@agency-lang/whisper-local 0.0.1

package/vendor/whisper.cpp/ggml/src/ggml-cuda/unary.cu

@@ -0,0 +1,289 @@
+#include "unary.cuh"
+
+static __device__ __forceinline__ float op_abs(float x) {
+    return fabsf(x);
+}
+
+static __device__ __forceinline__ float op_sgn(float x) {
+    return (x > 0.f ? 1.f : ((x < 0.f ? -1.f : 0.f)));
+}
+
+static __device__ __forceinline__ float op_neg(float x) {
+    return -x;
+}
+
+static __device__ __forceinline__ float op_step(float x) {
+    return x > 0.0f;
+}
+
+static __device__ __forceinline__ float op_gelu(float x) {
+    const float GELU_COEF_A    = 0.044715f;
+    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+
+    return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+static __device__ __forceinline__ float op_gelu_erf(float x) {
+    const float SQRT_2_INV = 0.70710678118654752440084436210484f;
+
+    return 0.5f*x*(1.0f + erff(x*SQRT_2_INV));
+}
+
+static __device__ __forceinline__ float op_gelu_quick(float x) {
+    const float GELU_QUICK_COEF = -1.702f;
+
+    return x * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x)));
+}
+
+static __device__ __forceinline__ float op_silu(float x) {
+    return x / (1.0f + expf(-x));
+}
+
+static __device__ __forceinline__ float op_tanh(float x) {
+    return tanhf(x);
+}
+
+static __device__ __forceinline__ float op_relu(float x) {
+    return fmaxf(x, 0);
+}
+
+static __device__ __forceinline__ float op_sigmoid(float x) {
+    return 1.0f / (1.0f + expf(-x));
+}
+
+static __device__ __forceinline__ float op_hardsigmoid(float x) {
+    return fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f));
+}
+
+static __device__ __forceinline__ float op_hardswish(float x) {
+    return x * fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f));
+}
+
+static __device__ __forceinline__ float op_exp(float x) {
+    return expf(x);
+}
+
+static __device__ __forceinline__ float op_sqr(float x) {
+    return x * x;
+}
+
+static __device__ __forceinline__ float op_sqrt(float x) {
+    return sqrtf(x);
+}
+
+static __device__ __forceinline__ float op_sin(float x) {
+    return sinf(x);
+}
+
+static __device__ __forceinline__ float op_cos(float x) {
+    return cosf(x);
+}
+
+static __device__ __forceinline__ float op_log(float x) {
+    return logf(x);
+}
+
+template <float (*op)(float), typename T>
+static __global__ void unary_op_kernel(const T * x, T * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = (T)op((float)x[i]);
+}
+
+template <float (*op)(float), typename T>
+static void unary_cuda(const T * x, T * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_NEG_BLOCK_SIZE - 1) / CUDA_NEG_BLOCK_SIZE;
+    unary_op_kernel<op><<<num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+template <float (*op)(float)>
+void ggml_cuda_op_unary(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const void * src0_d = src0->data;
+    void * dst_d = dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == dst->type);
+
+    if (src0->type == GGML_TYPE_F16) {
+        unary_cuda<op>((const half *)src0_d, (half *)dst_d, ggml_nelements(src0), stream);
+    } else {
+        unary_cuda<op>((const float *)src0_d, (float *)dst_d, ggml_nelements(src0), stream);
+    }
+}
+
+void ggml_cuda_op_abs(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_abs>(ctx, dst);
+}
+
+void ggml_cuda_op_sgn(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_sgn>(ctx, dst);
+}
+
+void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_neg>(ctx, dst);
+}
+
+void ggml_cuda_op_step(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_step>(ctx, dst);
+}
+
+void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_gelu>(ctx, dst);
+}
+
+void ggml_cuda_op_gelu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_gelu_erf>(ctx, dst);
+}
+
+void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_gelu_quick>(ctx, dst);
+}
+
+void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_silu>(ctx, dst);
+}
+
+void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_tanh>(ctx, dst);
+}
+
+void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_relu>(ctx, dst);
+}
+
+void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_sigmoid>(ctx, dst);
+}
+
+void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_hardsigmoid>(ctx, dst);
+}
+
+void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_hardswish>(ctx, dst);
+}
+
+void ggml_cuda_op_exp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_exp>(ctx, dst);
+}
+
+void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_sqr>(ctx, dst);
+}
+
+void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_sqrt>(ctx, dst);
+}
+
+void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_sin>(ctx, dst);
+}
+
+void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_cos>(ctx, dst);
+}
+
+void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_log>(ctx, dst);
+}
+
+/* silu_back */
+
+static __device__ __forceinline__ float op_silu_back(float grad, float x) {
+    const float s = 1.0f / (1.0f + expf(-x));
+    return grad * s * (1.0f + x * (1.0f - s));
+}
+
+template <class T>
+static __global__ void silu_back_kernel(const T * grad, const T * xf, T * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = (T)op_silu_back((float)grad[i], (float)xf[i]);
+}
+
+template <class T>
+static void silu_back_cuda(const T * grad, const T * x, T * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SILU_BACK_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE;
+    silu_back_kernel<<<num_blocks, CUDA_SILU_BACK_BLOCK_SIZE, 0, stream>>>(grad, x, dst, k);
+}
+
+void ggml_cuda_op_silu_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0]; // input from forward pass
+    const ggml_tensor * src1 = dst->src[1]; // grads of forward pass output
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       * dst_d  = (float       *) dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == dst->type);
+
+    if (src0->type == GGML_TYPE_F16) {
+        silu_back_cuda((const half *)src0_d, (const half *)src1_d, (half *)dst_d, ggml_nelements(src0), stream);
+    } else {
+        silu_back_cuda((const float*)src0_d, (const float*)src1_d, (float *)dst_d, ggml_nelements(src0), stream);
+    }
+}
+
+/* leaky relu */
+
+static __device__ __forceinline__ float op_leaky_relu(float x, const float negative_slope) {
+    return fmaxf(x, 0) + fminf(x, 0.0f) * negative_slope;
+}
+
+template <class T>
+static __global__ void leaky_relu_kernel(const T * x, T * dst, const int k, const float negative_slope) {
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = (T)op_leaky_relu((float)x[i], negative_slope);
+}
+
+template <class T>
+static void leaky_relu_cuda(const T * x, T * dst, const int k, const float negative_slope, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
+    leaky_relu_kernel<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
+}
+
+void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const void * src0_d = src0->data;
+    void * dst_d = dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == dst->type);
+
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+
+    if (src0->type == GGML_TYPE_F16) {
+        leaky_relu_cuda((const half *)src0_d, (half *)dst_d, ggml_nelements(src0), negative_slope, stream);
+    } else {
+        leaky_relu_cuda((const float *)src0_d, (float *)dst_d, ggml_nelements(src0), negative_slope, stream);
+    }
+}

