metal : better var names [no ci]
This commit is contained in:
Georgi Gerganov
parent
3ee077a7c8
commit
a408f51906
1 changed files with 182 additions and 182 deletions
|
|
@ -36,18 +36,18 @@ static struct ggml_backend_metal_device_context {
|
||||||
id<MTLDevice> mtl_device;
|
id<MTLDevice> mtl_device;
|
||||||
int mtl_device_ref_count;
|
int mtl_device_ref_count;
|
||||||
|
|
||||||
bool support_simdgroup_reduction;
|
bool has_simdgroup_reduction;
|
||||||
bool support_simdgroup_mm;
|
bool has_simdgroup_mm;
|
||||||
bool support_bfloat;
|
bool has_bfloat;
|
||||||
|
|
||||||
char name[128];
|
char name[128];
|
||||||
} g_ggml_ctx_dev_main = {
|
} g_ggml_ctx_dev_main = {
|
||||||
/*.mtl_device =*/ nil,
|
/*.mtl_device =*/ nil,
|
||||||
/*.mtl_device_ref_count =*/ 0,
|
/*.mtl_device_ref_count =*/ 0,
|
||||||
/*.support_simdgroup_reduction =*/ false,
|
/*.has_simdgroup_reduction =*/ false,
|
||||||
/*.support_simdgroup_mm =*/ false,
|
/*.has_simdgroup_mm =*/ false,
|
||||||
/*.support_bfloat =*/ false,
|
/*.has_bfloat =*/ false,
|
||||||
/*.name =*/ "",
|
/*.name =*/ "",
|
||||||
};
|
};
|
||||||
|
|
||||||
// acquire
|
// acquire
|
||||||
|
|
@ -57,13 +57,13 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
|
||||||
if (ctx->mtl_device == nil) {
|
if (ctx->mtl_device == nil) {
|
||||||
ctx->mtl_device = MTLCreateSystemDefaultDevice();
|
ctx->mtl_device = MTLCreateSystemDefaultDevice();
|
||||||
|
|
||||||
ctx->support_simdgroup_reduction = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
|
ctx->has_simdgroup_reduction = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
|
||||||
ctx->support_simdgroup_reduction |= [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
|
ctx->has_simdgroup_reduction |= [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
|
||||||
|
|
||||||
ctx->support_simdgroup_mm = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
|
ctx->has_simdgroup_mm = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
|
||||||
|
|
||||||
ctx->support_bfloat = [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
|
ctx->has_bfloat = [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
|
||||||
ctx->support_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
|
ctx->has_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
|
||||||
|
|
||||||
strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1);
|
strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1);
|
||||||
}
|
}
|
||||||
|
|
@ -546,9 +546,9 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
GGML_LOG_INFO("%s: simdgroup reduction = %s\n", __func__, ctx_dev->support_simdgroup_reduction ? "true" : "false");
|
GGML_LOG_INFO("%s: simdgroup reduction = %s\n", __func__, ctx_dev->has_simdgroup_reduction ? "true" : "false");
|
||||||
GGML_LOG_INFO("%s: simdgroup matrix mul. = %s\n", __func__, ctx_dev->support_simdgroup_mm ? "true" : "false");
|
GGML_LOG_INFO("%s: simdgroup matrix mul. = %s\n", __func__, ctx_dev->has_simdgroup_mm ? "true" : "false");
|
||||||
GGML_LOG_INFO("%s: bfloat = %s\n", __func__, ctx_dev->support_bfloat ? "true" : "false");
|
GGML_LOG_INFO("%s: bfloat = %s\n", __func__, ctx_dev->has_bfloat ? "true" : "false");
|
||||||
GGML_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx_dev->mtl_device.hasUnifiedMemory ? "true" : "false");
|
GGML_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx_dev->mtl_device.hasUnifiedMemory ? "true" : "false");
|
||||||
|
|
||||||
ctx->capture_next_compute = false;
|
ctx->capture_next_compute = false;
|
||||||
|
|
@ -595,9 +595,9 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
|
||||||
GGML_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \
|
GGML_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \
|
||||||
}
|
}
|
||||||
|
|
||||||
const bool support_simdgroup_mm = ctx_dev->support_simdgroup_mm;
|
const bool has_simdgroup_mm = ctx_dev->has_simdgroup_mm;
|
||||||
const bool support_simdgroup_reduction = ctx_dev->support_simdgroup_reduction;
|
const bool has_simdgroup_reduction = ctx_dev->has_simdgroup_reduction;
|
||||||
const bool support_bfloat = ctx_dev->support_bfloat;
|
const bool has_bfloat = ctx_dev->has_bfloat;
|
||||||
|
|
||||||
// simd_sum and simd_max requires MTLGPUFamilyApple7
|
// simd_sum and simd_max requires MTLGPUFamilyApple7
|
||||||
|
|
||||||
|
|
@ -625,15 +625,15 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true);
|
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, silu, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, 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, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, has_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_F16_4, soft_max_f16_4, has_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, soft_max_f32, has_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_SOFT_MAX_F32_4, soft_max_f32_4, has_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, 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_DIAG_MASK_INF_8, diag_mask_inf_8, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_BF16, get_rows_bf16, support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_BF16, get_rows_bf16, has_bfloat);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true);
|
||||||
|
|
@ -654,108 +654,108 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
|
||||||
