Enable non-contiguous support for simple ops
This commit is contained in:
0cc4m
parent
c19c59812e
commit
0caf8dc906
3 changed files with 2163 additions and 1154 deletions
File diff suppressed because it is too large
Load diff
|
|
@ -217,17 +217,19 @@ struct vk_op_push_constants {
|
|||
|
||||
struct vk_op_unary_push_constants {
|
||||
uint32_t ne;
|
||||
uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
|
||||
uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
|
||||
uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
|
||||
uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
|
||||
uint32_t d_offset;
|
||||
float param1; float param2;
|
||||
};
|
||||
|
||||
struct vk_op_binary_push_constants {
|
||||
uint32_t ne;
|
||||
uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
|
||||
uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
|
||||
uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t nb20; uint32_t nb21; uint32_t nb22; uint32_t nb23;
|
||||
uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
|
||||
uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
|
||||
uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23; uint32_t nb20; uint32_t nb21; uint32_t nb22; uint32_t nb23;
|
||||
uint32_t d_offset;
|
||||
float param1; float param2;
|
||||
};
|
||||
|
||||
struct vk_op_diag_mask_push_constants {
|
||||
|
|
@ -1039,15 +1041,15 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
|
|||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
|
||||
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
|
||||
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
|
||||
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
|
||||
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
|
||||
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
|
||||
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_gelu_f32, "gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
||||
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_silu_f32, "silu_f32", silu_f32_len, silu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
||||
|
|
@ -2217,9 +2219,10 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context
|
|||
|
||||
const vk_op_unary_push_constants pc = {
|
||||
(uint32_t)ne,
|
||||
(uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->nb[0] / tensor_type_size, (uint32_t)tensor->nb[1] / tensor_type_size, (uint32_t)tensor->nb[2] / tensor_type_size, (uint32_t)tensor->nb[3] / tensor_type_size,
|
||||
(uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], 1 , (uint32_t)tensor->ne[0] , (uint32_t)(tensor->ne[0] * tensor->ne[1]) , (uint32_t)(tensor->ne[0] * tensor->ne[1] * tensor->ne[2]),
|
||||
(uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], (uint32_t)tensor->nb[0] / tensor_type_size, (uint32_t)tensor->nb[1] / tensor_type_size, (uint32_t)tensor->nb[2] / tensor_type_size, (uint32_t)tensor->nb[3] / tensor_type_size,
|
||||
(uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], 1 , (uint32_t)tensor->ne[0] , (uint32_t)(tensor->ne[0] * tensor->ne[1]) , (uint32_t)(tensor->ne[0] * tensor->ne[1] * tensor->ne[2]),
|
||||
0,
|
||||
0.0f, 0.0f,
|
||||
};
|
||||
ggml_vk_sync_buffers(subctx);
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, { ne, 1, 1 });
|
||||
|
|
@ -3003,7 +3006,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
|
|||
#endif
|
||||
GGML_ASSERT(!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))); // NOLINT
|
||||
GGML_ASSERT(op == GGML_OP_CPY || ggml_vk_dim01_contiguous(src0)); // NOLINT
|
||||
GGML_ASSERT(src1 == nullptr || ggml_vk_dim01_contiguous(src1)); // NOLINT
|
||||
GGML_ASSERT(dst->extra != nullptr);
|
||||
const uint64_t ne00 = src0->ne[0];
|
||||
const uint64_t ne01 = src0->ne[1];
|
||||
|
|
@ -3166,6 +3168,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
|
|||
ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, d_sz);
|
||||
}
|
||||
} else {
|
||||
std::cerr << "oopsie" << std::endl;
|
||||
ggml_pipeline_allocate_descriptor_sets(ctx, pipeline, ne02 * ne03);
|
||||
|
||||
switch (dst->op) {
|
||||
|
|
@ -3220,37 +3223,86 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context * subctx,
|
|||
}
|
||||
|
||||
static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
||||
ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_ADD, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
|
||||
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, dst, GGML_OP_ADD, {
|
||||
(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,
|
||||
});
|
||||
}
|
||||
|
||||
static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
||||
ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_MUL, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
|
||||
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, dst, GGML_OP_MUL, {
|
||||
(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,
|
||||
});
|
||||
}
|
||||
|
||||
static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
||||
float * op_params = (float *)dst->op_params;
|
||||
ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_SCALE, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f });
|
||||
const uint32_t src0_type_size = ggml_type_size(src0->type);
|
||||
const uint32_t dst_type_size = ggml_type_size(dst->type);
|
||||
|
||||
ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_SCALE, {
|
||||
(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) 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,
|
||||
op_params[0], 0.0f
|
||||
});
|
||||
}
|
||||
|
||||
static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
||||
ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_SQR, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
|
||||
const uint32_t src0_type_size = ggml_type_size(src0->type);
|
||||
const uint32_t dst_type_size = ggml_type_size(dst->type);
|
||||
|
||||
ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_SQR, {
|
||||
(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) 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,
|
||||
});
|
||||
}
|
||||
|
||||
static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
||||
float * op_params = (float *)dst->op_params;
|
||||
ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_CLAMP, { (uint32_t)ggml_nelements(src0), 0, op_params[0], op_params[1] });
|
||||
const uint32_t src0_type_size = ggml_type_size(src0->type);
|
||||
const uint32_t dst_type_size = ggml_type_size(dst->type);
|
||||
|
||||
ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_CLAMP, {
|
||||
(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) 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,
|
||||
op_params[0], op_params[1],
|
||||
});
|
||||
}
|
||||
|
||||
static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
||||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
|
||||
const int src0_type_size = ggml_type_size(src0->type);
|
||||
const int dst_type_size = ggml_type_size(dst->type);
|
||||
const uint32_t src0_type_size = ggml_type_size(src0->type);
|
||||
const uint32_t dst_type_size = ggml_type_size(dst->type);
|
||||
const uint32_t d_offset = (extra->offset % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
|
||||
|
||||
ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_CPY, {
|
||||
(uint32_t)ggml_nelements(src0),
|
||||
(uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (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) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (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,
|
||||
(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) 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,
|
||||
d_offset,
|
||||
0.0f, 0.0f,
|
||||
});
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -1516,9 +1516,10 @@ generic_unary_op_head = """#version 450
|
|||
layout (push_constant) uniform parameter
|
||||
{
|
||||
uint ne;
|
||||
uint ne00; uint ne01; uint ne02; uint nb00; uint nb01; uint nb02; uint nb03;
|
||||
uint ne10; uint ne11; uint ne12; uint nb10; uint nb11; uint nb12; uint nb13;
|
||||
uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
|
||||
uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
|
||||
uint d_offset;
|
||||
float param1; float param2;
|
||||
} p;
|
||||
|
||||
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
|
||||
|
|
@ -1559,16 +1560,18 @@ generic_binary_op_head = """#version 450
|
|||
layout (push_constant) uniform parameter
|
||||
{
|
||||
uint ne;
|
||||
uint ne00; uint ne01; uint ne02; uint nb00; uint nb01; uint nb02; uint nb03;
|
||||
uint ne10; uint ne11; uint ne12; uint nb10; uint nb11; uint nb12; uint nb13;
|
||||
uint ne20; uint ne21; uint ne22; uint nb20; uint nb21; uint nb22; uint nb23;
|
||||
uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
|
||||
uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
|
||||
uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
|
||||
uint d_offset;
|
||||
uint param1; uint param2;
|
||||
} p;
|
||||
|
||||
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
|
||||
|
||||
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
|
||||
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
|
||||
layout (binding = 1) readonly buffer B {A_TYPE data_b[];};
|
||||
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
|
||||
|
||||
uint src0_idx(uint idx) {
|
||||
const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
|
||||
|
|
@ -1581,13 +1584,14 @@ uint src0_idx(uint idx) {
|
|||
}
|
||||
|
||||
uint src1_idx(uint idx) {
|
||||
const uint i13 = idx / (p.ne12*p.ne11*p.ne10);
|
||||
const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
|
||||
const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10);
|
||||
const uint i12_offset = i12*p.ne11*p.ne10;
|
||||
const uint i11 = (idx - i13_offset - i12_offset) / p.ne10;
|
||||
const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
|
||||
return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10;
|
||||
const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
|
||||
const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
|
||||
