diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp index 45d6a830c..afa109262 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp @@ -5,11 +5,12 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -shared FLOAT_TYPE sccache[BLOCK_SIZE/16][2][16]; +shared FLOAT_TYPE sccache1[BLOCK_SIZE/16][16]; +shared FLOAT_TYPE sccache2[BLOCK_SIZE/16][16]; FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; -void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint itid8, const uint v_im, const uint ix, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint v_im, const uint ix, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { const uint y_idx = i * QUANT_K + y_offset; [[unroll]] for (uint n = 0; n < num_rows; ++n) { @@ -17,16 +18,18 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, if (!all_threads) { // when we don't have enough blocks to use all threads if (i < num_blocks_per_row) { - sccache[ix][0][itid] = FLOAT_TYPE((data_a[ib0 + i].scales[itid8] >> v_im*4) & 0xF); // lower 8 bytes - sccache[ix][1][itid] = FLOAT_TYPE((data_a[ib0 + i].scales[itid8+8] >> v_im*4) & 0xF); // upper 8 bytes + const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]); + sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF); + sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); } barrier(); if (i >= num_blocks_per_row) continue; } else { - sccache[ix][0][itid] = FLOAT_TYPE((data_a[ib0 + i].scales[itid8] >> v_im*4) & 0xF); - sccache[ix][1][itid] = FLOAT_TYPE((data_a[ib0 + i].scales[itid8+8] >> v_im*4) & 0xF); + const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]); + sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF); + sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); barrier(); } @@ -53,22 +56,22 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, FLOAT_TYPE sum1 = FLOAT_TYPE(0.0); FLOAT_TYPE sum2 = FLOAT_TYPE(0.0); [[unroll]] for (int l = 0; l < 2; ++l) { - sum1 = fma(FLOAT_TYPE(b0[l]), sccache[ix][v_im][0] * qs_u32_0[l ], - fma(FLOAT_TYPE(b16[l]), sccache[ix][v_im][1] * qs_u32_0[l+2], - fma(FLOAT_TYPE(b32[l]), sccache[ix][v_im][2] * qs_u32_2[l ], - fma(FLOAT_TYPE(b48[l]), sccache[ix][v_im][3] * qs_u32_2[l+2], - fma(FLOAT_TYPE(b64[l]), sccache[ix][v_im][4] * qs_u32_4[l ], - fma(FLOAT_TYPE(b80[l]), sccache[ix][v_im][5] * qs_u32_4[l+2], - fma(FLOAT_TYPE(b96[l]), sccache[ix][v_im][6] * qs_u32_6[l ], - fma(FLOAT_TYPE(b112[l]), sccache[ix][v_im][7] * qs_u32_6[l+2], sum1)))))))); - sum2 = fma(FLOAT_TYPE(b0[l]), sccache[ix][v_im][ 8], - fma(FLOAT_TYPE(b16[l]), sccache[ix][v_im][ 9], - fma(FLOAT_TYPE(b32[l]), sccache[ix][v_im][10], - fma(FLOAT_TYPE(b48[l]), sccache[ix][v_im][11], - fma(FLOAT_TYPE(b64[l]), sccache[ix][v_im][12], - fma(FLOAT_TYPE(b80[l]), sccache[ix][v_im][13], - fma(FLOAT_TYPE(b96[l]), sccache[ix][v_im][14], - fma(FLOAT_TYPE(b112[l]), sccache[ix][v_im][15], sum2)))))))); + sum1 = fma(FLOAT_TYPE(b0[l]), sccache1[ix][ 8*v_im] * qs_u32_0[l ], + fma(FLOAT_TYPE(b16[l]), sccache1[ix][1 + 8*v_im] * qs_u32_0[l+2], + fma(FLOAT_TYPE(b32[l]), sccache1[ix][2 + 8*v_im] * qs_u32_2[l ], + fma(FLOAT_TYPE(b48[l]), sccache1[ix][3 + 8*v_im] * qs_u32_2[l+2], + fma(FLOAT_TYPE(b64[l]), sccache1[ix][4 + 8*v_im] * qs_u32_4[l ], + fma(FLOAT_TYPE(b80[l]), sccache1[ix][5 + 8*v_im] * qs_u32_4[l+2], + fma(FLOAT_TYPE(b96[l]), sccache1[ix][6 + 8*v_im] * qs_u32_6[l ], + fma(FLOAT_TYPE(b112[l]), sccache1[ix][7 + 8*v_im] * qs_u32_6[l+2], sum1)))))))); + sum2 = fma(FLOAT_TYPE(b0[l]), sccache2[ix][ 8*v_im], + fma(FLOAT_TYPE(b16[l]), sccache2[ix][1 + 8*v_im], + fma(FLOAT_TYPE(b32[l]), sccache2[ix][2 + 8*v_im], + fma(FLOAT_TYPE(b48[l]), sccache2[ix][3 + 8*v_im], + fma(FLOAT_TYPE(b64[l]), sccache2[ix][4 + 8*v_im], + fma(FLOAT_TYPE(b80[l]), sccache2[ix][5 + 8*v_im], + fma(FLOAT_TYPE(b96[l]), sccache2[ix][6 + 8*v_im], + fma(FLOAT_TYPE(b112[l]), sccache2[ix][7 + 8*v_im], sum2)))))))); } temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n])); } @@ -86,7 +89,6 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { const uint tid = gl_LocalInvocationID.x; const uint itid = tid%16; // 0...15 const uint ix = tid/16; - const uint itid8 = itid%8; const uint v_im = itid/8; // 0 or 1. 0 computes 0..., 1 computes 128... const uint v_in = itid - 8*v_im; // 0...7 @@ -105,8 +107,8 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { const uint nbr_all_th = num_blocks_per_row - nbr_par_th; uint i0 = 0; [[unroll]] for (; i0 < nbr_all_th; i0 += it_size) - calc_superblock(a_offset, b_offset, itid, itid8, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, true); - calc_superblock(a_offset, b_offset, itid, itid8, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, false); + calc_superblock(a_offset, b_offset, itid, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, true); + calc_superblock(a_offset, b_offset, itid, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, false); reduce_result(temp, d_offset, first_row, num_rows, tid); }