From e696d060a1f56371acb7256c20ca210362264568 Mon Sep 17 00:00:00 2001 From: Akarshan Biswas Date: Sat, 1 Mar 2025 19:40:56 +0530 Subject: [PATCH 1/3] SYCL: refactor and move cpy kernels to a separate file --- ggml/src/ggml-sycl/backend.hpp | 1 + ggml/src/ggml-sycl/common.hpp | 1 + ggml/src/ggml-sycl/cpy.cpp | 395 ++++++++++++++++++++++++++ ggml/src/ggml-sycl/cpy.hpp | 11 + ggml/src/ggml-sycl/ggml-sycl.cpp | 467 +------------------------------ 5 files changed, 409 insertions(+), 466 deletions(-) create mode 100644 ggml/src/ggml-sycl/cpy.cpp create mode 100644 ggml/src/ggml-sycl/cpy.hpp diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp index b1df4e5db1753..577ff51fde5a8 100644 --- a/ggml/src/ggml-sycl/backend.hpp +++ b/ggml/src/ggml-sycl/backend.hpp @@ -29,6 +29,7 @@ #include "wkv6.hpp" #include "outprod.hpp" #include "element_wise.hpp" +#include "cpy.hpp" #include "gla.hpp" #endif // GGML_SYCL_BACKEND_HPP diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp index 7c503a1b10e4a..a92988b7dbd24 100644 --- a/ggml/src/ggml-sycl/common.hpp +++ b/ggml/src/ggml-sycl/common.hpp @@ -34,6 +34,7 @@ #pragma clang diagnostic ignored "-Wnested-anon-types" #include "ggml-common.h" #pragma clang diagnostic pop +#include "ggml-impl.h" void* ggml_sycl_host_malloc(size_t size); void ggml_sycl_host_free(void* ptr); diff --git a/ggml/src/ggml-sycl/cpy.cpp b/ggml/src/ggml-sycl/cpy.cpp new file mode 100644 index 0000000000000..132c1fc3f3c48 --- /dev/null +++ b/ggml/src/ggml-sycl/cpy.cpp @@ -0,0 +1,395 @@ +#include "cpy.hpp" +#include + +static void cpy_1_f32_f32(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + float * dsti = (float *) cdsti; + + *dsti = *xi; +} + +static void cpy_1_f32_f16(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + sycl::half * dsti = (sycl::half *) cdsti; + + *dsti = sycl::vec(*xi).convert()[0]; +} + +static void cpy_1_f16_f16(const char * cxi, char * cdsti) { + const sycl::half * xi = (const sycl::half *) cxi; + sycl::half * dsti = (sycl::half *) cdsti; + + *dsti = *xi; +} + +static void cpy_1_f16_f32(const char * cxi, char * cdsti) { + const sycl::half * xi = (const sycl::half *) cxi; + float * dsti = (float *) cdsti; + + *dsti = *xi; +} + +static void cpy_1_i16_i16(const char * cxi, char * cdsti) { + const int16_t * xi = (const int16_t *) cxi; + int16_t * dsti = (int16_t *) cdsti; + + *dsti = *xi; +} + +static void cpy_1_i32_i32(const char * cxi, char * cdsti) { + const int32_t * xi = (const int32_t *) cxi; + int32_t * dsti = (int32_t *) cdsti; + + *dsti = *xi; +} + +template +static void cpy_f32_f16(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, + const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, + const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, + const sycl::nd_item<3> & item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); + + if (i >= ne) { + return; + } + + // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor + // then combine those indices with the corresponding byte offsets to get the total offsets + const int i03 = i / (ne00 * ne01 * ne02); + const int i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01); + const int i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00; + const int i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00 - i01 * ne00; + const int x_offset = i00 * nb00 + i01 * nb01 + i02 * nb02 + i03 * nb03; + + const int i13 = i / (ne10 * ne11 * ne12); + const int i12 = (i - i13 * ne10 * ne11 * ne12) / (ne10 * ne11); + const int i11 = (i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11) / ne10; + const int i10 = i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11 - i11 * ne10; + const int dst_offset = i10 * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13; + + cpy_1(cx + x_offset, cdst + dst_offset); +} + +static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q8_0 * dsti = (block_q8_0 *) cdsti; + + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + const float v = xi[j]; + amax = sycl::fmax(amax, sycl::fabs((float) v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f / d : 0.0f; + + dsti->d = d; + + for (int j = 0; j < QK8_0; ++j) { + const float x0 = xi[j] * id; + + dsti->qs[j] = sycl::round((float) x0); + } +} + +static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q4_0 * dsti = (block_q4_0 *) cdsti; + + float amax = 0.0f; + float vmax = 0.0f; + + for (int j = 0; j < QK4_0; ++j) { + const float v = xi[j]; + if (amax < sycl::fabs((float) v)) { + amax = sycl::fabs((float) v); + vmax = v; + } + } + + const float d = vmax / -8; + const float id = d ? 