END_TEST
START_TEST(test_gemmBatch_3d_F) {
GpuArray A;
GpuArray B;
GpuArray C;
size_t dims[3] = {2, 3, 3};
float data[] = {1, 2, 3, 4, 5, 6, 7, 8, 9,
1, 2, 3, 4, 5, 6, 7, 8, 9};
const float res[] = {42, 78, 78, 60, 114, 114, 51, 69, 96,
66, 39, 111, 54, 54, 90, 78, 78, 132};
ga_assert_ok(GpuArray_empty(&A, ctx, GA_FLOAT, 3, dims, GA_F_ORDER));
ga_assert_ok(GpuArray_empty(&B, ctx, GA_FLOAT, 3, dims, GA_F_ORDER));
ga_assert_ok(GpuArray_empty(&C, ctx, GA_FLOAT, 3, dims, GA_C_ORDER));
ga_assert_ok(GpuArray_write(&A, data, sizeof(data)));
ga_assert_ok(GpuArray_write(&B, data, sizeof(data)));
ga_assert_ok(GpuArray_rgemmBatch_3d(cb_no_trans, cb_no_trans, 1, &A, &B, 0, &C, 0));
ga_assert_ok(GpuArray_read(data, sizeof(data), &C));
ck_assert_fbuf_eq(data, res, sizeof(res)/sizeof(float));
}
END_TEST
START_TEST(test_take1_offset) {
const uint32_t data[4] = {0, 1, 2, 3};
const size_t data_dims[1] = {4};
const size_t out_dims[1] = {2};
const uint32_t idx[4] = {20, 3, 3, 2};
GpuArray v;
GpuArray i;
GpuArray r;
ga_assert_ok(GpuArray_empty(&v, ctx, GA_UINT, 1, data_dims, GA_C_ORDER));
ga_assert_ok(GpuArray_write(&v, data, sizeof(data)));
ga_assert_ok(GpuArray_empty(&i, ctx, GA_UINT, 1, data_dims, GA_C_ORDER));
ga_assert_ok(GpuArray_write(&i, idx, sizeof(idx)));
ga_assert_ok(GpuArray_empty(&r, ctx, GA_UINT, 1, out_dims, GA_C_ORDER));
/* Fake subtensor for offset */
i.offset = 8;
i.dimensions[0] = 2;
ga_assert_ok(GpuArray_take1(&r, &v, &i, 1));
/* The actual results are not important, this is just to check that
we don't trigger the out of bounds check */
}
END_TEST
START_TEST(test_basic_simple)
{
GpuArray a;
GpuArray b;
GpuArray c;
GpuElemwise *ge;
static const uint32_t data1[3] = {1, 2, 3};
static const uint32_t data2[3] = {4, 5, 6};
uint32_t data3[3] = {0};
size_t dims[2];
gpuelemwise_arg args[3] = {{0}};
void *rargs[3];
dims[0] = 1;
dims[1] = 3;
ga_assert_ok(GpuArray_empty(&a, ctx, GA_UINT, 2, dims, GA_C_ORDER));
ga_assert_ok(GpuArray_write(&a, data1, sizeof(data1)));
ga_assert_ok(GpuArray_empty(&b, ctx, GA_UINT, 2, dims, GA_F_ORDER));
ga_assert_ok(GpuArray_write(&b, data2, sizeof(data2)));
ga_assert_ok(GpuArray_empty(&c, ctx, GA_UINT, 2, dims, GA_C_ORDER));
args[0].name = "a";
args[0].typecode = GA_UINT;
args[0].flags = GE_READ;
args[1].name = "b";
args[1].typecode = GA_UINT;
args[1].flags = GE_READ;
args[2].name = "c";
args[2].typecode = GA_UINT;
args[2].flags = GE_WRITE;
ge = GpuElemwise_new(ctx, "", "c = a + b", 3, args, 2, 0);
ck_assert_ptr_ne(ge, NULL);
rargs[0] = &a;
rargs[1] = &b;
rargs[2] = &c;
ga_assert_ok(GpuElemwise_call(ge, rargs, GE_NOCOLLAPSE));
ga_assert_ok(GpuArray_read(data3, sizeof(data3), &c));
ck_assert_int_eq(data3[0], 5);
ck_assert_int_eq(data3[1], 7);
ck_assert_int_eq(data3[2], 9);
}
END_TEST
START_TEST(test_gemmBatch_3d_S) {
GpuArray A;
GpuArray B;
GpuArray C;
ssize_t t;
size_t dims[3] = {2, 3, 3};
float data[] = {1, 2, 3, 4, 5, 6, 7, 8, 9,
1, 2, 3, 4, 5, 6, 7, 8, 9};
const float res[] = {14, 32, 50, 50, 122, 194, 32, 77, 122,
26, 62, 98, 17, 53, 89, 44, 107, 170};
ga_assert_ok(GpuArray_empty(&A, ctx, GA_FLOAT, 3, dims, GA_F_ORDER));
ga_assert_ok(GpuArray_empty(&B, ctx, GA_FLOAT, 3, dims, GA_C_ORDER));
ga_assert_ok(GpuArray_empty(&C, ctx, GA_FLOAT, 3, dims, GA_C_ORDER));
