whisper.rn 0.4.0-rc.9 → 0.4.0

package/cpp/ggml-cpu/amx/amx.cpp

@@ -0,0 +1,221 @@
+#include "amx.h"
+#include "common.h"
+#include "mmq.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "traits.h"
+
+#if defined(__gnu_linux__)
+#include <sys/syscall.h>
+#include <unistd.h>
+#endif
+
+#include <cstdlib>
+#include <cstring>
+#include <memory>
+
+#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
+
+// AMX type_trais
+namespace ggml::cpu::amx {
+class tensor_traits : public ggml::cpu::tensor_traits {
+    bool work_size(int /* n_threads */, const struct wsp_ggml_tensor * op, size_t & size) override {
+        size = wsp_ggml_backend_amx_desired_wsize(op);
+        return true;
+    }
+
+    bool compute_forward(struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * op) override {
+        if (op->op == WSP_GGML_OP_MUL_MAT) {
+            wsp_ggml_backend_amx_mul_mat(params, op);
+            return true;
+        }
+        return false;
+    }
+};
+
+static ggml::cpu::tensor_traits * get_tensor_traits(wsp_ggml_backend_buffer_t, struct wsp_ggml_tensor *) {
+    static tensor_traits traits;
+    return &traits;
+}
+}  // namespace ggml::cpu::amx
+
+// AMX buffer interface
+static void wsp_ggml_backend_amx_buffer_free_buffer(wsp_ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+}
+
+static void * wsp_ggml_backend_amx_buffer_get_base(wsp_ggml_backend_buffer_t buffer) {
+    return (void *) (buffer->context);
+}
+
+static enum wsp_ggml_status wsp_ggml_backend_amx_buffer_init_tensor(wsp_ggml_backend_buffer_t buffer, struct wsp_ggml_tensor * tensor) {
+    tensor->extra = (void *) ggml::cpu::amx::get_tensor_traits(buffer, tensor);
+
+    WSP_GGML_UNUSED(buffer);
+    return WSP_GGML_STATUS_SUCCESS;
+}
+
+static void wsp_ggml_backend_amx_buffer_memset_tensor(wsp_ggml_backend_buffer_t buffer, struct wsp_ggml_tensor * tensor,
+                                                  uint8_t value, size_t offset, size_t size) {
+    memset((char *) tensor->data + offset, value, size);
+
+    WSP_GGML_UNUSED(buffer);
+}
+
+static void wsp_ggml_backend_amx_buffer_set_tensor(wsp_ggml_backend_buffer_t buffer, struct wsp_ggml_tensor * tensor,
+                                               const void * data, size_t offset, size_t size) {
+    if (qtype_has_amx_kernels(tensor->type)) {
+        WSP_GGML_LOG_DEBUG("%s: amx repack tensor %s of type %s\n", __func__, tensor->name, wsp_ggml_type_name(tensor->type));
+        wsp_ggml_backend_amx_convert_weight(tensor, data, offset, size);
+    } else {
+        memcpy((char *) tensor->data + offset, data, size);
+    }
+
+    WSP_GGML_UNUSED(buffer);
+}
+
+/*
+// need to figure what we need to do with buffer->extra.
+static void wsp_ggml_backend_amx_buffer_get_tensor(wsp_ggml_backend_buffer_t buffer, const struct wsp_ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    WSP_GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    WSP_GGML_UNUSED(buffer);
+}
+
+static bool wsp_ggml_backend_amx_buffer_cpy_tensor(wsp_ggml_backend_buffer_t buffer, const struct wsp_ggml_tensor * src, struct wsp_ggml_tensor * dst) {
+    if (wsp_ggml_backend_buffer_is_host(src->buffer)) {
+        if (qtype_has_amx_kernels(src->type)) {
+            wsp_ggml_backend_amx_convert_weight(dst, src->data, 0, wsp_ggml_nbytes(dst));
+        } else {
+            memcpy(dst->data, src->data, wsp_ggml_nbytes(src));
+        }
+        return true;
+    }
+    return false;
+
+    WSP_GGML_UNUSED(buffer);
+}
+*/
+
+static void wsp_ggml_backend_amx_buffer_clear(wsp_ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static wsp_ggml_backend_buffer_i wsp_ggml_backend_amx_buffer_interface = {
+    /* .free_buffer     = */ wsp_ggml_backend_amx_buffer_free_buffer,
+    /* .get_base        = */ wsp_ggml_backend_amx_buffer_get_base,
+    /* .init_tensor     = */ wsp_ggml_backend_amx_buffer_init_tensor,
+    /* .memset_tensor   = */ wsp_ggml_backend_amx_buffer_memset_tensor,
+    /* .set_tensor      = */ wsp_ggml_backend_amx_buffer_set_tensor,
+    /* .get_tensor      = */ nullptr,
+    /* .cpy_tensor      = */ nullptr,
+    /* .clear           = */ wsp_ggml_backend_amx_buffer_clear,
