diff --git a/ggml/include/ggml-metal.h b/ggml/include/ggml-metal.h index ae926c95a..b8d3f678b 100644 --- a/ggml/include/ggml-metal.h +++ b/ggml/include/ggml-metal.h @@ -43,7 +43,9 @@ GGML_API ggml_backend_t ggml_backend_metal_init(void); GGML_API bool ggml_backend_is_metal(ggml_backend_t backend); -GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size); +GGML_DEPRECATED( + GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size), + "obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713"); GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data); diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m index 9c01f0c44..d901b0518 100644 --- a/ggml/src/ggml-metal.m +++ b/ggml/src/ggml-metal.m @@ -29,44 +29,55 @@ static const NSInteger MTLGPUFamilyMetal3_GGML = 5001; static struct ggml_backend_reg g_ggml_backend_metal_reg; static struct ggml_backend_device g_ggml_backend_metal_device; -// global information about the Metal device +// information about a Metal device // note: assumes single GPU device - the default one -static struct ggml_metal_state { +// TODO: support multiple GPU devices +static struct ggml_backend_metal_device_context { id mtl_device; int mtl_device_ref_count; bool support_simdgroup_reduction; bool support_simdgroup_mm; + + char name[128]; } g_state = { - /*.mtl_device =*/ nil, - /*.mtl_device_ref_count =*/ 0, - /*.support_simdgroup_reduction =*/ false, - /*.support_simdgroup_mm =*/ false, + /*.mtl_device =*/ nil, + /*.mtl_device_ref_count =*/ 0, + /*.support_simdgroup_reduction =*/ false, + /*.support_simdgroup_mm =*/ false, + /*.name =*/ "", }; -static id ggml_backend_metal_get_device(void) { - if (g_state.mtl_device == nil) { - g_state.mtl_device = MTLCreateSystemDefaultDevice(); +// acquire +static id ggml_backend_metal_device_acq(struct ggml_backend_metal_device_context * ctx) { + assert(ctx != NULL); - g_state.support_simdgroup_reduction = [g_state.mtl_device supportsFamily:MTLGPUFamilyApple7]; - g_state.support_simdgroup_reduction |= [g_state.mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML]; + if (ctx->mtl_device == nil) { + ctx->mtl_device = MTLCreateSystemDefaultDevice(); - g_state.support_simdgroup_mm = [g_state.mtl_device supportsFamily:MTLGPUFamilyApple7]; + ctx->support_simdgroup_reduction = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7]; + ctx->support_simdgroup_reduction |= [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML]; + + ctx->support_simdgroup_mm = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7]; + + strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name)); } - g_state.mtl_device_ref_count++; + ctx->mtl_device_ref_count++; - return g_state.mtl_device; + return ctx->mtl_device; } -static void ggml_backend_metal_free_device(void) { - assert(g_state.mtl_device_ref_count > 0); +// release +static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_context * ctx) { + assert(ctx != NULL); + assert(ctx->mtl_device_ref_count > 0); - g_state.mtl_device_ref_count--; + ctx->mtl_device_ref_count--; - if (g_state.mtl_device_ref_count == 0) { - [g_state.mtl_device release]; - g_state.mtl_device = nil; + if (ctx->mtl_device_ref_count == 0) { + [ctx->mtl_device release]; + ctx->mtl_device = nil; } } @@ -266,7 +277,8 @@ enum ggml_metal_kernel_type { }; struct ggml_backend_metal_context { - id device; + struct ggml_backend_metal_device_context ctx_dev; + id queue; dispatch_queue_t d_queue; @@ -343,14 +355,13 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) { [devices release]; // since it was created by a *Copy* C method #endif - // Pick and show default Metal device - id device = ggml_backend_metal_get_device(); + // init context + struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context)); + + id device = ggml_backend_metal_device_acq(&ctx->ctx_dev); GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]); - // Configure context - struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context)); - ctx->device = device; - ctx->queue = [ctx->device newCommandQueue]; + ctx->queue = [device newCommandQueue]; ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT); id metal_library; @@ -383,7 +394,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) { NSURL * libURL = [NSURL fileURLWithPath:path_lib]; GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]); - metal_library = [ctx->device newLibraryWithURL:libURL error:&error]; + metal_library = [device newLibraryWithURL:libURL error:&error]; if (error) { GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); return NULL; @@ -433,7 +444,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) { //[options setFastMathEnabled:false]; - metal_library = [ctx->device newLibraryWithSource:src options:options error:&error]; + metal_library = [device newLibraryWithSource:src options:options error:&error]; if (error) { GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); return NULL; @@ -443,37 +454,37 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) { } // print MTL GPU family: - GGML_LOG_INFO("%s: GPU name: %s\n", __func__, [[ctx->device name] UTF8String]); + GGML_LOG_INFO("%s: GPU name: %s\n", __func__, [[device name] UTF8String]); // determine max supported GPU family // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf { for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) { - if ([ctx->device supportsFamily:i]) { + if ([device supportsFamily:i]) { GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i); break; } } for (int i = MTLGPUFamilyCommon1 + 5; i >= MTLGPUFamilyCommon1; --i) { - if ([ctx->device supportsFamily:i]) { + if ([device supportsFamily:i]) { GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyCommon%d (%d)\n", __func__, i - (int) MTLGPUFamilyCommon1 + 1, i); break; } } for (int i = MTLGPUFamilyMetal3_GGML + 5; i >= MTLGPUFamilyMetal3_GGML; --i) { - if ([ctx->device supportsFamily:i]) { + if ([device supportsFamily:i]) { GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyMetal%d (%d)\n", __func__, i - (int) MTLGPUFamilyMetal3_GGML + 3, i); break; } } } - GGML_LOG_INFO("%s: simdgroup reduction support = %s\n", __func__, g_state.support_simdgroup_reduction ? "true" : "false"); - GGML_LOG_INFO("%s: simdgroup matrix mul. support = %s\n", __func__, g_state.support_simdgroup_mm ? "true" : "false"); - GGML_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, g_state.mtl_device.hasUnifiedMemory ? "true" : "false"); + GGML_LOG_INFO("%s: simdgroup reduction support = %s\n", __func__, ctx->ctx_dev.support_simdgroup_reduction ? "true" : "false"); + GGML_LOG_INFO("%s: simdgroup matrix mul. support = %s\n", __func__, ctx->ctx_dev.support_simdgroup_mm ? "true" : "false"); + GGML_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx->ctx_dev.mtl_device.hasUnifiedMemory ? "true" : "false"); ctx->capture_next_compute = false; ctx->capture_started = false; @@ -487,11 +498,11 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) { #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15) if (@available(macOS 10.12, iOS 16.0, *)) { - GGML_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1e6); + GGML_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, device.recommendedMaxWorkingSetSize / 1e6); } #elif TARGET_OS_OSX - if (ctx->device.maxTransferRate != 0) { - GGML_LOG_INFO("%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1e6); + if (device.maxTransferRate != 0) { + GGML_LOG_INFO("%s: maxTransferRate = %8.2f MB/s\n", __func__, device.maxTransferRate / 1e6); } else { GGML_LOG_INFO("%s: maxTransferRate = built-in GPU\n", __func__); } @@ -514,7 +525,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) { if (supported) { \ struct ggml_metal_kernel * kernel = &ctx->kernels[e]; \ id metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \ - kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:metal_function error:&error]; \ + kernel->pipeline = [device newComputePipelineStateWithFunction:metal_function error:&error]; \ [metal_function release]; \ if (error) { \ GGML_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ @@ -525,6 +536,9 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) { GGML_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \ } + const bool support_simdgroup_mm = ctx->ctx_dev.support_simdgroup_mm; + const bool support_simdgroup_reduction = ctx->ctx_dev.support_simdgroup_reduction; + // simd_sum and simd_max requires MTLGPUFamilyApple7 GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true); @@ -551,10 +565,10 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, soft_max_f16_4, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, soft_max_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, soft_max_f32_4, g_state.support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, soft_max_f16_4, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, soft_max_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, soft_max_f32_4, support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true); @@ -579,101 +593,101 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, g_state.support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32, ssm_conv_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32, ssm_scan_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, g_state.support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, g_state.support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, g_state.support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, g_state.support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, g_state.support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, support_simdgroup_mm); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32, rope_norm_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16, rope_norm_f16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32, rope_neox_f32, true); @@ -687,14 +701,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, g_state.support_simdgroup_mm); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, g_state.support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, g_state.support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, g_state.support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, support_simdgroup_mm); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true); @@ -726,7 +740,7 @@ static void ggml_metal_free(struct ggml_backend_metal_context * ctx) { } [ctx->queue release]; - ggml_backend_metal_free_device(); + ggml_backend_metal_device_rel(&ctx->ctx_dev); dispatch_release(ctx->d_queue); @@ -785,14 +799,15 @@ static id ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs } static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx, const struct ggml_tensor * op) { - GGML_UNUSED(ctx); - for (size_t i = 0, n = 3; i < n; ++i) { if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) { return false; } } + const bool support_simdgroup_mm = ctx->ctx_dev.support_simdgroup_mm; + const bool support_simdgroup_reduction = ctx->ctx_dev.support_simdgroup_reduction; + switch (op->op) { case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { @@ -830,7 +845,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx case GGML_OP_SOFT_MAX: case GGML_OP_RMS_NORM: case GGML_OP_GROUP_NORM: - return g_state.support_simdgroup_reduction; + return support_simdgroup_reduction; case GGML_OP_NORM: case GGML_OP_ROPE: return true; @@ -856,13 +871,13 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx if (op->src[0]->ne[0] == 256) { return false; } - return g_state.support_simdgroup_mm; // TODO: over-restricted for vec-kernels + return support_simdgroup_mm; // TODO: over-restricted for vec-kernels case GGML_OP_SSM_CONV: case GGML_OP_SSM_SCAN: return true; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: - return g_state.support_simdgroup_reduction && + return support_simdgroup_reduction && (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32); case GGML_OP_CPY: case GGML_OP_DUP: @@ -1011,6 +1026,8 @@ static void ggml_metal_encode_node( // dst->name); //} + id device = ctx->ctx_dev.mtl_device; + switch (dst->op) { case GGML_OP_CONCAT: { @@ -1719,7 +1736,7 @@ static void ggml_metal_encode_node( // the numbers below are measured on M2 Ultra for 7B and 13B models // these numbers do not translate to other devices or model sizes // TODO: need to find a better approach - if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) { + if ([device.name isEqualToString:@"Apple M2 Ultra"]) { switch (src0t) { case GGML_TYPE_F16: ne11_mm_min = 2; break; case GGML_TYPE_Q8_0: ne11_mm_min = 7; break; @@ -1739,7 +1756,7 @@ static void ggml_metal_encode_node( // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel - if ([ctx->device supportsFamily:MTLGPUFamilyApple7] && + if ([device supportsFamily:MTLGPUFamilyApple7] && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1t == GGML_TYPE_F32 && @@ -2034,7 +2051,7 @@ static void ggml_metal_encode_node( // ne21 = n_rows const int dst_rows = ne20*ne21; const int dst_rows_min = n_as; - const int dst_rows_max = (ctx->device.maxThreadgroupMemoryLength - 32 - 8192)/4; + const int dst_rows_max = (device.maxThreadgroupMemoryLength - 32 - 8192)/4; // max size of the rowids array in the kernel shared buffer GGML_ASSERT(dst_rows <= dst_rows_max); @@ -2045,7 +2062,7 @@ static void ggml_metal_encode_node( // TODO: for now, always use mat-vec kernels until we figure out how to improve the // indirect matrix multiplication // !!! - if ([ctx->device supportsFamily:MTLGPUFamilyApple7] && + if ([device supportsFamily:MTLGPUFamilyApple7] && ne00 % 32 == 0 && ne00 >= 64 && dst_rows > dst_rows_min) { @@ -2884,7 +2901,7 @@ static void ggml_metal_encode_node( while (true) { const size_t smem = nqptg*(ne00 + 2*nsgmax*(ncpsg + nqptg))*(sizeof(float)/2); - if (smem > ctx->device.maxThreadgroupMemoryLength) { + if (smem > device.maxThreadgroupMemoryLength) { break; } nsgmax *= 2; @@ -2896,8 +2913,8 @@ static void ggml_metal_encode_node( const size_t smem = nqptg*(ne00 + 2*nsg*(ncpsg + nqptg))*(sizeof(float)/2); - //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength); - GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength); + //printf("smem: %zu, max: %zu\n", smem, device.maxThreadgroupMemoryLength); + GGML_ASSERT(smem <= device.maxThreadgroupMemoryLength); [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0]; @@ -2922,8 +2939,8 @@ static void ggml_metal_encode_node( const size_t smem = (nqptg*(ne00 + 2*nsg*(ncpsg + nqptg)) + nsg*ne00)*(sizeof(float)/2); - //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength); - GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength); + //printf("smem: %zu, max: %zu\n", smem, device.maxThreadgroupMemoryLength); + GGML_ASSERT(smem <= device.maxThreadgroupMemoryLength); [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; @@ -3027,7 +3044,7 @@ static enum ggml_status ggml_metal_graph_compute( if (!ctx->capture_started) { // create capture scope - ctx->capture_scope = [[MTLCaptureManager sharedCaptureManager] newCaptureScopeWithDevice:ctx->device]; + ctx->capture_scope = [[MTLCaptureManager sharedCaptureManager] newCaptureScopeWithDevice:ctx->ctx_dev.mtl_device]; MTLCaptureDescriptor * descriptor = [MTLCaptureDescriptor new]; descriptor.captureObject = ctx->capture_scope; @@ -3183,7 +3200,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) for (int i = 0; i < ctx->n_buffers; i++) { [ctx->buffers[i].metal release]; } - ggml_backend_metal_free_device(); + ggml_backend_metal_device_rel(buffer->buft->device->context); if (ctx->owned) { #if TARGET_OS_OSX @@ -3286,7 +3303,7 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba size_aligned += (size_page - (size_aligned % size_page)); } - id device = ggml_backend_metal_get_device(); + id device = ggml_backend_metal_device_acq(buft->device->context); ctx->all_data = ggml_metal_host_malloc(size_aligned); ctx->all_size = size_aligned; @@ -3300,16 +3317,16 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba if (size_aligned > 0) { ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data - length:size_aligned - options:MTLResourceStorageModeShared - deallocator:nil]; + length:size_aligned + options:MTLResourceStorageModeShared + deallocator:nil]; } } if (size_aligned > 0 && (ctx->all_data == NULL || ctx->buffers[0].metal == nil)) { GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0); free(ctx); - ggml_backend_metal_free_device(); + ggml_backend_metal_device_rel(buft->device->context); return NULL; } @@ -3324,9 +3341,9 @@ static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_t } static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { - id device = ggml_backend_metal_get_device(); - size_t max_size = device.maxBufferLength; - ggml_backend_metal_free_device(); + id device = ggml_backend_metal_device_acq(buft->device->context); + const size_t max_size = device.maxBufferLength; + ggml_backend_metal_device_rel(buft->device->context); return max_size; @@ -3356,6 +3373,7 @@ ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) { return &ggml_backend_buffer_type_metal; } +// TODO: obsoleted by ggml_backend_metal_device_buffer_from_ptr ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) { struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context)); @@ -3378,7 +3396,7 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz size_aligned += (size_page - (size_aligned % size_page)); } - id device = ggml_backend_metal_get_device(); + id device = ggml_backend_metal_device_acq(&g_state); // the buffer fits into the max buffer size allowed by the device if (size_aligned <= device.maxBufferLength) { @@ -3543,7 +3561,7 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) { struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; - return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)]; + return [ctx->ctx_dev.mtl_device supportsFamily:(MTLGPUFamilyApple1 + family - 1)]; } void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) { @@ -3562,23 +3580,27 @@ static const char * ggml_backend_metal_device_get_name(ggml_backend_dev_t dev) { } static const char * ggml_backend_metal_device_get_description(ggml_backend_dev_t dev) { - return [[g_state.mtl_device name] UTF8String]; + // acq/rel just to populate ctx->name in case it hasn't been done yet + struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context; + ggml_backend_metal_device_acq(ctx_dev); + ggml_backend_metal_device_rel(ctx_dev); - GGML_UNUSED(dev); + return ctx_dev->name; } static void ggml_backend_metal_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) { if (@available(macOS 10.12, iOS 16.0, *)) { - id device = ggml_backend_metal_get_device(); + struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context; + id device = ggml_backend_metal_device_acq(ctx_dev); + *total = device.recommendedMaxWorkingSetSize; *free = *total - device.currentAllocatedSize; - ggml_backend_metal_free_device(); + + ggml_backend_metal_device_rel(ctx_dev); } else { *free = 1; *total = 1; } - - GGML_UNUSED(dev); } static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backend_dev_t dev) { @@ -3614,10 +3636,82 @@ static ggml_backend_buffer_type_t ggml_backend_metal_device_get_buffer_type(ggml } static ggml_backend_buffer_t ggml_backend_metal_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) { - return ggml_backend_metal_buffer_from_ptr(ptr, size, max_tensor_size); + struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context)); - GGML_UNUSED(dev); - GGML_UNUSED(max_tensor_size); + ctx->all_data = ptr; + ctx->all_size = size; + ctx->owned = false; + ctx->n_buffers = 0; + + const size_t size_page = sysconf(_SC_PAGESIZE); + + // page-align the data ptr + { + const uintptr_t offs = (uintptr_t) ptr % size_page; + ptr = (void *) ((char *) ptr - offs); + size += offs; + } + + size_t size_aligned = size; + if ((size_aligned % size_page) != 0) { + size_aligned += (size_page - (size_aligned % size_page)); + } + + struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context; + id device = ggml_backend_metal_device_acq(ctx_dev); + + // the buffer fits into the max buffer size allowed by the device + if (size_aligned <= device.maxBufferLength) { + ctx->buffers[ctx->n_buffers].data = ptr; + ctx->buffers[ctx->n_buffers].size = size; + ctx->buffers[ctx->n_buffers].metal = nil; + + if (size_aligned > 0) { + ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:ptr length:size_aligned options:MTLResourceStorageModeShared deallocator:nil]; + + if (ctx->buffers[ctx->n_buffers].metal == nil) { + GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0); + return false; + } + } + + ggml_backend_metal_log_allocated_size(device, size_aligned); + + ++ctx->n_buffers; + } else { + // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into + // one of the views + const size_t size_ovlp = ((max_tensor_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case + const size_t size_step = device.maxBufferLength - size_ovlp; + const size_t size_view = device.maxBufferLength; + + for (size_t i = 0; i < size; i += size_step) { + const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i); + + ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) ptr + i); + ctx->buffers[ctx->n_buffers].size = size_step_aligned; + ctx->buffers[ctx->n_buffers].metal = nil; + + if (size_step_aligned > 0) { + ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) ptr + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil]; + + if (ctx->buffers[ctx->n_buffers].metal == nil) { + GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0); + return false; + } + } + + ggml_backend_metal_log_allocated_size(device, size_step_aligned); + + if (i + size_step < size) { + GGML_LOG_INFO("\n"); + } + + ++ctx->n_buffers; + } + } + + return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size); } static bool ggml_backend_metal_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) { @@ -3698,7 +3792,7 @@ ggml_backend_reg_t ggml_backend_metal_reg(void) { g_ggml_backend_metal_device = (struct ggml_backend_device) { /* .iface = */ ggml_backend_metal_device_i, /* .reg = */ &g_ggml_backend_metal_reg, - /* .context = */ NULL, + /* .context = */ &g_state, }; }