ga_assert_ok(GpuArray_write(&A, data, sizeof(data)));
ga_assert_ok(GpuArray_write(&B, data, sizeof(data)));
A.strides[0] = 8;
A.strides[1] = 24;
A.strides[2] = 4;
GpuArray_fix_flags(&A);
t = B.strides[1];
B.strides[1] = B.strides[2];
B.strides[2] = t;
GpuArray_fix_flags(&B);
ga_assert_ok(GpuArray_rgemmBatch_3d(cb_no_trans, cb_no_trans, 1, &A, &B, 0, &C, 1));
ga_assert_ok(GpuArray_read(data, sizeof(data), &C));
ck_assert_fbuf_eq(data, res, sizeof(res)/sizeof(float));
}
END_TEST
START_TEST(test_basic_neg_strides)
{
GpuArray a;
GpuArray b;
GpuArray c;
GpuElemwise *ge;
static const uint32_t data1[6] = {1, 2, 3, 4, 5, 6};
static const uint32_t data2[6] = {7, 8, 9, 10, 11, 12};
uint32_t data3[6] = {0};
size_t dims[1];
gpuelemwise_arg args[3] = {{0}};
void *rargs[3];
ssize_t starts[1];
ssize_t stops[1];
ssize_t steps[1];
dims[0] = 6;
ga_assert_ok(GpuArray_empty(&a, ctx, GA_UINT, 1, dims, GA_C_ORDER));
ga_assert_ok(GpuArray_write(&a, data1, sizeof(data1)));
ga_assert_ok(GpuArray_empty(&b, ctx, GA_UINT, 1, dims, GA_C_ORDER));
ga_assert_ok(GpuArray_write(&b, data2, sizeof(data2)));
starts[0] = 5;
stops[0] = -1;
steps[0] = -1;
ga_assert_ok(GpuArray_index_inplace(&b, starts, stops, steps));
ga_assert_ok(GpuArray_empty(&c, ctx, GA_UINT, 1, dims, GA_C_ORDER));
args[0].name = "a";
args[0].typecode = GA_UINT;
args[0].flags = GE_READ;
args[1].name = "b";
args[1].typecode = GA_UINT;
args[1].flags = GE_READ;
args[2].name = "c";
args[2].typecode = GA_UINT;
args[2].flags = GE_WRITE;
ge = GpuElemwise_new(ctx, "", "c = a + b", 3, args, 1, 0);
ck_assert_ptr_ne(ge, NULL);
rargs[0] = &a;
rargs[1] = &b;
rargs[2] = &c;
ga_assert_ok(GpuElemwise_call(ge, rargs, 0));
ga_assert_ok(GpuArray_read(data3, sizeof(data3), &c));
ck_assert_int_eq(data3[0], 13);
ck_assert_int_eq(data3[1], 13);
ck_assert_int_eq(data3[2], 13);
ck_assert_int_eq(data3[3], 13);
ck_assert_int_eq(data3[4], 13);
ck_assert_int_eq(data3[5], 13);
}
END_TEST
START_TEST(test_basic_remove1)
{
GpuArray a;
GpuArray b;
GpuArray c;
GpuElemwise *ge;
static const uint32_t data1[6] = {1, 2, 3, 4, 5, 6};
static const uint32_t data2[6] = {7, 8, 9, 10, 11, 12};
uint32_t data3[6] = {0};
size_t dims[4];
gpuelemwise_arg args[3] = {{0}};
void *rargs[3];
dims[0] = 1;
dims[1] = 3;
dims[2] = 2;
dims[3] = 1;
ga_assert_ok(GpuArray_empty(&a, ctx, GA_UINT, 4, dims, GA_C_ORDER));
ga_assert_ok(GpuArray_write(&a, data1, sizeof(data1)));
ga_assert_ok(GpuArray_empty(&b, ctx, GA_UINT, 4, dims, GA_F_ORDER));
ga_assert_ok(GpuArray_write(&b, data2, sizeof(data2)));
ga_assert_ok(GpuArray_empty(&c, ctx, GA_UINT, 4, dims, GA_C_ORDER));
args[0].name = "a";
args[0].typecode = GA_UINT;
args[0].flags = GE_READ;
args[1].name = "b";
args[1].typecode = GA_UINT;
args[1].flags = GE_READ;
args[2].name = "c";
args[2].typecode = GA_UINT;
args[2].flags = GE_WRITE;
ge = GpuElemwise_new(ctx, "", "c = a + b", 3, args, 0, 0);
ck_assert_ptr_ne(ge, NULL);
rargs[0] = &a;
rargs[1] = &b;
rargs[2] = &c;
ga_assert_ok(GpuElemwise_call(ge, rargs, 0));
ga_assert_ok(GpuArray_read(data3, sizeof(data3), &c));
ck_assert_int_eq(data3[0], 8);
ck_assert_int_eq(data3[1], 12);
ck_assert_int_eq(data3[2], 11);
ck_assert_int_eq(data3[3], 15);
ck_assert_int_eq(data3[4], 14);
ck_assert_int_eq(data3[5], 18);
}