llama_cpp 0.9.2 → 0.9.3

data/ext/llama_cpp/src/ggml-metal.m

@@ -1,5 +1,6 @@
 #import "ggml-metal.h"
 
+#import "ggml-backend-impl.h"
 #import "ggml.h"
 
 #import <Foundation/Foundation.h>
@@ -23,7 +24,7 @@
 
 #define UNUSED(x) (void)(x)
 
-#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
+#define GGML_MAX_CONCUR (2*GGML_DEFAULT_GRAPH_SIZE)
 
 struct ggml_metal_buffer {
     const char * name;
@@ -85,6 +86,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(norm);
     GGML_METAL_DECL_KERNEL(mul_mv_f32_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_f16_f16);
     GGML_METAL_DECL_KERNEL(mul_mv_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_1row);
     GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_l4);
@@ -113,6 +115,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(rope_f32);
     GGML_METAL_DECL_KERNEL(rope_f16);
     GGML_METAL_DECL_KERNEL(alibi_f32);
+    GGML_METAL_DECL_KERNEL(im2col_f16);
     GGML_METAL_DECL_KERNEL(cpy_f32_f16);
     GGML_METAL_DECL_KERNEL(cpy_f32_f32);
     GGML_METAL_DECL_KERNEL(cpy_f16_f16);
@@ -125,7 +128,7 @@ struct ggml_metal_context {
 // MSL code
 // TODO: move the contents here when ready
 //       for now it is easier to work in a separate file
-static NSString * const msl_library_source = @"see metal.metal";
+//static NSString * const msl_library_source = @"see metal.metal";
 
 // Here to assist with NSBundle Path Hack
 @interface GGMLMetalClass : NSObject
@@ -141,7 +144,8 @@ void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_dat
     ggml_metal_log_user_data = user_data;
 }
 
-static void ggml_metal_log(enum ggml_log_level level, const char* format, ...){
+GGML_ATTRIBUTE_FORMAT(2, 3)
+static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
     if (ggml_metal_log_callback != NULL) {
         va_list args;
         va_start(args, format);
@@ -209,7 +213,13 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         } else {
             GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
 
-            NSString * sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+            NSString * sourcePath;
+            NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+            if (ggmlMetalPathResources) {
+                sourcePath = [ggmlMetalPathResources stringByAppendingPathComponent:@"ggml-metal.metal"];
+            } else {
+                sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+            }
             if (sourcePath == nil) {
                 GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
                 sourcePath = @"ggml-metal.metal";
@@ -280,6 +290,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(norm);
         GGML_METAL_ADD_KERNEL(mul_mv_f32_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_f16_f16);
         GGML_METAL_ADD_KERNEL(mul_mv_f16_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_1row);
         GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_l4);
@@ -310,6 +321,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(rope_f32);
         GGML_METAL_ADD_KERNEL(rope_f16);
         GGML_METAL_ADD_KERNEL(alibi_f32);
+        GGML_METAL_ADD_KERNEL(im2col_f16);
         GGML_METAL_ADD_KERNEL(cpy_f32_f16);
         GGML_METAL_ADD_KERNEL(cpy_f32_f32);
         GGML_METAL_ADD_KERNEL(cpy_f16_f16);
@@ -328,15 +340,15 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
     // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
     for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
         if ([ctx->device supportsFamily:i]) {
-            GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - MTLGPUFamilyApple1 + 1, i);
+            GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
             break;
         }
     }
 
-    GGML_METAL_LOG_INFO("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
-    GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+    GGML_METAL_LOG_INFO("%s: hasUnifiedMemory              = %s\n",        __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MiB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
     if (ctx->device.maxTransferRate != 0) {
-        GGML_METAL_LOG_INFO("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+        GGML_METAL_LOG_INFO("%s: maxTransferRate               = %8.2f MiB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
     } else {
         GGML_METAL_LOG_INFO("%s: maxTransferRate               = built-in GPU\n", __func__);
     }
@@ -379,6 +391,7 @@ 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_mv_f32_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_f16_f16);
     GGML_METAL_DEL_KERNEL(mul_mv_f16_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_1row);
     GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_l4);
@@ -409,6 +422,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(rope_f32);
     GGML_METAL_DEL_KERNEL(rope_f16);
     GGML_METAL_DEL_KERNEL(alibi_f32);
+    GGML_METAL_DEL_KERNEL(im2col_f16);
     GGML_METAL_DEL_KERNEL(cpy_f32_f16);
     GGML_METAL_DEL_KERNEL(cpy_f32_f32);
     GGML_METAL_DEL_KERNEL(cpy_f16_f16);
@@ -466,6 +480,10 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
 
     const int64_t tsize = ggml_nbytes(t);
 
+    if (t->buffer && t->buffer->backend && t->buffer->backend->context) {
+        ctx = t->buffer->backend->context;
+    }
+
     // find the view that contains the tensor fully
     for (int i = 0; i < ctx->n_buffers; ++i) {
         const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
@@ -523,11 +541,11 @@ bool ggml_metal_add_buffer(
             ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
             if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
+                GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MiB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
                 return false;
             }
 
-            GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
+            GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MiB", __func__, name, size_aligned / 1024.0 / 1024.0);
 
             ++ctx->n_buffers;
         } else {
@@ -547,11 +565,11 @@ bool ggml_metal_add_buffer(
                 ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
                 if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                    GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
+                    GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MiB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
                     return false;
                 }
 
-                GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
+                GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MiB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
                 if (i + size_step < size) {
                     GGML_METAL_LOG_INFO("\n");
                 }
@@ -566,7 +584,7 @@ bool ggml_metal_add_buffer(
                 ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
 
         if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
-            GGML_METAL_LOG_WARN(", warning: current allocated size is greater than the recommended max working set size\n", __func__);
+            GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
         } else {
             GGML_METAL_LOG_INFO("\n");
         }
@@ -744,6 +762,20 @@ void ggml_metal_graph_compute(
                 struct ggml_tensor * src1 = gf->nodes[i]->src[1];
                 struct ggml_tensor * dst  = gf->nodes[i];
 
+                switch (dst->op) {
+                    case GGML_OP_NONE:
+                    case GGML_OP_RESHAPE:
+                    case GGML_OP_VIEW:
+                    case GGML_OP_TRANSPOSE:
+                    case GGML_OP_PERMUTE:
+                        {
+                            // noop -> next node
+                        } continue;
+                    default:
+                        {
+                        } break;
+                }
+
                 const int64_t  ne00 = src0 ? src0->ne[0] : 0;
                 const int64_t  ne01 = src0 ? src0->ne[1] : 0;
                 const int64_t  ne02 = src0 ? src0->ne[2] : 0;
@@ -797,14 +829,6 @@ void ggml_metal_graph_compute(
                 //}
 
                 switch (dst->op) {
-                    case GGML_OP_NONE:
-                    case GGML_OP_RESHAPE:
-                    case GGML_OP_VIEW:
-                    case GGML_OP_TRANSPOSE:
-                    case GGML_OP_PERMUTE:
-                        {
-                            // noop
-                        } break;
                     case GGML_OP_CONCAT:
                         {
                             const int64_t nb = ne00;
@@ -1017,7 +1041,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
                             [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
                             [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:MAX(16, nth/32*sizeof(float)) atIndex:0];
+                            [encoder setThreadgroupMemoryLength:GGML_PAD(nth/32*sizeof(float), 16) atIndex:0];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
@@ -1126,6 +1150,7 @@ void ggml_metal_graph_compute(
                                 switch (src0t) {
                                     case GGML_TYPE_F32:
                                         {
+                                            GGML_ASSERT(src1t == GGML_TYPE_F32);
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_f32_f32];
                                             nrows = 4;
                                         } break;
@@ -1133,13 +1158,18 @@ void ggml_metal_graph_compute(
                                         {
                                             nth0 = 32;
                                             nth1 = 1;
-                                            if (ne11 * ne12 < 4) {
-                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_1row];
-                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
-                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_l4];
-                                                nrows = ne11;
+                                            if (src1t == GGML_TYPE_F32) {
+                                                if (ne11 * ne12 < 4) {
+                                                    [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_1row];
+                                                } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                                    [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_l4];
+                                                    nrows = ne11;
+                                                } else {
+                                                    [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32];
+                                                    nrows = 4;
+                                                }
                                             } else {
-                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32];
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f16];
                                                 nrows = 4;
                                             }
                                         } break;
@@ -1329,7 +1359,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
                             [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
                             [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:nth/32*sizeof(float) atIndex:0];
+                            [encoder setThreadgroupMemoryLength:GGML_PAD(nth/32*sizeof(float), 16) atIndex:0];
 
                             const int64_t nrows = ggml_nrows(src0);
 
@@ -1348,7 +1378,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
                             [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
                             [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:MAX(16, nth*sizeof(float)) atIndex:0];
+                            [encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0];
 
                             const int64_t nrows = ggml_nrows(src0);
 
@@ -1403,8 +1433,7 @@ void ggml_metal_graph_compute(
                             const int n_past     = ((int32_t *) dst->op_params)[0];
                             const int n_dims     = ((int32_t *) dst->op_params)[1];
                             const int mode       = ((int32_t *) dst->op_params)[2];
-                            // skip 3, n_ctx, used in GLM RoPE, unimplemented in metal
-                            const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
+                            const int n_orig_ctx = ((int32_t *) dst->op_params)[3];
 
                             float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
                             memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
@@ -1452,6 +1481,58 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
+                    case GGML_OP_IM2COL:
+                        {
+                            GGML_ASSERT(src0->type == GGML_TYPE_F16);
+                            GGML_ASSERT(src1->type == GGML_TYPE_F32);
+                            GGML_ASSERT( dst->type == GGML_TYPE_F16);
+
+                            const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
+                            const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
+                            const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
+                            const int32_t p1 = ((const int32_t *)(dst->op_params))[3];
+                            const int32_t d0 = ((const int32_t *)(dst->op_params))[4];
+                            const int32_t d1 = ((const int32_t *)(dst->op_params))[5];
+                            const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1;
+
+                            const int32_t N  = src1->ne[is_2D ? 3 : 2];
+                            const int32_t IC = src1->ne[is_2D ? 2 : 1];
+                            const int32_t IH = is_2D ? src1->ne[1] : 1;
+                            const int32_t IW =         src1->ne[0];
+
+                            const int32_t KH = is_2D ? src0->ne[1] : 1;
+                            const int32_t KW =         src0->ne[0];
+
+                            const int32_t OH = is_2D ? dst->ne[2] : 1;
+                            const int32_t OW =         dst->ne[1];
+
+                            const int32_t CHW = IC * KH * KW;
+
+                            const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4;
+                            const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4;
+
+                            switch (src0->type) {
+                                case GGML_TYPE_F32: GGML_ASSERT(false && "not implemented"); break;
+                                case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_im2col_f16]; break;
+                                default: GGML_ASSERT(false);
+                            };
+
+                            [encoder setBuffer:id_src1 offset:offs_src1        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                            [encoder setBytes:&ofs0    length:sizeof( int32_t) atIndex:2];
+                            [encoder setBytes:&ofs1    length:sizeof( int32_t) atIndex:3];
+                            [encoder setBytes:&IW      length:sizeof( int32_t) atIndex:4];
+                            [encoder setBytes:&IH      length:sizeof( int32_t) atIndex:5];
+                            [encoder setBytes:&CHW     length:sizeof( int32_t) atIndex:6];
+                            [encoder setBytes:&s0      length:sizeof( int32_t) atIndex:7];
+                            [encoder setBytes:&s1      length:sizeof( int32_t) atIndex:8];
+                            [encoder setBytes:&p0      length:sizeof( int32_t) atIndex:9];
+                            [encoder setBytes:&p1      length:sizeof( int32_t) atIndex:10];
+                            [encoder setBytes:&d0      length:sizeof( int32_t) atIndex:11];
+                            [encoder setBytes:&d1      length:sizeof( int32_t) atIndex:12];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
+                        } break;
                     case GGML_OP_DUP:
                     case GGML_OP_CPY:
                     case GGML_OP_CONT:

data/ext/llama_cpp/src/ggml-metal.metal

@@ -792,7 +792,7 @@ kernel void kernel_mul_mv_f32_f32(
         constant   int64_t & ne0,
         constant   int64_t & ne1,
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]]) {
 
     const int64_t r0 = tgpig.x;
     const int64_t rb = tgpig.y*N_F32_F32;
@@ -844,6 +844,79 @@ kernel void kernel_mul_mv_f32_f32(
     }
 }
 
+#define N_F16_F16 4
+
+kernel void kernel_mul_mv_f16_f16(
+        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 rb = tgpig.y*N_F16_F16;
+    const int64_t im = tgpig.z;
+
+    device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F16; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const half * y = (device const half *) (src1 + r1*nb11 + im*nb12);
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00; i += 32) {
+                sumf += (half) x[i] * (half) y[i];
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        device const half4 * x4 = (device const half4 *)x;
+        for (int row = 0; row < N_F16_F16; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const half  * y  = (device const half  *) (src1 + r1*nb11 + im*nb12);
+            device const half4 * y4 = (device const half4 *) y;
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00/4; i += 32) {
+                for (int k = 0; k < 4; ++k) sumf += (half) x4[i][k] * y4[i][k];
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (half) x[i] * y[i];
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}
+
 kernel void kernel_mul_mv_f16_f32_1row(
         device const  char * src0,
         device const  char * src1,
@@ -1229,6 +1302,39 @@ kernel void kernel_rope(
 template [[host_name("kernel_rope_f32")]] kernel rope_t kernel_rope<float>;
 template [[host_name("kernel_rope_f16")]] kernel rope_t kernel_rope<half>;
 
+kernel void kernel_im2col_f16(
+        device const float * x,
+        device       half * dst,
+        constant   int32_t & ofs0,
+        constant   int32_t & ofs1,
+        constant   int32_t & IW,
+        constant   int32_t & IH,
+        constant   int32_t & CHW,
+        constant   int32_t & s0,
+        constant   int32_t & s1,
+        constant   int32_t & p0,
+        constant   int32_t & p1,
+        constant   int32_t & d0,
+        constant   int32_t & d1,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int32_t iiw = tgpig[2] * s0 + tpitg[2] * d0 - p0;
+    const int32_t iih = tgpig[1] * s1 + tpitg[1] * d1 - p1;
+
+    const int32_t offset_dst =
+        (tpitg[0] * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW +
+        (tgpig[0] * (ntg[1] * ntg[2]) + tpitg[1] * ntg[2] + tpitg[2]);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        const int32_t offset_src = tpitg[0] * ofs0 + tgpig[0] * ofs1;
+        dst[offset_dst] = x[offset_src + iih * IW + iiw];
+    }
+}
+
 kernel void kernel_cpy_f16_f16(
         device const half * src0,
         device       half * dst,