CANN: RoPE operator optimization (#10563)

* [cann] RoPE operator optimization

* [CANN]Code Formatting

---------

Co-authored-by: noemotiovon <noemotiovon@gmail.com>

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -2965,7 +2965,7 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
                              aclTensor* acl_cos_repeat_tensor,
                              aclTensor* acl_sin_repeat_tensor,
                              float theta_scale, float freq_scale,
-                             bool is_neox) {
+                             float attn_factor, bool is_neox) {
     // int sin/cos cache, cache has different repeat method depond on
     // @param.is_neox
 
@@ -3017,6 +3017,7 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
             ggml_type_size(src2->type), arange_ne, arange_nb, GGML_MAX_DIMS);
         aclnn_div_tensor(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor,
                          nullptr, true);
+        ACL_CHECK(aclDestroyTensor(acl_freq_factors_tensor));
     }
 
     // position
@@ -3047,16 +3048,6 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
               acl_theta_tensor);
 
-    // // power[] * position[] * freq_scale / freq_factors[]
-    // ggml_cann_pool_alloc theta_final_allocator(ctx.pool(),
-    //                                            theta_length *
-    //                                            sizeof(float_t));
-    // aclTensor* acl_theat_final_tensor = aclnn_zero(
-    //     ctx, theta_final_allocator.get(), sizeof(float_t) * theta_length,
-    //     theta_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t));
-    // aclnn_inplace_addcdiv(ctx, acl_theat_final_tensor, acl_theta_tensor,
-    //                       acl_freq_factors_tensor, freq_scale);
-
     // permute: [0,1,2,3]->[0,2,1,3]
     int64_t permute_ne[] = {arange_length, 1, position_length, 1};
     size_t permute_nb[GGML_MAX_DIMS];
@@ -3092,6 +3083,12 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclnn_cos(ctx, acl_permute_tensor, acl_cos_tensor);
 
+    // attn_factor
+    if (attn_factor != 1) {
+        aclnn_muls(ctx, acl_sin_tensor, attn_factor, nullptr, true);
+        aclnn_muls(ctx, acl_cos_tensor, attn_factor, nullptr, true);
+    }
+
     // repeat
     if (is_neox) {
         int64_t repeatsArray[] = {1, 1, 1, 2};
@@ -3155,15 +3152,11 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     memcpy(&beta_fast, (int32_t*)dst->op_params + 9, sizeof(float));
     memcpy(&beta_slow, (int32_t*)dst->op_params + 10, sizeof(float));
 
-    // TODO: attn_factor != 1
-    GGML_ASSERT(attn_factor == 1);
     // TODO: n_dims <= ne0
     GGML_ASSERT(n_dims == ne0);
     GGML_ASSERT(n_dims % 2 == 0);
     // TODO: ext_factor != 0
     GGML_ASSERT(ext_factor == 0);
-    // TODO: type == GGML_TYPE_F16
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
@@ -3194,7 +3187,217 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float_t),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
     aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor,
-                     theta_scale, freq_scale, is_neox);
+                     theta_scale, freq_scale, attn_factor, is_neox);
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src0);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+#ifdef ASCEND_310P
+    // Special ROPE operation for 310P
+
+    // roll input
+    void* input_roll_buffer;
+    aclTensor* acl_minus_one_tensor;
+    void* minus_one_scale_buffer = nullptr;
+    ggml_cann_pool_alloc roll_allocator(ctx.pool(), ggml_nbytes(src0));
+    ggml_cann_pool_alloc minus_one_scale_allocator(
+        ctx.pool(), sizeof(float_t) * src0->ne[0]);
+    if (!is_neox) {
+        // roll input: [q0,q1,q2,q3,...] -> [q1,q0,q3,q2,...]
+        input_roll_buffer = roll_allocator.get();
+        int64_t input_roll_ne[4] = {2, src0->ne[1] * (src0->ne[0] / 2),
+                                    src0->ne[2], src0->ne[3]};
+        size_t input_roll_nb[GGML_MAX_DIMS];
+        input_roll_nb[0] = ggml_type_size(src0->type);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            input_roll_nb[i] = input_roll_nb[i - 1] * input_roll_ne[i - 1];
+        }
+        aclTensor* acl_input_roll_tensor = ggml_cann_create_tensor(
+            input_roll_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), input_roll_ne, input_roll_nb,
+            GGML_MAX_DIMS);
+        aclTensor* acl_input_tensor = ggml_cann_create_tensor(
+            src0->data, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), input_roll_ne, input_roll_nb,
+            GGML_MAX_DIMS);
+
+        int64_t shifts[] = {1};
+        int64_t dims[] = {3};
+        aclnn_roll(ctx, acl_input_tensor, acl_input_roll_tensor, shifts, dims);
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+
+        // init [-1, 1, -1, 1, ...]
+        minus_one_scale_buffer = minus_one_scale_allocator.get();
+
+        int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
+        size_t minus_one_nb[GGML_MAX_DIMS];
+        minus_one_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
+        }
+        acl_minus_one_tensor = aclnn_values(
+            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
+            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
+        int64_t dim = 3;
+        int64_t* index = new int64_t[src0->ne[0]];
+        for (int i = 0; i < src0->ne[0]; i++) {
+            index[i] = i / 2 * 2;
+        }
+        int64_t index_num = src0->ne[0];
+        float value = -1;
+        aclnn_index_fill_tensor(ctx, acl_minus_one_tensor, dim, index,
+                                index_num, value);
+    } else {
+        // roll input: [q0,q1,q2,...] ->
+        // [q_half,q_half+1,...,q_end,q0,q1,...q_half-1]
+        input_roll_buffer = roll_allocator.get();
+        aclTensor* acl_input_roll_tensor = ggml_cann_create_tensor(
+            input_roll_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), src0->ne, src0->nb, GGML_MAX_DIMS);
+        aclTensor* acl_input_tensor = ggml_cann_create_tensor(src0);
+
+        int64_t shifts[] = {src0->ne[0] / 2};
+        int64_t dims[] = {3};
+        aclnn_roll(ctx, acl_input_tensor, acl_input_roll_tensor, shifts, dims);
+
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+        // init [-1, -1, -1, 1, 1，1，...]
+        minus_one_scale_buffer = minus_one_scale_allocator.get();
+        int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
+        size_t minus_one_nb[GGML_MAX_DIMS];
+        minus_one_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
+        }
+        acl_minus_one_tensor = aclnn_values(
+            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
+            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
+        // -1 * first half
+        int64_t first_half_ne[4] = {src0->ne[0] / 2, 1, 1, 1};
+        size_t first_half_nb[GGML_MAX_DIMS];
+        first_half_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            first_half_nb[i] = first_half_nb[i - 1] * first_half_ne[i - 1];
+        }
+        aclTensor* acl_first_half_tensor = ggml_cann_create_tensor(
+            minus_one_scale_buffer, ACL_FLOAT, sizeof(float_t), first_half_ne,
+            first_half_nb, GGML_MAX_DIMS);
+        bool inplace = true;
+        float scale = -1;
+        aclnn_muls(ctx, acl_first_half_tensor, scale, nullptr, inplace);
+        ACL_CHECK(aclDestroyTensor(acl_first_half_tensor));
+    }
+
+    // TODO: n_dims < ne0
+    GGML_ASSERT(n_dims == src0->ne[0]);
+
+    // input * scale
+    ggml_cann_pool_alloc roll_mul_scale_allocator(ctx.pool(),
+                                                  ggml_nbytes(src0));
+    void* input_roll_mul_scale_buffer = roll_mul_scale_allocator.get();
+    size_t input_nb[GGML_MAX_DIMS];
+    input_nb[0] = ggml_type_size(src0->type);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        input_nb[i] = input_nb[i - 1] * src0->ne[i - 1];
+    }
+    aclTensor* acl_input_roll_mul_scale_tensor = ggml_cann_create_tensor(
+        input_roll_mul_scale_buffer, ggml_cann_type_mapping(src0->type),
+        ggml_type_size(src0->type), src0->ne, input_nb, GGML_MAX_DIMS);
+    aclTensor* acl_input_roll_reshape_tensor = ggml_cann_create_tensor(
+        input_roll_buffer, ggml_cann_type_mapping(src0->type),
+        ggml_type_size(src0->type), src0->ne, input_nb, GGML_MAX_DIMS);
+
+    aclnn_mul(ctx, acl_input_roll_reshape_tensor, acl_minus_one_tensor,
+              acl_input_roll_mul_scale_tensor);
+
+    // output
+    void* output_fp32_buffer;
+    if (src0->type == GGML_TYPE_F32) {
+        aclnn_inplace_mul(ctx, acl_src, acl_cos_reshape_tensor);
+        aclnn_inplace_mul(ctx, acl_input_roll_mul_scale_tensor,
+                          acl_sin_reshape_tensor);
+        aclnn_add(ctx, acl_src, acl_input_roll_mul_scale_tensor, acl_dst);
+        // TODO: ne0 != n_dims in mode2
+    } else if (src0->type == GGML_TYPE_F16) {
+        size_t input_fp32_nb[GGML_MAX_DIMS];
+        input_fp32_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            input_fp32_nb[i] = input_fp32_nb[i - 1] * dst->ne[i - 1];
+        }
+        ggml_cann_pool_alloc fp32_allocator1(
+            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+        void* input_fp32_buffer1 = fp32_allocator1.get();
+        aclTensor* input_fp32_tensor1 = ggml_cann_create_tensor(
+            input_fp32_buffer1, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_nb, GGML_MAX_DIMS);
+        ggml_cann_pool_alloc fp32_allocator2(
+            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+        void* input_fp32_buffer2 = fp32_allocator2.get();
+        aclTensor* input_fp32_tensor2 = ggml_cann_create_tensor(
+            input_fp32_buffer2, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_nb, GGML_MAX_DIMS);
+
+        ggml_cann_pool_alloc fp32_allocator(
+            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+        output_fp32_buffer = fp32_allocator.get();
+        aclTensor* output_fp32_tensor = ggml_cann_create_tensor(
+            output_fp32_buffer, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_nb, GGML_MAX_DIMS);
+        aclnn_mul(ctx, acl_src, acl_cos_reshape_tensor, input_fp32_tensor1);
+        aclnn_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor,
+                  input_fp32_tensor2);
+        aclnn_add(ctx, input_fp32_tensor1, input_fp32_tensor2,
+                  output_fp32_tensor);
+        aclnn_cast(ctx, output_fp32_tensor, acl_dst, ACL_FLOAT16);
+
+        ACL_CHECK(aclDestroyTensor(input_fp32_tensor1));
+        ACL_CHECK(aclDestroyTensor(input_fp32_tensor2));
+        ACL_CHECK(aclDestroyTensor(output_fp32_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_src));
+    }
+    return;
+#endif
+
+    // src0 == GGML_TYPE_F16
+    // TODO: optimization this `if` code
+    if (src0->type == GGML_TYPE_F16) {
+        ggml_cann_pool_alloc sin_final_allocator(
+            ctx.pool(), src0->ne[0] * src0->ne[2] * ggml_type_size(src0->type));
+        ggml_cann_pool_alloc cos_final_allocator(
+            ctx.pool(), src0->ne[0] * src0->ne[2] * ggml_type_size(src0->type));
+        void* sin_final_buffer = sin_final_allocator.get();
+        void* cos_final_buffer = cos_final_allocator.get();
+
+        int64_t sin_final_ne[4] = {src0->ne[0], 1, src0->ne[2], 1};
+        size_t sin_final_nb[GGML_MAX_DIMS];
+        sin_final_nb[0] = ggml_type_size(src0->type);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            sin_final_nb[i] = sin_final_nb[i - 1] * sin_final_ne[i - 1];
+        }
+        aclTensor* acl_sin_final_tensor = ggml_cann_create_tensor(
+            sin_final_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), sin_final_ne, sin_final_nb,
+            GGML_MAX_DIMS);
+        aclTensor* acl_cos_final_tensor = ggml_cann_create_tensor(
+            cos_final_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), sin_final_ne, sin_final_nb,
+            GGML_MAX_DIMS);
+
+        aclnn_cast(ctx, acl_sin_reshape_tensor, acl_sin_final_tensor,
+                   ggml_cann_type_mapping(src0->type));
+        aclnn_cast(ctx, acl_cos_reshape_tensor, acl_cos_final_tensor,
+                   ggml_cann_type_mapping(src0->type));
+        ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+        acl_sin_reshape_tensor = acl_sin_final_tensor;
+        acl_cos_reshape_tensor = acl_cos_final_tensor;
+    }
 