1.0f / d : 0.0f; + + dsti->d = d; + + for (int j = 0; j < QK4_0 / 2; ++j) { + const float x0 = xi[0 + j] * id; + const float x1 = xi[QK4_0 / 2 + j] * id; + + const uint8_t xi0 = dpct::min(15, (int8_t) (x0 + 8.5f)); + const uint8_t xi1 = dpct::min(15, (int8_t) (x1 + 8.5f)); + + dsti->qs[j] = xi0; + dsti->qs[j] |= xi1 << 4; + } +} + +static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q4_1 * dsti = (block_q4_1 *) cdsti; + + float vmin = FLT_MAX; + float vmax = -FLT_MAX; + + for (int j = 0; j < QK4_1; ++j) { + const float v = xi[j]; + + if (v < vmin) { + vmin = v; + } + if (v > vmax) { + vmax = v; + } + } + + const float d = (vmax - vmin) / ((1 << 4) - 1); + const float id = d ? 1.0f / d : 0.0f; + + dsti->dm.x() = d; + dsti->dm.y() = vmin; + + for (int j = 0; j < QK4_1 / 2; ++j) { + const float x0 = (xi[0 + j] - vmin) * id; + const float x1 = (xi[QK4_1 / 2 + j] - vmin) * id; + + const uint8_t xi0 = dpct::min(15, (int8_t) (x0 + 0.5f)); + const uint8_t xi1 = dpct::min(15, (int8_t) (x1 + 0.5f)); + + dsti->qs[j] = xi0; + dsti->qs[j] |= xi1 << 4; + } +} + +template +static void cpy_f32_q(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, + const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, + const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, + const sycl::nd_item<3> & item_ct1) { + const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2)) * qk; + + if (i >= ne) { + return; + } + + const int i03 = i / (ne00 * ne01 * ne02); + const int i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01); + const int i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00; + const int i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00 - i01 * ne00; + const int x_offset = i00 * nb00 + i01 * nb01 + i02 * nb02 + i03 * nb03; + + const int i13 = i / (ne10 * ne11 * ne12); + const int i12 = (i - i13 * ne10 * ne11 * ne12) / (ne10 * ne11); + const int i11 = (i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11) / ne10; + const int i10 = i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11 - i11 * ne10; + const int dst_offset = (i10 / qk) * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13; + + cpy_blck(cx + x_offset, cdst + dst_offset); +} + +static void ggml_cpy_f16_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +static void ggml_cpy_f32_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +static void ggml_cpy_f32_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +static void ggml_cpy_f32_q8_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK8_0 == 0); + const int num_blocks = ne / QK8_0; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK4_0 == 0); + const int num_blocks = ne / QK4_0; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_f32_q4_1_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK4_1 == 0); + const int num_blocks = ne / QK4_1; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +static void ggml_cpy_i16_i16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + // dpct::has_capability_or_fail(stream->get_device(), + // {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + // dpct::has_capability_or_fail(stream->get_device(), + // {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try { + const int64_t ne = ggml_nelements(src0); + GGML_ASSERT(ne == ggml_nelements(src1)); + + GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX); + GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX); + + GGML_TENSOR_BINARY_OP_LOCALS01; + + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + queue_ptr main_stream = ctx.stream(); + + char * src0_ddc = (char *) src0->data; + char * src1_ddc = (char *) src1->data; + + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { + ggml_cpy_f32_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { + ggml_cpy_f32_f16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) { + ggml_cpy_f32_q8_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) { + ggml_cpy_f32_q4_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) { + ggml_cpy_f32_q4_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) { + ggml_cpy_f16_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { + ggml_cpy_f16_f16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16) { + ggml_cpy_i16_i16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) { + ggml_cpy_i32_i32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else { + GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type), + ggml_type_name(src1->type)); + GGML_ABORT("fatal error"); + } +} catch (const sycl::exception & exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { + // TODO: why do we pass dst as src1 here? + ggml_sycl_cpy(ctx, dst->src[0], dst); +} \ No newline at end of file diff --git a/ggml/src/ggml-sycl/cpy.hpp b/ggml/src/ggml-sycl/cpy.hpp new file mode 100644 index 0000000000000..93fda420098f8 --- /dev/null +++ b/ggml/src/ggml-sycl/cpy.hpp @@ -0,0 +1,11 @@ +#ifndef GGML_SYCL_CPY_HPP +#define GGML_SYCL_CPY_HPP + +#include "common.hpp" + +typedef void (*cpy_kernel_t)(const char * cx, char * cdst); + +void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1); +void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst); + +#endif // GGML_SYCL_CPY_HPP \ No newline at end of file diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index d804e66061721..ece7888c57dda 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -1285,8 +1285,6 @@ std::unique_ptr ggml_backend_sycl_context::new_pool_for_device(q // struct ggml_sycl_pool_vmm : public ggml_sycl_pool /// kernels - -typedef void (*cpy_kernel_t)(const char * cx, char * cdst); typedef void (*ggml_sycl_op_mul_mat_t)( ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, @@ -1468,193 +1466,6 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous } } -static void cpy_1_f32_f32(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - float * dsti = (float *) cdsti; - - *dsti = *xi; -} - -static void cpy_1_f32_f16(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - sycl::half *dsti = (sycl::half *)cdsti; - - *dsti = sycl::vec(*xi) - .convert()[0]; -} - -static void cpy_1_f16_f16(const char * cxi, char * cdsti) { - const sycl::half *xi = (const sycl::half *)cxi; - sycl::half *dsti = (sycl::half *)cdsti; - - *dsti = *xi; -} - -static void cpy_1_f16_f32(const char * cxi, char * cdsti) { - const sycl::half *xi = (const sycl::half *)cxi; - float * dsti = (float *) cdsti; - - *dsti = *xi; -} - -static void cpy_1_i16_i16(const char * cxi, char * cdsti) { - const int16_t *xi = (const int16_t *)cxi; - int16_t *dsti = (int16_t *)cdsti; - - *dsti = *xi; -} - -static void cpy_1_i32_i32(const char * cxi, char * cdsti) { - const int32_t *xi = (const int32_t *)cxi; - int32_t *dsti = (int32_t *)cdsti; - - *dsti = *xi; -} - -template -static void cpy_f32_f16(const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13, const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - - if (i >= ne) { - return; - } - - // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor - // then combine those indices with the corresponding byte offsets to get the total offsets - const int i03 = i/(ne00 * ne01 * ne02); - const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); - const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; - const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; - const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - - const int i13 = i/(ne10 * ne11 * ne12); - const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); - const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; - const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; - const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13; - - cpy_1(cx + x_offset, cdst + dst_offset); -} - -static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q8_0 * dsti = (block_q8_0 *) cdsti; - - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - const float v = xi[j]; - amax = sycl::fmax(amax, sycl::fabs((float)v)); - } - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - dsti->d = d; - - for (int j = 0; j < QK8_0; ++j) { - const float x0 = xi[j]*id; - - dsti->qs[j] = sycl::round((float)x0); - } -} - -static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q4_0 * dsti = (block_q4_0 *) cdsti; - - float amax = 0.0f; - float vmax = 0.0f; - - for (int j = 0; j < QK4_0; ++j) { - const float v = xi[j]; - if (amax < sycl::fabs((float)v)) { - amax = sycl::fabs((float)v); - vmax = v; - } - } - - const float d = vmax / -8; - const float id = d ? 