tests : add non-cont unary tests (#7857)

* tests : add non-cont unary tests

* ggml : update unary asserts and "supports_op"

ggml-ci

ggml-cuda.cu

@@ -2740,7 +2740,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
-                    return true;
+                    return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
             }

ggml-cuda/unary.cu

@@ -148,6 +148,8 @@ void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -160,6 +162,8 @@ void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -172,6 +176,8 @@ void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -184,6 +190,8 @@ void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -196,6 +204,8 @@ void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -208,6 +218,8 @@ void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -220,6 +232,8 @@ void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -232,6 +246,8 @@ void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -244,6 +260,8 @@ void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -259,6 +277,8 @@ void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 

ggml-kompute.cpp

@@ -1340,7 +1340,7 @@ static bool ggml_vk_supports_op(const struct ggml_tensor * op) {
                 case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
-                    return true;
+                    return ggml_is_contiguous(op->src[0]);
                 default:
                     ;
             }

ggml-metal.m

@@ -744,7 +744,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
-                    return true;
+                    return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
             }

ggml-sycl.cpp

@@ -17190,7 +17190,7 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
-                    return true;
+                    return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
             }

ggml-vulkan.cpp

@@ -6439,7 +6439,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
-                    return true;
+                    return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
             }

ggml.c

@@ -3212,35 +3212,42 @@ GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor) {
     return tensor->nb[0] > tensor->nb[1];
 }
 
-GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+static bool ggml_is_contiguous_n(const struct ggml_tensor * tensor, int n) {
+    size_t next_nb = ggml_type_size(tensor->type);
+    if (tensor->ne[0] != ggml_blck_size(tensor->type) && tensor->nb[0] != next_nb) {
+        return false;
+    }
+    next_nb *= tensor->ne[0]/ggml_blck_size(tensor->type);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        if (tensor->ne[i] != 1) {
+            if (i > n) {
+                if (tensor->nb[i] != next_nb) {
+                    return false;
+                }
+                next_nb *= tensor->ne[i];
+            } else {
+                // this dimension does not need to be contiguous
+                next_nb = tensor->ne[i]*tensor->nb[i];
+            }
+        }
+    }
+    return true;
+}
 
-    return
-        tensor->nb[0] == ggml_type_size(tensor->type) &&
-        tensor->nb[1] == (tensor->nb[0]*tensor->ne[0])/ggml_blck_size(tensor->type) &&
-        tensor->nb[2] == tensor->nb[1]*tensor->ne[1] &&
-        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
+GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor) {
+    return ggml_is_contiguous_0(tensor);
 }
 
 GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor) {
-    return ggml_is_contiguous(tensor);
+    return ggml_is_contiguous_n(tensor, 0);
 }
 
 GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
-
-    return
-        tensor->nb[0] == ggml_type_size(tensor->type) &&
-        tensor->nb[2] == tensor->nb[1]*tensor->ne[1] &&
-        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
+    return ggml_is_contiguous_n(tensor, 1);
 }
 
 GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
-
-    return
-        tensor->nb[0] == ggml_type_size(tensor->type) &&
-        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
+    return ggml_is_contiguous_n(tensor, 2);
 }
 
 GGML_CALL bool ggml_is_permuted(const struct ggml_tensor * tensor) {
@@ -3272,20 +3279,20 @@ bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return
-        (t0->ne[0] == t1->ne[0] ) &&
-        (t0->ne[1] == t1->ne[1] ) &&
-        (t0->ne[2] == t1->ne[2] ) &&
-        (t0->ne[3] == t1->ne[3] );
+        (t0->ne[0] == t1->ne[0]) &&
+        (t0->ne[1] == t1->ne[1]) &&
+        (t0->ne[2] == t1->ne[2]) &&
+        (t0->ne[3] == t1->ne[3]);
 }
 
 bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return
-        (t0->nb[0] == t1->nb[0] ) &&
-        (t0->nb[1] == t1->nb[1] ) &&
-        (t0->nb[2] == t1->nb[2] ) &&
-        (t0->nb[3] == t1->nb[3] );
+        (t0->nb[0] == t1->nb[0]) &&
+        (t0->nb[1] == t1->nb[1]) &&
+        (t0->nb[2] == t1->nb[2]) &&
+        (t0->nb[3] == t1->nb[3]);
 }
 