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
@@ -3206,10 +3409,8 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         acl_mode = 1;
     }
 
-    aclTensor* acl_x = ggml_cann_create_tensor(src0);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
     ACL_CHECK(aclnnRotaryPositionEmbeddingGetWorkspaceSize(
-        acl_x, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode,
+        acl_src, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode,
         acl_dst, &workspaceSize, &executor));
     if (workspaceSize > 0) {
         ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
@@ -3219,7 +3420,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ACL_CHECK(aclnnRotaryPositionEmbedding(workspaceAddr, workspaceSize,
                                            executor, ctx.stream()));
 
-    ACL_CHECK(aclDestroyTensor(acl_x));
+    ACL_CHECK(aclDestroyTensor(acl_src));
     ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
     ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
     ACL_CHECK(aclDestroyTensor(acl_dst));

ggml/src/ggml-cann/ggml-cann.cpp

@@ -1739,7 +1739,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_ROPE: {
             // TODO: with ops-test v == 1
             float * ext_factor = (float*)((int32_t*)op->op_params + 7);
-            float * attn_factor = (float*)((int32_t*)op->op_params + 8);
             // TODO: n_dims <= ne0
             if (op->src[0]->ne[0] != op->op_params[1]) {
                 return false;
@@ -1748,17 +1747,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if (*ext_factor != 0) {
                 return false;
             }
-            // TODO: attn_factor != 1
+            return true;
-            if (*attn_factor != 1) {
-                return false;
-            }
-            //TODO: type == GGML_TYPE_F16
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                    return true;
-                default:
-                    return false;
-            }
         }
         case GGML_OP_UPSCALE: {
             // aclnnUpsampleNearest2dGetWorkspaceSize not support