+    /* .reset           = */ nullptr,
+};
+
+static const char * wsp_ggml_backend_amx_buffer_type_get_name(wsp_ggml_backend_buffer_type_t buft) {
+    return "AMX";
+
+    WSP_GGML_UNUSED(buft);
+}
+
+static wsp_ggml_backend_buffer_t wsp_ggml_backend_amx_buffer_type_alloc_buffer(wsp_ggml_backend_buffer_type_t buft, size_t size) {
+    void * data = wsp_ggml_aligned_malloc(size);
+    if (data == NULL) {
+        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return wsp_ggml_backend_buffer_init(buft, wsp_ggml_backend_amx_buffer_interface, data, size);
+}
+
+static size_t wsp_ggml_backend_amx_buffer_type_get_alignment(wsp_ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    WSP_GGML_UNUSED(buft);
+}
+
+namespace ggml::cpu::amx {
+class extra_buffer_type : ggml::cpu::extra_buffer_type {
+    bool supports_op(wsp_ggml_backend_dev_t, const struct wsp_ggml_tensor * op) override {
+        // handle only 2d gemm for now
+        auto is_contiguous_2d = [](const struct wsp_ggml_tensor * t) {
+            return wsp_ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
+        };
+
+        if (op->op == WSP_GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) &&  // src0 must be contiguous
+            is_contiguous_2d(op->src[1]) &&                               // src1 must be contiguous
+            op->src[0]->buffer && op->src[0]->buffer->buft == wsp_ggml_backend_amx_buffer_type() &&
+            op->ne[0] % (TILE_N * 2) == 0 &&                              // out_features is 32x
+            (qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == WSP_GGML_TYPE_F16))) {
+            // src1 must be host buffer
+            if (op->src[1]->buffer && !wsp_ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
+                return false;
+            }
+            // src1 must be float32
+            if (op->src[1]->type == WSP_GGML_TYPE_F32) {
+                return true;
+            }
+        }
+        return false;
+    }
+
+    ggml::cpu::tensor_traits * get_tensor_traits(const struct wsp_ggml_tensor * op) override {
+        if (op->op == WSP_GGML_OP_MUL_MAT && op->src[0]->buffer &&
+            op->src[0]->buffer->buft == wsp_ggml_backend_amx_buffer_type()) {
+            return (ggml::cpu::tensor_traits *) op->src[0]->extra;
+        }
+
+        return nullptr;
+    }
+};
+}  // namespace ggml::cpu::amx
+
+static size_t wsp_ggml_backend_amx_buffer_type_get_alloc_size(wsp_ggml_backend_buffer_type_t buft, const wsp_ggml_tensor * tensor) {
+    return wsp_ggml_backend_amx_get_alloc_size(tensor);
+
+    WSP_GGML_UNUSED(buft);
+}
+
+#define ARCH_GET_XCOMP_PERM     0x1022
+#define ARCH_REQ_XCOMP_PERM     0x1023
+#define XFEATURE_XTILECFG       17
+#define XFEATURE_XTILEDATA      18
+
+static bool wsp_ggml_amx_init() {
+#if defined(__gnu_linux__)
+    if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
+        fprintf(stderr, "AMX is not ready to be used!\n");
+        return false;
+    }
+    return true;
+#elif defined(_WIN32)
+    return true;
+#endif
+}
+
+wsp_ggml_backend_buffer_type_t wsp_ggml_backend_amx_buffer_type() {
+    static struct wsp_ggml_backend_buffer_type wsp_ggml_backend_buffer_type_amx = {
+        /* .iface = */ {
+                        /* .get_name         = */ wsp_ggml_backend_amx_buffer_type_get_name,
+                        /* .alloc_buffer     = */ wsp_ggml_backend_amx_buffer_type_alloc_buffer,
+                        /* .get_alignment    = */ wsp_ggml_backend_amx_buffer_type_get_alignment,
+                        /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
+                        /* .get_alloc_size   = */ wsp_ggml_backend_amx_buffer_type_get_alloc_size,
+                        /* .is_host          = */ nullptr,
+                        },
+        /* .device  = */ wsp_ggml_backend_reg_dev_get(wsp_ggml_backend_cpu_reg(), 0),
+        /* .context = */ new ggml::cpu::amx::extra_buffer_type(),
+    };
+
+    if (!wsp_ggml_amx_init()) {
+        return nullptr;
+    }
+
+    return &wsp_ggml_backend_buffer_type_amx;
+}
+
+#endif  // defined(__AMX_INT8__) && defined(__AVX512VNNI__)

