metal : make the NSG a function constant in mul_mv kernels

ggml-ci

Author: ggerganov

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -615,7 +615,13 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv(ggml_metal_library_
         return res;
     }
 
-    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+    ggml_metal_cv_set_int16(cv, nsg, FC_MUL_MV + 0);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);
 
     ggml_metal_pipeline_set_nr0 (res, nr0);
     ggml_metal_pipeline_set_nr1 (res, nr1);
@@ -825,7 +831,13 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_id(ggml_metal_libra
         return res;
     }
 
-    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+    ggml_metal_cv_set_int16(cv, nsg, FC_MUL_MV + 0);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);
 
     ggml_metal_pipeline_set_nr0 (res, nr0);
     ggml_metal_pipeline_set_nr1 (res, nr1);

ggml/src/ggml-metal/ggml-metal-device.h

@@ -22,6 +22,7 @@ typedef struct ggml_metal_cv * ggml_metal_cv_t;
 ggml_metal_cv_t ggml_metal_cv_init(void);
 void ggml_metal_cv_free(ggml_metal_cv_t cv);
 
+void ggml_metal_cv_set_int16(ggml_metal_cv_t cv, int16_t value, int32_t idx);
 void ggml_metal_cv_set_int32(ggml_metal_cv_t cv, int32_t value, int32_t idx);
 void ggml_metal_cv_set_bool (ggml_metal_cv_t cv, bool    value, int32_t idx);
 

ggml/src/ggml-metal/ggml-metal-device.m

@@ -51,6 +51,10 @@ void ggml_metal_cv_free(ggml_metal_cv_t cv) {
     free(cv);
 }
 
+void ggml_metal_cv_set_int16(ggml_metal_cv_t cv, int16_t value, int32_t idx) {
+    [cv->obj setConstantValue:&value type:MTLDataTypeShort atIndex:idx];
+}
+
 void ggml_metal_cv_set_int32(ggml_metal_cv_t cv, int32_t value, int32_t idx) {
     [cv->obj setConstantValue:&value type:MTLDataTypeInt atIndex:idx];
 }

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -75,6 +75,7 @@
 #define FC_FLASH_ATTN_EXT              100
 #define FC_FLASH_ATTN_EXT_VEC          200
 #define FC_FLASH_ATTN_EXT_VEC_REDUCE   300
+#define FC_MUL_MV                      400
 
 // kernel argument structs
 //