Sun Mar 3 01:09:37 AM EST 2024

diff --git a/src/gpu/gpu.c b/src/gpu/gpu.c
index 9ece44894b0f9253288ce0a7d872c6348a037c36..02463a3e326cc1aa3c4cec47459d296303df7b5e 100644 (file)
--- a/src/gpu/gpu.c
+++ b/src/gpu/gpu.c
@@ -3,6 +3,8 @@ cl_platform_id GPU_platform_id;
 cl_device_id GPU_device_id;
 cl_context GPU_context;
 cl_command_queue GPU_command_queue;
+unsigned char* GPU_mmul_cache = NULL;
+size_t GPU_mmul_cache_len = 0;
 
 uint8_t GPU_init()
 {
@@ -102,31 +104,43 @@ void GPU_mmul(float* ptrR, float* ptrA, float* ptrB, size_t rowsA, size_t colsB,
                fprintf(stderr, "GPU fatal error: clEnqueueWriteBuffer() failed.\n");
                exit(1);
        }
-       //Load and compile program
-       //printf("------------------------------\n%s\n------------------------------\n", src_gpu_mmul_cl); 
-       char* tmp = malloc(src_gpu_mmul_cl_len);
-       memcpy(tmp, src_gpu_mmul_cl, src_gpu_mmul_cl_len);
-       const char* ptr = (const char*)src_gpu_mmul_cl;
-       cl_program program = clCreateProgramWithSource(GPU_context, 1, (const char**)&tmp, NULL, &err);
-       if (err != CL_SUCCESS)
-       {
+
+       cl_program program;
+       if (GPU_mmul_cache == NULL)
+       {
+               //Load and compile program
+               char* tmp = malloc(src_gpu_mmul_cl_len);
+               memcpy(tmp, src_gpu_mmul_cl, src_gpu_mmul_cl_len);
+               const char* ptr = (const char*)src_gpu_mmul_cl;
+               cl_program program = clCreateProgramWithSource(GPU_context, 1, (const char**)&tmp, NULL, &err);
+               if (err != CL_SUCCESS)
+               {
+                       free(tmp);
+                       fprintf(stderr, "GPU fatal error: clCreateProgramWithSource() failed.\n");
+                       exit(1);
+               }
+               err = clBuildProgram(program, 1, &GPU_device_id, NULL, NULL, NULL);
                free(tmp);
-               fprintf(stderr, "GPU fatal error: clCreateProgramWithSource() failed.\n");
-               exit(1);
-       }
-       err = clBuildProgram(program, 1, &GPU_device_id, NULL, NULL, NULL);
-       free(tmp);
-       if (err != CL_SUCCESS)
-       {
-               fprintf(stderr, "GPU fatal error: clBuildProgram() failed.\n");
-               size_t log_size;
-               clGetProgramBuildInfo(program, GPU_device_id, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
-               char* log = malloc(log_size);
-               clGetProgramBuildInfo(program, GPU_device_id, CL_PROGRAM_BUILD_LOG, log_size, log, NULL);
-               printf("%s", log);
-               free(log);
-               exit(1);
+               if (err != CL_SUCCESS)
+               {
+                       fprintf(stderr, "GPU fatal error: clBuildProgram() failed.\n");
+                       size_t log_size;
+                       clGetProgramBuildInfo(program, GPU_device_id, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
+                       char* log = malloc(log_size);
+                       clGetProgramBuildInfo(program, GPU_device_id, CL_PROGRAM_BUILD_LOG, log_size, log, NULL);
+                       printf("%s", log);
+                       free(log);
+                       exit(1);
+               }
+               clGetProgramInfo(program, CL_PROGRAM_BINARY_SIZES, sizeof(size_t), &GPU_mmul_cache_len, NULL);
+               GPU_mmul_cache = malloc(GPU_mmul_cache_len);
+               clGetProgramInfo(program, CL_PROGRAM_BINARIES, sizeof(unsigned char*), &GPU_mmul_cache, NULL);
+       }
+       else
+       {
+               program = clCreateProgramWithBinary(GPU_context, 1, &GPU_device_id, &GPU_mmul_cache_len, (const unsigned char**)&GPU_mmul_cache, NULL, &err);
        }
+
        //Setup kernel
        cl_kernel kernel = clCreateKernel(program, "gpu_mmul", &err);
        if (err != CL_SUCCESS)