Merge "Collapse code paths for single- and multi-input kernels."

3 files changed, 130 insertions, 226 deletions

diff --git a/rs/java/android/renderscript/RenderScript.java b/rs/java/android/renderscript/RenderScript.java
index f2b1380d5ac4..bd272bcc2e7b 100644
--- a/rs/java/android/renderscript/RenderScript.java
+++ b/rs/java/android/renderscript/RenderScript.java
@@ -590,52 +590,14 @@ public class RenderScript {
         validate();
         rsnScriptInvoke(mContext, id, slot);
     }
-    native void rsnScriptForEach(long con, long id, int slot, long ain, long aout, byte[] params);
-    native void rsnScriptForEach(long con, long id, int slot, long ain, long aout);
-    native void rsnScriptForEachClipped(long con, long id, int slot, long ain, long aout, byte[] params,
-                                        int xstart, int xend, int ystart, int yend, int zstart, int zend);
-    native void rsnScriptForEachClipped(long con, long id, int slot, long ain, long aout,
-                                        int xstart, int xend, int ystart, int yend, int zstart, int zend);
-    synchronized void nScriptForEach(long id, int slot, long ain, long aout, byte[] params) {
-        validate();
-        if (params == null) {
-            rsnScriptForEach(mContext, id, slot, ain, aout);
-        } else {
-            rsnScriptForEach(mContext, id, slot, ain, aout, params);
-        }
-    }
 
-    synchronized void nScriptForEachClipped(long id, int slot, long ain, long aout, byte[] params,
-                                            int xstart, int xend, int ystart, int yend, int zstart, int zend) {
-        validate();
-        if (params == null) {
-            rsnScriptForEachClipped(mContext, id, slot, ain, aout, xstart, xend, ystart, yend, zstart, zend);
-        } else {
-            rsnScriptForEachClipped(mContext, id, slot, ain, aout, params, xstart, xend, ystart, yend, zstart, zend);
-        }
-    }
+    native void rsnScriptForEach(long con, long id, int slot, long[] ains,
+                                 long aout, byte[] params, int[] limits);
 
-    /**
-     * Multi-input code.
-     *
-     */
-
-    // @hide
-    native void rsnScriptForEachMultiClipped(long con, long id, int slot, long[] ains, long aout, byte[] params,
-                                             int xstart, int xend, int ystart, int yend, int zstart, int zend);
-    // @hide
-    native void rsnScriptForEachMultiClipped(long con, long id, int slot, long[] ains, long aout,
-                                             int xstart, int xend, int ystart, int yend, int zstart, int zend);
-
-    // @hide
-    synchronized void nScriptForEachMultiClipped(long id, int slot, long[] ains, long aout, byte[] params,
-                                                 int xstart, int xend, int ystart, int yend, int zstart, int zend) {
-        validate();
-        if (params == null) {
-            rsnScriptForEachMultiClipped(mContext, id, slot, ains, aout, xstart, xend, ystart, yend, zstart, zend);
-        } else {
-            rsnScriptForEachMultiClipped(mContext, id, slot, ains, aout, params, xstart, xend, ystart, yend, zstart, zend);
-        }
+    synchronized void nScriptForEach(long id, int slot, long[] ains, long aout,
+                                     byte[] params, int[] limits) {
+        validate();
+        rsnScriptForEach(mContext, id, slot, ains, aout, params, limits);
     }
 
