metal : add Q8_0 mul_mm kernel

Author: ggerganov

ggml-metal.m

@@ -83,6 +83,7 @@
     GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q8_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
@@ -209,6 +210,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
@@ -751,9 +753,10 @@ void ggml_metal_graph_compute(
                                 ne00%32 == 0 &&
                                 ne11 > 1) {
                                 switch (src0->type) {
-                                    case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
+                                    case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32];  break;
                                     case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
                                     case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
+                                    case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q8_0_f32]; break;
                                     case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
                                     case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
                                     case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;

ggml-metal.metal

@@ -2041,6 +2041,7 @@ typedef void (mat_mm_t)(device const uchar *, device const float *, device float
 template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1, dequantize_f16>;
 template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2, dequantize_q4_0>;
 template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2, dequantize_q8_0>;
 template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
 template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;