1.0f/d : 0.0f; - - dsti->d = d; - - for (int j = 0; j < QK4_0/2; ++j) { - const float x0 = xi[0 + j]*id; - const float x1 = xi[QK4_0/2 + j]*id; - - const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 8.5f)); - const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 8.5f)); - - dsti->qs[j] = xi0; - dsti->qs[j] |= xi1 << 4; - } -} - -static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q4_1 * dsti = (block_q4_1 *) cdsti; - - float vmin = FLT_MAX; - float vmax = -FLT_MAX; - - for (int j = 0; j < QK4_1; ++j) { - const float v = xi[j]; - - if (v < vmin) vmin = v; - if (v > vmax) vmax = v; - } - - const float d = (vmax - vmin) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - dsti->dm.x() = d; - dsti->dm.y() = vmin; - - for (int j = 0; j < QK4_1/2; ++j) { - const float x0 = (xi[0 + j] - vmin)*id; - const float x1 = (xi[QK4_1/2 + j] - vmin)*id; - - const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 0.5f)); - const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 0.5f)); - - dsti->qs[j] = xi0; - dsti->qs[j] |= xi1 << 4; - } -} - -template -static void cpy_f32_q(const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13, const sycl::nd_item<3> &item_ct1) { - const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2)) * - qk; - - if (i >= ne) { - return; - } - - const int i03 = i/(ne00 * ne01 * ne02); - const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); - const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; - const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; - const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - - const int i13 = i/(ne10 * ne11 * ne12); - const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); - const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; - const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; - const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13; - - cpy_blck(cx + x_offset, cdst + dst_offset); -} - static void k_sum_rows_f32(const float * x, float * dst, const int ncols, const sycl::nd_item<3> &item_ct1) { const int row = item_ct1.get_group(1); @@ -1903,231 +1714,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl( } } -static void -ggml_cpy_f16_f32_sycl(const char *cx, char *cdst, const int ne, const int ne00, - const int ne01, const int ne02, const int nb00, - const int nb01, const int nb02, const int nb03, - const int ne10, const int ne11, const int ne12, - const int nb10, const int nb11, const int nb12, - const int nb13, queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, - nb01, nb02, nb03, ne10, ne11, ne12, - nb10, nb11, nb12, nb13, item_ct1); - }); - } -} - -static void ggml_cpy_f32_f32_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); - } -} - -static void ggml_cpy_f32_f16_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); - } -} - -static void ggml_cpy_f32_q8_0_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - GGML_ASSERT(ne % QK8_0 == 0); - const int num_blocks = ne / QK8_0; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), - sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q( - cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); -} - -static void ggml_cpy_f32_q4_0_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - GGML_ASSERT(ne % QK4_0 == 0); - const int num_blocks = ne / QK4_0; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), - sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q( - cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); -} - -static void ggml_cpy_f32_q4_1_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - GGML_ASSERT(ne % QK4_1 == 0); - const int num_blocks = ne / QK4_1; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), - sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q( - cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); -} - -static void ggml_cpy_f16_f16_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); - } -} - -static void ggml_cpy_i16_i16_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - // dpct::has_capability_or_fail(stream->get_device(), - // {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); - } -} - -static void ggml_cpy_i32_i32_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - // dpct::has_capability_or_fail(stream->get_device(), - // {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); - } -} static void scale_f32_sycl(const float *x, float *dst, const float scale, const int k, queue_ptr stream) { @@ -3645,58 +3232,6 @@ static void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_clamp); } -static void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst) try { - const int64_t ne = ggml_nelements(src0); - GGML_ASSERT(ne == ggml_nelements(src1)); - - GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX); - GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX); - - GGML_TENSOR_BINARY_OP_LOCALS01; - - SYCL_CHECK(ggml_sycl_set_device(ctx.device)); - queue_ptr main_stream = ctx.stream(); - - char * src0_ddc = (char *) src0->data; - char * src1_ddc = (char *) src1->data; - - if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { - ggml_cpy_f32_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { - ggml_cpy_f32_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) { - ggml_cpy_f32_q8_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) { - ggml_cpy_f32_q4_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) { - ggml_cpy_f32_q4_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) { - ggml_cpy_f16_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { - ggml_cpy_f16_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16) { - ggml_cpy_i16_i16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) { - ggml_cpy_i32_i32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else { - GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__, - ggml_type_name(src0->type), ggml_type_name(src1->type)); - GGML_ABORT("fatal error"); - } - GGML_UNUSED(dst); -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} - -static void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - // TODO: why do we pass dst as src1 here? - ggml_sycl_cpy(ctx, dst->src[0], dst, nullptr); -} - static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_diag_mask_inf); } @@ -3893,7 +3428,7 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens ggml_sycl_clamp(ctx, dst); break; case GGML_OP_CPY: - ggml_sycl_cpy(ctx, dst->src[0], dst->src[1], dst); + ggml_sycl_cpy(ctx, dst->src[0], dst->src[1]); break; case GGML_OP_CONT: ggml_sycl_dup(ctx, dst); From 4da9281b04c0c5a0d047b46cec69d0531e9908e7 Mon Sep 17 00:00:00 2001 From: Akarshan Biswas Date: Sat, 1 Mar 2025 21:08:37 +0530 Subject: [PATCH 2/3] Add few missing cpy kernels --- ggml/src/ggml-sycl/cpy.cpp | 304 ++++++++++++++++++++++++++++++- ggml/src/ggml-sycl/cpy.hpp | 2 +- ggml/src/ggml-sycl/ggml-sycl.cpp | 24 +++ 3 files changed, 328 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-sycl/cpy.cpp b/ggml/src/ggml-sycl/cpy.cpp index 132c1fc3f3c48..7b89da25cf553 100644 --- a/ggml/src/ggml-sycl/cpy.cpp +++ b/ggml/src/ggml-sycl/cpy.cpp @@ -1,6 +1,28 @@ #include "cpy.hpp" + #include +#include "dequantize.hpp" + +static __dpct_inline__ int best_index_int8(int n, const int8_t * val, float x) { + if (x <= val[0]) { + return 0; + } + if (x >= val[n - 1]) { + return n - 1; + } + int ml = 0, mu = n - 1; + while (mu - ml > 1) { + int mav = (ml + mu) / 2; + if (x < val[mav]) { + mu = mav; + } else { + ml = mav; + } + } + return x - val[mu - 1] < val[mu] - x ? mu - 1 : mu; +} + static void cpy_1_f32_f32(const char * cxi, char * cdsti) { const float * xi = (const float *) cxi; float * dsti = (float *) cdsti; @@ -94,6 +116,17 @@ static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { } } +static void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) { + float * cdstf = (float *) (cdsti); + + for (int j = 0; j < QK8_0; j += 2) { + dfloat2 dq; + dequantize_q8_0(cxi, 0, j, dq); + *(cdstf + j) = dq.x(); + *(cdstf + j + 1) = dq.y(); + } +} + static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) { const float * xi = (const float *) cxi; block_q4_0 * dsti = (block_q4_0 *) cdsti; @@ -162,6 +195,122 @@ static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) { } } +static void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q5_0 * dsti = (block_q5_0 *) cdsti; + + float amax = 0.0f; + float vmax = 0.0f; + + for (int j = 0; j < QK5_0; ++j) { + const float v = xi[j]; + if (amax < sycl::fabs((float) v)) { + amax = sycl::fabs((float) v); + vmax = v; + } + } + + const float d = vmax / -16; + const float id = d ? 1.0f / d : 0.0f; + + dsti->d = d; + + uint32_t qh = 0; + for (int j = 0; j < QK5_0 / 2; ++j) { + const float x0 = xi[0 + j] * id; + const float x1 = xi[QK5_0 / 2 + j] * id; + + const uint8_t xi0 = dpct::min(31, (int8_t) (x0 + 16.5f)); + const uint8_t xi1 = dpct::min(31, (int8_t) (x1 + 16.5f)); + + dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4); + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0 / 2); + } + memcpy(dsti->qh, &qh, sizeof(qh)); +} + +static void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q5_1 * dsti = (block_q5_1 *) cdsti; + + float min = xi[0]; + float max = xi[0]; + + for (int j = 1; j < QK5_1; ++j) { + const float v = xi[j]; + min = v < min ? v : min; + max = v > max ? v : max; + } + + const float d = (max - min) / 31; + const float id = d ? 