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_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_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_GET_ROWS_I32, get_rows_i32, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, has_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_GROUP_NORM, group_norm, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true);
|
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_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_SSM_SCAN_F32, ssm_scan_f32, true);
|
||||||
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_F32_F32, mul_mv_f32_f32, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32, mul_mv_bf16_f32, support_simdgroup_reduction && support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32, mul_mv_bf16_f32, has_simdgroup_reduction && has_bfloat);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW, mul_mv_bf16_f32_1row, support_simdgroup_reduction && support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW, mul_mv_bf16_f32_1row, has_simdgroup_reduction && has_bfloat);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4, mul_mv_bf16_f32_l4, support_simdgroup_reduction && support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4, mul_mv_bf16_f32_l4, has_simdgroup_reduction && has_bfloat);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16, mul_mv_bf16_bf16, support_simdgroup_reduction && support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16, mul_mv_bf16_bf16, has_simdgroup_reduction && has_bfloat);
|
||||||
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, mul_mv_f16_f32, has_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_1ROW, mul_mv_f16_f32_1row, has_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_F16_F32_L4, mul_mv_f16_f32_l4, has_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_F16, mul_mv_f16_f16, has_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_0_F32, mul_mv_q4_0_f32, has_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_Q4_1_F32, mul_mv_q4_1_f32, has_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_0_F32, mul_mv_q5_0_f32, has_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_Q5_1_F32, mul_mv_q5_1_f32, has_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_Q8_0_F32, mul_mv_q8_0_f32, has_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_Q2_K_F32, mul_mv_q2_K_f32, has_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_Q3_K_F32, mul_mv_q3_K_f32, has_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_Q4_K_F32, mul_mv_q4_K_f32, has_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_Q5_K_F32, mul_mv_q5_K_f32, has_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_Q6_K_F32, mul_mv_q6_K_f32, has_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_XXS_F32, mul_mv_iq2_xxs_f32, has_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_IQ2_XS_F32, mul_mv_iq2_xs_f32, has_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_XXS_F32, mul_mv_iq3_xxs_f32, has_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_IQ3_S_F32, mul_mv_iq3_s_f32, has_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_IQ2_S_F32, mul_mv_iq2_s_f32, has_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_S_F32, mul_mv_iq1_s_f32, has_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_IQ1_M_F32, mul_mv_iq1_m_f32, has_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_NL_F32, mul_mv_iq4_nl_f32, has_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_IQ4_XS_F32, mul_mv_iq4_xs_f32, has_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_F32_F32, mul_mv_id_f32_f32, has_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, mul_mv_id_f16_f32, has_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_1ROW, mul_mv_id_f16_f32_1row, has_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_F16_F32_L4, mul_mv_id_f16_f32_l4, has_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_F16, mul_mv_id_f16_f16, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_BF16_F32, mul_mv_id_bf16_f32, support_simdgroup_reduction && support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_BF16_F32, mul_mv_id_bf16_f32, has_simdgroup_reduction && has_bfloat);
|
||||||
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_0_F32, mul_mv_id_q4_0_f32, has_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_Q4_1_F32, mul_mv_id_q4_1_f32, has_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_0_F32, mul_mv_id_q5_0_f32, has_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_Q5_1_F32, mul_mv_id_q5_1_f32, has_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_Q8_0_F32, mul_mv_id_q8_0_f32, has_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_Q2_K_F32, mul_mv_id_q2_K_f32, has_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_Q3_K_F32, mul_mv_id_q3_K_f32, has_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_Q4_K_F32, mul_mv_id_q4_K_f32, has_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_Q5_K_F32, mul_mv_id_q5_K_f32, has_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_Q6_K_F32, mul_mv_id_q6_K_f32, has_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_XXS_F32, mul_mv_id_iq2_xxs_f32, has_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_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, has_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_XXS_F32, mul_mv_id_iq3_xxs_f32, has_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_IQ3_S_F32, mul_mv_id_iq3_s_f32, has_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_IQ2_S_F32, mul_mv_id_iq2_s_f32, has_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_S_F32, mul_mv_id_iq1_s_f32, has_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_IQ1_M_F32, mul_mv_id_iq1_m_f32, has_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_NL_F32, mul_mv_id_iq4_nl_f32, has_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_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, has_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_F32_F32, mul_mm_f32_f32, has_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_F16_F32, mul_mm_f16_f32, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32, mul_mm_bf16_f32, support_simdgroup_mm && support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32, mul_mm_bf16_f32, has_simdgroup_mm && has_bfloat);
|
||||||
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_0_F32, mul_mm_q4_0_f32, has_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_Q4_1_F32, mul_mm_q4_1_f32, has_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_0_F32, mul_mm_q5_0_f32, has_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_Q5_1_F32, mul_mm_q5_1_f32, has_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_Q8_0_F32, mul_mm_q8_0_f32, has_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_Q2_K_F32, mul_mm_q2_K_f32, has_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_Q3_K_F32, mul_mm_q3_K_f32, has_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_Q4_K_F32, mul_mm_q4_K_f32, has_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_Q5_K_F32, mul_mm_q5_K_f32, has_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_Q6_K_F32, mul_mm_q6_K_f32, has_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_XXS_F32, mul_mm_iq2_xxs_f32, has_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_IQ2_XS_F32, mul_mm_iq2_xs_f32, has_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_XXS_F32, mul_mm_iq3_xxs_f32, has_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_IQ3_S_F32, mul_mm_iq3_s_f32, has_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_IQ2_S_F32, mul_mm_iq2_s_f32, has_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_S_F32, mul_mm_iq1_s_f32, has_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_IQ1_M_F32, mul_mm_iq1_m_f32, has_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_NL_F32, mul_mm_iq4_nl_f32, has_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_IQ4_XS_F32, mul_mm_iq4_xs_f32, has_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_F32_F32, mul_mm_id_f32_f32, has_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_F16_F32, mul_mm_id_f16_f32, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32, mul_mm_id_bf16_f32, support_simdgroup_mm && support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32, mul_mm_id_bf16_f32, has_simdgroup_mm && has_bfloat);
|
||||||
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_0_F32, mul_mm_id_q4_0_f32, has_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_Q4_1_F32, mul_mm_id_q4_1_f32, has_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_0_F32, mul_mm_id_q5_0_f32, has_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_Q5_1_F32, mul_mm_id_q5_1_f32, has_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_Q8_0_F32, mul_mm_id_q8_0_f32, has_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_Q2_K_F32, mul_mm_id_q2_K_f32, has_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_Q3_K_F32, mul_mm_id_q3_K_f32, has_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_Q4_K_F32, mul_mm_id_q4_K_f32, has_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_Q5_K_F32, mul_mm_id_q5_K_f32, has_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_Q6_K_F32, mul_mm_id_q6_K_f32, has_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_XXS_F32, mul_mm_id_iq2_xxs_f32, has_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_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, has_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_XXS_F32, mul_mm_id_iq3_xxs_f32, has_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_IQ3_S_F32, mul_mm_id_iq3_s_f32, has_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_IQ2_S_F32, mul_mm_id_iq2_s_f32, has_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_S_F32, mul_mm_id_iq1_s_f32, has_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_IQ1_M_F32, mul_mm_id_iq1_m_f32, has_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_NL_F32, mul_mm_id_iq4_nl_f32, has_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_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, has_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_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_NORM_F16, rope_norm_f16, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32, rope_neox_f32, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32, rope_neox_f32, true);
|
||||||
|
|
@ -771,61 +771,61 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
|
||||||
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_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_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_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, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, has_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_H80, flash_attn_ext_f16_h80, has_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_H96, flash_attn_ext_f16_h96, has_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_H112, flash_attn_ext_f16_h112, has_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_H128, flash_attn_ext_f16_h128, has_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_F16_H256, flash_attn_ext_f16_h256, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64, flash_attn_ext_q4_0_h64, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64, flash_attn_ext_q4_0_h64, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80, flash_attn_ext_q4_0_h80, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80, flash_attn_ext_q4_0_h80, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96, flash_attn_ext_q4_0_h96, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96, flash_attn_ext_q4_0_h96, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H112, flash_attn_ext_q4_0_h112, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H112, flash_attn_ext_q4_0_h112, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H128, flash_attn_ext_q4_0_h128, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H128, flash_attn_ext_q4_0_h128, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256, flash_attn_ext_q4_0_h256, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256, flash_attn_ext_q4_0_h256, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64, flash_attn_ext_q4_1_h64, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64, flash_attn_ext_q4_1_h64, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80, flash_attn_ext_q4_1_h80, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80, flash_attn_ext_q4_1_h80, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96, flash_attn_ext_q4_1_h96, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96, flash_attn_ext_q4_1_h96, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H112, flash_attn_ext_q4_1_h112, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H112, flash_attn_ext_q4_1_h112, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H128, flash_attn_ext_q4_1_h128, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H128, flash_attn_ext_q4_1_h128, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256, flash_attn_ext_q4_1_h256, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256, flash_attn_ext_q4_1_h256, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64, flash_attn_ext_q5_0_h64, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64, flash_attn_ext_q5_0_h64, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80, flash_attn_ext_q5_0_h80, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80, flash_attn_ext_q5_0_h80, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96, flash_attn_ext_q5_0_h96, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96, flash_attn_ext_q5_0_h96, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H112, flash_attn_ext_q5_0_h112, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H112, flash_attn_ext_q5_0_h112, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H128, flash_attn_ext_q5_0_h128, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H128, flash_attn_ext_q5_0_h128, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256, flash_attn_ext_q5_0_h256, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256, flash_attn_ext_q5_0_h256, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64, flash_attn_ext_q5_1_h64, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64, flash_attn_ext_q5_1_h64, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80, flash_attn_ext_q5_1_h80, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80, flash_attn_ext_q5_1_h80, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96, flash_attn_ext_q5_1_h96, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96, flash_attn_ext_q5_1_h96, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H112, flash_attn_ext_q5_1_h112, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H112, flash_attn_ext_q5_1_h112, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H128, flash_attn_ext_q5_1_h128, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H128, flash_attn_ext_q5_1_h128, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256, flash_attn_ext_q5_1_h256, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256, flash_attn_ext_q5_1_h256, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64, flash_attn_ext_q8_0_h64, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64, flash_attn_ext_q8_0_h64, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80, flash_attn_ext_q8_0_h80, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80, flash_attn_ext_q8_0_h80, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96, flash_attn_ext_q8_0_h96, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96, flash_attn_ext_q8_0_h96, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H112, flash_attn_ext_q8_0_h112, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H112, flash_attn_ext_q8_0_h112, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H128, flash_attn_ext_q8_0_h128, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H128, flash_attn_ext_q8_0_h128, has_simdgroup_mm);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256, flash_attn_ext_q8_0_h256, support_simdgroup_mm);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256, flash_attn_ext_q8_0_h256, has_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_H128, flash_attn_ext_vec_f16_h128, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128, flash_attn_ext_vec_q4_0_h128, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128, flash_attn_ext_vec_q4_0_h128, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H128, flash_attn_ext_vec_q4_1_h128, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H128, flash_attn_ext_vec_q4_1_h128, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H128, flash_attn_ext_vec_q5_0_h128, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H128, flash_attn_ext_vec_q5_0_h128, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H128, flash_attn_ext_vec_q5_1_h128, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H128, flash_attn_ext_vec_q5_1_h128, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H128, flash_attn_ext_vec_q8_0_h128, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H128, flash_attn_ext_vec_q8_0_h128, has_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_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H256, flash_attn_ext_vec_q4_0_h256, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H256, flash_attn_ext_vec_q4_0_h256, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H256, flash_attn_ext_vec_q4_1_h256, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H256, flash_attn_ext_vec_q4_1_h256, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256, flash_attn_ext_vec_q5_0_h256, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256, flash_attn_ext_vec_q5_0_h256, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256, flash_attn_ext_vec_q5_1_h256, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256, flash_attn_ext_vec_q5_1_h256, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256, flash_attn_ext_vec_q8_0_h256, support_simdgroup_reduction);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256, flash_attn_ext_vec_q8_0_h256, has_simdgroup_reduction);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, 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_F32_F16, cpy_f32_f16, true);
|
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_BF16, cpy_f32_bf16, support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_BF16, cpy_f32_bf16, has_bfloat);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_F32, cpy_bf16_f32, support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_F32, cpy_bf16_f32, has_bfloat);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_BF16, cpy_bf16_bf16, support_bfloat);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_BF16, cpy_bf16_bf16, has_bfloat);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true);
|
||||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true);
|
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true);
|
||||||
|
|
@ -915,11 +915,11 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs
|
||||||
}
|
}
|
||||||
|
|
||||||
static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_context * ctx_dev, const struct ggml_tensor * op) {
|
static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_context * ctx_dev, const struct ggml_tensor * op) {
|
||||||
const bool support_simdgroup_mm = ctx_dev->support_simdgroup_mm;
|
const bool has_simdgroup_mm = ctx_dev->has_simdgroup_mm;
|
||||||
const bool support_simdgroup_reduction = ctx_dev->support_simdgroup_reduction;
|
const bool has_simdgroup_reduction = ctx_dev->has_simdgroup_reduction;
|
||||||
const bool support_bfloat = ctx_dev->support_bfloat;
|
const bool has_bfloat = ctx_dev->has_bfloat;
|
||||||
|
|
||||||
if (!support_bfloat) {
|
if (!has_bfloat) {
|
||||||
for (size_t i = 0, n = 3; i < n; ++i) {
|
for (size_t i = 0, n = 3; i < n; ++i) {
|
||||||
if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
|
if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
|
||||||
return false;
|
return false;
|
||||||
|
|
@ -964,7 +964,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
|
||||||
case GGML_OP_SOFT_MAX:
|
case GGML_OP_SOFT_MAX:
|
||||||
case GGML_OP_RMS_NORM:
|
case GGML_OP_RMS_NORM:
|
||||||
case GGML_OP_GROUP_NORM:
|
case GGML_OP_GROUP_NORM:
|
||||||
return support_simdgroup_reduction;
|
return has_simdgroup_reduction;
|
||||||
case GGML_OP_NORM:
|
case GGML_OP_NORM:
|
||||||
case GGML_OP_ROPE:
|
case GGML_OP_ROPE:
|
||||||
return true;
|
return true;
|
||||||
|
|
@ -984,13 +984,13 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
|
||||||
if (op->src[1]->type != op->src[2]->type) {
|
if (op->src[1]->type != op->src[2]->type) {
|
||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
return support_simdgroup_mm; // TODO: over-restricted for vec-kernels
|
return has_simdgroup_mm; // TODO: over-restricted for vec-kernels
|
||||||
case GGML_OP_SSM_CONV:
|
case GGML_OP_SSM_CONV:
|
||||||
case GGML_OP_SSM_SCAN:
|
case GGML_OP_SSM_SCAN:
|
||||||
return true;
|
return true;
|
||||||
case GGML_OP_MUL_MAT:
|
case GGML_OP_MUL_MAT:
|
||||||
case GGML_OP_MUL_MAT_ID:
|
case GGML_OP_MUL_MAT_ID:
|
||||||
return support_simdgroup_reduction &&
|
return has_simdgroup_reduction &&
|
||||||
(op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32);
|
(op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32);
|
||||||
case GGML_OP_CPY:
|
case GGML_OP_CPY:
|
||||||
case GGML_OP_DUP:
|
case GGML_OP_DUP:
|
||||||
|
|
|
||||||
Loading…
Add table
Add a link
Reference in a new issue