public_git
/
whisper.cpp
mirror of https://github.com/ggml-org/whisper.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vulkan: implement ADD1, ARANGE, FILL, SOFTPLUS, STEP, ROUND, CEIL, FLOOR, TRUNC (llama/17319) 

* vulkan: initialize array

* vulkan: implement ADD1

* vulkan: implement ARANGE

* vulkan: implement FILL

* vulkan: implement SOFTPLUS

* vulkan: implement STEP

* vulkan: implement ROUND

* vulkan: implement CEIL

* vulkan: implement FLOOR

* vulkan: implement TRUNC

* docs: update Vulkan ops

Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
This commit is contained in:
Giuseppe Scrivano 2025-11-19 17:29:45 +01:00 committed by Georgi Gerganov
parent 95d0b0b0cf
commit 24b14cad87
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
11 changed files with 421 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

196
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -669,6 +669,20 @@ struct vk_device_struct {
vk_pipeline pipeline_hardsigmoid[2];
vk_pipeline pipeline_hardswish[2];
vk_pipeline pipeline_abs[2];
vk_pipeline pipeline_softplus[2];
vk_pipeline pipeline_step[2];
vk_pipeline pipeline_round[2];
vk_pipeline pipeline_ceil[2];
vk_pipeline pipeline_floor[2];
vk_pipeline pipeline_trunc[2];
vk_pipeline pipeline_add1_f16_f16;
vk_pipeline pipeline_add1_f16_f32;
vk_pipeline pipeline_add1_f32_f32;
vk_pipeline pipeline_arange_f32;
vk_pipeline pipeline_fill_f32;
vk_pipeline pipeline_geglu[2];
vk_pipeline pipeline_reglu[2];
@ -3841,6 +3855,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
CREATE_UNARY(hardsigmoid)
CREATE_UNARY(hardswish)
CREATE_UNARY(abs)
CREATE_UNARY(softplus)
CREATE_UNARY(step)
CREATE_UNARY(round)
CREATE_UNARY(ceil)
CREATE_UNARY(floor)
CREATE_UNARY(trunc)
#undef CREATE_UNARY
#define CREATE_UNARY_RTE(name) \
@ -3854,6 +3874,14 @@ static void ggml_vk_load_shaders(vk_device& device) {
CREATE_UNARY_RTE(exp)
#undef CREATE_UNARY_RTE
ggml_vk_create_pipeline(device, device->pipeline_add1_f16_f16, "add1_f16_f16", add1_f16_f16_len, add1_f16_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_add1_f16_f32, "add1_f16_f32", add1_f16_f32_len, add1_f16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_add1_f32_f32, "add1_f32_f32", add1_f32_f32_len, add1_f32_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_arange_f32, "arange_f32", arange_f32_len, arange_f32_data, "main", 1, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_fill_f32, "fill_f32", fill_f32_len, fill_f32_data, "main", 1, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
#define CREATE_GLU(name) \
if (device->float_controls_rte_fp16) { \
ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32_rte", name ## _f32_rte_len, name ## _f32_rte_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true); \
@ -8279,6 +8307,18 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return ctx->device->pipeline_hardswish[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_ABS:
return ctx->device->pipeline_abs[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_SOFTPLUS:
return ctx->device->pipeline_softplus[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_STEP:
return ctx->device->pipeline_step[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_ROUND:
return ctx->device->pipeline_round[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_CEIL:
return ctx->device->pipeline_ceil[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_FLOOR:
return ctx->device->pipeline_floor[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_TRUNC:
return ctx->device->pipeline_trunc[dst->type == GGML_TYPE_F16];
default:
break;
}
@ -8473,7 +8513,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
case GGML_OP_CONV_TRANSPOSE_2D:
if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
std::array<uint32_t, 3> elements;
std::array<uint32_t, 3> elements{};
if (op == GGML_OP_CONV_2D) elements = ggml_vk_get_conv_elements(dst);
else if (op == GGML_OP_CONV_TRANSPOSE_2D) elements = ggml_vk_get_conv_transpose_2d_elements(dst);
vk_conv_shapes shape;
@ -8551,6 +8591,27 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
}
}
return nullptr;
case GGML_OP_ADD1:
if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
return ctx->device->pipeline_add1_f16_f16;
}
if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
return ctx->device->pipeline_add1_f16_f32;
}
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_add1_f32_f32;
}
return nullptr;
case GGML_OP_ARANGE:
if (dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_arange_f32;
}
return nullptr;
case GGML_OP_FILL:
if (dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_fill_f32;
}
return nullptr;
default:
return nullptr;
}
@ -8840,6 +8901,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
case GGML_OP_SUB:
case GGML_OP_DIV:
case GGML_OP_MUL:
case GGML_OP_ADD1:
case GGML_OP_ARANGE:
case GGML_OP_FILL:
case GGML_OP_SCALE:
case GGML_OP_SQR:
case GGML_OP_SQRT:
@ -9457,6 +9521,63 @@ static void ggml_vk_sqrt(ggml_backend_vk_context * ctx, vk_context& subctx, cons
ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SQRT, vk_op_unary_push_constants_init(src0, dst));
}
static void ggml_vk_add1(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const uint32_t src0_type_size = ggml_type_size(src0->type);
const uint32_t src1_type_size = ggml_type_size(src1->type);
const uint32_t dst_type_size = ggml_type_size(dst->type);
ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_ADD1, {
(uint32_t)ggml_nelements(src0),
(uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
(uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
(uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size,
0,
0.0f, 0.0f, 0,
});
}
static void ggml_vk_arange(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
VK_LOG_DEBUG("ggml_vk_arange(dst=" << dst << ", ne=" << ggml_nelements(dst) << ")");
vk_op_push_constants pc = {
(uint32_t)ggml_nelements(dst),
1,
ggml_get_op_params_f32(dst, 0),
ggml_get_op_params_f32(dst, 2),
};
vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, nullptr, nullptr, nullptr, dst, GGML_OP_ARANGE);
GGML_ASSERT(pipeline != nullptr);
ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, false);
std::array<uint32_t, 3> elements = { (uint32_t)ggml_nelements(dst), 1, 1 };
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { dst_buf }, pc, elements);
}
static void ggml_vk_fill(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
VK_LOG_DEBUG("ggml_vk_fill(dst=" << dst << ", ne=" << ggml_nelements(dst) << ")");
vk_op_push_constants pc = {
(uint32_t)ggml_nelements(dst),
1,
ggml_get_op_params_f32(dst, 0),
0.0f,
};
vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, nullptr, nullptr, nullptr, dst, GGML_OP_FILL);
GGML_ASSERT(pipeline != nullptr);
ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, false);
std::array<uint32_t, 3> elements = { (uint32_t)ggml_nelements(dst), 1, 1 };
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { dst_buf }, pc, elements);
}
static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst));
}
@ -11289,6 +11410,12 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
case GGML_UNARY_OP_HARDSIGMOID:
case GGML_UNARY_OP_HARDSWISH:
case GGML_UNARY_OP_ABS:
case GGML_UNARY_OP_SOFTPLUS:
case GGML_UNARY_OP_STEP:
case GGML_UNARY_OP_ROUND:
case GGML_UNARY_OP_CEIL:
case GGML_UNARY_OP_FLOOR:
case GGML_UNARY_OP_TRUNC:
break;
default:
return false;
@ -11330,6 +11457,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
case GGML_OP_SUB:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_ADD1:
case GGML_OP_ARANGE:
case GGML_OP_FILL:
case GGML_OP_CONCAT:
case GGML_OP_UPSCALE:
case GGML_OP_SCALE:
