Skip to content

Commit

Permalink
[TEMP]
Browse files Browse the repository at this point in the history 
Signed-off-by: Min, Byung-il <byungil.min@intel.com>
  • Loading branch information
@byungilm
byungilm committed Oct 17, 2024
1 parent cf43e5e commit 0b0e7ef
Show file tree
Hide file tree
Showing 3 changed files with 156 additions and 196 deletions.
170 changes: 58 additions & 112 deletions src/plugins/intel_gpu/src/kernel_selector/cl_kernels/fully_connected_gpu_bf_tiled.cl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -214,11 +214,10 @@ inline void FUNC(fc_bf_tiled_kernel_default)(
ACCUMULATOR_VEC_TYPE acc[TILE_B] = { };
INPUT_VEC_TYPE in_0[TILE_B] = { };

#if !USE_SLM
#if !USE_SLM || !COMPRESSED_WEIGHTS_INT4
FILTER_VEC_TYPE wei = 0;
#endif


#if OUTPUT_3D
uint out_b0 = out_b / OUTPUT_FEATURE_NUM;
uint out_b1 = out_b % OUTPUT_FEATURE_NUM;
Expand Down Expand Up @@ -372,6 +371,9 @@ inline void FUNC(fc_bf_tiled_kernel_default)(
#else
SLM_FILTER_PACKED_VEC wei_packed = BLOCK_READN(FILTER_TYPE, FILTER_LOAD_BLOCK_SIZE/*4*/, weights, weights_idx);
SLM_FILTER_UNPACKED_VEC wei_unpacked = UNPACK_INT4(ACCUMULATOR_TYPE, *((INT4_PACKED_TYPE_PRELOAD*)&wei_packed));
// comapare
// FILTER_PACKED_VEC_TYPE wei_packed = BLOCK_READN(FILTER_TYPE, TILE_K_OFM_PACKED, weights, weights_offset)
// wei = UNPACK_INT4(ACCUMULATOR_TYPE, *((INT4_PACKED_TYPE*)&wei_packed));
#endif
ACCUMULATOR_TYPE* w = (ACCUMULATOR_TYPE*)(&wei_unpacked);
unroll_for(uint fi = 0; fi < TILE_OFM; ++fi) {
Expand Down Expand Up @@ -751,7 +753,8 @@ inline void FUNC(fc_bf_tiled_kernel_default)(


// Dyc Quantize
//#if USE_SLM && DYNAMIC_QUANTIZE
#if USE_SLM && DYNAMIC_QUANTIZE

#define PACKED_DQ_TYPE int
#define DQ_VEC_TYPE MAKE_VECTOR_TYPE(DQ_TYPE, TILE_IFM)
#define DQ_SLM_FILTER_VEC MAKE_VECTOR_TYPE(DQ_TYPE, 4)
Expand All @@ -768,6 +771,7 @@ inline void FUNC(fc_bf_tiled_kernel_default)(
#define AS_TYPE_N(type, n, x) AS_TYPE_N_(type, n, x)
#define AS_DQ_TYPE_4(x) AS_TYPE_N(DQ_TYPE, INPUT_LOAD_SIZE, x)

#if 0
inline void FUNC(fc_bf_tiled_kernel_dyn_quan_none_slm)(
OPTIONAL_SHAPE_INFO_ARG
const __global INPUT0_TYPE* input,
Expand Down Expand Up @@ -926,9 +930,8 @@ inline void FUNC(fc_bf_tiled_kernel_dyn_quan_none_slm)(
// SLM_FILTER_PACKED_VEC wei_packed = BLOCK_READN(FILTER_TYPE, FILTER_LOAD_BLOCK_SIZE, weights, weights_idx);
// DQ_SLM_FILTER_UNPACKED_VEC dq_wei_unpacked = UNPACK_TRANSPOSED_INT4(DQ_TYPE, *((INT4_PACKED_TYPE_PRELOAD *)&wei_packed));
#else
// wei = TO_FILTER_VEC_TYPE(FILTER_BLOCK_READ(weights, weights_offset));
DQ_FILTER_VEC_TYPE wei_packed = TO_DQ_FILTER_VEC_TYPE(FILTER_BLOCK_READ(weights, weights_offset));
// DQ_SLM_FILTER_UNPACKED_VEC wei_packed = TO_DQ_SLM_FILTER_UNPACKED_VEC(BLOCK_READN(FILTER_TYPE, TILE_K_OFM_PACKED, weights, weights_offset));
// DQ_FILTER_VEC_TYPE wei_packed = TO_DQ_FILTER_VEC_TYPE(FILTER_BLOCK_READ(weights, weights_offset)); // BLOCK_READN(FILTER_TYPE, TILE_K_OFM_PACKED, weights, weights_offset)
DQ_FILTER_VEC_TYPE wei_packed = TO_DQ_FILTER_VEC_TYPE(BLOCK_READN(FILTER_TYPE, TILE_K_OFM_PACKED, weights, weights_offset));
wei.s0123 = wei_packed.s0246;
wei.s4567 = wei_packed.s1357;
#endif
Expand Down Expand Up @@ -1108,8 +1111,7 @@ inline void FUNC(fc_bf_tiled_kernel_dyn_quan_none_slm)(
}
// =====================================================================================================================================
}

