@fugood/llama.node 0.3.0 → 0.3.2

package/src/llama.cpp/ggml/src/llamafile/sgemm.cpp

@@ -50,6 +50,7 @@
 
 #include "sgemm.h"
 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
 
 #ifdef _MSC_VER
@@ -235,6 +236,14 @@ template <> inline __m512 load(const ggml_fp16_t *p) {
 }
 #endif // __AVX512F__
 
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// CONSTANTS
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // FLOATING POINT MATRIX MULTIPLICATION
 
@@ -606,17 +615,29 @@ class tinyBLAS_Q0_AVX {
         case 0x44:
             mc = 4;
             nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<4>(m0, m, n0, n);
+#else
             gemm<4, 4>(m0, m, n0, n);
+#endif
             break;
         case 0x43:
             mc = 4;
             nc = 3;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<3>(m0, m, n0, n);
+#else
             gemm<4, 3>(m0, m, n0, n);
+#endif
             break;
         case 0x34:
             mc = 3;
             nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<3>(m0, m, n0, n);
+#else
             gemm<3, 4>(m0, m, n0, n);
+#endif
             break;
         case 0x33:
             mc = 3;
@@ -626,12 +647,20 @@ class tinyBLAS_Q0_AVX {
         case 0x42:
             mc = 4;
             nc = 2;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<2>(m0, m, n0, n);
+#else
             gemm<4, 2>(m0, m, n0, n);
+#endif
             break;
         case 0x24:
             mc = 2;
             nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<2>(m0, m, n0, n);
+#else
             gemm<2, 4>(m0, m, n0, n);
+#endif
             break;
 #else
         case 0x44:
@@ -639,13 +668,21 @@ class tinyBLAS_Q0_AVX {
         case 0x42:
             mc = 4;
             nc = 2;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<2>(m0, m, n0, n);
+#else
             gemm<4, 2>(m0, m, n0, n);
+#endif
             break;
         case 0x34:
         case 0x24:
             mc = 2;
             nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<2>(m0, m, n0, n);
+#else
             gemm<2, 4>(m0, m, n0, n);
+#endif
             break;
         case 0x33:
 #endif
@@ -662,7 +699,11 @@ class tinyBLAS_Q0_AVX {
         case 0x41:
             mc = 4;
             nc = 1;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<1>(m0, m, n0, n);
+#else
             gemm<4, 1>(m0, m, n0, n);
+#endif
             break;
         case 0x22:
             mc = 2;
@@ -672,7 +713,11 @@ class tinyBLAS_Q0_AVX {
         case 0x14:
             mc = 1;
             nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<1>(m0, m, n0, n);
+#else
             gemm<1, 4>(m0, m, n0, n);
+#endif
             break;
         case 0x31:
             mc = 3;
@@ -708,6 +753,119 @@ class tinyBLAS_Q0_AVX {
         mnpack(m0, m, np, n);
     }
 
+#if defined(__AVX2__) && defined(__F16C__)
+// Templated functions for gemm of dimensions 4xN
+    template <int RN>
+    NOINLINE void gemm4xN(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / 4;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * 4;
+            int64_t jj = n0 + job % xtiles * RN;
+            __m256 Cv[RN][4] = {};
+            for (int64_t l = 0; l < k; ++l) {
+                uint64_t a_delta = ((uint64_t)A[lda * (ii + 3) + l].d << 48) | ((uint64_t)A[lda * (ii + 2) + l].d << 32) | ((uint64_t)A[lda * (ii + 1) + l].d << 16) | (A[lda * (ii + 0) + l].d);
+                // Convert delta values for four blocks to float values
+                __m128 da = _mm_cvtph_ps(_mm_set_epi64x(0, a_delta));
+                __m256i avec0 = load(A + lda * (ii + 0) + l);
+                __m256i avec1 = load(A + lda * (ii + 1) + l);
+                __m256i avec2 = load(A + lda * (ii + 2) + l);
+                __m256i avec3 = load(A + lda * (ii + 3) + l);
+                for (int64_t j = 0; j < RN; ++j) {
+                        __m128 db = _mm_set1_ps(unhalf(B[ldb * (jj + j) + l].d));
+                        // Computation of product of delta values for four blocks and replicate it across 256 bit lane
+                        __m256 dvec =  _mm256_castps128_ps256(_mm_mul_ps(da, db));
+                        dvec = _mm256_permute2f128_ps(dvec ,dvec, 0);
+                        // Computation of dot product and multiplication with appropriate delta value products
+                        Cv[j][0] = madd(_mm256_shuffle_ps(dvec, dvec, 0),
+                                    updot(_mm256_sign_epi8(avec0, avec0),
+                                          _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec0)),
+                                    Cv[j][0]);
+                        Cv[j][1] = madd(_mm256_shuffle_ps(dvec, dvec, 85),
+                                    updot(_mm256_sign_epi8(avec1, avec1),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec1)),
+                                    Cv[j][1]);
+                        Cv[j][2] = madd(_mm256_shuffle_ps(dvec, dvec, 170),
+                                    updot(_mm256_sign_epi8(avec2, avec2),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec2)),
+                                    Cv[j][2]);
+                        Cv[j][3] = madd(_mm256_shuffle_ps(dvec, dvec, 255),
+                                    updot(_mm256_sign_epi8(avec3, avec3),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec3)),
+                                    Cv[j][3]);
+                }
+            }
+
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < 4; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    // Templated functions for gemm of dimensions Mx4
+    template <int RM>
+    NOINLINE void gemmMx4(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / 4;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * 4;
+            __m256 Cv[4][RM] = {};
+            for (int64_t l = 0; l < k; ++l) {
+                uint64_t b_delta = ((uint64_t)B[ldb * (jj + 3) + l].d << 48) | ((uint64_t)B[ldb * (jj + 2) + l].d << 32) | ((uint64_t)B[ldb * (jj + 1) + l].d << 16) | (B[ldb * (jj + 0) + l].d);
+                // Convert delta values for four blocks to float values
+                __m128 db = _mm_cvtph_ps(_mm_set_epi64x(0, b_delta));
+                __m256i bvec0 = load(B + ldb * (jj + 0) + l);
+                __m256i bvec1 = load(B + ldb * (jj + 1) + l);
+                __m256i bvec2 = load(B + ldb * (jj + 2) + l);
+                __m256i bvec3 = load(B + ldb * (jj + 3) + l);
+                for (int64_t i = 0; i < RM; ++i) {
+                    __m128 da = _mm_set1_ps(unhalf((A[lda * (ii + i) + l].d)));
+                    // Computation of product of delta values for four blocks and replicate it across 256 bit lane
+                    __m256 dvec =  _mm256_castps128_ps256(_mm_mul_ps(da, db));
+                    dvec = _mm256_permute2f128_ps(dvec ,dvec, 0);
+                    // Computation of dot product and multiplication with appropriate delta value products
+                    Cv[0][i] = madd(_mm256_shuffle_ps(dvec, dvec, 0),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec0, load(A + lda * (ii + i) + l))),
+                                    Cv[0][i]);
+                    Cv[1][i] = madd(_mm256_shuffle_ps(dvec, dvec, 85),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec1, load(A + lda * (ii + i) + l))),
+                                    Cv[1][i]);
+                    Cv[2][i] = madd(_mm256_shuffle_ps(dvec, dvec, 170),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec2, load(A + lda * (ii + i) + l))),
+                                    Cv[2][i]);
+                    Cv[3][i] = madd(_mm256_shuffle_ps(dvec, dvec, 255),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec3, load(A + lda * (ii + i) + l))),
+                                    Cv[3][i]);
+                }
+            }
+            for (int64_t j = 0; j < 4; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+#endif
+
     template <int RM, int RN>
     NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int64_t ytiles = (m - m0) / RM;
@@ -784,6 +942,20 @@ class tinyBLAS_Q0_AVX {
         return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
     }
 
+    inline __m256i load(const block_iq4_nl *b) {
+        return MM256_SET_M128I(load1(b), load0(b));
+    }
+
+    inline __m128i load0(const block_iq4_nl *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), x));
+    }
+
+    inline __m128i load1(const block_iq4_nl *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)));
+    }
+
     inline __m256 updot(__m256i u, __m256i s) {
         __m256i res;
 #if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
@@ -857,6 +1029,10 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
     assert(nth > 0);
     assert(ith < nth);
 
+    // only enable sgemm for prompt processing
+    if (n < 2)
+        return false;
+
     if (Ctype != GGML_TYPE_F32)
         return false;
 
@@ -1006,6 +1182,22 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
 #endif
     }
 
+    case GGML_TYPE_IQ4_NL: {
+        if (Btype != GGML_TYPE_Q8_0)
+            return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_iq4_nl, block_q8_0, float> tb{
+            k, (const block_iq4_nl *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
     default:
         return false;
     }

package/src/llama.cpp/ggml/src/vulkan-shaders/CMakeLists.txt

@@ -1,5 +1,7 @@
+find_package (Threads REQUIRED)
 
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)

package/src/llama.cpp/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -22,6 +22,7 @@
 #ifdef _WIN32
     #include <windows.h>
     #include <direct.h> // For _mkdir on Windows
+    #include <algorithm> // For std::replace on w64devkit
 #else
     #include <unistd.h>
     #include <sys/wait.h>
@@ -30,20 +31,6 @@
 
 #define ASYNCIO_CONCURRENCY 64
 
-// define prototypes
-void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str);
-bool directory_exists(const std::string& path);
-bool create_directory(const std::string& path);
-std::string to_uppercase(const std::string& input);
-bool string_ends_with(const std::string& str, const std::string& suffix);
-std::string join_paths(const std::string& path1, const std::string& path2);
-std::string basename(const std::string &path);
-void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16);
-std::map<std::string, std::string> merge_maps(const std::map<std::string, std::string>& a, const std::map<std::string, std::string>& b);
-void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmul_id);
-void process_shaders(std::vector<std::future<void>>& tasks);
-void write_output_files();
-
 std::mutex lock;
 std::vector<std::pair<std::string, std::string>> shader_fnames;
 
@@ -52,7 +39,7 @@ std::string input_dir = "vulkan-shaders";
 std::string output_dir = "/tmp";
 std::string target_hpp = "ggml-vulkan-shaders.hpp";
 std::string target_cpp = "ggml-vulkan-shaders.cpp";
-bool clean = true;
+bool no_clean = false;
 
 const std::vector<std::string> type_names = {
     "f32",
@@ -193,11 +180,7 @@ bool string_ends_with(const std::string& str, const std::string& suffix) {
     return std::equal(suffix.rbegin(), suffix.rend(), str.rbegin());
 }
 
-#ifdef _WIN32
-    static const char path_separator = '\\';
-#else
-    static const char path_separator = '/';
-#endif
+static const char path_separator = '/';
 
 std::string join_paths(const std::string& path1, const std::string& path2) {
     return path1 + path_separator + path2;
@@ -212,7 +195,16 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const
     std::string out_fname = join_paths(output_dir, name + ".spv");
     std::string in_path = join_paths(input_dir, in_fname);
 
-    std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o", out_fname};
+    #ifdef _WIN32
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
+    #else
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o",  out_fname};
+    #endif
+
+    #ifdef GGML_VULKAN_SHADER_DEBUG_INFO
+        cmd.push_back("-g");
+    #endif
+
     for (const auto& define : defines) {
         cmd.push_back("-D" + define.first + "=" + define.second);
     }
@@ -283,9 +275,12 @@ void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmu
 
     for (const auto& tname : type_names) {
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+        // For unaligned, load one at a time for f32/f16, or two at a time for quants
+        std::string load_vec_a_unaligned = (tname == "f32" || tname == "f16") ? "1" : "2";
+        // For aligned matmul loads
         std::string load_vec_a = (tname == "f32" || tname == "f16") ? load_vec : "2";
         tasks.push_back(std::async(std::launch::async, [=] {
-            string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
+            string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
         }));
         tasks.push_back(std::async(std::launch::async, [=] {
             string_to_spv(shader_name + "_" + tname + "_f32_aligned", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}}), fp16);
@@ -354,6 +349,9 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
     tasks.push_back(std::async(std::launch::async, [=] {
         string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
     tasks.push_back(std::async(std::launch::async, [=] {
         string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     }));
@@ -371,6 +369,13 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("acc_f32", "acc.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
 
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
@@ -384,6 +389,10 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
         string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));
 
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));
@@ -392,19 +401,54 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
         string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));
 
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));
 
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_f32", "concat.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_f16", "concat.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     }));
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
 
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -438,6 +482,17 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
     tasks.push_back(std::async(std::launch::async, [=] {
         string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}));
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
 }
 
 void write_output_files() {
@@ -449,10 +504,16 @@ void write_output_files() {
 
     for (const auto& pair : shader_fnames) {
         const std::string& name = pair.first;
-        const std::string& path = pair.second;
+        #ifdef _WIN32
+            std::string path = pair.second;
+            std::replace(path.begin(), path.end(), '/', '\\' );
+        #else
+            const std::string& path = pair.second;
+        #endif
+
         FILE* spv = fopen(path.c_str(), "rb");
         if (!spv) {
-            std::cerr << "Error opening SPIR-V file: " << path << "\n";
+            std::cerr << "Error opening SPIR-V file: " << path << " (" << strerror(errno) << ")\n";
             continue;
         }
 
@@ -464,7 +525,7 @@ void write_output_files() {
         size_t read_size = fread(data.data(), 1, size, spv);
         fclose(spv);
         if (read_size != size) {
-            std::cerr << "Error reading SPIR-V file: " << path << "\n";
+            std::cerr << "Error reading SPIR-V file: " << path << " (" << strerror(errno) << ")\n";
             continue;
         }
 
@@ -478,9 +539,8 @@ void write_output_files() {
         }
         fprintf(src, "\n};\n\n");
 
-        if (clean) {
+        if (!no_clean) {
             std::remove(path.c_str());
-            // fprintf(stderr, "Removed: %s\n", path.c_str());
         }
     }
 
@@ -496,18 +556,6 @@ int main(int argc, char** argv) {
         }
     }
 
-    if (argc <= 1 || args.find("--help") != args.end()) {
-        std::cout << "Usage:\n"
-                     "\tvulkan-shaders-gen [options]\n\n"
-                     "Options:\n"
-                     "\t--glslc <path>        Path to glslc executable (default: /usr/bin/glslc)\n"
-                     "\t--input-dir           Directory containing shader sources (required)\n"
-                     "\t--output-dir          Output directory for generated SPIR-V files and optional C++ headers\n"
-                     "\t--target-hpp <path>   Path to generate a header file with shader declarations in C++ format\n"
-                     "\t--target-cpp <path>   Path to generate a source code file implementing the declared shaders (optional)\n"
-                     "\t--no-clean            Keep temporary SPIR-V files after build (default: remove them)\n";
-        return EXIT_SUCCESS;
-    }
     if (args.find("--glslc") != args.end()) {
         GLSLC = args["--glslc"]; // Path to glslc
     }
@@ -524,7 +572,7 @@ int main(int argc, char** argv) {
         target_cpp = args["--target-cpp"]; // Path to generated cpp file
     }
     if (args.find("--no-clean") != args.end()) {
-        clean = false; // Keep temporary SPIR-V files in output-dir after build
+        no_clean = true; // Keep temporary SPIR-V files in output-dir after build
     }
 
     if (!directory_exists(input_dir)) {