Another attempt

Author: Kawrakow

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);
@@ -205,6 +206,7 @@ @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);
@@ -270,6 +272,7 @@ 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);
@@ -854,7 +857,11 @@ void ggml_metal_graph_compute(
                                         {
                                             nth0 = 32;
                                             nth1 = 1;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                            if (ne11 * ne12 < 2) {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_1row];
+                                            } else {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                            }
                                         } break;
                                     case GGML_TYPE_Q4_0:
                                         {
@@ -974,8 +981,8 @@ 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 + 3)/4, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    int64_t ny = (ne11 + 3)/4;
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                             }
                         } break;

ggml-metal.metal

@@ -505,6 +505,45 @@ kernel void kernel_mul_mat_q8_0_f32(
     }
 }
 
+kernel void kernel_mul_mat_f16_f32_1row(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]]) {
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+    const int64_t im = tgpig.z;
+
+    device const half  * x = (device const half  *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+    device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
+
+    float sumf = 0;
+    for (int i = tiisg; i < ne00; i += 32) {
+        sumf += (float) x[i] * (float) y[i];
+    }
+
+    float all_sum = simd_sum(sumf);
+    if (tiisg == 0) {
+        dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+    }
+}
+
 #define N_F16_F32 4
 
 kernel void kernel_mul_mat_f16_f32(