation *>(va);
-    a->elementData(rsc, x, y, data, eoff, sizeBytes);
+    a->elementData(rsc, x, 0, 0, data, eoff, sizeBytes);
 }
 
-void rsi_Allocation1DElementData(Context *rsc, RsAllocation va, uint32_t x, uint32_t lod,
-                                 const void *data, size_t sizeBytes, size_t eoff) {
+void rsi_AllocationElementData(Context *rsc, RsAllocation va, uint32_t x, uint32_t y, uint32_t z,
+                               uint32_t lod, const void *data, size_t sizeBytes, size_t eoff) {
     Allocation *a = static_cast<Allocation *>(va);
-    a->elementData(rsc, x, data, eoff, sizeBytes);
+    a->elementData(rsc, x, y, z, data, eoff, sizeBytes);
 }
 
 void rsi_Allocation2DData(Context *rsc, RsAllocation va, uint32_t xoff, uint32_t yoff, uint32_t lod, RsAllocationCubemapFace face,
@@ -681,7 +695,10 @@ void rsi_Allocation3DData(Context *rsc, RsAllocation va, uint32_t xoff, uint32_t
 void rsi_AllocationRead(Context *rsc, RsAllocation va, void *data, size_t sizeBytes) {
     Allocation *a = static_cast<Allocation *>(va);
     const Type * t = a->getType();
-    if(t->getDimY()) {
+    if(t->getDimZ()) {
+        a->read(rsc, 0, 0, 0, 0, t->getDimX(), t->getDimY(), t->getDimZ(),
+                data, sizeBytes, 0);
+    } else if(t->getDimY()) {
         a->read(rsc, 0, 0, 0, RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X,
                 t->getDimX(), t->getDimY(), data, sizeBytes, 0);
     } else {
@@ -840,6 +857,12 @@ void rsi_Allocation1DRead(Context *rsc, RsAllocation va, uint32_t xoff, uint32_t
     rsc->mHal.funcs.allocation.read1D(rsc, a, xoff, lod, count, data, sizeBytes);
 }
 
+void rsi_AllocationElementRead(Context *rsc, RsAllocation va, uint32_t x, uint32_t y, uint32_t z,
+                                 uint32_t lod, void *data, size_t sizeBytes, size_t eoff) {
+    Allocation *a = static_cast<Allocation *>(va);
+    a->elementRead(rsc, x, y, z, data, eoff, sizeBytes);
+}
+
 void rsi_Allocation2DRead(Context *rsc, RsAllocation va, uint32_t xoff, uint32_t yoff,
                           uint32_t lod, RsAllocationCubemapFace face, uint32_t w,
                           uint32_t h, void *data, size_t sizeBytes, size_t stride) {
@@ -847,6 +870,14 @@ void rsi_Allocation2DRead(Context *rsc, RsAllocation va, uint32_t xoff, uint32_t
     a->read(rsc, xoff, yoff, lod, face, w, h, data, sizeBytes, stride);
 }
 
+void rsi_Allocation3DRead(Context *rsc, RsAllocation va,
+                          uint32_t xoff, uint32_t yoff, uint32_t zoff,
+                          uint32_t lod, uint32_t w, uint32_t h, uint32_t d,
+                          void *data, size_t sizeBytes, size_t stride) {
+    Allocation *a = static_cast<Allocation *>(va);
+    a->read(rsc, xoff, yoff, zoff, lod, w, h, d, data, sizeBytes, stride);
+}
+
 RsAllocation rsi_AllocationAdapterCreate(Context *rsc, RsType vwindow, RsAllocation vbase) {
 
 
diff --git a/rsAllocation.h b/rsAllocation.h
index 10009575..a06e28d9 100644
--- a/rsAllocation.h
+++ b/rsAllocation.h
@@ -138,11 +138,12 @@ public:
     void read(Context *rsc, uint32_t xoff, uint32_t yoff, uint32_t zoff, uint32_t lod,
               uint32_t w, uint32_t h, uint32_t d, void *data, size_t sizeBytes, size_t stride);
 
-    void elementData(Context *rsc, uint32_t x,
-                     const void *data, uint32_t elementOff, size_t sizeBytes);
-    void elementData(Context *rsc, uint32_t x, uint32_t y,
+    void elementData(Context *rsc, uint32_t x, uint32_t y, uint32_t z,
                      const void *data, uint32_t elementOff, size_t sizeBytes);
 
+    void elementRead(Context *rsc, uint32_t x, uint32_t y, uint32_t z,
+                     void *data, uint32_t elementOff, size_t sizeBytes);
+
     void addProgramToDirty(const Program *);
     void removeProgramToDirty(const Program *);
 
diff --git a/rs_hal.h b/rs_hal.h
index 9a4e9a50..5e8fee8f 100644
--- a/rs_hal.h
+++ b/rs_hal.h
@@ -236,10 +236,12 @@ typedef struct {
                             uint32_t srcXoff, uint32_t srcYoff, uint32_t srcZoff,
                             uint32_t srcLod);
 
-        void (*elementData1D)(const Context *rsc, const Allocation *alloc, uint32_t x,
-                              const void *data, uint32_t elementOff, size_t sizeBytes);
-        void (*elementData2D)(const Context *rsc, const Allocation *alloc, uint32_t x, uint32_t y,
-                              const void *data, uint32_t elementOff, size_t sizeBytes);
+        void (*elementData)(const Context *rsc, const Allocation *alloc,
+                            uint32_t x, uint32_t y, uint32_t z,
+                            const void *data, uint32_t elementOff, size_t sizeBytes);
+        void (*elementRead)(const Context *rsc, const Allocation *alloc,
+                            uint32_t x, uint32_t y, uint32_t z,
+                            void *data, uint32_t elementOff, size_t sizeBytes);
 
         void (*generateMipmaps)(const Context *rsc, const Allocation *alloc);