 // check if t1 can be represented as a repeatition of t0
@@ -4078,32 +4085,26 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) {
     switch (tensor->type) {
         case GGML_TYPE_I8:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int8_t));
                 return ((int8_t *)(tensor->data))[i];
             }
         case GGML_TYPE_I16:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int16_t));
                 return ((int16_t *)(tensor->data))[i];
             }
         case GGML_TYPE_I32:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int32_t));
                 return ((int32_t *)(tensor->data))[i];
             }
         case GGML_TYPE_F16:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
                 return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
             }
         case GGML_TYPE_BF16:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(ggml_bf16_t));
                 return GGML_BF16_TO_FP32(((ggml_bf16_t *)(tensor->data))[i]);
             }
         case GGML_TYPE_F32:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(float));
                 return ((float *)(tensor->data))[i];
             }
         default:
@@ -4125,32 +4126,26 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
     switch (tensor->type) {
         case GGML_TYPE_I8:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int8_t));
                 ((int8_t *)(tensor->data))[i] = value;
             } break;
         case GGML_TYPE_I16:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int16_t));
                 ((int16_t *)(tensor->data))[i] = value;
             } break;
         case GGML_TYPE_I32:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int32_t));
                 ((int32_t *)(tensor->data))[i] = value;
             } break;
         case GGML_TYPE_F16:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
                 ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(ggml_bf16_t));
                 ((ggml_bf16_t *)(tensor->data))[i] = GGML_FP32_TO_BF16(value);
             } break;
         case GGML_TYPE_F32:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(float));
                 ((float *)(tensor->data))[i] = value;
             } break;
         default:
@@ -7343,13 +7338,15 @@ struct ggml_tensor * ggml_add_rel_pos_inplace(
     return ggml_add_rel_pos_impl(ctx, a, pw, ph, true);
 }
 
-// gmml_unary
+// ggml_unary
 
 static struct ggml_tensor * ggml_unary_impl(
         struct ggml_context * ctx,
         struct ggml_tensor * a,
         enum ggml_unary_op op,
         bool inplace) {
+    GGML_ASSERT(ggml_is_contiguous_1(a));
+
     bool is_node = false;
 
     if (!inplace && (a->grad)) {
@@ -11014,6 +11011,8 @@ static void ggml_compute_forward_abs_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11023,9 +11022,6 @@ static void ggml_compute_forward_abs_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_abs_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -11060,6 +11056,8 @@ static void ggml_compute_forward_sgn_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11069,9 +11067,6 @@ static void ggml_compute_forward_sgn_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_sgn_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -11106,6 +11101,8 @@ static void ggml_compute_forward_neg_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11115,9 +11112,6 @@ static void ggml_compute_forward_neg_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_neg_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -11152,6 +11146,8 @@ static void ggml_compute_forward_step_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11161,9 +11157,6 @@ static void ggml_compute_forward_step_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_step_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -11198,6 +11191,8 @@ static void ggml_compute_forward_tanh_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11207,9 +11202,6 @@ static void ggml_compute_forward_tanh_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_tanh_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -11244,6 +11236,8 @@ static void ggml_compute_forward_elu_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11253,9 +11247,6 @@ static void ggml_compute_forward_elu_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_elu_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -11290,6 +11281,8 @@ static void ggml_compute_forward_relu_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11299,9 +11292,6 @@ static void ggml_compute_forward_relu_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_relu_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -11336,6 +11326,8 @@ static void ggml_compute_forward_sigmoid_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11345,9 +11337,6 @@ static void ggml_compute_forward_sigmoid_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_sigmoid_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -11381,9 +11370,9 @@ static void ggml_compute_forward_gelu_f32(
 
     const struct ggml_tensor * src0 = dst->src[0];
 
-    GGML_ASSERT(ggml_is_contiguous_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_1(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
         return;
@@ -11444,9 +11433,9 @@ static void ggml_compute_forward_gelu_quick_f32(
 
     const struct ggml_tensor * src0 = dst->src[0];
 
