vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

metal : set kernel family requirements

Browse source
This commit is contained in:
Georgi Gerganov 2024-01-06 13:47:08 +02:00
parent 9bc3618de8
commit 8437ff5ee2
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 37 additions and 35 deletions
Download patch file Download diff file Expand all files Collapse all files

72
ggml-metal.m
View file

@ -371,6 +371,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_LOG_INFO("%s: skipping %-32s\n", __func__, "kernel_"#name); \
}
// simd_sum and simd_max requires MTLGPUFamilyApple7
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW, add_row, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL, mul, MTLGPUFamilyCommon1);
@ -384,8 +386,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX, soft_max, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4, soft_max_4, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX, soft_max, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4, soft_max_4, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, MTLGPUFamilyCommon1);
@ -401,39 +403,39 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K, get_rows_q5_K, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K, get_rows_q6_K, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, MTLGPUFamilyCommon1);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, MTLGPUFamilyCommon1);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, MTLGPUFamilyCommon1);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, MTLGPUFamilyCommon1);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, MTLGPUFamilyApple7);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, MTLGPUFamilyApple7);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, MTLGPUFamilyApple7);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, MTLGPUFamilyApple7);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, MTLGPUFamilyApple7);