metal : more optimizations (#2959)

* Very minor speedup via simd-group synchronization in f16 x f32

* Another very minor speedup on metal

* Quite significant PP speedup on metal

* Another attempt

* Minor

* Massive improvement for TG for fp16

* ~4-5% improvement for Q8_0 TG on metal

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

Author: 3 people

ggml-metal.m

@@ -76,6 +76,7 @@
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(norm);
     GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_1row);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32);
@@ -219,6 +220,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(norm);
         GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_1row);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32);
@@ -284,6 +286,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(rms_norm);
     GGML_METAL_DEL_KERNEL(norm);
     GGML_METAL_DEL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_1row);
     GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32);
@@ -868,7 +871,11 @@ void ggml_metal_graph_compute(
                                         {
                                             nth0 = 32;
                                             nth1 = 1;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                            if (ne11 * ne12 < 4) {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_1row];
+                                            } else {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                            }
                                         } break;
                                     case GGML_TYPE_Q4_0:
                                         {
@@ -920,8 +927,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 2;
-                                            nth1 = 32;
+                                            nth0 = 4; //1;
+                                            nth1 = 8; //32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
                                         } break;
                                     case GGML_TYPE_Q5_K:
@@ -969,9 +976,12 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&gqa  length:sizeof(gqa)  atIndex:17];
 
                                 if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
-                                    src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
+                                    src0t == GGML_TYPE_Q2_K) {// || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
+                                else if (src0t == GGML_TYPE_Q4_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
                                 else if (src0t == GGML_TYPE_Q3_K) {
 #ifdef GGML_QKK_64
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -985,8 +995,8 @@ void ggml_metal_graph_compute(
                                 else if (src0t == GGML_TYPE_Q6_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
-                                    [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    int64_t ny = (ne11 + 3)/4;
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                             }
                         } break;