feat: implement backward computation for more operators

include/ggml.h

@@ -488,6 +488,7 @@ extern "C" {
         GGML_OP_ROPE,
         GGML_OP_ROPE_BACK,
         GGML_OP_CLAMP,
+        GGML_OP_CLAMP_BACK,
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
         GGML_OP_CONV_TRANSPOSE_2D,
@@ -535,10 +536,13 @@ extern "C" {
         GGML_UNARY_OP_STEP,
         GGML_UNARY_OP_TANH,
         GGML_UNARY_OP_ELU,
+        GGML_UNARY_OP_ELU_BACK,
         GGML_UNARY_OP_RELU,
         GGML_UNARY_OP_SIGMOID,
         GGML_UNARY_OP_GELU,
+        GGML_UNARY_OP_GELU_BACK,
         GGML_UNARY_OP_GELU_QUICK,
+        GGML_UNARY_OP_GELU_QUICK_BACK,
         GGML_UNARY_OP_SILU,
         GGML_UNARY_OP_HARDSWISH,
         GGML_UNARY_OP_HARDSIGMOID,
@@ -1074,6 +1078,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_elu_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_elu_back_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_relu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1102,6 +1114,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_gelu_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_back_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_gelu_quick(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1110,6 +1130,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_gelu_quick_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_quick_back_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_silu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1575,13 +1603,30 @@ extern "C" {
             float                 beta_slow);
 
     // clamp
-    // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_clamp(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             float                 min,
             float                 max);
 
+    GGML_API struct ggml_tensor * ggml_clamp_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
+
+    GGML_API struct ggml_tensor * ggml_clamp_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
+
+    GGML_API struct ggml_tensor * ggml_clamp_back_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
+
     GGML_API struct ggml_tensor * ggml_im2col(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

src/ggml-cuda.cu

@@ -2276,6 +2276,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_CLAMP:
             ggml_cuda_op_clamp(ctx, dst);
             break;
+        case GGML_OP_CLAMP_BACK:
+            ggml_cuda_op_clamp_back(ctx, dst);
+            break;
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
@@ -2868,6 +2871,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
+        case GGML_OP_CLAMP_BACK:
         case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:

src/ggml-cuda/clamp.cu

@@ -15,6 +15,21 @@ static void clamp_f32_cuda(const float * x, float * dst, const float min, const
     clamp_f32<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
 }
 
+static __global__ void clamp_back_f32(const float * x, float * dst, const float min, const float max, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = x[i] < min || x[i] > max ? 0.0f : 1.0f;
+}
+
+static void clamp_back_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_CLAMP_BACK_BLOCK_SIZE - 1) / CUDA_CLAMP_BACK_BLOCK_SIZE;
+    clamp_back_f32<<<num_blocks, CUDA_CLAMP_BACK_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
+}
+
 
 void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
@@ -32,3 +47,20 @@ void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     clamp_f32_cuda(src0_d, dst_d, min, max, ggml_nelements(src0), stream);
 }
+
+void ggml_cuda_op_clamp_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float min;
+    float max;
+    memcpy(&min, dst->op_params, sizeof(float));
+    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
+
+    clamp_back_f32_cuda(src0_d, dst_d, min, max, ggml_nelements(src0), stream);
+}

src/ggml-cuda/clamp.cuh

@@ -1,5 +1,7 @@
 #include "common.cuh"
 
 #define CUDA_CLAMP_BLOCK_SIZE 256
+#define CUDA_CLAMP_BACK_BLOCK_SIZE 256
 
 void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_clamp_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);