diff --git a/examples/cli/cli.cpp b/examples/cli/cli.cpp index f202ebd5259..fccfd13eb0d 100644 --- a/examples/cli/cli.cpp +++ b/examples/cli/cli.cpp @@ -9,7 +9,6 @@ #include #include #include -#include #include #include @@ -1019,10 +1018,12 @@ int main(int argc, char ** argv) { bool open(const char * ext, const char * function) { if (is_stdout) { - if (std::exchange(used_stdout, true)) { + if (used_stdout) { fprintf(stderr, "warning: Not appending multiple file formats to stdout\n"); return false; } + + used_stdout = true; #ifdef _WIN32 fout = std::ofstream{"CON"}; #else @@ -1032,6 +1033,7 @@ int main(int argc, char ** argv) { // Also assuming /dev is mounted return true; } + fname_out.resize(basename_length); fname_out += ext; fout = std::ofstream{fname_out}; diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c index a3d3f690133..5fd379f6a94 100644 --- a/ggml/src/ggml-alloc.c +++ b/ggml/src/ggml-alloc.c @@ -816,7 +816,10 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) { size_t node_size = 0; if (!node->data && !node->view_src) { - GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API + // If we previously had data but don't now then reallocate + if (talloc->buffer_id < 0) { + return false; + } node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node); } return talloc->size_max >= node_size; diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp index f6374f7894a..1d46158f928 100644 --- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp +++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp @@ -1054,6 +1054,493 @@ class tinyBLAS_Q0_AVX { } \ } \ +template +class tinyBLAS_BF16_PPC { + public: + tinyBLAS_BF16_PPC(int64_t k, + const TA *A, int64_t lda, + const TB *B, int64_t ldb, + TC *C, int64_t ldc, + int ith, int nth) + : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) { + } + + void matmul(int64_t m, int64_t n) { + mnpack(0, m, 0, n); + } + + private: + void vector_permute_store(vec_t *c, int numVec, unsigned char *vecOffset) { + vec_t t[8], s[8]; + vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23}; + vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31}; + vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23}; + vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31}; + + if (numVec == 2) { + t[0] = vec_perm(c[0], c[1], swiz1); + t[1] = vec_perm(c[2], c[3], swiz1); + s[0] = vec_perm(t[0], t[1], swiz3); + s[1] = vec_perm(t[0], t[1], swiz4); + vec_xst(s[0], 0, (vec_t*)vecOffset); + vec_xst(s[1], 0, (vec_t*)(vecOffset + 16)); + } else if (numVec == 4) { + t[0] = vec_perm(c[0], c[1], swiz1); + t[1] = vec_perm(c[0], c[1], swiz2); + t[2] = vec_perm(c[2], c[3], swiz1); + t[3] = vec_perm(c[2], c[3], swiz2); + s[0] = vec_perm(t[0], t[2], swiz3); + s[1] = vec_perm(t[0], t[2], swiz4); + s[2] = vec_perm(t[1], t[3], swiz3); + s[3] = vec_perm(t[1], t[3], swiz4); + for (int i = 0; i < 4; ++i) + vec_xst(s[i], 0, (vec_t*)(vecOffset + i * 16)); + } else if (numVec == 8) { + for (int i = 0; i < 4; i += 2) { + t[i+0] = vec_perm(c[i+0], c[i+1], swiz1); + t[i+1] = vec_perm(c[i+0], c[i+1], swiz2); + } + for (int i = 4; i < 8; i += 2) { + t[i+0] = vec_perm(c[i+0], c[i+1], swiz1); + t[i+1] = vec_perm(c[i+0], c[i+1], swiz2); + } + s[0] = vec_perm(t[0], t[2], swiz3); + s[1] = vec_perm(t[0], t[2], swiz4); + s[2] = vec_perm(t[1], t[3], swiz3); + s[3] = vec_perm(t[1], t[3], swiz4); + s[4] = vec_perm(t[4], t[6], swiz3); + s[5] = vec_perm(t[4], t[6], swiz4); + s[6] = vec_perm(t[5], t[7], swiz3); + s[7] = vec_perm(t[5], t[7], swiz4); + for (int i = 0; i < 8; ++i) + vec_xst(s[i], 0, (vec_t*)(vecOffset + i * 16)); + } + } + + void packNormal(const TA* a, int64_t lda, int rows, int cols, unsigned char* vec) { + int64_t i, j; + TA *aoffset = NULL; + unsigned char *vecOffset = NULL; + TA * aoffsets[8]; + vector unsigned char c_arr[8]; + aoffset = const_cast(a); + vecOffset = vec; + j = (rows >> 3); + if (j > 0) { + do { + if (cols == 4) { + aoffsets[0] = aoffset; + for (int it = 1; it < 4; ++it) + aoffsets[it] = aoffsets[it-1] + lda; + aoffset += 4 * lda; + for (int i = 0; i < 4; ++i) + c_arr[i] = vec_xl(0, (vector unsigned char*)aoffsets[i]); + vector_permute_store(c_arr, 4, vecOffset); + for (int i = 0; i<4; i++) + aoffsets[i] = aoffsets[i]+lda; + vecOffset +=64; + } + i = (cols >> 3); + if (i > 0) { + aoffsets[0] = aoffset; + for (int it = 1; it < 8; ++it) { + aoffsets[it] = aoffsets[it-1] + lda; + } + aoffset += 8 * lda; + do { + for (int it = 0; it < 8; ++it) + c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]); + vector_permute_store(c_arr, 8, vecOffset); + for (int it = 0; it < 8; ++it) + aoffsets[it] = aoffsets[it] + 8*lda; + vecOffset += 128; + i--; + } while(i > 0); + } + j--; + } while(j > 0); + } + if (rows & 4) { + aoffsets[0] = aoffset; + for (int it = 1; it < 4; ++it) + aoffsets[it] = aoffsets[it-1] + lda; + aoffset += 4 * lda; + if (cols == 4) { + for (int it = 0; it < 4; ++it) + c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]); + vector_permute_store(c_arr, 2, vecOffset); + for (int it = 0; it< 4; it++) + aoffsets[it] = aoffsets[it] + lda; + vecOffset += 32; + } + i = (cols >> 3); + if (i > 0) { + do { + for (int it = 0; it < 4; ++it) + c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]); + vector_permute_store(c_arr, 4, vecOffset); + for (int it = 0; it< 4; it++) + aoffsets[it] = aoffsets[it] + 8*lda; + vecOffset += 64; + i--; + } while(i > 0); + } + } + if (rows & 3) { + aoffsets[0] = aoffset; + for (int it = 1; it < 4; ++it) + aoffsets[it] = aoffsets[it-1] + lda; + if (cols == 4) { + switch(rows) { + case 3: c_arr[2] = vec_xl(0, (vector unsigned char*)aoffsets[2]); + case 2: c_arr[1] = vec_xl(0, (vector unsigned char*)aoffsets[1]); + case 1: c_arr[0] = vec_xl(0, (vector unsigned char*)aoffsets[0]); + break; + } + vector_permute_store(c_arr, 2, vecOffset); + for (int it = 0; it< 4; it++) + aoffsets[it] = aoffsets[it] + lda; + vecOffset += 32; + } + i = (cols >> 3); + if (i > 0) { + do { + switch(rows) { + case 3: c_arr[2] = vec_xl(0, (vector unsigned char*)aoffsets[2]); + case 2: c_arr[1] = vec_xl(0, (vector unsigned char*)aoffsets[1]); + case 1: c_arr[0] = vec_xl(0, (vector unsigned char*)aoffsets[0]); + break; + } + vector_permute_store(c_arr, 4, vecOffset); + for (int it = 0; it <4; it++) + aoffsets[it] = aoffsets[it] + 8* lda; + vecOffset += 64; + i--; + } while(i > 0); + } + } + } + + void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t mc, nc, mp, np; + int m_rem = MIN(m - m0, 8); + int n_rem = MIN(n - n0, 8); + + if (m_rem >= 8 && n_rem >= 8) { + mc = 8; + nc = 8; + gemm<8,8>(m0, m, n0, n); + } else if (m_rem >= 4 && n_rem >= 8) { + mc = 4; + nc = 8; + gemm<4,8>(m0, m, n0, n); + } else if (m_rem >=8 && n_rem >=4){ + mc = 8; + nc = 4; + gemm<8,4>(m0, m, n0, n); + } else if ((m_rem < 4) && (n_rem >= 8)) { + nc = 8; + switch(m_rem) { + case 1: + mc = 1; + gemm_Mx8<1>(m0, m, n0, n); + break; + case 2: + mc = 2; + gemm_Mx8<2>(m0, m, n0, n); + break; + case 3: + mc = 3; + gemm_Mx8<3>(m0, m, n0, n); + break; + default: + return; + } + } else if (m_rem >= 4 && n_rem >= 4) { + mc = 4; + nc = 4; + gemm_small<4, 4>(m0, m, n0, n); + } else if ((m_rem > 4) && (n_rem < 4)) { + mc = 4; + switch(n_rem) { + case 1: + nc = 1; + gemm_small<4, 1>(m0, m, n0, n); + break; + case 2: + nc = 2; + gemm_small<4, 2>(m0, m, n0, n); + break; + case 3: + nc = 3; + gemm_small<4, 3>(m0, m, n0, n); + break; + + default: + return; + } + } else { + switch((m_rem << 4) | n_rem) { + case 0x43: + mc = 4; + nc = 3; + gemm_small<4, 3>(m0, m, n0, n); + break; + case 0x42: + mc = 4; + nc = 2; + gemm_small<4, 2>(m0, m, n0, n); + break; + case 0x41: + mc = 4; + nc = 1; + gemm_small<4, 1>(m0, m, n0, n); + break; + case 0x34: + mc = 3; + nc = 4; + gemm_small<3, 4>(m0, m, n0, n); + break; + case 0x33: + mc = 3; + nc = 3; + gemm_small<3, 3>(m0, m, n0, n); + break; + case 0x32: + mc = 3; + nc = 2; + gemm_small<3, 2>(m0, m, n0, n); + break; + case 0x31: + mc = 3; + nc = 1; + gemm_small<3, 1>(m0, m, n0, n); + break; + case 0x24: + mc = 2; + nc = 4; + gemm_small<2,4>(m0, m, n0, n); + break; + case 0x23: + mc = 2; + nc = 3; + gemm_small<2, 3>(m0, m, n0, n); + break; + case 0x22: + mc = 2; + nc = 2; + gemm_small<2, 2>(m0, m, n0, n); + break; + case 0x21: + mc = 2; + nc = 1; + gemm_small<2, 1>(m0, m, n0, n); + break; + case 0x14: + mc = 1; + nc = 4; + gemm_small<1, 4>(m0, m, n0, n); + break; + case 0x13: + mc = 1; + nc = 3; + gemm_small<1, 3>(m0, m, n0, n); + break; + case 0x12: + mc = 1; + nc = 2; + gemm_small<1, 2>(m0, m, n0, n); + break; + case 0x11: + mc = 1; + nc = 1; + gemm_small<1, 1>(m0, m, n0, n); + break; + default: + return; + } + } + mp = m0 + (m - m0) / mc * mc; + np = n0 + (n - n0) / nc * nc; + mnpack(mp, m, n0, np); + mnpack(m0, m, np, n); + } + + void KERNEL_4x8(int64_t ii, int64_t jj) { + vec_t vec_A[4], vec_B[8] , vec_C[4]; + acc_t acc_0, acc_1; + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + for (int l = 0; l < k; l+=8) { + packNormal((A+(ii*lda)+l), lda, 4, 8, (uint8_t*)vec_A); + packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B); + for (int x = 0; x < 4; x++) { + __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]); + __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]); + } + } + SAVE_ACC(&acc_0, ii, jj); + SAVE_ACC(&acc_1, ii, jj+4); + } + + void KERNEL_8x4(int64_t ii, int64_t jj) { + vec_t vec_A[8], vec_B[4] , vec_C[4]; + acc_t acc_0, acc_1; + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + for (int l = 0; l < k; l+=8) { + packNormal((A+(ii*lda)+l), lda, 8, 8, (uint8_t*)vec_A); + packNormal((B+(jj*ldb)+l), ldb, 8, 4, (uint8_t*)vec_B); + for (int x = 0; x < 4; x++) { + __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]); + __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x+4], vec_B[x]); + } + } + SAVE_ACC(&acc_0, ii, jj); + SAVE_ACC(&acc_1, ii+4, jj); + } + + + void KERNEL_8x8(int64_t ii, int64_t jj) { + vec_t vec_A[8], vec_B[8], vec_C[4]; + acc_t acc_0, acc_1, acc_2, acc_3; + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + __builtin_mma_xxsetaccz(&acc_2); + __builtin_mma_xxsetaccz(&acc_3); + for (int l = 0; l < k; l+=8) { + packNormal(A+(ii*lda)+l, lda, 8, 8, (uint8_t*)vec_A); + packNormal(B+(jj*ldb)+l, ldb, 8, 8, (uint8_t*)vec_B); + for (int x = 0; x < 4; x++) { + __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]); + __builtin_mma_xvbf16ger2pp(&acc_1, (vec_t)vec_A[x], (vec_t)vec_B[x+4]); + __builtin_mma_xvbf16ger2pp(&acc_2, (vec_t)vec_A[x+4], (vec_t)vec_B[x]); + __builtin_mma_xvbf16ger2pp(&acc_3, (vec_t)vec_A[x+4], (vec_t)vec_B[x+4]); + } + } + + SAVE_ACC(&acc_0, ii, jj); + SAVE_ACC(&acc_1, ii, jj+4); + SAVE_ACC(&acc_2, ii+4, jj); + SAVE_ACC(&acc_3, ii+4, jj+4); + } + + template + void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t ytiles = (m - m0) / RM; + 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 * RM; + int64_t jj = n0 + job % xtiles * RN; + vec_t vec_C[4]; + acc_t acc_0; + __builtin_mma_xxsetaccz(&acc_0); + vec_t vec_A[2], vec_B[2]; + for (int l=0; l + void gemm_Mx8(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int RN = 8; + int64_t ytiles = (m - m0) / RM; + 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 * RM; + int64_t jj = n0 + job % xtiles * RN; + vec_t vec_C[4]; + acc_t acc_0, acc_1; + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + vec_t vec_A[4], vec_B[8]; + for (int l=0; l + inline void kernel(int64_t ii, int64_t jj) { + if constexpr(RM == 4 && RN == 8) { + KERNEL_4x8(ii,jj); + } else if constexpr(RM == 8 && RN == 8) { + KERNEL_8x8(ii,jj); + } else if constexpr(RM == 8 && RN == 4) { + KERNEL_8x4(ii,jj); + } else { + static_assert(false, "RN/RM values not supported"); + } + } + + template + NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t ytiles = (m - m0) / RM; + 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 * RM; + int64_t jj = n0 + job % xtiles * RN; + kernel(ii, jj); + } + } + + const TA *const A; + const TB *const B; + TC *C; + const int64_t k; + const int64_t lda; + const int64_t ldb; + const int64_t ldc; + const int ith; + const int nth; +}; + template class tinyBLAS_Q0_PPC { public: @@ -2202,6 +2689,7 @@ class tinyBLAS_PPC { boffset = vec; j = (rows >> 3); if (j > 0) { + do { aoffset1 = aoffset; aoffset2 = aoffset1 + lda; @@ -2875,9 +3363,22 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 (float *)C, ldc}; return tb.matmul(m, n); } +#elif defined(__MMA__) + if ((k % 8)) + return false; + if(Btype == GGML_TYPE_BF16) { + tinyBLAS_BF16_PPC tb{ k, + (const ggml_bf16_t *)A, lda, + (const ggml_bf16_t *)B, ldb, + (float *)C, ldc, + params->ith, params->nth}; + tb.matmul(m, n); + return true; + } #endif return false; } + case GGML_TYPE_F16: { #if defined(__AVX512F__) if (Btype == GGML_TYPE_F16) { diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt index 8623214c78a..f3cfdeaef51 100644 --- a/ggml/src/ggml-cuda/CMakeLists.txt +++ b/ggml/src/ggml-cuda/CMakeLists.txt @@ -133,6 +133,7 @@ if (CUDAToolkit_FOUND) COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion" OUTPUT_VARIABLE CUDA_CCVER ERROR_QUIET + OUTPUT_STRIP_TRAILING_WHITESPACE ) else() if (CUDA_CCFULLVER MATCHES Apple) @@ -143,7 +144,7 @@ if (CUDAToolkit_FOUND) string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER}) endif() - message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}") + message(STATUS "CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}") ggml_get_flags(${CUDA_CCID} ${CUDA_CCVER}) list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS}) # This is passed to -Xcompiler later diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp index 140a775f980..e662cc6eb3f 100644 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp @@ -518,6 +518,11 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) { result.view_src = reinterpret_cast(tensor->view_src); result.view_offs = tensor->view_offs; result.data = reinterpret_cast(tensor->data); + + // Avoid sending uninitialized data over the wire + memset(result.name, 0, sizeof(result.name)); + memset(result.padding, 0, sizeof(result.padding)); + snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name); return result; } diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt index 9d028f718d0..31816219c06 100644 --- a/ggml/src/ggml-vulkan/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/CMakeLists.txt @@ -71,6 +71,22 @@ if (Vulkan_FOUND) add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) endif() + # Compile a test shader to determine whether GL_EXT_bfloat16 is supported. + # If it's not, there will be an error to stderr. + # If it's supported, set a define to indicate that we should compile those shaders + execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp" + OUTPUT_VARIABLE glslc_output + ERROR_VARIABLE glslc_error) + + if (${glslc_error} MATCHES ".