const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
|
||||
const uint i02_offset = i02*p.ne01*p.ne00;
|
||||
const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
|
||||
const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
|
||||
|
||||
return (i03 % p.ne13)*p.nb13 + (i02 % p.ne12)*p.nb12 + (i01 % p.ne11)*p.nb11 + (i00 % p.ne10)*p.nb10;
|
||||
}
|
||||
|
||||
uint dst_idx(uint idx) {
|
||||
|
|
@ -1607,92 +1611,34 @@ void main() {
|
|||
"""
|
||||
|
||||
# MUL F32
|
||||
mul_body = """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) readonly buffer Y {B_TYPE data_b[];};
|
||||
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
|
||||
|
||||
void main() {
|
||||
const uint idx = gl_GlobalInvocationID.x;
|
||||
|
||||
if (idx >= p.KX) {
|
||||
return;
|
||||
}
|
||||
|
||||
data_d[idx] = D_TYPE(FLOAT_TYPE(data_a[idx]) * FLOAT_TYPE(data_b[idx % p.KY]));
|
||||
mul_body = """
|
||||
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
|
||||
}
|
||||
"""
|
||||
|
||||
# ADD
|
||||
add_body = """
|
||||
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) readonly buffer Y {B_TYPE data_b[];};
|
||||
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
|
||||
|
||||
void main() {
|
||||
const uint idx = gl_GlobalInvocationID.x;
|
||||
|
||||
if (idx >= p.KX) {
|
||||
return;
|
||||
}
|
||||
|
||||
data_d[idx] = D_TYPE(FLOAT_TYPE(data_a[idx]) + FLOAT_TYPE(data_b[idx % p.KY]));
|
||||
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) + FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
|
||||
}
|
||||
"""
|
||||
|
||||
# SCALE
|
||||
scale_body = """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 idx = gl_GlobalInvocationID.x;
|
||||
|
||||
if (idx >= p.KX) {
|
||||
return;
|
||||
}
|
||||
|
||||
data_d[idx] = D_TYPE(FLOAT_TYPE(data_a[idx]) * FLOAT_TYPE(p.param1));
|
||||
scale_body = """
|
||||
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) + FLOAT_TYPE(p.param1));
|
||||
}
|
||||
"""
|
||||
|
||||
# SQR
|
||||
sqr_body = """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 idx = gl_GlobalInvocationID.x;
|
||||
|
||||
if (idx >= p.KX) {
|
||||
return;
|
||||
}
|
||||
|
||||
const FLOAT_TYPE val = FLOAT_TYPE(data_a[idx]);
|
||||
data_d[idx] = D_TYPE(val * val);
|
||||
sqr_body = """
|
||||
const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
|
||||
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val * val);
|
||||
}
|
||||
"""
|
||||
|
||||
# CLAMP
|
||||
clamp_body = """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 idx = gl_GlobalInvocationID.x;
|
||||
|
||||
if (idx >= p.KX) {
|
||||
return;
|
||||
}
|
||||
|
||||
const FLOAT_TYPE val = FLOAT_TYPE(data_a[idx]);
|
||||
data_d[idx] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
|
||||
clamp_body = """
|
||||
const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
|
||||
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
|
||||
}
|
||||
"""
|
||||
|
||||
|
|
@ -2367,16 +2313,16 @@ async def main():
|
|||
tasks.append(string_to_spv("cpy_f32_f16", f"{generic_unary_op_head}\n{cpy_end}", {"A_TYPE": "float", "D_TYPE": "float16_t"}))
|
||||
tasks.append(string_to_spv("cpy_f16_f16", f"{generic_unary_op_head}\n{cpy_f16_f16_end}", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
|
||||
|
||||
tasks.append(string_to_spv("add_f32", f"{generic_head}\n{shader_f32}\n{add_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}))
|
||||
tasks.append(string_to_spv("add_f32", f"{generic_binary_op_head}\n{add_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
|
||||
|
||||
tasks.append(string_to_spv("split_k_reduce", mulmat_split_k_reduce_src, {}))
|
||||
tasks.append(string_to_spv("mul_f32", f"{generic_head}\n{shader_f32}\n{mul_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}))
|
||||
tasks.append(string_to_spv("mul_f32", f"{generic_binary_op_head}\n{mul_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
|
||||
|
||||
tasks.append(string_to_spv("scale_f32", f"{generic_head}\n{shader_f32}\n{scale_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
|
||||
tasks.append(string_to_spv("scale_f32", f"{generic_unary_op_head}\n{scale_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
|
||||
|
||||
tasks.append(string_to_spv("sqr_f32", f"{generic_head}\n{shader_f32}\n{sqr_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
|
||||
tasks.append(string_to_spv("sqr_f32", f"{generic_unary_op_head}\n{sqr_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
|
||||
|
||||
tasks.append(string_to_spv("clamp_f32", f"{generic_head}\n{shader_f32}\n{clamp_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
|
||||
tasks.append(string_to_spv("clamp_f32", f"{generic_unary_op_head}\n{clamp_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
|
||||
|
||||
tasks.append(string_to_spv("gelu_f32", f"{generic_head}\n{shader_f32}\n{gelu_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
|
||||
tasks.append(string_to_spv("silu_f32", f"{generic_head}\n{shader_f32}\n{silu_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue