File size: 4,042 Bytes
1d30d42 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 |
//------------------------------------------------------------------------------
// This file is contains additional kernels for data conversion.
// These kernels are used when loading the model, so its performance is less
// important.
//------------------------------------------------------------------------------
#ifdef cl_khr_fp16
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#elif defined(cl_amd_fp16)
#pragma OPENCL EXTENSION cl_amd_fp16 : enable
#else
#error "Half precision floating point not supportedby OpenCL implementation on your device."
#endif
#ifdef cl_khr_subgroups
#pragma OPENCL EXTENSION cl_khr_subgroups : enable
#elif defined(cl_intel_subgroups)
#pragma OPENCL EXTENSION cl_intel_subgroups : enable
#else
#error "Subgroup not supported on your device."
#endif
#ifdef cl_intel_required_subgroup_size
// Always use subgroup size of 32 on Intel.
#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
#define INTEL_GPU 1
#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
#elif defined(cl_qcom_reqd_sub_group_size)
// Always use subgroups size of 64 on Adreno.
#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
#define ADRENO_GPU 1
#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
#else
// TODO: do not know how to choose subgroup size on other GPUs.
#error "Selecting subgroup size is not supported on your device."
#endif
#define QK4_0 32
#define QR4_0 2
#define QK4_1 32
#define QR4_1 2
#define QK5_0 32
#define QR5_0 2
#define QK5_1 32
#define QR5_1 2
#define QK8_0 32
#define QR8_0 1
#define QK_K 256
#define K_QUANTS_PER_ITERATION 2
typedef char int8_t;
typedef uchar uint8_t;
typedef short int16_t;
typedef ushort uint16_t;
typedef int int32_t;
typedef uint uint32_t;
//------------------------------------------------------------------------------
// block_q4_0
//------------------------------------------------------------------------------
struct block_q4_0
{
half d;
uint8_t qs[QK4_0 / 2];
};
//------------------------------------------------------------------------------
// mul_vec_q_n_f32_flat_noshuffle
//
// This variation uses flat arrays (struct of arrays, SOA) representation for
// quant tensors. It also uses non shuffled bit order for weights.
//
// The shuffled version is kept in the original file because moving it here
// seems to result in worse performance for adreno.
//------------------------------------------------------------------------------
kernel void kernel_convert_block_q4_0_noshuffle(
global struct block_q4_0 * src0,
global uchar * dst_q,
global half * dst_d
) {
global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0);
global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0);
global half * d = (global half *) dst_d + get_global_id(0);
*d = b->d;
for (int i = 0; i < QK4_0/4; ++i) {
uchar x0 = b->qs[2*i + 0];
uchar x1 = b->qs[2*i + 1];
q[i + 0 ] = convert_uchar(x0 & 0x0F) | convert_uchar((x1 & 0x0F) << 4);
q[i + QK4_0/4] = convert_uchar((x0 & 0xF0) >> 4) | convert_uchar(x1 & 0xF0);
#ifdef ADRENO_GPU
// Workaround for adreno - must have the following printf statement for
// the kernel to work properly. Otherwise it produces incorrect result.
// convert_uchar above also seems necessary.
// Compare against a large number so that it does not print anything.
// get_sub_group_local_id() also works.
if (get_global_id(0) == 65536*4096) {
printf("%04x - %02x\n", *(global ushort*)d, ((x0 & 0xF0) >> 4) | (x1 & 0xF0));
}
#endif
}
}
|