Add stable-diffusion.cpp

Makefile

@@ -13,11 +13,16 @@ include build/rules.mk
 
 include llamafile/BUILD.mk
 include llama.cpp/BUILD.mk
+include stable-diffusion.cpp/BUILD.mk
+include stb/BUILD.mk
 
 # the root package is `o//` by default
 # building a package also builds its sub-packages
 .PHONY: o/$(MODE)/
-o/$(MODE)/: o/$(MODE)/llama.cpp o/$(MODE)/llamafile o/$(MODE)/depend.test
+o/$(MODE)/:	o/$(MODE)/llamafile					\
+		o/$(MODE)/stable-diffusion.cpp				\
+		o/$(MODE)/stb						\
+		o/$(MODE)/depend.test
 
 # for installing to `make PREFIX=/usr/local`
 .PHONY: install

llama.cpp/BUILD.mk

@@ -37,7 +37,6 @@ o/$(MODE)/llama.cpp/ggml-backend.o			\
 o/$(MODE)/llama.cpp/grammar-parser.o			\
 o/$(MODE)/llama.cpp/json-schema-to-grammar.o		\
 o/$(MODE)/llama.cpp/llama.o				\
-o/$(MODE)/llama.cpp/stb_image.o				\
 o/$(MODE)/llama.cpp/unicode.o				\
 o/$(MODE)/llama.cpp/sampling.o				\
 o/$(MODE)/llama.cpp/ggml-alloc.o			\

llama.cpp/ggml.c

@@ -11848,6 +11848,7 @@ IQK_MulMat_Not_Available:;
 
 // ggml_compute_forward_out_prod
 
+__target_clones("avx2") // [jart]
 static void ggml_compute_forward_out_prod_f32(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
@@ -12266,6 +12267,7 @@ static void ggml_compute_forward_scale(
 
 // ggml_compute_forward_set
 
+__target_clones("avx2") // [jart]
 static void ggml_compute_forward_set_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -12684,6 +12686,7 @@ static void ggml_compute_forward_get_rows(
 
 // ggml_compute_forward_get_rows_back
 
+__target_clones("avx2") // [jart]
 static void ggml_compute_forward_get_rows_back_f32_f16(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
@@ -12723,6 +12726,7 @@ static void ggml_compute_forward_get_rows_back_f32_f16(
     }
 }
 
+__target_clones("avx2") // [jart]
 static void ggml_compute_forward_get_rows_back_f32(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
@@ -12866,6 +12870,7 @@ static void ggml_compute_forward_diag(
 
 // ggml_compute_forward_diag_mask_inf
 
+__target_clones("avx2") // [jart]
 static void ggml_compute_forward_diag_mask_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst,
@@ -13954,6 +13959,7 @@ static void ggml_compute_forward_conv_transpose_1d(
 // src0: kernel [OC, IC, KH, KW]
 // src1: image [N, IC, IH, IW]
 // dst:  result [N, OH, OW, IC*KH*KW]
+__target_clones("arch=znver4,avx2") // [jart]
 static void ggml_compute_forward_im2col_f32(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
@@ -14148,6 +14154,7 @@ static void ggml_compute_forward_im2col(
 
 // ggml_compute_forward_conv_transpose_2d
 
+__target_clones("avx2") // [jart]
 static void ggml_compute_forward_conv_transpose_2d(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
@@ -15947,6 +15954,7 @@ static void ggml_compute_forward_ssm_conv(
 
 // ggml_compute_forward_ssm_scan
 
+__target_clones("avx2") // [jart]
 static void ggml_compute_forward_ssm_scan_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -21249,7 +21257,7 @@ size_t ggml_quantize_chunk(
             assert(false);
     }
 
-    GGML_ASSERT(result == nrows * row_size);
+    // GGML_ASSERT(result == nrows * row_size); // [jart] should return error
 
     return result;
 }

llama.cpp/ggml.h

@@ -233,7 +233,7 @@
 #define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            10
 #ifndef GGML_MAX_NAME
-#define GGML_MAX_NAME           64
+#define GGML_MAX_NAME           128 // [jart] for stable diffusion
 #endif
 #define GGML_MAX_OP_PARAMS      64
 #define GGML_DEFAULT_N_THREADS  4

llama.cpp/llava/BUILD.mk

@@ -17,7 +17,8 @@ o/$(MODE)/llama.cpp/llava/llava-quantize:				\
 		o/$(MODE)/llama.cpp/llava/llava-quantize.o		\
 		o/$(MODE)/llama.cpp/llava/llava-quantize.1.asc.zip.o	\
 		o/$(MODE)/llama.cpp/llava/llava.a			\
-		o/$(MODE)/llama.cpp/llama.cpp.a
+		o/$(MODE)/llama.cpp/llama.cpp.a				\
+		o/$(MODE)/stb/stb.a
 
 o/$(MODE)/llama.cpp/llava/clip.o					\
 o/$(MODE)/llama.cpp/llava/llava.o					\

llama.cpp/llava/clip.cpp

@@ -23,8 +23,8 @@
 #include "llama.cpp/llava/clip.h"
 #include "llama.cpp/ggml-metal.h"
 #include "llama.cpp/ggml-cuda.h"
-#include "llama.cpp/stb_image.h"
 #include "llama.cpp/ggml.h"
+#include "stb/stb_image.h"
 