@ -11542,6 +11672,18 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
case GGML_OP_UPSCALE:
ggml_vk_upscale(ctx, compute_ctx, src0, node);
break;
case GGML_OP_ADD1:
ggml_vk_add1(ctx, compute_ctx, src0, src1, node);
break;
case GGML_OP_ARANGE:
ggml_vk_arange(ctx, compute_ctx, node);
break;
case GGML_OP_FILL:
ggml_vk_fill(ctx, compute_ctx, node);
break;
case GGML_OP_SCALE:
ggml_vk_scale(ctx, compute_ctx, src0, node);
@ -11626,6 +11768,12 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
case GGML_UNARY_OP_HARDSIGMOID:
case GGML_UNARY_OP_HARDSWISH:
case GGML_UNARY_OP_ABS:
case GGML_UNARY_OP_SOFTPLUS:
case GGML_UNARY_OP_STEP:
case GGML_UNARY_OP_ROUND:
case GGML_UNARY_OP_CEIL:
case GGML_UNARY_OP_FLOOR:
case GGML_UNARY_OP_TRUNC:
ggml_vk_unary(ctx, compute_ctx, src0, node);
break;
default:
@ -11828,6 +11976,9 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
case GGML_OP_SUB:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_ADD1:
case GGML_OP_ARANGE:
case GGML_OP_FILL:
case GGML_OP_ADD_ID:
case GGML_OP_CONCAT:
case GGML_OP_UPSCALE:
@ -11899,6 +12050,12 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
case GGML_UNARY_OP_HARDSIGMOID:
case GGML_UNARY_OP_HARDSWISH:
case GGML_UNARY_OP_ABS:
case GGML_UNARY_OP_SOFTPLUS:
case GGML_UNARY_OP_STEP:
case GGML_UNARY_OP_ROUND:
case GGML_UNARY_OP_CEIL:
case GGML_UNARY_OP_FLOOR:
case GGML_UNARY_OP_TRUNC:
buf = tensor->buffer;
break;
default:
@ -13501,6 +13658,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_UNARY_OP_HARDSIGMOID:
case GGML_UNARY_OP_HARDSWISH:
case GGML_UNARY_OP_ABS:
case GGML_UNARY_OP_SOFTPLUS:
case GGML_UNARY_OP_STEP:
case GGML_UNARY_OP_ROUND:
case GGML_UNARY_OP_CEIL:
case GGML_UNARY_OP_FLOOR:
case GGML_UNARY_OP_TRUNC:
return ggml_is_contiguous(op->src[0]) &&
(op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
(op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
@ -13818,6 +13981,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_UPSCALE:
case GGML_OP_ACC:
case GGML_OP_CONCAT:
case GGML_OP_ADD1:
case GGML_OP_ARANGE:
case GGML_OP_FILL:
case GGML_OP_SCALE:
case GGML_OP_PAD:
case GGML_OP_ROLL:
@ -14300,6 +14466,16 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
} else if (tensor->op == GGML_OP_SCALE) {
const float * params = (const float *)tensor->op_params;
tensor_clone = ggml_scale_bias(ggml_ctx, src_clone[0], params[0], params[1]);
} else if (tensor->op == GGML_OP_ADD1) {
tensor_clone = ggml_add1(ggml_ctx, src_clone[0], src_clone[1]);
} else if (tensor->op == GGML_OP_ARANGE) {
const float start = ggml_get_op_params_f32(tensor, 0);
const float stop = ggml_get_op_params_f32(tensor, 1);
const float step = ggml_get_op_params_f32(tensor, 2);
tensor_clone = ggml_arange(ggml_ctx, start, stop, step);
} else if (tensor->op == GGML_OP_FILL) {
const float value = ggml_get_op_params_f32(tensor, 0);
tensor_clone = ggml_fill(ggml_ctx, tensor_clone, value);
} else if (tensor->op == GGML_OP_SQR) {
tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]);
} else if (tensor->op == GGML_OP_SQRT) {
@ -14413,6 +14589,24 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
case GGML_UNARY_OP_ABS:
tensor_clone = ggml_abs(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_SOFTPLUS:
tensor_clone = ggml_softplus(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_STEP:
tensor_clone = ggml_step(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_ROUND:
tensor_clone = ggml_round(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_CEIL:
tensor_clone = ggml_ceil(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_FLOOR:
tensor_clone = ggml_floor(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_TRUNC:
tensor_clone = ggml_trunc(ggml_ctx, src_clone[0]);
break;
default:
std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
GGML_ABORT("fatal error");

