@@ -59,7 +59,7 @@ module VX_tcu_bbuf import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
5959 input wire req_is_sparse,
6060 input wire [3 : 0 ] req_step_m,
6161 input wire [3 : 0 ] req_step_k,
62- input wire [3 : 0 ] req_step_n,
62+ input wire [TCU_TBUF_N_W - 1 : 0 ] req_step_n,
6363 input wire [1 : 0 ] req_cd_nregs,
6464 input wire [NCTA_WIDTH - 1 : 0 ] req_cta_id,
6565 input wire [`VX_CFG_XLEN - 1 : 0 ] req_desc_b,
@@ -116,8 +116,8 @@ module VX_tcu_bbuf import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
116116 localparam LG_B_BLOCK_WORDS = $clog2 (B_BLOCK_WORDS );
117117 localparam KM_CTR_W = $clog2 (TCU_TC_N + 1 );
118118 // 32-bit-word offset width for K-major address arithmetic.
119- // step_n × TCU_TC_N + j has ≤ 4 +4 bits; × ldm_words adds LDM_W.
120- localparam KM_OFF_W = LDM_W + 4 + 4 ;
119+ // step_n × TCU_TC_N + j has ≤ N_W +4 bits; × ldm_words adds LDM_W.
120+ localparam KM_OFF_W = LDM_W + TCU_TBUF_N_W + 4 ;
121121 // Tensor ops reuse slots across descriptors and prefetch the next N block.
122122 localparam bit TMEM_EN = (`VX_CFG_TCU_TMEM_ENABLED != 0 );
123123
@@ -126,14 +126,23 @@ module VX_tcu_bbuf import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
126126 // K_STEPS=2 always; N_STEPS=4/8/16 for cd_nregs=0/1/2 (NRC=8/16/32).
127127 // -----------------------------------------------------------------------
128128
129- logic [4 : 0 ] block_index;
129+ // A tensor op is one K step (step_k = 0) over up to 2^N_W N blocks, so
130+ // its block index is step_n itself.
131+ localparam BI_W = 1 + TCU_TBUF_N_W ;
132+ logic [BI_W - 1 : 0 ] block_index;
130133 always_comb begin
131- case (req_cd_nregs)
132- 2'd0 : block_index = { 2'b0 , req_step_k[0 ], req_step_n[1 : 0 ]} ; // N_STEPS=4
133- 2'd1 : block_index = { 1'b0 , req_step_k[0 ], req_step_n[2 : 0 ]} ; // N_STEPS=8
134- default : block_index = { req_step_k[0 ], req_step_n[3 : 0 ]} ; // N_STEPS=16
135- endcase
134+ if (TMEM_EN ) begin
135+ block_index = BI_W ' (req_step_n);
136+ end else begin
137+ case (req_cd_nregs)
138+ 2'd0 : block_index = BI_W ' ({ req_step_k[0 ], req_step_n[1 : 0 ]} ); // N_STEPS=4
139+ 2'd1 : block_index = BI_W ' ({ req_step_k[0 ], req_step_n[2 : 0 ]} ); // N_STEPS=8
140+ default : block_index = BI_W ' ({ req_step_k[0 ], req_step_n[3 : 0 ]} ); // N_STEPS=16
141+ endcase
142+ end
136143 end
144+ `RUNTIME_ASSERT (~ (TMEM_EN && req_valid && (req_step_k != 4'd0 )),
145+ (" %s tensor-op tile-buffer request with step_k %0d " , INSTANCE_ID , req_step_k))
137146 if (4 > 1 ) begin : g_step_k_upper_unused
138147 `UNUSED_VAR (req_step_k[3 : 1 ])
139148 end
@@ -147,19 +156,17 @@ module VX_tcu_bbuf import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
147156 // Sparse blocks use the flat producer layout and can span one or two
148157 // logical 32-bit bank rows.
149158 localparam TOTAL_SHIFT = LG_B_SUB_BLOCKS + LG_XLEN_RATIO ;
150- wire [4 : 0 ] dense_offset = (TOTAL_SHIFT == 0 )
151- ? block_index
152- : 5 '(block_index >> TOTAL_SHIFT );
159+ wire [BANK_ADDR_WIDTH - 1 : 0 ] dense_offset = BANK_ADDR_WIDTH ' (block_index >> TOTAL_SHIFT );
153160
154161 // Dense within-physical-bank-row selector (XLEN>32 only). Picks which
155162 // of the XLEN_RATIO logical 32-bit bank-rows to copy into slot A.