 //#define CLIP_DEBUG_FUNCTIONS
 

llama.cpp/main/BUILD.mk

@@ -14,6 +14,7 @@ o/$(MODE)/llama.cpp/main/main:					\
 		o/$(MODE)/llama.cpp/server/server.a		\
 		o/$(MODE)/llama.cpp/llava/llava.a		\
 		o/$(MODE)/llama.cpp/llama.cpp.a			\
+		o/$(MODE)/stb/stb.a				\
 		o/$(MODE)/llama.cpp/main/main.1.asc.zip.o	\
 		$(LLAMA_CPP_SERVER_ASSETS:%=o/$(MODE)/%.zip.o)
 

llama.cpp/server/server.cpp

@@ -5,7 +5,7 @@
 #include "llama.cpp/llama.h"
 #include "llama.cpp/grammar-parser.h"
 #include "llama.cpp/llava/llava.h"
-#include "llama.cpp/stb_image.h"
+#include "stb/stb_image.h"
 #include "utils.h"
 #include "oai.h"
 #include "llamafile/micros.h"

stable-diffusion.cpp/BUILD.mk

@@ -0,0 +1,40 @@
+#-*-mode:makefile-gmake;indent-tabs-mode:t;tab-width:8;coding:utf-8-*-┐
+#── vi: set noet ft=make ts=8 sw=8 fenc=utf-8 :vi ────────────────────┘
+
+PKGS += STABLE_DIFFUSION_CPP
+
+STABLE_DIFFUSION_CPP_FILES := $(wildcard stable-diffusion.cpp/*.*)
+STABLE_DIFFUSION_CPP_HDRS = $(filter %.h,$(STABLE_DIFFUSION_CPP_FILES))	\
+			    $(filter %.hpp,$(STABLE_DIFFUSION_CPP_FILES))
+STABLE_DIFFUSION_CPP_INCS = $(filter %.inc,$(STABLE_DIFFUSION_CPP_FILES))
+STABLE_DIFFUSION_CPP_SRCS_C = $(filter %.c,$(STABLE_DIFFUSION_CPP_FILES))
+STABLE_DIFFUSION_CPP_SRCS_CPP = $(filter %.cpp,$(STABLE_DIFFUSION_CPP_FILES))
+STABLE_DIFFUSION_CPP_SRCS = $(STABLE_DIFFUSION_CPP_SRCS_C) $(STABLE_DIFFUSION_CPP_SRCS_CPP)
+
+STABLE_DIFFUSION_CPP_OBJS =						\
+	$(LLAMAFILE_OBJS)						\
+	$(STABLE_DIFFUSION_CPP_SRCS_C:%.c=o/$(MODE)/%.o)		\
+	$(STABLE_DIFFUSION_CPP_SRCS_CPP:%.cpp=o/$(MODE)/%.o)		\
+	$(STABLE_DIFFUSION_CPP_FILES:%=o/$(MODE)/%.zip.o)
+
+o/$(MODE)/stable-diffusion.cpp/stable-diffusion.cpp.a: $(STABLE_DIFFUSION_CPP_OBJS)
+
+$(STABLE_DIFFUSION_CPP_OBJS): private					\
+		CCFLAGS +=						\
+			-DGGML_MULTIPLATFORM
+
+$(STABLE_DIFFUSION_CPP_OBJS): private					\
+		CXXFLAGS +=						\
+			-frtti
+
+o/$(MODE)/stable-diffusion.cpp/main:					\
+		o/$(MODE)/stable-diffusion.cpp/main.o			\
+		o/$(MODE)/stable-diffusion.cpp/stable-diffusion.cpp.a	\
+		o/$(MODE)/llama.cpp/llama.cpp.a				\
+		o/$(MODE)/stb/stb.a
+
+$(STABLE_DIFFUSION_CPP_OBJS): stable-diffusion.cpp/BUILD.mk
+
+.PHONY: o/$(MODE)/stable-diffusion.cpp
+o/$(MODE)/stable-diffusion.cpp:						\
+		o/$(MODE)/stable-diffusion.cpp/main

stable-diffusion.cpp/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 leejet
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

stable-diffusion.cpp/README.llamafile

@@ -0,0 +1,19 @@
+DESCRIPTION
+
+  stable-diffusion.cpp is an image file generator
+
+LICENSE
+
+  MIT
+
+ORIGIN
+
+  git@github.com:leejet/stable-diffusion.cpp.git
+  1d2af5ca3fc30dec9657a6cca8f87c20dfd83f1c
+  Tue May 14 23:06:52 2024 +0800
+
+LOCAL MODIFICATIONS
+
+  - Made crc32 go faster
+  - Make work with llama.cpp flavor of ggml
+  - Remove sd_type_t (error prone intended to be ggml_type)