-    GGML_ASSERT(ggml_is_contiguous_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_1(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
         return;
@@ -11507,9 +11496,9 @@ static void ggml_compute_forward_silu_f32(
 
     const struct ggml_tensor * src0 = dst->src[0];
 
-    GGML_ASSERT(ggml_is_contiguous_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_1(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
         return;
@@ -11570,6 +11559,8 @@ static void ggml_compute_forward_leaky_relu_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11619,11 +11610,11 @@ static void ggml_compute_forward_silu_back_f32(
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * grad = dst->src[1];
 
-    GGML_ASSERT(ggml_is_contiguous_1(grad));
-    GGML_ASSERT(ggml_is_contiguous_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_1(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, grad));
+    assert(ggml_is_contiguous_1(grad));
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+    assert(ggml_are_same_shape(src0, grad));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
         return;
@@ -11685,6 +11676,8 @@ static void ggml_compute_forward_hardswish_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11694,9 +11687,6 @@ static void ggml_compute_forward_hardswish_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_hardswish_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -11728,6 +11718,8 @@ static void ggml_compute_forward_hardsigmoid_f32(
     const struct ggml_tensor * src0 = dst->src[0];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -11737,9 +11729,6 @@ static void ggml_compute_forward_hardsigmoid_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_hardsigmoid_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -16686,7 +16675,10 @@ static void ggml_compute_forward_map_unary_f32(
 
     const struct ggml_tensor * src0 = dst->src[0];
 
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+    assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
         return;
@@ -16695,9 +16687,6 @@ static void ggml_compute_forward_map_unary_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert( dst->nb[0] == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         fun(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -16735,6 +16724,9 @@ static void ggml_compute_forward_map_binary_f32(
     const struct ggml_tensor * src1 = dst->src[1];
 
     assert(params->ith == 0);
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(src1));
+    assert(ggml_is_contiguous_1(dst));
     assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@@ -16744,10 +16736,6 @@ static void ggml_compute_forward_map_binary_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert( dst->nb[0] == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-    assert(src1->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         fun(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),

tests/test-backend-ops.cpp

@@ -642,20 +642,29 @@ struct test_case {
 struct test_unary : public test_case {
     const ggml_unary_op op;
     const ggml_type type;
-    const std::array<int64_t, 4> ne;
+    const std::array<int64_t, 4> ne_a;
+    int v; // view (1 : non-contiguous a)
 
     std::string vars() override {
-        return VARS_TO_STR2(type, ne);
+        return VARS_TO_STR3(type, ne_a, v);
     }
 
     test_unary(ggml_unary_op op,
             ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {128, 10, 10, 10})
-        : op(op), type(type), ne(ne) {}
+            std::array<int64_t, 4> ne_a = {128, 10, 10, 10},
+            int v = 0)
+        : op(op), type(type), ne_a(ne_a), v(v) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * in = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_tensor * out = ggml_unary(ctx, in, op);
+        ggml_tensor * a;
+        if (v & 1) {
+            auto ne = ne_a; ne[0] *= 3;
+            a = ggml_new_tensor(ctx, type, 4, ne.data());
+            a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0);
+        } else {
+            a = ggml_new_tensor(ctx, type, 4, ne_a.data());
+        }
+        ggml_tensor * out = ggml_unary(ctx, a, op);
         return out;
     }
 
@@ -2016,9 +2025,11 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     };
 
     // unary ops
-    for (int op = 0; op < GGML_UNARY_OP_COUNT; op++) {
-        test_cases.emplace_back(new test_unary((ggml_unary_op) op));
-        test_cases.emplace_back(new test_unary((ggml_unary_op) op, GGML_TYPE_F32, { 7, 13, 19, 23 }));
+    for (int v : {0, 1}) {
+        for (int op = 0; op < GGML_UNARY_OP_COUNT; op++) {
+            test_cases.emplace_back(new test_unary((ggml_unary_op) op, GGML_TYPE_F32, { 128, 10, 10, 10 }, v));
+            test_cases.emplace_back(new test_unary((ggml_unary_op) op, GGML_TYPE_F32, { 7, 13, 19, 23 }, v));
+        }
     }
 
     test_cases.emplace_back(new test_get_rows(GGML_TYPE_F32, 1, 8, 2, 1, false));