156163 localparam SUB_HALF_W = (LG_XLEN_RATIO == 0 ) ? 1 : LG_XLEN_RATIO ;
157164 wire [SUB_HALF_W - 1 : 0 ] dense_sub_half =
158165 (LG_XLEN_RATIO == 0 )
159166 ? '0
160- : SUB_HALF_W ' (({ 27'b0 , block_index} >> LG_B_SUB_BLOCKS ) & ((1 << LG_XLEN_RATIO ) - 1 ));
167+ : SUB_HALF_W ' ((32 '( block_index) >> LG_B_SUB_BLOCKS ) & ((1 << LG_XLEN_RATIO ) - 1 ));
161168
162- localparam SP_LOGROW_W = 6 ;
169+ localparam SP_LOGROW_W = BI_W + 1 ;
163170 wire [SP_LOGROW_W - 1 : 0 ] sp_logrow_a =
164171 SP_LOGROW_W ' (block_index) * SP_LOGROW_W ' (LOGICAL_ROWS_PER_BLK );
165172 wire [SP_LOGROW_W - 1 : 0 ] sp_logrow_b = sp_logrow_a + SP_LOGROW_W ' (1 );
@@ -214,7 +221,7 @@ module VX_tcu_bbuf import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
214221 logic slot_row_major_r;
215222 logic [LDM_W - 1 : 0 ] slot_ldm_words_r;
216223 logic [3 : 0 ] slot_step_k_r;
217- logic [3 : 0 ] slot_step_n_r;
224+ logic [TCU_TBUF_N_W - 1 : 0 ] slot_step_n_r;
218225 logic [NCTA_WIDTH - 1 : 0 ] slot_cta_id_r;
219226 // K-major descriptors can point to storage rewritten between WGMMA instructions.
220227 logic refetched_for_first_uop_r;
@@ -231,7 +238,7 @@ module VX_tcu_bbuf import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
231238
232239 // Per-mode fetch addresses.
233240 wire [BANK_ADDR_WIDTH - 1 : 0 ] fetch_addr_dense =
234- desc_b_row_base + BANK_ADDR_WIDTH ' ( dense_offset) ;
241+ desc_b_row_base + dense_offset;
235242 wire [BANK_ADDR_WIDTH - 1 : 0 ] fetch_addr_a_sparse =
236243 desc_b_row_base + BANK_ADDR_WIDTH ' (sp_logrow_a >> LG_XLEN_RATIO );
237244 wire [BANK_ADDR_WIDTH - 1 : 0 ] fetch_addr_b_sparse =
@@ -279,7 +286,8 @@ module VX_tcu_bbuf import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
279286 // slot B (unused by dense B) while slot A fetches the next N block.
280287 wire req_dense_km = req_wants_kmajor && ! req_is_sparse;
281288 logic w1_valid_r;
282- logic [3 : 0 ] w1_step_k_r, w1_step_n_r;
289+ logic [3 : 0 ] w1_step_k_r;
290+ logic [TCU_TBUF_N_W - 1 : 0 ] w1_step_n_r;
283291 logic [NCTA_WIDTH - 1 : 0 ] w1_cta_id_r;
284292 logic [BANK_ADDR_WIDTH - 1 : 0 ] w1_row_base_r;
285293 logic [BANK_ROW_WORDS_LOG2 - 1 : 0 ] w1_inrow_r;
@@ -299,8 +307,8 @@ module VX_tcu_bbuf import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
299307 wire need_fetch = req_valid && ! bank_row_resident;
300308 wire alloc_en = need_fetch && ! slot_fetching_r;
301309 wire pf_en = TMEM_EN && req_valid && req_dense_km && req_pf_ok && bank_row_resident_kmajor
302- && ! slot_fetching_r && ! req_setup && (req_step_n != 4'd15 );
303- wire [3 : 0 ] fetch_step_n = pf_en ? (req_step_n + 4'd1 ) : req_step_n;
310+ && ! slot_fetching_r && ! req_setup && (req_step_n != '1 );
311+ wire [TCU_TBUF_N_W - 1 : 0 ] fetch_step_n = pf_en ? (req_step_n + TCU_TBUF_N_W ' ( 1 ) ) : req_step_n;
304312 wire rd_w1 = bank_row_resident_w1 && ! bank_row_resident_kmajor;
305313
306314 assign bbuf_ready = ! req_valid || bank_row_resident;
0 commit comments