     native void rsnScriptInvokeV(long con, long id, int slot, byte[] params);
diff --git a/rs/java/android/renderscript/Script.java b/rs/java/android/renderscript/Script.java
index c49ef948db97..eb1687a37a49 100644
--- a/rs/java/android/renderscript/Script.java
+++ b/rs/java/android/renderscript/Script.java
@@ -48,7 +48,8 @@ public class Script extends BaseObj {
     /**
      * Only to be used by generated reflected classes.
      */
-    protected KernelID createKernelID(int slot, int sig, Element ein, Element eout) {
+    protected KernelID createKernelID(int slot, int sig, Element ein,
+                                      Element eout) {
         KernelID k = mKIDs.get(slot);
         if (k != null) {
             return k;
@@ -127,59 +128,56 @@ public class Script extends BaseObj {
      * Only intended for use by generated reflected code.
      *
      */
-    protected void forEach(int slot, Allocation ain, Allocation aout, FieldPacker v) {
-        mRS.validate();
-        mRS.validateObject(ain);
-        mRS.validateObject(aout);
-        if (ain == null && aout == null) {
-            throw new RSIllegalArgumentException(
-                "At least one of ain or aout is required to be non-null.");
-        }
-        long in_id = 0;
-        if (ain != null) {
-            in_id = ain.getID(mRS);
-        }
-        long out_id = 0;
-        if (aout != null) {
-            out_id = aout.getID(mRS);
-        }
-        byte[] params = null;
-        if (v != null) {
-            params = v.getData();
-        }
-        mRS.nScriptForEach(getID(mRS), slot, in_id, out_id, params);
+    protected void forEach(int slot, Allocation ain, Allocation aout,
+                           FieldPacker v) {
+        forEach(slot, ain, aout, v, null);
     }
 
     /**
      * Only intended for use by generated reflected code.
      *
      */
-    protected void forEach(int slot, Allocation ain, Allocation aout, FieldPacker v, LaunchOptions sc) {
+    protected void forEach(int slot, Allocation ain, Allocation aout,
+                           FieldPacker v, LaunchOptions sc) {
+        // TODO: Is this necessary if nScriptForEach calls validate as well?
         mRS.validate();
         mRS.validateObject(ain);
         mRS.validateObject(aout);
+
         if (ain == null && aout == null) {
             throw new RSIllegalArgumentException(
                 "At least one of ain or aout is required to be non-null.");
         }
 
-        if (sc == null) {
-            forEach(slot, ain, aout, v);
-            return;
-        }
-        long in_id = 0;
+        long[] in_ids = null;
         if (ain != null) {
-            in_id = ain.getID(mRS);
+            in_ids    = mInIdsBuffer;
+            in_ids[0] = ain.getID(mRS);
         }
+
         long out_id = 0;
         if (aout != null) {
             out_id = aout.getID(mRS);
         }
+
         byte[] params = null;
         if (v != null) {
             params = v.getData();
         }
-        mRS.nScriptForEachClipped(getID(mRS), slot, in_id, out_id, params, sc.xstart, sc.xend, sc.ystart, sc.yend, sc.zstart, sc.zend);
+
+        int[] limits = null;
+        if (sc != null) {
+            limits = new int[6];
+
+            limits[0] = sc.xstart;
+            limits[1] = sc.xend;
+            limits[2] = sc.ystart;
+            limits[3] = sc.yend;
+            limits[4] = sc.zstart;
+            limits[5] = sc.zend;
+        }
+
+        mRS.nScriptForEach(getID(mRS), slot, in_ids, out_id, params, limits);
     }
 
     /**
@@ -187,8 +185,9 @@ public class Script extends BaseObj {
      *
      * @hide
      */
-    protected void forEach(int slot, Allocation[] ains, Allocation aout, FieldPacker v) {
-        forEach(slot, ains, aout, v, new LaunchOptions());
+    protected void forEach(int slot, Allocation[] ains, Allocation aout,
+                           FieldPacker v) {
+        forEach(slot, ains, aout, v, null);
     }
 
     /**
@@ -196,24 +195,20 @@ public class Script extends BaseObj {
      *
      * @hide
      */
-    protected void forEach(int slot, Allocation[] ains, Allocation aout, FieldPacker v, LaunchOptions sc) {
+    protected void forEach(int slot, Allocation[] ains, Allocation aout,
+                           FieldPacker v, LaunchOptions sc) {
+        // TODO: Is this necessary if nScriptForEach calls validate as well?
         mRS.validate();
-
         for (Allocation ain : ains) {
           mRS.validateObject(ain);
         }
-
         mRS.validateObject(aout);
+
         if (ains == null && aout == null) {
             throw new RSIllegalArgumentException(
                 "At least one of ain or aout is required to be non-null.");
         }
 