28
ggml/src/ggml-vulkan/vulkan-shaders/add1.comp Normal file
View File

@ -0,0 +1,28 @@
#version 450
#extension GL_EXT_shader_16bit_storage : require
#include "types.glsl"
#include "generic_binary_head.glsl"
const uint num_threads = 256;
layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
void main() {
uint idx = get_idx();
const uint num_iter = 2;
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
if (idx >= p.ne) {
continue;
}
uint i00, i01, i02, i03;
get_indices(idx, i00, i01, i02, i03);
data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset()]));
idx += num_threads;
}
}

20
ggml/src/ggml-vulkan/vulkan-shaders/arange.comp Normal file
View File

@ -0,0 +1,20 @@
#version 450
#include "generic_head.glsl"
#include "types.glsl"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
// p.param1 = start, p.param2 = step
float value = p.param1 + p.param2 * float(i);
data_d[i] = D_TYPE(value);
}

22
ggml/src/ggml-vulkan/vulkan-shaders/ceil.comp Normal file
View File

@ -0,0 +1,22 @@
#version 450
#include "generic_head.glsl"
#include "types.glsl"
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
const float x = float(data_a[i]);
data_d[i] = D_TYPE(ceil(x));
}

19
ggml/src/ggml-vulkan/vulkan-shaders/fill.comp Normal file
View File

@ -0,0 +1,19 @@
#version 450
#include "generic_head.glsl"
#include "types.glsl"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
// p.param1 = fill value
data_d[i] = D_TYPE(p.param1);
}

22
ggml/src/ggml-vulkan/vulkan-shaders/floor.comp Normal file
View File

@ -0,0 +1,22 @@
#version 450
#include "generic_head.glsl"
#include "types.glsl"
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
const float x = float(data_a[i]);
data_d[i] = D_TYPE(floor(x));
}

29
ggml/src/ggml-vulkan/vulkan-shaders/round.comp Normal file
View File

@ -0,0 +1,29 @@
#version 450
#include "generic_head.glsl"
#include "types.glsl"
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
const float x = float(data_a[i]);
float result;
// Round halfway cases away from zero as roundf does.
if (x >= 0.0) {
result = floor(x + 0.5);
} else {
result = ceil(x - 0.5);
}
data_d[i] = D_TYPE(result);
}

23
ggml/src/ggml-vulkan/vulkan-shaders/softplus.comp Normal file
View File

@ -0,0 +1,23 @@
#version 450
#include "generic_head.glsl"
#include "types.glsl"
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
const float x = float(data_a[i]);
const float result = (x > 20.0f) ? x : log(1.0f + exp(x));
data_d[i] = D_TYPE(result);
}

22
ggml/src/ggml-vulkan/vulkan-shaders/step.comp Normal file
View File

@ -0,0 +1,22 @@
#version 450
#include "generic_head.glsl"
#include "types.glsl"
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
const float x = float(data_a[i]);
data_d[i] = D_TYPE(x >= 0.0f ? 1.0f : 0.0f);
}

22
ggml/src/ggml-vulkan/vulkan-shaders/trunc.comp Normal file
View File

@ -0,0 +1,22 @@
#version 450
#include "generic_head.glsl"
#include "types.glsl"
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
const float x = float(data_a[i]);
data_d[i] = D_TYPE(trunc(x));
}

19
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View File

@ -846,6 +846,25 @@ void process_shaders() {
string_to_spv("abs_f16", "abs.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("abs_f32", "abs.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("softplus_f16", "softplus.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("softplus_f32", "softplus.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("add1_f16_f16", "add1.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
string_to_spv("add1_f16_f32", "add1.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
string_to_spv("add1_f32_f32", "add1.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("arange_f32", "arange.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("fill_f32", "fill.comp", {{"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("step_f16", "step.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("step_f32", "step.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("round_f16", "round.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("round_f32", "round.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("ceil_f16", "ceil.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("ceil_f32", "ceil.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("floor_f16", "floor.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("floor_f32", "floor.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("trunc_f16", "trunc.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("trunc_f32", "trunc.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
for (auto rte : {false, true}) {
std::string suffix = rte ? "_rte" : "";
string_to_spv("geglu_f16" + suffix, "geglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});