1.0f / d : 0.0f; + + dsti->dm.x() = d; + dsti->dm.y() = min; + + uint32_t qh = 0; + for (int j = 0; j < QK5_1 / 2; ++j) { + const float x0 = (xi[0 + j] - min) * id; + const float x1 = (xi[QK5_1 / 2 + j] - min) * id; + + const uint8_t xi0 = (uint8_t) (x0 + 0.5f); + const uint8_t xi1 = (uint8_t) (x1 + 0.5f); + + dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4); + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1 / 2); + } + memcpy(dsti->qh, &qh, sizeof(qh)); +} + +static void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_iq4_nl * dsti = (block_iq4_nl *) cdsti; + + float amax = 0.0f; + float vmax = 0.0f; + + for (int j = 0; j < QK4_NL; ++j) { + const float v = xi[j]; + if (amax < sycl::fabs((float) v)) { + amax = sycl::fabs((float) v); + vmax = v; + } + } + + float d = vmax / kvalues_iq4nl[0]; + const float id = d ? 1.0f / d : 0.0f; + + float sumqx = 0, sumq2 = 0; + for (int j = 0; j < QK4_NL / 2; ++j) { + const float x0 = xi[0 + j] * id; + const float x1 = xi[QK4_NL / 2 + j] * id; + const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0); + const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1); + dsti->qs[j] = xi0 | (xi1 << 4); + const float v0 = kvalues_iq4nl[xi0]; + const float v1 = kvalues_iq4nl[xi1]; + const float w0 = xi[0 + j] * xi[0 + j]; + const float w1 = xi[QK4_NL / 2 + j] * xi[QK4_NL / 2 + j]; + sumqx += w0 * v0 * xi[j] + w1 * v1 * xi[QK4_NL / 2 + j]; + sumq2 += w0 * v0 * v0 + w1 * v1 * v1; + } + + dsti->d = sumq2 > 0 ? sumqx / sumq2 : d; +} + +template static void cpy_blck_q_f32(const char * cxi, char * cdsti) { + float * cdstf = (float *) (cdsti); + + for (int j = 0; j < qk / 2; j++) { + dfloat2 dq; + dequant(cxi, 0, j, dq); + *(cdstf + j) = dq.x(); + *(cdstf + j + qk / 2) = dq.y(); + } +} + template static void cpy_f32_q(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, @@ -188,6 +337,32 @@ static void cpy_f32_q(const char * cx, char * cdst, const int ne, const int ne00 cpy_blck(cx + x_offset, cdst + dst_offset); } +template +static void cpy_q_f32(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, + const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, + const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, + const sycl::nd_item<3> & item_ct1) { + const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2)) * qk; + + if (i >= ne) { + return; + } + + const int i03 = i / (ne00 * ne01 * ne02); + const int i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01); + const int i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00; + const int i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00 - i01 * ne00; + const int x_offset = (i00 / qk) * nb00 + i01 * nb01 + i02 * nb02 + i03 * nb03; + + const int i13 = i / (ne10 * ne11 * ne12); + const int i12 = (i - i13 * ne10 * ne11 * ne12) / (ne10 * ne11); + const int i11 = (i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11) / ne10; + const int i10 = i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11 - i11 * ne10; + const int dst_offset = i10 * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13; + + cpy_blck(cx + x_offset, cdst + dst_offset); +} + static void ggml_cpy_f16_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, @@ -255,6 +430,18 @@ static void ggml_cpy_f32_q8_0_sycl(const char * cx, char * cdst, const int ne, c }); } +static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = ne; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, @@ -268,6 +455,19 @@ static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, c }); } +static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = ne; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, + nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, + item_ct1); + }); +} + static void ggml_cpy_f32_q4_1_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, @@ -281,6 +481,84 @@ static void ggml_cpy_f32_q4_1_sycl(const char * cx, char * cdst, const int ne, c }); } +static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = ne; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, + nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, + item_ct1); + }); +} + +static void ggml_cpy_f32_q5_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK5_0 == 0); + const int