package/cpp/ggml-cpu/amx/amx.h

@@ -0,0 +1,8 @@
+#include "ggml-backend.h"
+#include "ggml-cpu-impl.h"
+
+// GGML internal header
+
+#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
+wsp_ggml_backend_buffer_type_t wsp_ggml_backend_amx_buffer_type(void);
+#endif

package/cpp/ggml-cpu/amx/common.h

@@ -0,0 +1,91 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-cpu-impl.h"
+
+#include <algorithm>
+#include <memory>
+#include <type_traits>
+
+#if defined(WSP_GGML_USE_OPENMP)
+#include <omp.h>
+#endif
+
+#define TILE_M 16
+#define TILE_N 16
+#define TILE_K 32
+#define VNNI_BLK 4
+
+#define AMX_BLK_SIZE 32
+
+#define TMM0 0
+#define TMM1 1
+#define TMM2 2
+#define TMM3 3
+#define TMM4 4
+#define TMM5 5
+#define TMM6 6
+#define TMM7 7
+
+// parallel routines
+template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
+inline T div_up(T x, T y) { return (x + y - 1) / y; }
+
+template <typename T>
+inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
+#if 0
+    // onednn partition pattern
+    T& n_my = n_end;
+    if (nth <= 1 || n == 0) {
+        n_start = 0;
+        n_my = n;
+    } else {
+        T n1 = div_up(n, nth);
+        T n2 = n1 - 1;
+        T T1 = n - n2 * nth;
+        n_my = ith < T1 ? n1 : n2;
+        n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2;
+    }
+    n_end += n_start;
+#else
+    // pytorch aten partition pattern
+    T n_my = div_up(n, nth);
+    n_start = ith * n_my;
+    n_end = std::min(n_start + n_my, n);
+#endif
+}
+
+template <typename func_t>
+inline void parallel_for(int n, const func_t& f) {
+#if defined(WSP_GGML_USE_OPENMP)
+#pragma omp parallel
+{
+    int nth = omp_get_num_threads();
+    int ith = omp_get_thread_num();
+    int tbegin, tend;
+    balance211(n, nth, ith, tbegin, tend);
+    f(tbegin, tend);
+}
+#else
+    f(0, n);
+#endif
+}
+
+template <typename func_t>
+inline void parallel_for_ggml(const wsp_ggml_compute_params * params, int n, const func_t & f) {
+    int tbegin, tend;
+    balance211(n, params->nth, params->ith, tbegin, tend);
+    f(tbegin, tend);
+}
+
+// quantized types that have AMX support
+inline bool qtype_has_amx_kernels(const enum wsp_ggml_type type) {
+    // TODO: fix padding for vnni format
+    return (type == WSP_GGML_TYPE_Q4_0) ||
+        (type == WSP_GGML_TYPE_Q4_1) ||
+        (type == WSP_GGML_TYPE_Q8_0) ||
+        (type == WSP_GGML_TYPE_Q4_K) ||
+        (type == WSP_GGML_TYPE_Q5_K) ||
+        (type == WSP_GGML_TYPE_Q6_K) ||
+        (type == WSP_GGML_TYPE_IQ4_XS);
+}