-        if (sc == null) {
-            forEach(slot, ains, aout, v);
-            return;
-        }
-
         long[] in_ids = new long[ains.length];
         for (int index = 0; index < ains.length; ++index) {
             in_ids[index] = ains[index].getID(mRS);
@@ -223,15 +218,33 @@ public class Script extends BaseObj {
         if (aout != null) {
             out_id = aout.getID(mRS);
         }
+
         byte[] params = null;
         if (v != null) {
             params = v.getData();
         }
-        mRS.nScriptForEachMultiClipped(getID(mRS), slot, in_ids, out_id, params, sc.xstart, sc.xend, sc.ystart, sc.yend, sc.zstart, sc.zend);
+
+        int[] limits = null;
+        if (sc != null) {
+            limits = new int[6];
+
+            limits[0] = sc.xstart;
+            limits[1] = sc.xend;
+            limits[2] = sc.ystart;
+            limits[3] = sc.yend;
+            limits[4] = sc.zstart;
+            limits[5] = sc.zend;
+        }
+
+        mRS.nScriptForEach(getID(mRS), slot, in_ids, out_id, params, limits);
     }
 
+    long[] mInIdsBuffer;
+
     Script(long id, RenderScript rs) {
         super(id, rs);
+
+        mInIdsBuffer = new long[1];
     }
 
 
@@ -243,11 +256,17 @@ public class Script extends BaseObj {
         mRS.validate();
         mRS.validateObject(va);
         if (va != null) {
-            if (mRS.getApplicationContext().getApplicationInfo().targetSdkVersion >= 20) {
+
+            android.content.Context context = mRS.getApplicationContext();
+
+            if (context.getApplicationInfo().targetSdkVersion >= 20) {
                 final Type t = va.mType;
-                if (t.hasMipmaps() || t.hasFaces() || (t.getY() != 0) || (t.getZ() != 0)) {
+                if (t.hasMipmaps() || t.hasFaces() || (t.getY() != 0) ||
+                    (t.getZ() != 0)) {
+
                     throw new RSIllegalArgumentException(
-                        "API 20+ only allows simple 1D allocations to be used with bind.");
+                        "API 20+ only allows simple 1D allocations to be " +
+                        "used with bind.");
                 }
             }
             mRS.nScriptBindAllocation(getID(mRS), va.getID(mRS), slot);
@@ -378,11 +397,14 @@ public class Script extends BaseObj {
         protected Allocation mAllocation;
 
         protected void init(RenderScript rs, int dimx) {
-            mAllocation = Allocation.createSized(rs, mElement, dimx, Allocation.USAGE_SCRIPT);
+            mAllocation = Allocation.createSized(rs, mElement, dimx,
+                                                 Allocation.USAGE_SCRIPT);
         }
 
         protected void init(RenderScript rs, int dimx, int usages) {
-            mAllocation = Allocation.createSized(rs, mElement, dimx, Allocation.USAGE_SCRIPT | usages);
+            mAllocation =
+                Allocation.createSized(rs, mElement, dimx,
+                                       Allocation.USAGE_SCRIPT | usages);
         }
 
         protected FieldBase() {
diff --git a/rs/jni/android_renderscript_RenderScript.cpp b/rs/jni/android_renderscript_RenderScript.cpp
index 7133a212c481..bd7925ffd10a 100644
--- a/rs/jni/android_renderscript_RenderScript.cpp
+++ b/rs/jni/android_renderscript_RenderScript.cpp
@@ -1094,159 +1094,82 @@ nScriptInvokeV(JNIEnv *_env, jobject _this, jlong con, jlong script, jint slot,
 }
 
 static void
-nScriptForEach(JNIEnv *_env, jobject _this, jlong con,
-               jlong script, jint slot, jlong ain, jlong aout)
+nScriptForEach(JNIEnv *_env, jobject _this, jlong con, jlong script, jint slot,
+               jlongArray ains, jlong aout, jbyteArray params,
+               jintArray limits)
 {
-    LOG_API("nScriptForEach, con(%p), s(%p), slot(%i)", (RsContext)con, (void *)script, slot);
-    rsScriptForEach((RsContext)con, (RsScript)script, slot, (RsAllocation)ain, (RsAllocation)aout, NULL, 0, NULL, 0);
-}
-static void
-nScriptForEachV(JNIEnv *_env, jobject _this, jlong con,
-                jlong script, jint slot, jlong ain, jlong aout, jbyteArray params)
-{
-    LOG_API("nScriptForEach, con(%p), s(%p), slot(%i)", (RsContext)con, (void *)script, slot);
-    jint len = _env->GetArrayLength(params);
-    jbyte *ptr = _env->GetByteArrayElements(params, NULL);
-    rsScriptForEach((RsContext)con, (RsScript)script, slot, (RsAllocation)ain, (RsAllocation)aout, ptr, len, NULL, 0);
-    _env->ReleaseByteArrayElements(params, ptr, JNI_ABORT);
-}
-
-static void
-nScriptForEachClipped(JNIEnv *_env, jobject _this, jlong con,
-                      jlong script, jint slot, jlong ain, jlong aout,
-                      jint xstart, jint xend,
-                      jint ystart, jint yend, jint zstart, jint zend)
-{
-    LOG_API("nScriptForEachClipped, con(%p), s(%p), slot(%i)", (RsContext)con, (void *)script, slot);
-    RsScriptCall sc;
-    sc.xStart = xstart;
-    sc.xEnd = xend;
-    sc.yStart = ystart;
-    sc.yEnd = yend;
-    sc.zStart = zstart;
-    sc.zEnd = zend;
-    sc.strategy = RS_FOR_EACH_STRATEGY_DONT_CARE;
-    sc.arrayStart = 0;
-    sc.arrayEnd = 0;
-    rsScriptForEach((RsContext)con, (RsScript)script, slot, (RsAllocation)ain, (RsAllocation)aout, NULL, 0, &sc, sizeof(sc));
-}
+    LOG_API("nScriptForEach, con(%p), s(%p), slot(%i)", (RsContext)con,
+            (void *)script, slot);
 
-static void
-nScriptForEachClippedV(JNIEnv *_env, jobject _this, jlong con,
-                       jlong script, jint slot, jlong ain, jlong aout,
-                       jbyteArray params, jint xstart, jint xend,
-                       jint ystart, jint yend, jint zstart, jint zend)
-{
-    LOG_API("nScriptForEachClipped, con(%p), s(%p), slot(%i)", (RsContext)con, (void *)script, slot);
-    jint len = _env->GetArrayLength(params);
-    jbyte *ptr = _env->GetByteArrayElements(params, NULL);
-    RsScriptCall sc;
-    sc.xStart = xstart;
-    sc.xEnd = xend;
-    sc.yStart = ystart;
-    sc.yEnd = yend;
-    sc.zStart = zstart;
-    sc.zEnd = zend;
-    sc.strategy = RS_FOR_EACH_STRATEGY_DONT_CARE;
-    sc.arrayStart = 0;
-    sc.arrayEnd = 0;
-    rsScriptForEach((RsContext)con, (RsScript)script, slot, (RsAllocation)ain, (RsAllocation)aout, ptr, len, &sc, sizeof(sc));
-    _env->ReleaseByteArrayElements(params, ptr, JNI_ABORT);
-}
-
-static void
-nScriptForEachMultiClipped(JNIEnv *_env, jobject _this, jlong con,
-                      jlong script, jint slot, jlongArray ains, jlong aout,
-                      jint xstart, jint xend,
-                      jint ystart, jint yend, jint zstart, jint zend)
-{
-    LOG_API("nScriptForEachMultiClipped, con(%p), s(%p), slot(%i)", (RsContext)con, (void *)script, slot);
-
-    jint   in_len = _env->GetArrayLength(ains);
-    jlong* in_ptr = _env->GetLongArrayElements(ains, NULL);
+    jint   in_len = 0;
+    jlong *in_ptr = NULL;
 
     RsAllocation *in_allocs = NULL;
 
-    if (sizeof(RsAllocation) == sizeof(jlong)) {
-      in_allocs = (RsAllocation*)in_ptr;
+    if (ains != NULL) {
+        in_len = _env->GetArrayLength(ains);
+        in_ptr = _env->GetLongArrayElements(ains, NULL);
 
-    } else {
-      // Convert from 64-bit jlong types to the native pointer type.
+        if (sizeof(RsAllocation) == sizeof(jlong)) {
+            in_allocs = (RsAllocation*)in_ptr;
 
-      in_allocs = new RsAllocation[in_len];
+        } else {
+            // Convert from 64-bit jlong types to the native pointer type.
 
-      for (int index = in_len; --index >= 0;) {
-        in_allocs[index] = (RsAllocation)in_ptr[index];
-      }
+            in_allocs = (RsAllocation*)alloca(in_len * sizeof(RsAllocation));
+
+            for (int index = in_len; --index >= 0;) {
+                in_allocs[index] = (RsAllocation)in_ptr[index];
+            }
+        }
     }
 
-    RsScriptCall sc;
-    sc.xStart = xstart;
-    sc.xEnd = xend;
-    sc.yStart = ystart;
-    sc.yEnd = yend;
-    sc.zStart = zstart;
-    sc.zEnd = zend;
-    sc.strategy = RS_FOR_EACH_STRATEGY_DONT_CARE;
-    sc.arrayStart = 0;
-    sc.arrayEnd = 0;
-
-    rsScriptForEachMulti((RsContext)con, (RsScript)script, slot, in_allocs, in_len, (RsAllocation)aout, NULL, 0, &sc, sizeof(sc));
-
-    if (sizeof(RsAllocation) != sizeof(jlong)) {
-      delete[] in_allocs;
+    jint   param_len = 0;
+    jbyte *param_ptr = NULL;
+
+    if (params != NULL) {
+        param_len = _env->GetArrayLength(params);
+        param_ptr = _env->GetByteArrayElements(params, NULL);
     }
 
-    _env->ReleaseLongArrayElements(ains, in_ptr, JNI_ABORT);
-}
+    RsScriptCall sc, *sca = NULL;
+    uint32_t sc_size = 0;
 
-static void
-nScriptForEachMultiClippedV(JNIEnv *_env, jobject _this, jlong con,
-                       jlong script, jint slot, jlongArray ains, jlong aout,
-                       jbyteArray params, jint xstart, jint xend,
-                       jint ystart, jint yend, jint zstart, jint zend)
-{
-    LOG_API("nScriptForEachMultiClippedV, con(%p), s(%p), slot(%i)", (RsContext)con, (void *)script, slot);
+    jint  limit_len = 0;
+    jint *limit_ptr = NULL;
 
-    jint   in_len = _env->GetArrayLength(ains);
-    jlong* in_ptr = _env->GetLongArrayElements(ains, NULL);
+    if (limits != NULL) {
+        limit_len = _env->GetArrayLength(limits);
+        limit_ptr = _env->GetIntArrayElements(limits, NULL);
 
-    RsAllocation *in_allocs = NULL;
+        assert(limit_len == 6);
 
-    if (sizeof(RsAllocation) == sizeof(jlong)) {
-      in_allocs = (RsAllocation*)in_ptr;
+        sc.xStart     = limit_ptr[0];
+        sc.xEnd       = limit_ptr[1];
+        sc.yStart     = limit_ptr[2];
+        sc.yEnd       = limit_ptr[3];
+        sc.zStart     = limit_ptr[4];
+        sc.zEnd       = limit_ptr[5];
+        sc.strategy   = RS_FOR_EACH_STRATEGY_DONT_CARE;
 
-    } else {
-      // Convert from 64-bit jlong types to the native pointer type.
+        sca = &sc;
+    }
 
-      in_allocs = new RsAllocation[in_len];
+    rsScriptForEachMulti((RsContext)con, (RsScript)script, slot,
+                         in_allocs, in_len, (RsAllocation)aout,
+                         param_ptr, param_len, sca, sc_size);
 
-      for (int index = in_len; --index >= 0;) {
-        in_allocs[index] = (RsAllocation)in_ptr[index];
-      }
+    if (ains != NULL) {
+        _env->ReleaseLongArrayElements(ains, in_ptr, JNI_ABORT);
     }
 
-    jint   param_len = _env->GetArrayLength(params);
-    jbyte* param_ptr = _env->GetByteArrayElements(params, NULL);
-
-    RsScriptCall sc;
-    sc.xStart = xstart;
-    sc.xEnd = xend;
-    sc.yStart = ystart;
-    sc.yEnd = yend;
-    sc.zStart = zstart;
-    sc.zEnd = zend;
-    sc.strategy = RS_FOR_EACH_STRATEGY_DONT_CARE;
-    sc.arrayStart = 0;
-    sc.arrayEnd = 0;
-    rsScriptForEachMulti((RsContext)con, (RsScript)script, slot, in_allocs, in_len, (RsAllocation)aout, param_ptr, param_len, &sc, sizeof(sc));
-
-    if (sizeof(RsAllocation) != sizeof(jlong)) {
-      delete[] in_allocs;
+    if (params != NULL) {
+        _env->ReleaseByteArrayElements(params, param_ptr, JNI_ABORT);
     }
 
-    _env->ReleaseLongArrayElements(ains, in_ptr, JNI_ABORT);
-    _env->ReleaseByteArrayElements(params, param_ptr, JNI_ABORT);
+    if (limits != NULL) {
+        _env->ReleaseIntArrayElements(limits, limit_ptr, JNI_ABORT);
+    }
 }
 
 // -----------------------------------
@@ -1760,12 +1683,9 @@ static JNINativeMethod methods[] = {
 {"rsnScriptSetTimeZone",             "(JJ[B)V",                               (void*)nScriptSetTimeZone },
 {"rsnScriptInvoke",                  "(JJI)V",                                (void*)nScriptInvoke },
 {"rsnScriptInvokeV",                 "(JJI[B)V",                              (void*)nScriptInvokeV },
-{"rsnScriptForEach",                 "(JJIJJ)V",                              (void*)nScriptForEach },
-{"rsnScriptForEach",                 "(JJIJJ[B)V",                            (void*)nScriptForEachV },
-{"rsnScriptForEachClipped",          "(JJIJJIIIIII)V",                        (void*)nScriptForEachClipped },
-{"rsnScriptForEachClipped",          "(JJIJJ[BIIIIII)V",                      (void*)nScriptForEachClippedV },
-{"rsnScriptForEachMultiClipped",     "(JJI[JJIIIIII)V",                       (void*)nScriptForEachMultiClipped },
-{"rsnScriptForEachMultiClipped",     "(JJI[JJ[BIIIIII)V",                     (void*)nScriptForEachMultiClippedV },
+
+{"rsnScriptForEach",                 "(JJI[JJ[B[I)V",                         (void*)nScriptForEach },
+
 {"rsnScriptSetVarI",                 "(JJII)V",                               (void*)nScriptSetVarI },
 {"rsnScriptGetVarI",                 "(JJI)I",                                (void*)nScriptGetVarI },
 {"rsnScriptSetVarJ",                 "(JJIJ)V",                               (void*)nScriptSetVarJ },