Fixed single GPU performance regression

Author: JohannesGaessler

ggml-cuda.cu

@@ -934,6 +934,30 @@ bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_te
     return false;
 }
 
+bool ggml_cuda_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
+    size_t src0_sz = ggml_nbytes(src0);
+    size_t src1_sz = ggml_nbytes(src1);
+
+    // mul_mat_q: src0 is converted to fp32 on device
+    size_t mul_mat_q_transfer = src0_sz + src1_sz;
+
+    // mul_mat_f16: src1 is converted to fp16 on cpu
+    size_t mul_mat_f16_transfer = src0_sz + sizeof(half) * ggml_nelements(src1);
+
+    // choose the smaller one to transfer to the device
+    // TODO: this is not always the best choice due to the overhead of converting to fp16
+    return mul_mat_f16_transfer < mul_mat_q_transfer;
+}
+
+size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+    if (ggml_cuda_mul_mat_use_f16(src0, src1, dst)) {
+        return ggml_nelements(src1) * sizeof(ggml_fp16_t);
+    }
+    else {
+        return 0;
+    }
+}
+
 void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(ggml_cuda_can_mul_mat(src0, src1, dst));
 
@@ -950,6 +974,99 @@ void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_
     }
 }
 
+static void ggml_cuda_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t /* wsize */) {
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+
+    const int nb10 = src1->nb[0];
+    const int nb11 = src1->nb[1];
+    const int nb12 = src1->nb[2];
+    const int nb13 = src1->nb[3];
+
+    const int nb2  = dst->nb[2];
+    const int nb3  = dst->nb[3];
+
+    const float alpha = 1.0f;
+    const float beta = 0.0f;
+    const int x_ne = ne01 * ne00;
+    const int y_ne = ne11 * ne10;
+    const int d_ne = ne11 * ne01;
+    const int n_mm = ne03 * ne02;
+
+    size_t x_size, y_size, d_size;
+    half  * d_X =  (half *) ggml_cuda_pool_malloc(n_mm * sizeof(half) * x_ne, &x_size);
+    half  * d_Y =  (half *) ggml_cuda_pool_malloc(n_mm * sizeof(half) * y_ne, &y_size);
+    float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size);
+
+    bool src1_cont_rows = nb10 == sizeof(float);
+    bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            int i = i03*ne02 + i02;
+            cudaStream_t cudaStream = g_cudaStreams_main[0][i % GGML_CUDA_MAX_STREAMS];
+
+            half  * c_X = d_X + i * x_ne;
+            half  * c_Y = d_Y + i * y_ne;
+            float * c_D = d_D + i * d_ne;
+
+            // copy src0 to device
+            CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_X, src0, i03, i02, 0, ne01, cudaStream));
+
+            // convert src1 to fp16
+            // TODO: use multiple threads
+            ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i03 * ne02 + i02);
+            char * src1i = (char *) src1->data + i03*nb13 + i02*nb12;
+            if (src1_cont_rows) {
+                if (src1_cont_cols) {
+                    ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
+                }
+                else {
+                    for (int64_t i01 = 0; i01 < ne11; i01++) {
+                        ggml_fp32_to_fp16_row((float *) (src1i + i01*nb11), tmp + i01*ne10, ne10);
+                    }
+                }
+            }
+            else {
+                for (int64_t i01 = 0; i01 < ne11; i01++) {
+                    for (int64_t i00 = 0; i00 < ne10; i00++) {
+                        // very slow due to no inlining
+                        tmp[i01*ne10 + i00] = ggml_fp32_to_fp16(*(float *) (src1i + i01*nb11 + i00*nb10));
+                    }
+                }
+            }
+
+            // copy src1 to device
+            CUDA_CHECK(cudaMemcpyAsync(c_Y, tmp, sizeof(half) * y_ne, cudaMemcpyHostToDevice, cudaStream));
+
+            // compute
+            CUBLAS_CHECK(cublasSetStream(g_cublas_handles[0], cudaStream));
+            CUBLAS_CHECK(
+                cublasGemmEx(g_cublas_handles[0], CUBLAS_OP_T, CUBLAS_OP_N,
+                        ne01, ne11, ne10,
+                        &alpha, c_X, CUDA_R_16F, ne00,
+                                c_Y, CUDA_R_16F, ne10,
+                        &beta,  c_D, CUDA_R_32F, ne01,
+                        CUBLAS_COMPUTE_32F_FAST_16F,
+                        CUBLAS_GEMM_DEFAULT));
+
+            // copy dst to host
+            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            CUDA_CHECK(cudaMemcpyAsync(d, c_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream));
+        }
+    }
+
+    CUDA_CHECK(cudaDeviceSynchronize());
+    ggml_cuda_pool_free(d_X, x_size);
+    ggml_cuda_pool_free(d_Y, y_size);
+    ggml_cuda_pool_free(d_D, d_size);
+}
+
 void ggml_cuda_load_data(const char * fname, struct ggml_tensor * tensor, const size_t offset, int n_layer) {
     FILE * fp = fopen(fname, "rb");
     int nrows = ggml_nrows(tensor);
@@ -1054,6 +1171,19 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
             if (!ggml_cuda_can_mul_mat(tensor->src0, tensor->src1, tensor)) {
                 return false;
             }
+            if (g_device_count == 1 && tensor->src0->type == GGML_TYPE_F16 &&
+                ggml_cuda_mul_mat_use_f16(tensor->src0, tensor->src1, tensor)) {
+
+                if (params->ith != 0) {
+                    return true;
+                }
+                if (params->type == GGML_TASK_COMPUTE) {
+                    ggml_cuda_mul_mat_f16(tensor->src0, tensor->src1, tensor, params->wdata, params->wsize);
+                    return true;
+                }
+
+                return false;
+            }
             func = ggml_cuda_mul_mat;
             break;
         default:

ggml.c

@@ -14203,6 +14203,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                         if (ggml_cuda_can_mul_mat(node->src0, node->src1, node)) {
                             node->n_tasks = 1; // TODO: this actually is doing nothing
                                                 //       the threads are still spinning
+                            cur = ggml_cuda_mul_mat_get_wsize(node->src0, node->src1, node);
                         }
                         else
 #elif defined(GGML_USE_CLBLAST)