Merge branch 'launch3d' into nvptx

test_static.cpp

@@ -62,17 +62,20 @@ extern "C" {
     extern void f_di(float *result, double *a, int *b);
     extern void result(float *val);
 
-    void ISPCLaunch(void **handlePtr, void *f, void *d, int);
+    void ISPCLaunch(void **handlePtr, void *f, void *d, int,int,int);
     void ISPCSync(void *handle);
     void *ISPCAlloc(void **handlePtr, int64_t size, int32_t alignment);
 }
-
-void ISPCLaunch(void **handle, void *f, void *d, int count) {
+ 
+void ISPCLaunch(void **handle, void *f, void *d, int count0, int count1, int count2) {
     *handle = (void *)0xdeadbeef;
-    typedef void (*TaskFuncType)(void *, int, int, int, int);
+    typedef void (*TaskFuncType)(void *, int, int, int, int, int,int,int, int,int,int);
     TaskFuncType func = (TaskFuncType)f;
-    for (int i = 0; i < count; ++i)
-        func(d, 0, 1, i, count);
+    int count = count0*count1*count2, idx = 0;
+    for (int k = 0; k < count2; ++k)
+      for (int j = 0; j < count1; ++j)
+        for (int i = 0; i < count0; ++i)
+        func(d, 0, 1, idx++, count, i,j,k,count0,count1,count2);
 }
 
 void ISPCSync(void *) {

tests/launch-8.ispc

@@ -0,0 +1,42 @@
+
+export uniform int width() { return programCount; }
+
+
+#define N0 10
+#define N1 20
+#define N2 50
+static uniform float array[N2][N1][N0];
+
+task void x(const float f) {
+    uniform int j;
+
+    assert(taskCount  == N0*N1*N2);
+    assert(taskCount0 == N0);
+    assert(taskCount1 == N1);
+    assert(taskCount2 == N2);
+    assert(taskIndex  == taskIndex0 + N0*(taskIndex1 + N1*taskIndex2));
+    assert(taskIndex0 < N0);
+    assert(taskIndex1 < N1);
+    assert(taskIndex2 < N2);
+
+    const uniform int i0 = taskIndex0;
+    const uniform int i1 = taskIndex1;
+    const uniform int i2 = taskIndex2;
+    const uniform int i  = taskIndex;
+    array[i2][i1][i0] = i / 10000.;
+    cfor (j = 0; j < 10000; ++j)
+        array[i2][i1][i0] = sin(array[i2][i1][i0]);
+    if (array[i2][i1][i0] < .02)
+        array[i2][i1][i0] = i;
+}
+export void f_f(uniform float RET[], uniform float fFOO[]) { 
+    float f = fFOO[programIndex];
+    launch[N2][N1][N0] x(f);
+    sync;
+    RET[programIndex] = array[N2-1][N1-1][N0-1];
+}
+
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 9999.000000;
+}

tests/launch-9.ispc

@@ -0,0 +1,42 @@
+
+export uniform int width() { return programCount; }
+
+
+#define N0 10
+#define N1 20
+#define N2 50
+static uniform float array[N2][N1][N0];
+
+task void x(const float f) {
+    uniform int j;
+
+    assert(taskCount  == N0*N1*N2);
+    assert(taskCount0 == N0);
+    assert(taskCount1 == N1);
+    assert(taskCount2 == N2);
+    assert(taskIndex  == taskIndex0 + N0*(taskIndex1 + N1*taskIndex2));
+    assert(taskIndex0 < N0);
+    assert(taskIndex1 < N1);
+    assert(taskIndex2 < N2);
+
+    const uniform int i0 = taskIndex0;
+    const uniform int i1 = taskIndex1;
+    const uniform int i2 = taskIndex2;
+    const uniform int i  = taskIndex;
+    array[i2][i1][i0] = i / 10000.;
+    cfor (j = 0; j < 10000; ++j)
+        array[i2][i1][i0] = sin(array[i2][i1][i0]);
+    if (array[i2][i1][i0] < .02)
+        array[i2][i1][i0] = i;
+}
+export void f_f(uniform float RET[], uniform float fFOO[]) { 
+    float f = fFOO[programIndex];
+    launch[N0,N1,N2] x(f);
+    sync;
+    RET[programIndex] = array[N2-1][N1-1][N0-1];
+}
+
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 9999.000000;
+}