#if USE_SLM && DYNAMIC_QUANTIZE
#endif

inline void FUNC(fc_bf_tiled_kernel_dyn_quan)(
OPTIONAL_SHAPE_INFO_ARG
Expand Down Expand Up @@ -1240,15 +1242,6 @@ inline void FUNC(fc_bf_tiled_kernel_dyn_quan)(
// Next batch
in_offset += (TILE_IN_B_PITCH * 2);

// [TEST]
// // Load input.
//#define LOAD_IN_0(bi) do { \
// in_0[bi] = INPUT_BLOCK_READ(input, input_offset); \
// input_offset += TILE_IN_B_PITCH; \
// } while (false)
//CONST_LOOP(TILE_B, LOAD_IN_0);
//#undef LOAD_IN_0

#if NUM_LOOP_IN_DYN_QUAN_GROUP == 1
de_quantize_scale[bi * 2] = scale[scale_offset];
de_quantize_scale[bi * 2 + 1] = scale[scale_offset+ scale_pitch];
Expand Down Expand Up @@ -1281,10 +1274,14 @@ inline void FUNC(fc_bf_tiled_kernel_dyn_quan)(

__local int* char_slm_weight = (__local int*)wei_local_mem;

#if FILTER_LAYOUT_OS_IS_YX_OSV64_ISV2
uint weights_idx = weights_offset + local_id * SIMD * FILTER_LOAD_ITERS * FILTER_LOAD_BLOCK_SIZE * 2;
#if COMPRESSED_WEIGHTS_INT4
#if FILTER_LAYOUT_OS_IS_YX_OSV64_ISV2
uint weights_idx = weights_offset + local_id * SIMD * FILTER_LOAD_ITERS * FILTER_LOAD_BLOCK_SIZE * 2;
#else
uint weights_idx = weights_offset + local_id * SIMD * FILTER_LOAD_ITERS * FILTER_ACTUAL_LOAD_BLOCK_SIZE;
#endif
#else
uint weights_idx = weights_offset + local_id * SIMD * FILTER_LOAD_ITERS * FILTER_ACTUAL_LOAD_BLOCK_SIZE;
uint weights_idx = weights_offset + local_id * SIMD * FILTER_LOAD_ITERS * TILE_K_OFM_PACKED;
#endif
uint wei_local_idx = local_id * SIMD * FILTER_LOAD_ITERS * (FILTER_LOAD_BLOCK_SIZE/2) + sglid * 2;

Expand Down Expand Up @@ -1314,13 +1311,8 @@ inline void FUNC(fc_bf_tiled_kernel_dyn_quan)(
DQ_SLM_FILTER_UNPACKED_VEC dq_wei_unpacked = UNPACK_TRANSPOSED_INT4(DQ_TYPE, *((INT4_PACKED_TYPE_PRELOAD *)&wei_packed));
#endif
#else
// INT4
// FILTER_PACKED_VEC_TYPE wei_packed = BLOCK_READN(FILTER_TYPE, TILE_K_OFM_PACKED, weights, weights_offset);
// DQ_SLM_FILTER_UNPACKED_VEC dq_wei_unpacked = UNPACK_INT4(DQ_TYPE, *((INT4_PACKED_TYPE*)&wei_packed));
// INT8
// DQ_SLM_FILTER_UNPACKED_VEC dq_wei_unpacked = TO_DQ_SLM_FILTER_UNPACKED_VEC(BLOCK_READN(FILTER_TYPE, TILE_K_OFM_PACKED, weights, weights_offset));
DQ_SLM_FILTER_UNPACKED_VEC wei_packed = BLOCK_READN(FILTER_TYPE, TILE_K_OFM_PACKED, weights, weights_offset);
DQ_SLM_FILTER_UNPACKED_VEC dq_wei_unpacked;
DQ_FILTER_VEC_TYPE wei_packed = TO_DQ_FILTER_VEC_TYPE(FILTER_BLOCK_READ(weights, weights_idx));
dq_wei_unpacked.s0123 = wei_packed.s0246;
dq_wei_unpacked.s4567 = wei_packed.s1357;
#endif
Expand Down Expand Up @@ -1715,50 +1707,27 @@ KERNEL(fc)(
#endif
);
#else
#if 0
FUNC_CALL(fc_bf_tiled_kernel_default)(
OPTIONAL_SHAPE_INFO_TENSOR
input,
#if DECOMPRESSION_SCALE_TERM
decompression_scale,
#endif
#if DECOMPRESSION_ZP_TERM && !DECOMPRESSION_ZP_SCALAR
decompression_zp,
#endif
output,
weights
#if USE_SLM
, wei_local_mem
#endif
#if BIAS_TERM
, biases
#endif
#if HAS_FUSED_OPS_DECLS
, FUSED_OPS_ARGS
#endif
);
#else
FUNC_CALL(fc_bf_tiled_kernel_dyn_quan_none_slm)(
OPTIONAL_SHAPE_INFO_TENSOR
input,
quantized_input,
de_quan_scale,
#if DECOMPRESSION_SCALE_TERM
decompression_scale,
#endif
#if DECOMPRESSION_ZP_TERM && !DECOMPRESSION_ZP_SCALAR
decompression_zp,
#endif
output,
weights
#if BIAS_TERM
, biases
#endif
#if HAS_FUSED_OPS_DECLS
, FUSED_OPS_ARGS
#endif
);
FUNC_CALL(fc_bf_tiled_kernel_default)(
OPTIONAL_SHAPE_INFO_TENSOR
input,
#if DECOMPRESSION_SCALE_TERM
decompression_scale,
#endif
#if DECOMPRESSION_ZP_TERM && !DECOMPRESSION_ZP_SCALAR
decompression_zp,
#endif
output,
weights
#if USE_SLM
, wei_local_mem
#endif
#if BIAS_TERM
, biases
#endif
#if HAS_FUSED_OPS_DECLS
, FUSED_OPS_ARGS
#endif
);
#endif
}
#else
Expand All @@ -1785,50 +1754,27 @@ KERNEL(fc)(
#endif
);
#else
#if 0
FUNC_CALL(fc_bf_tiled_kernel_default)(
OPTIONAL_SHAPE_INFO_TENSOR
input,
#if DECOMPRESSION_SCALE_TERM
decompression_scale,
#endif
#if DECOMPRESSION_ZP_TERM && !DECOMPRESSION_ZP_SCALAR
decompression_zp,
#endif
output,
weights
#if USE_SLM
, wei_local_mem
#endif
#if BIAS_TERM
, biases
#endif
#if HAS_FUSED_OPS_DECLS
, FUSED_OPS_ARGS
#endif
);
#else
FUNC_CALL(fc_bf_tiled_kernel_dyn_quan_none_slm)(
OPTIONAL_SHAPE_INFO_TENSOR
input,
quantized_input,
de_quan_scale,
#if DECOMPRESSION_SCALE_TERM
decompression_scale,
#endif
#if DECOMPRESSION_ZP_TERM && !DECOMPRESSION_ZP_SCALAR
decompression_zp,
#endif
output,
weights
#if BIAS_TERM
, biases
#endif
#if HAS_FUSED_OPS_DECLS
, FUSED_OPS_ARGS
#endif
);
FUNC_CALL(fc_bf_tiled_kernel_default)(
OPTIONAL_SHAPE_INFO_TENSOR
input,
#if DECOMPRESSION_SCALE_TERM
decompression_scale,
#endif
#if DECOMPRESSION_ZP_TERM && !DECOMPRESSION_ZP_SCALAR
decompression_zp,
#endif
output,
weights
#if USE_SLM
, wei_local_mem
#endif
#if BIAS_TERM
, biases
#endif
#if HAS_FUSED_OPS_DECLS
, FUSED_OPS_ARGS
#endif
);
#endif
#endif
}
Expand Down
Loading

0 comments on commit 0b0e7ef

Please sign in to comment.