*extension not supported: GL_EXT_bfloat16.*") + message(STATUS "GL_EXT_bfloat16 not supported by glslc") + set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT OFF) + else() + message(STATUS "GL_EXT_bfloat16 supported by glslc") + set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT ON) + add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + endif() + target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan) target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR}) @@ -142,6 +158,7 @@ if (Vulkan_FOUND) -DGGML_VULKAN_COOPMAT_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT_GLSLC_SUPPORT} -DGGML_VULKAN_COOPMAT2_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT} -DGGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT=${GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT} + -DGGML_VULKAN_BFLOAT16_GLSLC_SUPPORT=${GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT} BUILD_COMMAND ${CMAKE_COMMAND} --build . INSTALL_COMMAND ${CMAKE_COMMAND} --install . INSTALL_DIR ${CMAKE_BINARY_DIR} diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index c0bdb9e17a7..c4df47abae4 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -51,6 +51,24 @@ #include "ggml-vulkan-shaders.hpp" +// remove this once it's more widely available in the SDK +#if !defined(VK_KHR_shader_bfloat16) + +#define VK_KHR_shader_bfloat16 1 +#define VK_KHR_SHADER_BFLOAT16_SPEC_VERSION 1 +#define VK_KHR_SHADER_BFLOAT16_EXTENSION_NAME "VK_KHR_shader_bfloat16" +#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR ((VkStructureType)1000141000) +#define VK_COMPONENT_TYPE_BFLOAT16_KHR ((VkComponentTypeKHR)1000141000) + +typedef struct VkPhysicalDeviceShaderBfloat16FeaturesKHR { + VkStructureType sType; + void* pNext; + VkBool32 shaderBFloat16Type; + VkBool32 shaderBFloat16DotProduct; + VkBool32 shaderBFloat16CooperativeMatrix; +} VkPhysicalDeviceShaderBfloat16FeaturesKHR; +#endif + #define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1)) #define CEIL_DIV(M, N) (((M) + (N)-1) / (N)) static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; } @@ -266,8 +284,9 @@ struct vk_device_struct { bool subgroup_require_full_support; bool coopmat_support; - bool coopmat_acc_f32_support; - bool coopmat_acc_f16_support; + bool coopmat_acc_f32_support {}; + bool coopmat_acc_f16_support {}; + bool coopmat_bf16_support {}; uint32_t coopmat_m; uint32_t coopmat_n; uint32_t coopmat_k; @@ -293,6 +312,7 @@ struct vk_device_struct { vk_matmul_pipeline pipeline_matmul_f32 {}; vk_matmul_pipeline pipeline_matmul_f32_f16 {}; + vk_matmul_pipeline pipeline_matmul_bf16 {}; vk_matmul_pipeline2 pipeline_matmul_f16; vk_matmul_pipeline2 pipeline_matmul_f16_f32; @@ -301,6 +321,7 @@ struct vk_device_struct { vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_COUNT]; vk_matmul_pipeline pipeline_matmul_id_f32 {}; + vk_matmul_pipeline pipeline_matmul_id_bf16 {}; vk_matmul_pipeline2 pipeline_matmul_id_f16; vk_matmul_pipeline2 pipeline_matmul_id_f16_f32; @@ -333,8 +354,8 @@ struct vk_device_struct { vk_pipeline pipeline_clamp_f32; vk_pipeline pipeline_pad_f32; vk_pipeline pipeline_repeat_f32, pipeline_repeat_back_f32; - vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16; - vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16; + vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16, pipeline_cpy_f32_bf16; + vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16, pipeline_contig_cpy_f32_bf16; vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT]; vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT]; vk_pipeline pipeline_norm_f32; @@ -368,6 +389,8 @@ struct vk_device_struct { vk_pipeline pipeline_rwkv_wkv6_f32; vk_pipeline pipeline_rwkv_wkv7_f32; vk_pipeline pipeline_opt_step_adamw_f32; + vk_pipeline pipeline_conv2d_dw_whcn_f32; + vk_pipeline pipeline_conv2d_dw_cwhn_f32; // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned} vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2]; @@ -680,6 +703,24 @@ struct vk_op_rwkv_wkv7_push_constants { uint32_t H; }; +struct vk_op_conv2d_dw_push_constants { + uint32_t ne; + uint32_t batches; + uint32_t channels; + uint32_t dst_w; + uint32_t dst_h; + uint32_t src_w; + uint32_t src_h; + uint32_t knl_w; + uint32_t knl_h; + int32_t stride_x; + int32_t stride_y; + int32_t pad_x; + int32_t pad_y; + int32_t dilation_x; + int32_t dilation_y; +}; + struct vk_op_upscale_push_constants { uint32_t ne; uint32_t a_offset; uint32_t d_offset; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03; @@ -1791,6 +1832,12 @@ static void ggml_vk_load_shaders(vk_device& device) { if (!device->pipeline_matmul_id_f32) { device->pipeline_matmul_id_f32 = std::make_shared(); } + if (!device->pipeline_matmul_bf16) { + device->pipeline_matmul_bf16 = std::make_shared(); + } + if (!device->pipeline_matmul_id_bf16) { + device->pipeline_matmul_id_bf16 = std::make_shared(); + } std::vector> compiles; auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, @@ -1900,6 +1947,11 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3) +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3) + } +#endif CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) @@ -1921,6 +1973,11 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) + } +#endif CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) @@ -1974,6 +2031,11 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ) + } +#endif if (device->coopmat_acc_f16_support) { CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); @@ -2022,6 +2084,11 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + } +#endif if (device->coopmat_acc_f16_support) { CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); @@ -2104,6 +2171,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); @@ -2139,6 +2208,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); @@ -2191,6 +2262,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); @@ -2226,6 +2299,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); @@ -2246,8 +2321,26 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); -#undef CREATE_MM } + // reusing CREATE_MM from the fp32 path + if ((device->coopmat2 || device->coopmat_support) +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + && !device->coopmat_bf16_support +#endif + ) { + // use scalar tile sizes + l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 }; + m_warptile = { 128, 64, 64, 16, subgroup_size_8, 32, 2, 4, 2, 1, subgroup_size_8 }; + s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, 2, 2, 1, subgroup_size_8 }; + + l_wg_denoms = {128, 128, 1 }; + m_wg_denoms = { 64, 64, 1 }; + s_wg_denoms = { 32, 32, 1 }; + + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id); + } +#undef CREATE_MM // mul mat vec @@ -2266,6 +2359,7 @@ static void ggml_vk_load_shaders(vk_device& device) { for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) { ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1), mul_mat_vec_f32_f32_f32_len, mul_mat_vec_f32_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1), mul_mat_vec_f16_f32_f32_len, mul_mat_vec_f16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f32_f32_len, mul_mat_vec_bf16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); @@ -2288,6 +2382,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1), mul_mat_vec_f32_f16_f32_len, mul_mat_vec_f32_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1), mul_mat_vec_f16_f16_f32_len, mul_mat_vec_f16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f16_f32_len, mul_mat_vec_bf16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); @@ -2311,6 +2406,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32", mul_mat_vec_id_f16_f32_len, mul_mat_vec_id_f16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_BF16], "mul_mat_vec_id_bf16_f32", mul_mat_vec_id_bf16_f32_len, mul_mat_vec_id_bf16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true); @@ -2356,6 +2452,7 @@ static void ggml_vk_load_shaders(vk_device& device) { // get_rows ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F32 ], "get_rows_f32", get_rows_f32_len, get_rows_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F16 ], "get_rows_f16", get_rows_f16_len, get_rows_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_BF16], "get_rows_bf16", get_rows_bf16_len, get_rows_bf16_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_0], "get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_1], "get_rows_q4_1", get_rows_q4_1_len, get_rows_q4_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); @@ -2373,6 +2470,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f32_f32", get_rows_f32_f32_len, get_rows_f32_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F16 ], "get_rows_f16_f32", get_rows_f16_f32_len, get_rows_f16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_BF16], "get_rows_bf16_f32", get_rows_bf16_f32_len, get_rows_bf16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_0], "get_rows_q4_0_f32", get_rows_q4_0_f32_len, get_rows_q4_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_1], "get_rows_q4_1_f32", get_rows_q4_1_f32_len, get_rows_q4_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); @@ -2399,7 +2497,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true); } } - ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 9 * sizeof(uint32_t), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); @@ -2410,10 +2508,13 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_bf16,"cpy_f32_bf16",cpy_f32_bf16_len,cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_bf16,"contig_cpy_f32_bf16",contig_cpy_f32_bf16_len,contig_cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + if (device->float_controls_rte_fp16) { ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); @@ -2529,6 +2630,9 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f32, "conv2d_dw_cwhn_f32", conv2d_dw_cwhn_f32_len, conv2d_dw_cwhn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1); + for (auto &c : compiles) { c.wait(); } @@ -2578,6 +2682,7 @@ static vk_device ggml_vk_get_device(size_t idx) { bool coopmat2_support = false; device->coopmat_support = false; device->integer_dot_product = false; + bool bfloat16_support = false; for (const auto& properties : ext_props) { if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) { @@ -2608,6 +2713,9 @@ static vk_device ggml_vk_get_device(size_t idx) { !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) { device->integer_dot_product = true; #endif + } else if (strcmp("VK_KHR_shader_bfloat16", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_BFLOAT16")) { + bfloat16_support = true; } } @@ -2794,6 +2902,17 @@ static vk_device ggml_vk_get_device(size_t idx) { } #endif +#if defined(VK_KHR_shader_bfloat16) + VkPhysicalDeviceShaderBfloat16FeaturesKHR bfloat16_features {}; + bfloat16_features.pNext = nullptr; + bfloat16_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR; + if (bfloat16_support) { + last_struct->pNext = (VkBaseOutStructure *)&bfloat16_features; + last_struct = (VkBaseOutStructure *)&bfloat16_features; + device_extensions.push_back("VK_KHR_shader_bfloat16"); + } +#endif + VkPhysicalDeviceMaintenance4Features maint4_features {}; maint4_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_FEATURES; if (maintenance4_support) { @@ -2991,6 +3110,25 @@ static vk_device ggml_vk_get_device(size_t idx) { device->coopmat_int_n = prop.NSize; device->coopmat_int_k = prop.KSize; } +#if defined(VK_KHR_shader_bfloat16) && defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (prop.AType == VK_COMPONENT_TYPE_BFLOAT16_KHR && + prop.BType == VK_COMPONENT_TYPE_BFLOAT16_KHR && + prop.CType == VK_COMPONENT_TYPE_FLOAT32_KHR && + prop.ResultType == VK_COMPONENT_TYPE_FLOAT32_KHR && + (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup + ) { + // coopmat sizes not set yet + if (device->coopmat_m == 0) { + device->coopmat_bf16_support = true; + device->coopmat_m = prop.MSize; + device->coopmat_n = prop.NSize; + device->coopmat_k = prop.KSize; + } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) { + // Only enable if shape is identical + device->coopmat_bf16_support = true; + } + } +#endif } if (device->coopmat_m == 0 || !device->coopmat_acc_f32_support) { @@ -2998,11 +3136,19 @@ static vk_device ggml_vk_get_device(size_t idx) { GGML_LOG_DEBUG("ggml_vulkan: WARNING: No suitable matrix core mode found. Disabling matrix cores.\n"); device->coopmat_support = false; } + if (getenv("GGML_VK_DISABLE_BFLOAT16")) { + device->coopmat_bf16_support = false; + } } if (device->coopmat_support) { device_extensions.push_back("VK_KHR_cooperative_matrix"); } +#if defined(VK_KHR_shader_bfloat16) + if (device->coopmat_bf16_support) { + device_extensions.push_back("VK_KHR_shader_bfloat16"); + } +#endif #endif device->name = GGML_VK_NAME + std::to_string(idx); @@ -3459,6 +3605,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) { return ctx->device->pipeline_matmul_f32_f16; } + if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) { + return ctx->device->pipeline_matmul_bf16; + } if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) { if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_f16_f32.f16acc; @@ -3530,6 +3679,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * switch (a_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -3562,6 +3712,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_id_f32; } + if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) { + return ctx->device->pipeline_matmul_id_bf16; + } if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) { if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_id_f16_f32.f16acc; @@ -3615,6 +3768,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context switch (a_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -4350,6 +4504,13 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_cpy_f16_f16; } } + if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_BF16) { + if (contig) { + return ctx->device->pipeline_contig_cpy_f32_bf16; + } else { + return ctx->device->pipeline_cpy_f32_bf16; + } + } if (src->type == GGML_TYPE_F32) { switch (to) { case GGML_TYPE_Q4_0: @@ -4477,8 +4638,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) || !ggml_vk_dim01_contiguous(src0); const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) || + (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) || !ggml_vk_dim01_contiguous(src1); + // If src0 is BF16, try to use a BF16 x BF16 multiply + ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16; + const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig; bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0; @@ -4488,25 +4653,25 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub if (mmp == nullptr) { // Fall back to f16 dequant mul mat - mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]); + mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]); quantize_y = false; } const bool qx_needs_dequant = mmp == nullptr || x_non_contig; - const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig); + const bool qy_needs_dequant = !quantize_y && ((src1->type != f16_type && !y_f32_kernel) || y_non_contig); if (qx_needs_dequant) { // Fall back to dequant + f16 mulmat - mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]); + mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]); } // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT - const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type))); + const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type))); const bool aligned = !quantize_y && ne10 == kpad && ne01 > 8 && ne11 > 8; - vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)); + vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)); // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11; @@ -4527,12 +4692,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub vk_pipeline to_q8_1 = nullptr; if (x_non_contig) { - to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16); + to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type); } else { to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type); } if (y_non_contig) { - to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16); + to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type); } else { to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); } @@ -4949,6 +5114,8 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con const uint64_t nb01 = src0->nb[1]; const uint64_t nb02 = src0->nb[2]; + const uint64_t nb12 = src1->nb[2]; + // const uint64_t ne10 = src1->ne[0]; const uint64_t ne11 = src1->ne[1]; const uint64_t ne12 = src1->ne[2]; @@ -4974,6 +5141,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t); const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t); + const uint32_t channel_stride_y = nb12 / sizeof(float); const uint64_t qx_sz = ggml_nbytes(src0); const uint64_t qy_sz = ggml_nbytes(src1); @@ -5004,7 +5172,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset; // compute - const std::array pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) }; + const std::array pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) }; ggml_vk_sync_buffers(subctx); ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 }); @@ -5029,7 +5197,7 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four) // when ne12 and ne13 are one. } else if ((dst->ne[1] == 1 || (dst->ne[1] <= mul_mat_vec_max_cols && src1->ne[2] * src1->ne[3] == 1)) && - (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) { + (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || ggml_is_quantized(src0->type))) { ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun); } else { ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, dryrun); @@ -5097,27 +5265,31 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) || !ggml_vk_dim01_contiguous(src0); const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) || + (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) || !ggml_vk_dim01_contiguous(src1); + // If src0 is BF16, try to use a BF16 x BF16 multiply + ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16; + const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig; - vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]); + vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]); const bool qx_needs_dequant = mmp == nullptr || x_non_contig; - const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig; + const bool qy_needs_dequant = (src1->type != f16_type && !y_f32_kernel) || y_non_contig; if (qx_needs_dequant) { // Fall back to dequant + f16 mulmat - mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]); + mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]); } // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT - const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? GGML_TYPE_F16 : src0->type)); + const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? f16_type : src0->type)); const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8; - vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type); + vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? f16_type : src0->type); // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11; @@ -5136,12 +5308,12 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& vk_pipeline to_fp16_vk_1 = nullptr; if (x_non_contig) { - to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16); + to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type); } else { to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type); } if (y_non_contig) { - to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16); + to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type); } else { to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); } @@ -5988,6 +6160,15 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_leaky_relu_f32; } return nullptr; + case GGML_OP_CONV_2D_DW: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + if (ggml_is_contiguous(src1)) { + return ctx->device->pipeline_conv2d_dw_whcn_f32; + } else if (ggml_is_contiguous_channels(src1)) { + return ctx->device->pipeline_conv2d_dw_cwhn_f32; + } + } + return nullptr; default: return nullptr; } @@ -6014,6 +6195,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) { case GGML_OP_REPEAT_BACK: case GGML_OP_ROPE: case GGML_OP_RMS_NORM: + case GGML_OP_CONV_2D_DW: return true; default: return false; @@ -6310,6 +6492,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co case GGML_OP_CONCAT: case GGML_OP_UPSCALE: case GGML_OP_UNARY: + case GGML_OP_CONV_2D_DW: { const uint32_t ne = ggml_nelements(dst); if (ne > 262144) { @@ -7096,6 +7279,30 @@ static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, c }, dryrun); } +static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + vk_op_conv2d_dw_push_constants p{}; + p.ne = ggml_nelements(dst); + p.channels = dst->ne[2]; + p.batches = dst->ne[3]; + p.dst_w = dst->ne[0]; + p.dst_h = dst->ne[1]; + p.src_w = src1->ne[0]; + p.src_h = src1->ne[1]; + p.knl_w = src0->ne[0]; + p.knl_h = src0->ne[1]; + p.stride_x = dst->op_params[0]; + p.stride_y = dst->op_params[1]; + p.pad_x = dst->op_params[2]; + p.pad_y = dst->op_params[3]; + p.dilation_x = dst->op_params[4]; + p.dilation_y = dst->op_params[5]; + + GGML_ASSERT(src0->ne[3] == p.channels); + GGML_ASSERT(src1->ne[3] == p.batches); + + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p), dryrun); +} + static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const float * op_params = (const float *)dst->op_params; ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun); @@ -8116,6 +8323,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_POOL_2D: + case GGML_OP_CONV_2D_DW: case GGML_OP_RWKV_WKV6: case GGML_OP_RWKV_WKV7: case GGML_OP_LEAKY_RELU: @@ -8179,6 +8387,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_POOL_2D: + case GGML_OP_CONV_2D_DW: case GGML_OP_LEAKY_RELU: { // These operations all go through ggml_vk_op_f32, so short-circuit and @@ -8352,6 +8561,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_POOL_2D: ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun); + break; + case GGML_OP_CONV_2D_DW: + ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node, dryrun); + break; case GGML_OP_LEAKY_RELU: ggml_vk_leaky_relu(ctx, compute_ctx, src0, node, dryrun); @@ -8473,6 +8686,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_POOL_2D: + case GGML_OP_CONV_2D_DW: case GGML_OP_RWKV_WKV6: case GGML_OP_RWKV_WKV7: case GGML_OP_LEAKY_RELU: @@ -9227,6 +9441,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm switch (src0_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -9262,10 +9477,15 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm if (a->ne[3] != b->ne[3]) { return false; } - if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) || + if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16) || !(ggml_vk_dim01_contiguous(op->src[1]) || op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16)) { return false; } + if (op->src[0]->type == GGML_TYPE_BF16 && op->src[1]->type == GGML_TYPE_F16) { + // We currently don't have a bf16 x f16 shader, or an fp16->bf16 copy shader. + // So don't support this combination for now. + return false; + } return true; } break; @@ -9338,6 +9558,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm switch (op->src[0]->type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -9368,6 +9589,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm switch (src1_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -9442,6 +9664,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm case GGML_OP_COUNT_EQUAL: case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: + case GGML_OP_CONV_2D_DW: case GGML_OP_POOL_2D: case GGML_OP_RWKV_WKV6: case GGML_OP_RWKV_WKV7: diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt index d6e0b2a5a5d..ad13f69b3fb 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt @@ -12,6 +12,9 @@ endif() if (GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) endif() +if (GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) +endif() set(TARGET vulkan-shaders-gen) add_executable(${TARGET} vulkan-shaders-gen.cpp) install(TARGETS ${TARGET} RUNTIME) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp b/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp index dd828c23262..6567a8c54cf 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp @@ -18,7 +18,11 @@ void main() { // fast path for when all four iterations are in-bounds if (idx + (num_iter-1)*num_threads < p.ne) { [[unroll]] for (uint i = 0; i < num_iter; ++i) { -#ifndef OPTIMIZATION_ERROR_WORKAROUND + +#if defined(DATA_D_BF16) + float f = float(data_a[get_aoffset() + idx]); + data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f)); +#elif !defined(OPTIMIZATION_ERROR_WORKAROUND) data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]); #else data_d[get_doffset() + idx] = data_a[get_aoffset() + idx]; @@ -31,7 +35,10 @@ void main() { continue; } -#ifndef OPTIMIZATION_ERROR_WORKAROUND +#if defined(DATA_D_BF16) + float f = float(data_a[get_aoffset() + idx]); + data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f)); +#elif !defined(OPTIMIZATION_ERROR_WORKAROUND) data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]); #else data_d[get_doffset() + idx] = data_a[get_aoffset() + idx]; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp b/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp new file mode 100644 index 00000000000..938c74da500 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp @@ -0,0 +1,105 @@ +#version 450 + +#include "types.comp" + +layout (push_constant) uniform parameter +{ + uint ne; + uint batches; + uint channels; + uint dst_w; + uint dst_h; + uint src_w; + uint src_h; + uint knl_w; + uint knl_h; + int stride_x; + int stride_y; + int pad_x; + int pad_y; + int dilation_x; + int dilation_y; +} p; + +layout (binding = 0) readonly buffer A {A_TYPE knl_data[];}; +layout (binding = 1) readonly buffer B {B_TYPE src_data[];}; +layout (binding = 2) writeonly buffer D {D_TYPE dst_data[];}; + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE conv_2d_dw_whcn(uint idx) { + uint i0 = idx / p.dst_w; + uint dst_x = idx - i0 * p.dst_w; + uint i1 = i0 / p.dst_h; + uint dst_y = i0 - i1 * p.dst_h; + uint n = i1 / p.channels; + uint c = i1 - n * p.channels; + + uint src_i = n * p.channels * p.src_h * p.src_w + c * p.src_h * p.src_w; + uint knl_i = c * p.knl_h * p.knl_w; + + FLOAT_TYPE sum = 0.0; + for (uint knl_y = 0; knl_y < p.knl_h; ++knl_y) { + uint src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y; + if (src_y >= p.src_h) { // src_y < 0 will wrap to a large unsigned int + continue; + } + for (uint knl_x = 0; knl_x < p.knl_w; ++knl_x) { + uint src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x; + if (src_x >= p.src_w) { // src_x < 0 will wrap to a large unsigned int + continue; + } + FLOAT_TYPE v = FLOAT_TYPE(src_data[src_i + src_y * p.src_w + src_x]); + FLOAT_TYPE k = FLOAT_TYPE(knl_data[knl_i + knl_y * p.knl_w + knl_x]); + sum = fma(v, k, sum); + } + } + return sum; +} + +FLOAT_TYPE conv_2d_dw_cwhn(uint idx) { + uint i0 = idx / p.channels; + uint c = idx - i0 * p.channels; + uint i1 = i0 / p.dst_w; + uint dst_x = i0 - i1 * p.dst_w; + uint n = i1 / p.dst_h; + uint dst_y = i1 - n * p.dst_h; + + uint src_i = n * p.channels * p.src_h * p.src_w; + uint src_row = p.src_w * p.channels; + uint knl_row = p.knl_w * p.channels; + + FLOAT_TYPE sum = 0.0; + for (uint knl_y = 0; knl_y < p.knl_h; ++knl_y) { + uint src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y; + if (src_y >= p.src_h) { // src_y < 0 will wrap to a large unsigned int + continue; + } + for (uint knl_x = 0; knl_x < p.knl_w; ++knl_x) { + uint src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x; + if (src_x >= p.src_w) { // src_x < 0 will wrap to a large unsigned int + continue; + } + FLOAT_TYPE v = FLOAT_TYPE(src_data[src_i + src_y * src_row + src_x * p.channels + c]); + FLOAT_TYPE k = FLOAT_TYPE(knl_data[ knl_y * knl_row + knl_x * p.channels + c]); + sum = fma(v, k, sum); + } + } + return sum; +} + +void main() { + uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; + if (idx >= p.ne) { + return; + } + + FLOAT_TYPE result = +#ifdef WHCN + conv_2d_dw_whcn(idx); +#else + conv_2d_dw_cwhn(idx); +#endif + dst_data[idx] = D_TYPE(result); +} + diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp index 29c9064942d..f476a2e3dd8 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp @@ -12,7 +12,10 @@ void main() { return; } -#ifndef OPTIMIZATION_ERROR_WORKAROUND +#if defined(DATA_D_BF16) + float f = float(data_a[get_aoffset() + src0_idx(idx)]); + data_d[get_doffset() + dst_idx(idx)] = D_TYPE(fp32_to_bf16(f)); +#elif !defined(OPTIMIZATION_ERROR_WORKAROUND) data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]); #else data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)]; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp index 2a162a2c815..0d9739d4060 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp @@ -23,6 +23,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { } #endif +#if defined(DATA_A_BF16) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + return vec2(bf16_to_fp32(data_a[a_offset + ib]), bf16_to_fp32(data_a[a_offset + ib + 1])); +} +#endif + #if defined(DATA_A_Q4_0) vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); @@ -428,7 +434,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) { } #endif -#if defined(DATA_A_F32) || defined(DATA_A_F16) +#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16) vec2 get_dm(uint ib, uint a_offset) { return vec2(0, 0); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp index e877ed7796a..ee6b86a18dd 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp @@ -20,9 +20,14 @@ void main() { const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03; const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23; +#if defined(DATA_A_BF16) + FLOAT_TYPE v = FLOAT_TYPE(bf16_to_fp32(data_a[a_offset + i00])); +#else + FLOAT_TYPE v = FLOAT_TYPE(data_a[a_offset + i00]); +#endif #ifndef OPTIMIZATION_ERROR_WORKAROUND - data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]); + data_d[d_offset + i00] = D_TYPE(v); #else - data_d[d_offset + i00] = data_a[a_offset + i00]; + data_d[d_offset + i00] = D_TYPE(v); #endif } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp index 775b48cd05e..bb429dd5945 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp @@ -6,7 +6,7 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -#if !defined(DATA_A_F32) && !defined(DATA_A_F16) +#if !defined(DATA_A_F32) && !defined(DATA_A_F16) && !defined(DATA_A_BF16) #define K_PER_ITER 8 #else #define K_PER_ITER 2 diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp index 48376637fb3..bc633369f9b 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp @@ -21,7 +21,9 @@ layout (push_constant) uniform parameter uint nrows_x; uint row_stride_x; uint channel_stride_x; + uint channel_stride_y; uint channel_x_divisor; + uint ne12; uint b_offset; uint d_offset; } p; @@ -33,6 +35,7 @@ void main() { const uint row_x = gl_GlobalInvocationID.y; const uint channel = gl_GlobalInvocationID.z; const uint channel_x = channel / p.channel_x_divisor; + const uint channel_y = channel % p.ne12; const uint nrows_y = p.ncols_x; const uint nrows_dst = p.nrows_x; @@ -56,7 +59,7 @@ void main() { const uint row_y = col_x; const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; - const uint iy = channel*nrows_y + row_y; + const uint iy = channel_y*p.channel_stride_y + row_y; const vec4 av4 = vec4(data_a_v4[ix / 4]); const vec4 bv4 = vec4(data_b_v4[iy / 4]); @@ -72,7 +75,7 @@ void main() { const uint row_y = col_x; const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; - const uint iy = channel*nrows_y + row_y; + const uint iy = channel_y*p.channel_stride_y + row_y; const vec4 av4 = vec4(data_a_v4[ix / 4]); const vec4 bv4 = vec4(data_b_v4[iy / 4]); @@ -89,7 +92,7 @@ void main() { const uint row_y = col_x; const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; - const uint iy = channel*nrows_y + row_y; + const uint iy = channel_y*p.channel_stride_y + row_y; const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp index 23ce8ceec33..529ac4d44fe 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp @@ -10,6 +10,10 @@ #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require #endif +#if defined(DATA_A_BF16) && defined(COOPMAT) +#extension GL_EXT_bfloat16 : enable +#endif + #ifdef COOPMAT #extension GL_KHR_cooperative_matrix : enable #extension GL_KHR_memory_scope_semantics : enable @@ -29,6 +33,10 @@ #define LOAD_VEC_B 1 #endif +#if !defined(TO_FLOAT_TYPE) +#define TO_FLOAT_TYPE FLOAT_TYPE +#endif + layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; @@ -202,8 +210,8 @@ void main() { #endif #ifdef COOPMAT - coopmat cache_a; - coopmat cache_b; + coopmat cache_a; + coopmat cache_b; coopmat sums[cms_per_row * cms_per_col]; [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) { @@ -248,6 +256,21 @@ void main() { buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(0.0f); } #endif +#elif defined(DATA_A_BF16) +#if LOAD_VEC_A == 4 + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + buf_a[buf_idx ] = TO_FLOAT_TYPE(data_a[idx].x); + buf_a[buf_idx + 1] = TO_FLOAT_TYPE(data_a[idx].y); + buf_a[buf_idx + 2] = TO_FLOAT_TYPE(data_a[idx].z); + buf_a[buf_idx + 3] = TO_FLOAT_TYPE(data_a[idx].w); +#else + if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) { + buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]); + } else { + buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(uint16_t(0)); + } +#endif #elif defined(DATA_A_Q4_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 4 * loadr_a; @@ -695,13 +718,13 @@ void main() { const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b; #endif const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B; - buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x); - buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y); - buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z); - buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w); + buf_b[buf_idx + 0] = TO_FLOAT_TYPE(data_b[idx].x); + buf_b[buf_idx + 1] = TO_FLOAT_TYPE(data_b[idx].y); + buf_b[buf_idx + 2] = TO_FLOAT_TYPE(data_b[idx].z); + buf_b[buf_idx + 3] = TO_FLOAT_TYPE(data_b[idx].w); #elif !MUL_MAT_ID if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) { - buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]); } else { buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f); } @@ -709,7 +732,7 @@ void main() { const uint row_i = ic * BN + loadc_b + l; if (row_i < _ne1) { const u16vec2 row_idx = row_ids[row_i]; - buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]); } else { buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp index 06b7ab09ea5..344b466101b 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp @@ -14,6 +14,9 @@ #extension GL_EXT_buffer_reference : enable #extension GL_KHR_shader_subgroup_ballot : enable #extension GL_KHR_shader_subgroup_vote : enable +#ifdef DATA_A_BF16 +#extension GL_EXT_bfloat16 : enable +#endif #include "types.comp" @@ -80,6 +83,12 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; #define store_scales(a) #endif +#if defined(DATA_A_BF16) +#define MAT_TYPE bfloat16_t +#else +#define MAT_TYPE FLOAT_TYPE +#endif + #ifdef MUL_MAT_ID layout (binding = 3) readonly buffer IDS {int data_ids[];}; @@ -271,8 +280,8 @@ void main() { // Manually partial unroll [[unroll]] for (uint j = 0; j < unroll_count; ++j) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose); @@ -286,8 +295,8 @@ void main() { store_scales(tid); } while (block_k < end_k) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose); @@ -310,8 +319,8 @@ void main() { // Manually partial unroll [[unroll]] for (uint j = 0; j < unroll_count; ++j) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose); @@ -325,8 +334,8 @@ void main() { store_scales(tid); } while (block_k < end_k) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose); @@ -350,8 +359,8 @@ void main() { // Manually partial unroll [[unroll]] for (uint j = 0; j < unroll_count; ++j) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); @@ -365,8 +374,8 @@ void main() { store_scales(tid); } while (block_k < end_k) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); @@ -405,8 +414,8 @@ void main() { fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false); } - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA); #ifdef MUL_MAT_ID diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp b/ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp new file mode 100644 index 00000000000..fd0ba401feb --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp @@ -0,0 +1,7 @@ +#version 460 + +#extension GL_EXT_bfloat16 : require + +void main() +{ +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp index f5b29bfb13a..3bde717832b 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp @@ -33,6 +33,19 @@ #endif #endif +#if defined(DATA_A_BF16) +#define QUANT_K 1 +#define QUANT_R 1 + +#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1 +#define A_TYPE uint16_t +#elif LOAD_VEC_A == 4 +#define A_TYPE u16vec4 +#elif LOAD_VEC_A == 8 +#error unsupported +#endif +#endif + #define QUANT_K_Q4_0 32 #define QUANT_R_Q4_0 2 @@ -1343,4 +1356,18 @@ void init_iq_shmem(uvec3 wgsize) } #endif +// returns the bfloat value in the low 16b. +// See ggml_compute_fp32_to_bf16 +uint32_t fp32_to_bf16(float f) +{ + uint32_t u = floatBitsToUint(f); + u = (u + (0x7fff + ((u >> 16) & 1))) >> 16; + return u; +} + +float bf16_to_fp32(uint32_t u) +{ + return uintBitsToFloat(u << 16); +} + #endif // !defined(GGML_TYPES_COMP) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index cf74625cc56..759112b9462 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -63,7 +63,8 @@ const std::vector type_names = { "iq3_xxs", "iq3_s", "iq4_xs", - "iq4_nl" + "iq4_nl", + "bf16", }; namespace { @@ -296,7 +297,6 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4"; std::map base_dict = { - {"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"}, {"FLOAT_TYPE_VEC2", (coopmat2 || fp16) ? "f16vec2" : "vec2"}, }; std::string shader_name = "matmul"; @@ -318,12 +318,45 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp"; + auto const &FLOAT_TYPE = [&](const std::string &t) -> std::string { + if (t == "bf16") { + // scalar path promotes to float + if (!coopmat && !coopmat2) { + return "float"; + } + return "bfloat16_t"; + } + if (coopmat2 || fp16) { + return "float16_t"; + } + return "float"; + }; + // Shaders with f16 B_TYPE - string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + + string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + + // bf16 + { + std::string load_vec_a_unaligned = "1"; + // For aligned matmul loads + std::string load_vec_a = coopmat2 ? "1" : "4"; - string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + // scalar path promotes to float + std::string to_float_type = (coopmat || coopmat2) ? "uintBitsToBFloat16EXT" : "bf16_to_fp32"; + + // If bfloat16 is not supported, then only compile the scalar (promote to fp32) shader +#if !defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (!(coopmat || coopmat2)) +#endif + { + string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_bf16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc); + } + } for (const auto& tname : type_names) { std::string load_vec_quant = "2"; @@ -332,26 +365,30 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq4_nl")) load_vec_quant = "4"; + if (tname == "bf16") { + continue; + } + 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 = (coopmat2 || tname == "f32" || tname == "f16") ? "1" : load_vec_quant; + std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant; // For aligned matmul loads - std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : load_vec_quant; + std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant; // don't generate f32 variants for coopmat2 if (!coopmat2) { - string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, 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"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); } if (tname != "f16" && tname != "f32") { - string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); } #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) { - string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc); } #endif } @@ -393,6 +430,7 @@ void process_shaders() { if (tname == "f32") { continue; } + if (tname == "bf16") continue; if (tname == "f16") { string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp", @@ -417,12 +455,12 @@ void process_shaders() { string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); // Dequant shaders - if (tname != "f16") { + if (tname != "f16" && tname != "bf16") { string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}})); } if (!string_ends_with(tname, "_k")) { - shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp"; + shader = (tname == "f32" || tname == "f16" || tname == "bf16") ? "get_rows.comp" : "get_rows_quant.comp"; if (tname == "f16") { string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}})); @@ -447,9 +485,11 @@ void process_shaders() { string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}); string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}}); string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("contig_cpy_f32_f16", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}); string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}}); for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) { string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); @@ -544,6 +584,9 @@ void process_shaders() { string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}})); + string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}})); + string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}})); + for (auto &c : compiles) { c.wait(); } diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index e5b94b2296b..54547843d88 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -5d22a79efe9cf2738da1ee80a596c970d629f44d +17733de6a7854b9696be7a563711c9aa4a34b2d3