num_blocks = ne / QK5_0; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = ne; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, + nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, + item_ct1); + }); +} + +static void ggml_cpy_f32_q5_1_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK5_1 == 0); + const int num_blocks = ne / QK5_1; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = ne; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, + nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, + item_ct1); + }); +} + +static void ggml_cpy_f32_iq4_nl_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK4_NL == 0); + const int num_blocks = ne / QK4_NL; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, @@ -379,6 +657,30 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) { ggml_cpy_i32_i32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q4_0_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q4_1_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q8_0_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) { + ggml_cpy_f32_q5_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q5_0_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) { + ggml_cpy_f32_q5_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q5_1_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) { + ggml_cpy_f32_iq4_nl_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, main_stream); } else { GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type), ggml_type_name(src1->type)); @@ -392,4 +694,4 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { // TODO: why do we pass dst as src1 here? ggml_sycl_cpy(ctx, dst->src[0], dst); -} \ No newline at end of file +} diff --git a/ggml/src/ggml-sycl/cpy.hpp b/ggml/src/ggml-sycl/cpy.hpp index 93fda420098f8..0a0f561d2309a 100644 --- a/ggml/src/ggml-sycl/cpy.hpp +++ b/ggml/src/ggml-sycl/cpy.hpp @@ -8,4 +8,4 @@ typedef void (*cpy_kernel_t)(const char * cx, char * cdst); void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1); void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst); -#endif // GGML_SYCL_CPY_HPP \ No newline at end of file +#endif // GGML_SYCL_CPY_HPP diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index ece7888c57dda..e7304947fbd59 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -3942,6 +3942,30 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { return true; } + if (src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_Q4_0 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_Q4_1 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_0) { + return true; + } + if (src0_type == GGML_TYPE_Q5_0 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_1) { + return true; + } + if (src0_type == GGML_TYPE_Q5_1 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) { + return true; + } return false; } break; case GGML_OP_CONCAT: From 38b88ba9cc5d0a50c734dce1eb3e2444078690af Mon Sep 17 00:00:00 2001 From: Akarshan Biswas Date: Sun, 2 Mar 2025 07:43:26 +0530 Subject: [PATCH 3/3] refactor and add debug logs --- ggml/src/ggml-sycl/cpy.cpp | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-sycl/cpy.cpp b/ggml/src/ggml-sycl/cpy.cpp index 7b89da25cf553..5a23145895f26 100644 --- a/ggml/src/ggml-sycl/cpy.cpp +++ b/ggml/src/ggml-sycl/cpy.cpp @@ -629,6 +629,8 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co char * src0_ddc = (char *) src0->data; char * src1_ddc = (char *) src1->data; + GGML_SYCL_DEBUG("[SYCL] %s: Tensor supplied: %s to %s\n", __func__, ggml_type_name(src0->type), + ggml_type_name(src1->type)); if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { ggml_cpy_f32_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, @@ -672,13 +674,13 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) { ggml_cpy_q5_0_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) { + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) { ggml_cpy_f32_q5_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) { ggml_cpy_q5_1_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) { + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) { ggml_cpy_f32_iq4_nl_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); } else { @@ -693,5 +695,7 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { // TODO: why do we pass dst as src1 here? + GGML_SYCL_DEBUG("[SYCL] call %s\n", __func__); ggml_sycl_cpy(ctx, dst->src[0], dst); + GGML_SYCL_DEBUG("[SYCL] call %s done\n", __func__); }