package/vendor/whisper.cpp/ggml/src/ggml-cuda/unary.cuh

@@ -0,0 +1,59 @@
+#include "common.cuh"
+
+#define CUDA_NEG_BLOCK_SIZE 256
+#define CUDA_STEP_BLOCK_SIZE 256
+#define CUDA_GELU_BLOCK_SIZE 256
+#define CUDA_SILU_BLOCK_SIZE 256
+#define CUDA_SILU_BACK_BLOCK_SIZE 256
+#define CUDA_TANH_BLOCK_SIZE 256
+#define CUDA_RELU_BLOCK_SIZE 256
+#define CUDA_SIGMOID_BLOCK_SIZE 256
+#define CUDA_HARDSIGMOID_BLOCK_SIZE 256
+#define CUDA_EXP_BLOCK_SIZE 256
+#define CUDA_HARDSWISH_BLOCK_SIZE 256
+#define CUDA_SQR_BLOCK_SIZE 256
+#define CUDA_SQRT_BLOCK_SIZE 256
+#define CUDA_SIN_BLOCK_SIZE 256
+#define CUDA_COS_BLOCK_SIZE 256
+
+void ggml_cuda_op_abs(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sgn(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_step(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_silu_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_gelu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_exp(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

package/vendor/whisper.cpp/ggml/src/ggml-cuda/upscale.cu

@@ -0,0 +1,51 @@
+#include "upscale.cuh"
+
+static __global__ void upscale_f32(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne10, const int ne11, const int ne12, const int ne13,
+        const float sf0, const float sf1, const float sf2, const float sf3) {
+    int index = threadIdx.x + blockIdx.x * blockDim.x;
+    if (index >= ne10 * ne11 * ne12 * ne13) {
+        return;
+    }
+
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
+
+    int i00 = i10 / sf0;
+    int i01 = i11 / sf1;
+    int i02 = i12 / sf2;
+    int i03 = i13 / sf3;
+
+    dst[index] = *( (const float *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00) );
+}
+
+static void upscale_f32_cuda(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne10, const int ne11, const int ne12, const int ne13,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        cudaStream_t stream) {
+    int dst_size = ne10 * ne11 * ne12 * ne13;
+    int num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
+
+    upscale_f32<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3);
+}
+
+void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const float sf0 = (float)dst->ne[0]/src0->ne[0];
+    const float sf1 = (float)dst->ne[1]/src0->ne[1];
+    const float sf2 = (float)dst->ne[2]/src0->ne[2];
+    const float sf3 = (float)dst->ne[3]/src0->ne[3];
+
+    upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
+}

package/vendor/whisper.cpp/ggml/src/ggml-cuda/upscale.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_UPSCALE_BLOCK_SIZE 256
+
+void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst);